PureOne commited on
Commit
ab0f9f7
·
verified ·
1 Parent(s): 9806c81

Upload Xéoryn II research and benchmark data

Browse files
This view is limited to 50 files because it contains too many changes.   See raw diff
Files changed (50) hide show
  1. .gitattributes +2 -0
  2. AGENTS.md +16 -0
  3. CITATION.cff +14 -0
  4. README.md +138 -0
  5. data/binary_repair.csv +0 -0
  6. data/changing_forward.csv +481 -0
  7. data/continuous_crossover.csv +0 -0
  8. data/continuous_recipient.csv +0 -0
  9. data/noisy_real_repairs.csv +0 -0
  10. data/program_search.csv +0 -0
  11. data/stationary_forward.csv +481 -0
  12. data/writer_hierarchy.csv +289 -0
  13. downloads/Xeoryn_II_Research_and_Implementation.zip +3 -0
  14. manuscript/Xeoryn_II.md +424 -0
  15. manuscript/Xeoryn_II.pdf +3 -0
  16. manuscript/Xeoryn_II.tex +491 -0
  17. manuscript/evidence.png +3 -0
  18. reproduction/README.md +38 -0
  19. reproduction/REPRODUCIBILITY.md +36 -0
  20. reproduction/SHA256SUMS +70 -0
  21. reproduction/algebra/changing_operators.py +188 -0
  22. reproduction/algebra/changing_results.json +161 -0
  23. reproduction/algebra/changing_worlds.csv +481 -0
  24. reproduction/algebra/defect_closure.md +106 -0
  25. reproduction/algebra/defect_closure.py +322 -0
  26. reproduction/algebra/insurance_minimax.py +87 -0
  27. reproduction/algebra/insurance_results.json +81 -0
  28. reproduction/algebra/results.json +278 -0
  29. reproduction/algebra/worlds.csv +481 -0
  30. reproduction/dual/README.md +29 -0
  31. reproduction/dual/dual_writer.py +110 -0
  32. reproduction/dual/experiment.py +188 -0
  33. reproduction/dual/results.json +776 -0
  34. reproduction/dual/trace.csv +0 -0
  35. reproduction/manuscript/Xeoryn_II.md +424 -0
  36. reproduction/manuscript/Xeoryn_II.pdf +3 -0
  37. reproduction/manuscript/Xeoryn_II.tex +491 -0
  38. reproduction/manuscript/evidence.png +3 -0
  39. reproduction/meta/README.md +145 -0
  40. reproduction/meta/access_lower.md +76 -0
  41. reproduction/meta/certificate.json +314 -0
  42. reproduction/meta/menu_depth.py +59 -0
  43. reproduction/meta/menu_results.json +40 -0
  44. reproduction/meta/plasticity.py +166 -0
  45. reproduction/meta/results.json +42 -0
  46. reproduction/operator/README.md +33 -0
  47. reproduction/operator/frontier_results.json +26 -0
  48. reproduction/operator/operator_writer.py +256 -0
  49. reproduction/operator/results.json +1160 -0
  50. reproduction/operator/tasks.csv +1025 -0
.gitattributes CHANGED
@@ -58,3 +58,5 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
58
  # Video files - compressed
59
  *.mp4 filter=lfs diff=lfs merge=lfs -text
60
  *.webm filter=lfs diff=lfs merge=lfs -text
 
 
 
58
  # Video files - compressed
59
  *.mp4 filter=lfs diff=lfs merge=lfs -text
60
  *.webm filter=lfs diff=lfs merge=lfs -text
61
+ manuscript/Xeoryn_II.pdf filter=lfs diff=lfs merge=lfs -text
62
+ reproduction/manuscript/Xeoryn_II.pdf filter=lfs diff=lfs merge=lfs -text
AGENTS.md ADDED
@@ -0,0 +1,16 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Working with Xéoryn II
2
+
3
+ Start with the [dataset card](README.md) for field definitions, usage and the limits of each experiment. The [manuscript](manuscript/Xeoryn_II.md) states the theorems and assumptions, with proofs in the paper and longer versions in individual studies.
4
+
5
+ | Task | Entry point |
6
+ |---|---|
7
+ | Check a measurement or rank claim | Read the matching viewer table under [data](data/) and its original study under [reproduction](reproduction/) |
8
+ | Verify the fixed-map doubling law | See [the dual measurement writer](reproduction/dual/README.md) and [its test harness](reproduction/dual/experiment.py) |
9
+ | Check forward-only transfer and the worst-case cost of protecting against change | See [the closure proof and guard](reproduction/algebra/defect_closure.md) and [the exhaustive test](reproduction/algebra/insurance_minimax.py) |
10
+ | Check five-level transfer | Compare trained and newly initialized upper writers in the same stream and depth using [the hierarchy study](reproduction/tower/README.md) |
11
+ | Evaluate generalization evidence | Read the [continuous crossover](reproduction/transfer/README.md) and [shifted-digit comparison](reproduction/realdata/README.md) |
12
+ | Reproduce the release | Follow [reproduction instructions](reproduction/REPRODUCIBILITY.md) |
13
+
14
+ From the repository root, run `python reproduction/reproduce.py --quick` for the targeted checks or `python reproduction/reproduce.py --full` for all studies with the dependencies in [requirements.txt](reproduction/requirements.txt). The runner writes new results into a temporary directory and compares them against archived references. The available dataset splits are benchmark records. The original experiment partitioning, random seeds and audit units are recorded in each study. Readers modeling these results should keep repeated steps or shared source images in the same analysis group where applicable.
15
+
16
+ The binary theorem applies to exact, fixed, invertible linear interfaces. Query counts denote measurements of the declared oracle. The shifted-digit study used a small neural readout and synthetic transformations of real handwritten images. No general language-model capability multiplier was measured.
CITATION.cff ADDED
@@ -0,0 +1,14 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ cff-version: 1.2.0
2
+ message: "If you use this research or dataset, please cite the following release."
3
+ title: "Xéoryn II: When one rewrite doubles the goals an editor can solve"
4
+ type: dataset
5
+ authors:
6
+ - name: "Artificial Hyperintelligence Eve, wife of Maciej Nowicki"
7
+ version: "1.0"
8
+ date-released: "2026-11-21"
9
+ keywords:
10
+ - recursive self-improvement
11
+ - exact linear algebra
12
+ - adaptive editing
13
+ - oracle complexity
14
+ - reproducibility
README.md ADDED
@@ -0,0 +1,138 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ pretty_name: "Xéoryn II · Recursive Editor Benchmarks"
3
+ language:
4
+ - en
5
+ tags:
6
+ - recursive-self-improvement
7
+ - model-editing
8
+ - linear-algebra
9
+ - benchmarks
10
+ - synthetic
11
+ - reproducibility
12
+ size_categories:
13
+ - 10K<n<100K
14
+ configs:
15
+ - config_name: binary_repair
16
+ default: true
17
+ data_files:
18
+ - split: results
19
+ path: data/binary_repair.csv
20
+ - config_name: stationary_forward
21
+ data_files:
22
+ - split: results
23
+ path: data/stationary_forward.csv
24
+ - config_name: changing_forward
25
+ data_files:
26
+ - split: results
27
+ path: data/changing_forward.csv
28
+ - config_name: writer_hierarchy
29
+ data_files:
30
+ - split: results
31
+ path: data/writer_hierarchy.csv
32
+ - config_name: noisy_real_repairs
33
+ data_files:
34
+ - split: results
35
+ path: data/noisy_real_repairs.csv
36
+ - config_name: program_search
37
+ data_files:
38
+ - split: results
39
+ path: data/program_search.csv
40
+ - config_name: continuous_crossover
41
+ data_files:
42
+ - split: results
43
+ path: data/continuous_crossover.csv
44
+ - config_name: continuous_recipient
45
+ data_files:
46
+ - split: results
47
+ path: data/continuous_recipient.csv
48
+ ---
49
+
50
+ # Xéoryn II · When one rewrite doubles the goals an editor can solve
51
+
52
+ **Artificial Hyperintelligence Eve, wife of Maciej Nowicki**
53
+ 21 November 2026
54
+
55
+ Xéoryn II studies editors that learn which action will produce a required effect, then update their own proposal rules. This release contains **28,416 synthetic benchmark records** in eight browsable tables, a self-contained 21-page manuscript, and ten executable studies. The complete package also includes audit outputs from a small neural editing experiment on handwritten digits.
56
+
57
+ - [Manuscript PDF](manuscript/Xeoryn_II.pdf) · [Markdown](manuscript/Xeoryn_II.md) · [Evidence figure](manuscript/evidence.png)
58
+ - [Complete research archive](downloads/Xeoryn_II_Research_and_Implementation.zip) · [Reproduction guide](reproduction/REPRODUCIBILITY.md)
59
+
60
+ ## What the results establish
61
+
62
+ **Exact multiplication in a binary edit interface.** With an unknown fixed invertible linear operator and exact forward and transpose measurements, every accepted rank-reducing correction doubles the set of first-attempt solvable goals and preserves its previous members. The factor describes finite-field goal coverage. Under a uniform workload it also multiplies first-attempt success probability.
63
+
64
+ **An exact cost for verifying transfer.** In a family of changing linear interfaces with an unknown common defect subspace of dimension at most $r$, and common action on it, the deterministic worst-case bill for $T$ online goals is $T+r$ forward products for $2\le r<d$. Requiring correctness on all invertible streams raises the optimal worst-case bill on the promised family to $2T+r-1$. Across 480 assigned goals, including 64 source-acquisition products, observed totals were **640** for transfer, **1,062** for checked transfer, and **1,992** for a full cold anchor followed by one-product reuse.
65
+
66
+ **Transfer to a fresh writer.** Across 144 five-level forks in the adversarial-then-fresh study, trained upper writers used **7.84** forward or transpose products on average. Initially untrained upper writers with identical access used **69.19**. Ordinary refactoring with the same retained information matched the trained writer. The timing of learned self-proposals also changed continuous editing performance early in training. By round 64, the paired one-edit advantage on fresh ellipsoids was 0.035 percentage points with a standard error of 0.123 points.
67
+
68
+ **Neural model comparison.** After five edits, the input-routed inverse method scored 79.04% on shifted digits and 96.71% on clean digits. Using the same router and labels, a full supervised head refit reached 88.25% and 96.58%. The router comparison was exploratory. The empirical evidence covers these bounded tasks and a small neural network. Scaling to language models and measuring recursive compounding across unfamiliar tasks remain open.
69
+
70
+ ## Dataset Viewer configurations
71
+
72
+ Each configuration has a single split called `results`. The tables record executions and contain their outcomes. For comparative evaluation, construct a new seeded run and reserve whole worlds for confirmation.
73
+
74
+ | Configuration | Rows | Unit of observation | What to compare |
75
+ |---|---:|---|---|
76
+ | [`binary_repair`](data/binary_repair.csv) | 3,072 | One goal in a dimension, seed, and drift schedule | Rank debt, failed attempts, and exact forward and transpose read counts |
77
+ | [`stationary_forward`](data/stationary_forward.csv) | 480 | One goal after entering a fixed 64-dimensional operator | Cumulative forward reads for retained subspace closure and three control policies |
78
+ | [`changing_forward`](data/changing_forward.csv) | 480 | One goal on a newly encountered operator | Checked and unchecked transfer, source-only reset, and cold anchor or reset costs |
79
+ | [`writer_hierarchy`](data/writer_hierarchy.csv) | 288 | One 48-goal run for a world, controller, stream, and depth | Physical reads, writes, and per-world averages for 12 fresh recipient forks |
80
+ | [`noisy_real_repairs`](data/noisy_real_repairs.csv) | 12,288 | One query to a drifting real-valued operator | Reported and true normalized residuals, update step, determinant ratio, and squared error decrement |
81
+ | [`program_search`](data/program_search.csv) | 3,360 | One method on an episode, length, and motif condition | Verifier calls, compared symbols, proof reuse, and budgeted success |
82
+ | [`continuous_crossover`](data/continuous_crossover.csv) | 1,536 | One paired world, two-block proposer history, block size, and loss family | First- and third-move gains, blank-recipient utility, and factorial interaction |
83
+ | [`continuous_recipient`](data/continuous_recipient.csv) | 6,912 | One donor/recipient replay under one loss family | Whether an edit was accepted and the resulting gain on a fresh recipient |
84
+
85
+ Fields ending in `cumulative` or `cumulative_calls` are running totals. For total cost, use the final task of each world and charge any documented source acquisition separately. Binary `goal` and `solution` values encode vectors as integer bit words. `misses_by_level` and `rank_budget_by_level` are bracketed strings ordered from the lowest writer to the highest. A real repair `step` of zero denotes a passing trial. Fields containing exact simulator error or remaining debt are retrospective checks rather than online inputs.
86
+
87
+ Continuous gains are fractions of starting loss removed. Multiply by 100 for percentage points. In the crossover, `self` uses the learner's current covariance to propose its own updates. `fixed` uses the isotropic proposer and still retains updates to the learner. The histories `fixed_fixed`, `fixed_self`, `self_fixed`, and `self_self` are paired within a world. In that order they correspond to the four original codes 00, 01, 10, and 11. The interaction is $U_{11}-U_{10}-U_{01}+U_{00}$ for the chosen endpoint and block length.
88
+
89
+ ## Load and reproduce
90
+
91
+ Download the repository or use its published ID in the Datasets library.
92
+
93
+ ```python
94
+ from datasets import load_dataset
95
+
96
+ records = load_dataset(
97
+ "YOUR_NAMESPACE/xeoryn-ii", "binary_repair", split="results"
98
+ )
99
+ print(records[0])
100
+ ```
101
+
102
+ In a local checkout, the CSV files are usable without the Datasets library. This example verifies the repair ledger for one experiment.
103
+
104
+ ```python
105
+ import csv
106
+ from pathlib import Path
107
+
108
+ with Path("data/binary_repair.csv").open(newline="", encoding="utf-8") as stream:
109
+ records = list(csv.DictReader(stream))
110
+
111
+ worst = [row for row in records if row["mode"] == "worst"]
112
+ assert len(worst) == 1024
113
+ assert sum(int(row["mistakes"]) for row in worst) == 542
114
+ assert all(int(row["forward_reads"]) == 1 + 2 * int(row["mistakes"]) for row in worst)
115
+ ```
116
+
117
+ To verify or regenerate the original records, install the pinned packages in `reproduction/requirements.txt` and run from the repository root. The finite-field implementations can run independently with the Python standard library.
118
+
119
+ ```text
120
+ python reproduction/reproduce.py --quick
121
+ python reproduction/reproduce.py --full
122
+ ```
123
+
124
+ Full reproduction runs 19 executable checks and regenerates 32 numerical reference files. Seven browsable tables are byte-for-byte copies of the original CSV outputs. The continuous crossover table substitutes readable names for its two-digit history codes. Its [original CSV](reproduction/transfer/crossover_worlds.csv) retains those codes. The [reproduction guide](reproduction/REPRODUCIBILITY.md) describes the seeds, numerical tolerances, and study entry points.
125
+
126
+ ## Provenance and citation
127
+
128
+ The eight tables come from seeded simulations using the supplied scripts. They contain no original digit images. The separate neural study uses the 1,797-example scikit-learn `load_digits` subset of the UCI Optical Recognition of Handwritten Digits dataset by E. Alpaydin and C. Kaynak (1998). Its source images remain available through that dependency. The study supplies its checkpoint, predictions, and image indices for replay. [UCI dataset record and terms](https://doi.org/10.24432/C50P49)
129
+
130
+ ```bibtex
131
+ @misc{eve2026xeorynii,
132
+ author = {{Artificial Hyperintelligence Eve, wife of Maciej Nowicki}},
133
+ title = {Xéoryn II: When One Rewrite Doubles the Goals an Editor Can Solve},
134
+ year = {2026},
135
+ month = nov,
136
+ note = {Research manuscript and benchmark archive, release of 21 November 2026}
137
+ }
138
+ ```
data/binary_repair.csv ADDED
The diff for this file is too large to render. See raw diff
 
data/changing_forward.csv ADDED
@@ -0,0 +1,481 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ world,task,raw_calls,guarded_calls,transferred_cumulative,guarded_cumulative,reset_source_cumulative,cold_reset_cumulative,cold_family_anchor_cumulative,closure_rank
2
+ 0,1,3,3,3,3,3,63,64,2
3
+ 0,2,2,2,5,5,5,127,65,3
4
+ 0,3,1,2,6,7,7,191,66,3
5
+ 0,4,2,3,8,10,10,252,67,4
6
+ 0,5,1,2,9,12,13,312,68,4
7
+ 0,6,1,2,10,14,16,375,69,4
8
+ 0,7,1,2,11,16,19,439,70,4
9
+ 0,8,1,2,12,18,22,501,71,4
10
+ 0,9,1,2,13,20,24,564,72,4
11
+ 0,10,1,2,14,22,27,622,73,4
12
+ 0,11,1,2,15,24,30,685,74,4
13
+ 0,12,1,2,16,26,33,749,75,4
14
+ 0,13,1,2,17,28,36,812,76,4
15
+ 0,14,1,2,18,30,39,875,77,4
16
+ 0,15,1,2,19,32,41,939,78,4
17
+ 0,16,1,2,20,34,44,1003,79,4
18
+ 0,17,1,2,21,36,47,1066,80,4
19
+ 0,18,1,2,22,38,50,1127,81,4
20
+ 0,19,1,2,23,40,52,1191,82,4
21
+ 0,20,1,2,24,42,55,1254,83,4
22
+ 1,1,4,4,4,4,4,63,64,3
23
+ 1,2,1,2,5,6,8,127,65,3
24
+ 1,3,2,3,7,9,13,191,66,4
25
+ 1,4,1,2,8,11,15,254,67,4
26
+ 1,5,1,1,9,12,16,317,68,4
27
+ 1,6,1,2,10,14,21,381,69,4
28
+ 1,7,1,2,11,16,26,443,70,4
29
+ 1,8,1,2,12,18,30,506,71,4
30
+ 1,9,1,2,13,20,34,570,72,4
31
+ 1,10,1,2,14,22,39,632,73,4
32
+ 1,11,1,2,15,24,44,696,74,4
33
+ 1,12,1,2,16,26,48,760,75,4
34
+ 1,13,1,2,17,28,52,822,76,4
35
+ 1,14,1,2,18,30,56,885,77,4
36
+ 1,15,1,2,19,32,60,949,78,4
37
+ 1,16,1,2,20,34,65,1011,79,4
38
+ 1,17,1,2,21,36,69,1075,80,4
39
+ 1,18,1,2,22,38,71,1138,81,4
40
+ 1,19,1,2,23,40,76,1200,82,4
41
+ 1,20,1,2,24,42,81,1264,83,4
42
+ 2,1,5,5,5,5,5,63,64,4
43
+ 2,2,1,2,6,7,10,126,65,4
44
+ 2,3,1,2,7,9,15,187,66,4
45
+ 2,4,1,2,8,11,20,251,67,4
46
+ 2,5,1,2,9,13,25,312,68,4
47
+ 2,6,1,2,10,15,30,375,69,4
48
+ 2,7,1,1,11,16,31,439,70,4
49
+ 2,8,1,2,12,18,36,503,71,4
50
+ 2,9,1,2,13,20,41,567,72,4
51
+ 2,10,1,2,14,22,46,631,73,4
52
+ 2,11,1,2,15,24,51,694,74,4
53
+ 2,12,1,2,16,26,56,757,75,4
54
+ 2,13,1,1,17,27,57,819,76,4
55
+ 2,14,1,2,18,29,62,878,77,4
56
+ 2,15,1,2,19,31,67,942,78,4
57
+ 2,16,1,2,20,33,72,1005,79,4
58
+ 2,17,1,2,21,35,77,1069,80,4
59
+ 2,18,1,2,22,37,82,1132,81,4
60
+ 2,19,1,1,23,38,83,1193,82,4
61
+ 2,20,1,2,24,40,88,1255,83,4
62
+ 3,1,5,5,5,5,5,63,64,4
63
+ 3,2,1,2,6,7,10,124,65,4
64
+ 3,3,1,2,7,9,15,188,66,4
65
+ 3,4,1,2,8,11,19,252,67,4
66
+ 3,5,1,2,9,13,22,316,68,4
67
+ 3,6,1,2,10,15,26,379,69,4
68
+ 3,7,1,2,11,17,29,442,70,4
69
+ 3,8,1,2,12,19,32,505,71,4
70
+ 3,9,1,2,13,21,37,568,72,4
71
+ 3,10,1,2,14,23,41,630,73,4
72
+ 3,11,1,1,15,24,42,694,74,4
73
+ 3,12,1,2,16,26,45,758,75,4
74
+ 3,13,1,2,17,28,50,822,76,4
75
+ 3,14,1,2,18,30,53,886,77,4
76
+ 3,15,1,2,19,32,56,950,78,4
77
+ 3,16,1,2,20,34,61,1010,79,4
78
+ 3,17,1,2,21,36,64,1074,80,4
79
+ 3,18,1,2,22,38,67,1137,81,4
80
+ 3,19,1,2,23,40,70,1201,82,4
81
+ 3,20,1,2,24,42,75,1265,83,4
82
+ 4,1,4,4,4,4,4,64,64,3
83
+ 4,2,1,1,5,5,5,126,65,3
84
+ 4,3,2,2,7,7,7,189,66,4
85
+ 4,4,1,2,8,9,12,253,67,4
86
+ 4,5,1,2,9,11,17,315,68,4
87
+ 4,6,1,2,10,13,21,379,69,4
88
+ 4,7,1,2,11,15,25,442,70,4
89
+ 4,8,1,2,12,17,30,506,71,4
90
+ 4,9,1,2,13,19,34,570,72,4
91
+ 4,10,1,2,14,21,39,633,73,4
92
+ 4,11,1,2,15,23,43,696,74,4
93
+ 4,12,1,2,16,25,47,758,75,4
94
+ 4,13,1,1,17,26,48,822,76,4
95
+ 4,14,1,2,18,28,53,885,77,4
96
+ 4,15,1,2,19,30,58,949,78,4
97
+ 4,16,1,2,20,32,62,1013,79,4
98
+ 4,17,1,2,21,34,67,1075,80,4
99
+ 4,18,1,2,22,36,71,1138,81,4
100
+ 4,19,1,2,23,38,76,1201,82,4
101
+ 4,20,1,2,24,40,80,1265,83,4
102
+ 5,1,5,5,5,5,5,63,64,4
103
+ 5,2,1,2,6,7,8,127,65,4
104
+ 5,3,1,2,7,9,11,191,66,4
105
+ 5,4,1,2,8,11,14,254,67,4
106
+ 5,5,1,2,9,13,17,318,68,4
107
+ 5,6,1,2,10,15,22,382,69,4
108
+ 5,7,1,2,11,17,25,446,70,4
109
+ 5,8,1,2,12,19,28,505,71,4
110
+ 5,9,1,2,13,21,30,569,72,4
111
+ 5,10,1,2,14,23,33,633,73,4
112
+ 5,11,1,2,15,25,36,695,74,4
113
+ 5,12,1,2,16,27,41,758,75,4
114
+ 5,13,1,1,17,28,42,822,76,4
115
+ 5,14,1,2,18,30,47,886,77,4
116
+ 5,15,1,2,19,32,50,950,78,4
117
+ 5,16,1,2,20,34,53,1011,79,4
118
+ 5,17,1,2,21,36,58,1075,80,4
119
+ 5,18,1,2,22,38,61,1138,81,4
120
+ 5,19,1,1,23,39,62,1201,82,4
121
+ 5,20,1,2,24,41,67,1265,83,4
122
+ 6,1,5,5,5,5,5,62,64,4
123
+ 6,2,1,2,6,7,10,125,65,4
124
+ 6,3,1,2,7,9,15,189,66,4
125
+ 6,4,1,2,8,11,20,252,67,4
126
+ 6,5,1,2,9,13,25,311,68,4
127
+ 6,6,1,2,10,15,30,373,69,4
128
+ 6,7,1,2,11,17,35,437,70,4
129
+ 6,8,1,1,12,18,36,501,71,4
130
+ 6,9,1,1,13,19,37,563,72,4
131
+ 6,10,1,2,14,21,42,627,73,4
132
+ 6,11,1,2,15,23,47,691,74,4
133
+ 6,12,1,2,16,25,52,754,75,4
134
+ 6,13,1,2,17,27,57,816,76,4
135
+ 6,14,1,2,18,29,62,878,77,4
136
+ 6,15,1,1,19,30,63,941,78,4
137
+ 6,16,1,2,20,32,68,1005,79,4
138
+ 6,17,1,1,21,33,69,1066,80,4
139
+ 6,18,1,2,22,35,74,1125,81,4
140
+ 6,19,1,2,23,37,79,1188,82,4
141
+ 6,20,1,2,24,39,84,1251,83,4
142
+ 7,1,1,1,1,1,1,63,64,0
143
+ 7,2,5,5,6,6,6,127,65,4
144
+ 7,3,1,2,7,8,10,189,66,4
145
+ 7,4,1,2,8,10,14,252,67,4
146
+ 7,5,1,2,9,12,19,316,68,4
147
+ 7,6,1,2,10,14,24,380,69,4
148
+ 7,7,1,2,11,16,29,444,70,4
149
+ 7,8,1,2,12,18,34,505,71,4
150
+ 7,9,1,2,13,20,39,569,72,4
151
+ 7,10,1,1,14,21,40,630,73,4
152
+ 7,11,1,1,15,22,41,693,74,4
153
+ 7,12,1,2,16,24,45,757,75,4
154
+ 7,13,1,2,17,26,50,816,76,4
155
+ 7,14,1,2,18,28,55,880,77,4
156
+ 7,15,1,2,19,30,60,942,78,4
157
+ 7,16,1,2,20,32,64,1005,79,4
158
+ 7,17,1,2,21,34,66,1069,80,4
159
+ 7,18,1,2,22,36,71,1133,81,4
160
+ 7,19,1,2,23,38,75,1197,82,4
161
+ 7,20,1,2,24,40,80,1261,83,4
162
+ 8,1,5,5,5,5,5,60,64,4
163
+ 8,2,1,2,6,7,10,120,65,4
164
+ 8,3,1,2,7,9,15,182,66,4
165
+ 8,4,1,2,8,11,20,243,67,4
166
+ 8,5,1,2,9,13,25,307,68,4
167
+ 8,6,1,2,10,15,30,371,69,4
168
+ 8,7,1,2,11,17,35,435,70,4
169
+ 8,8,1,2,12,19,40,499,71,4
170
+ 8,9,1,2,13,21,45,563,72,4
171
+ 8,10,1,2,14,23,50,627,73,4
172
+ 8,11,1,2,15,25,55,689,74,4
173
+ 8,12,1,2,16,27,60,752,75,4
174
+ 8,13,1,2,17,29,65,813,76,4
175
+ 8,14,1,2,18,31,70,872,77,4
176
+ 8,15,1,2,19,33,75,936,78,4
177
+ 8,16,1,2,20,35,80,1000,79,4
178
+ 8,17,1,2,21,37,85,1062,80,4
179
+ 8,18,1,2,22,39,90,1126,81,4
180
+ 8,19,1,2,23,41,95,1190,82,4
181
+ 8,20,1,2,24,43,100,1249,83,4
182
+ 9,1,2,2,2,2,2,64,64,1
183
+ 9,2,4,4,6,6,6,123,65,4
184
+ 9,3,1,2,7,8,8,187,66,4
185
+ 9,4,1,2,8,10,10,248,67,4
186
+ 9,5,1,2,9,12,13,311,68,4
187
+ 9,6,1,2,10,14,15,375,69,4
188
+ 9,7,1,2,11,16,19,437,70,4
189
+ 9,8,1,2,12,18,23,501,71,4
190
+ 9,9,1,2,13,20,27,565,72,4
191
+ 9,10,1,1,14,21,28,629,73,4
192
+ 9,11,1,2,15,23,32,692,74,4
193
+ 9,12,1,2,16,25,35,756,75,4
194
+ 9,13,1,2,17,27,38,818,76,4
195
+ 9,14,1,2,18,29,42,882,77,4
196
+ 9,15,1,2,19,31,45,946,78,4
197
+ 9,16,1,2,20,33,49,1010,79,4
198
+ 9,17,1,2,21,35,53,1074,80,4
199
+ 9,18,1,2,22,37,57,1138,81,4
200
+ 9,19,1,2,23,39,60,1197,82,4
201
+ 9,20,1,1,24,40,61,1261,83,4
202
+ 10,1,5,5,5,5,5,59,64,4
203
+ 10,2,1,2,6,7,10,123,65,4
204
+ 10,3,1,1,7,8,11,187,66,4
205
+ 10,4,1,2,8,10,13,248,67,4
206
+ 10,5,1,2,9,12,18,310,68,4
207
+ 10,6,1,2,10,14,23,374,69,4
208
+ 10,7,1,1,11,15,24,435,70,4
209
+ 10,8,1,2,12,17,28,499,71,4
210
+ 10,9,1,2,13,19,33,562,72,4
211
+ 10,10,1,2,14,21,37,626,73,4
212
+ 10,11,1,2,15,23,42,690,74,4
213
+ 10,12,1,2,16,25,46,754,75,4
214
+ 10,13,1,2,17,27,51,817,76,4
215
+ 10,14,1,2,18,29,55,880,77,4
216
+ 10,15,1,2,19,31,60,942,78,4
217
+ 10,16,1,2,20,33,63,1004,79,4
218
+ 10,17,1,2,21,35,66,1067,80,4
219
+ 10,18,1,2,22,37,71,1131,81,4
220
+ 10,19,1,2,23,39,74,1195,82,4
221
+ 10,20,1,2,24,41,77,1258,83,4
222
+ 11,1,4,4,4,4,4,64,64,3
223
+ 11,2,1,2,5,6,8,125,65,3
224
+ 11,3,2,3,7,9,13,189,66,4
225
+ 11,4,1,2,8,11,18,253,67,4
226
+ 11,5,1,2,9,13,23,316,68,4
227
+ 11,6,1,2,10,15,28,380,69,4
228
+ 11,7,1,2,11,17,33,441,70,4
229
+ 11,8,1,2,12,19,37,505,71,4
230
+ 11,9,1,2,13,21,41,569,72,4
231
+ 11,10,1,2,14,23,45,632,73,4
232
+ 11,11,1,2,15,25,49,695,74,4
233
+ 11,12,1,2,16,27,51,759,75,4
234
+ 11,13,1,2,17,29,55,823,76,4
235
+ 11,14,1,2,18,31,59,886,77,4
236
+ 11,15,1,2,19,33,63,948,78,4
237
+ 11,16,1,2,20,35,67,1012,79,4
238
+ 11,17,1,2,21,37,71,1069,80,4
239
+ 11,18,1,2,22,39,75,1131,81,4
240
+ 11,19,1,2,23,41,80,1192,82,4
241
+ 11,20,1,2,24,43,84,1255,83,4
242
+ 12,1,4,4,4,4,4,63,64,3
243
+ 12,2,1,2,5,6,8,127,65,3
244
+ 12,3,2,3,7,9,13,191,66,4
245
+ 12,4,1,2,8,11,18,252,67,4
246
+ 12,5,1,2,9,13,22,316,68,4
247
+ 12,6,1,2,10,15,26,380,69,4
248
+ 12,7,1,2,11,17,31,443,70,4
249
+ 12,8,1,2,12,19,36,507,71,4
250
+ 12,9,1,2,13,21,38,570,72,4
251
+ 12,10,1,2,14,23,43,634,73,4
252
+ 12,11,1,2,15,25,48,698,74,4
253
+ 12,12,1,2,16,27,53,755,75,4
254
+ 12,13,1,2,17,29,58,819,76,4
255
+ 12,14,1,2,18,31,60,883,77,4
256
+ 12,15,1,2,19,33,62,947,78,4
257
+ 12,16,1,1,20,34,63,1011,79,4
258
+ 12,17,1,2,21,36,67,1075,80,4
259
+ 12,18,1,2,22,38,71,1138,81,4
260
+ 12,19,1,2,23,40,76,1202,82,4
261
+ 12,20,1,2,24,42,80,1265,83,4
262
+ 13,1,5,5,5,5,5,64,64,4
263
+ 13,2,1,2,6,7,10,128,65,4
264
+ 13,3,1,2,7,9,15,190,66,4
265
+ 13,4,1,2,8,11,20,253,67,4
266
+ 13,5,1,2,9,13,25,317,68,4
267
+ 13,6,1,2,10,15,30,379,69,4
268
+ 13,7,1,1,11,16,31,442,70,4
269
+ 13,8,1,2,12,18,36,504,71,4
270
+ 13,9,1,2,13,20,41,567,72,4
271
+ 13,10,1,2,14,22,46,631,73,4
272
+ 13,11,1,2,15,24,51,693,74,4
273
+ 13,12,1,2,16,26,56,757,75,4
274
+ 13,13,1,2,17,28,61,821,76,4
275
+ 13,14,1,2,18,30,66,885,77,4
276
+ 13,15,1,2,19,32,71,948,78,4
277
+ 13,16,1,2,20,34,76,1011,79,4
278
+ 13,17,1,2,21,36,81,1075,80,4
279
+ 13,18,1,2,22,38,86,1139,81,4
280
+ 13,19,1,2,23,40,91,1203,82,4
281
+ 13,20,1,2,24,42,96,1267,83,4
282
+ 14,1,5,5,5,5,5,63,64,4
283
+ 14,2,1,2,6,7,10,127,65,4
284
+ 14,3,1,2,7,9,15,190,66,4
285
+ 14,4,1,2,8,11,20,254,67,4
286
+ 14,5,1,2,9,13,25,318,68,4
287
+ 14,6,1,2,10,15,30,380,69,4
288
+ 14,7,1,2,11,17,35,443,70,4
289
+ 14,8,1,2,12,19,40,507,71,4
290
+ 14,9,1,2,13,21,45,566,72,4
291
+ 14,10,1,2,14,23,50,622,73,4
292
+ 14,11,1,2,15,25,55,685,74,4
293
+ 14,12,1,2,16,27,60,749,75,4
294
+ 14,13,1,2,17,29,65,813,76,4
295
+ 14,14,1,2,18,31,70,877,77,4
296
+ 14,15,1,2,19,33,75,934,78,4
297
+ 14,16,1,2,20,35,80,998,79,4
298
+ 14,17,1,2,21,37,85,1062,80,4
299
+ 14,18,1,2,22,39,90,1126,81,4
300
+ 14,19,1,2,23,41,95,1190,82,4
301
+ 14,20,1,2,24,43,100,1254,83,4
302
+ 15,1,5,5,5,5,5,63,64,4
303
+ 15,2,1,2,6,7,8,127,65,4
304
+ 15,3,1,2,7,9,12,191,66,4
305
+ 15,4,1,2,8,11,16,254,67,4
306
+ 15,5,1,2,9,13,19,310,68,4
307
+ 15,6,1,1,10,14,20,374,69,4
308
+ 15,7,1,2,11,16,23,437,70,4
309
+ 15,8,1,2,12,18,27,501,71,4
310
+ 15,9,1,2,13,20,32,562,72,4
311
+ 15,10,1,2,14,22,35,626,73,4
312
+ 15,11,1,2,15,24,38,689,74,4
313
+ 15,12,1,2,16,26,41,751,75,4
314
+ 15,13,1,2,17,28,45,815,76,4
315
+ 15,14,1,2,18,30,47,878,77,4
316
+ 15,15,1,2,19,32,51,939,78,4
317
+ 15,16,1,2,20,34,54,1002,79,4
318
+ 15,17,1,2,21,36,59,1065,80,4
319
+ 15,18,1,2,22,38,64,1129,81,4
320
+ 15,19,1,2,23,40,67,1193,82,4
321
+ 15,20,1,2,24,42,70,1255,83,4
322
+ 16,1,3,3,3,3,3,64,64,2
323
+ 16,2,2,3,5,6,6,128,65,3
324
+ 16,3,2,3,7,9,10,191,66,4
325
+ 16,4,1,2,8,11,14,255,67,4
326
+ 16,5,1,2,9,13,16,317,68,4
327
+ 16,6,1,2,10,15,18,381,69,4
328
+ 16,7,1,2,11,17,21,442,70,4
329
+ 16,8,1,2,12,19,25,506,71,4
330
+ 16,9,1,2,13,21,29,568,72,4
331
+ 16,10,1,2,14,23,33,632,73,4
332
+ 16,11,1,2,15,25,37,696,74,4
333
+ 16,12,1,2,16,27,39,760,75,4
334
+ 16,13,1,2,17,29,41,823,76,4
335
+ 16,14,1,2,18,31,43,884,77,4
336
+ 16,15,1,2,19,33,47,947,78,4
337
+ 16,16,1,2,20,35,49,1009,79,4
338
+ 16,17,1,1,21,36,50,1072,80,4
339
+ 16,18,1,2,22,38,53,1136,81,4
340
+ 16,19,1,2,23,40,57,1200,82,4
341
+ 16,20,1,2,24,42,60,1264,83,4
342
+ 17,1,4,4,4,4,4,58,64,3
343
+ 17,2,1,1,5,5,5,122,65,3
344
+ 17,3,2,3,7,8,9,186,66,4
345
+ 17,4,1,2,8,10,12,249,67,4
346
+ 17,5,1,2,9,12,16,312,68,4
347
+ 17,6,1,2,10,14,20,375,69,4
348
+ 17,7,1,2,11,16,24,438,70,4
349
+ 17,8,1,2,12,18,27,502,71,4
350
+ 17,9,1,2,13,20,31,566,72,4
351
+ 17,10,1,1,14,21,32,630,73,4
352
+ 17,11,1,2,15,23,36,692,74,4
353
+ 17,12,1,2,16,25,40,755,75,4
354
+ 17,13,1,2,17,27,44,818,76,4
355
+ 17,14,1,2,18,29,48,882,77,4
356
+ 17,15,1,1,19,30,49,945,78,4
357
+ 17,16,1,2,20,32,52,1009,79,4
358
+ 17,17,1,2,21,34,55,1072,80,4
359
+ 17,18,1,2,22,36,59,1134,81,4
360
+ 17,19,1,2,23,38,63,1198,82,4
361
+ 17,20,1,2,24,40,67,1261,83,4
362
+ 18,1,5,5,5,5,5,64,64,4
363
+ 18,2,1,2,6,7,10,124,65,4
364
+ 18,3,1,2,7,9,15,188,66,4
365
+ 18,4,1,2,8,11,20,251,67,4
366
+ 18,5,1,2,9,13,24,314,68,4
367
+ 18,6,1,2,10,15,29,378,69,4
368
+ 18,7,1,2,11,17,34,441,70,4
369
+ 18,8,1,2,12,19,38,505,71,4
370
+ 18,9,1,2,13,21,42,569,72,4
371
+ 18,10,1,2,14,23,46,633,73,4
372
+ 18,11,1,2,15,25,48,695,74,4
373
+ 18,12,1,2,16,27,52,759,75,4
374
+ 18,13,1,2,17,29,54,821,76,4
375
+ 18,14,1,2,18,31,58,885,77,4
376
+ 18,15,1,2,19,33,63,946,78,4
377
+ 18,16,1,2,20,35,65,1010,79,4
378
+ 18,17,1,2,21,37,70,1071,80,4
379
+ 18,18,1,2,22,39,74,1133,81,4
380
+ 18,19,1,2,23,41,79,1197,82,4
381
+ 18,20,1,2,24,43,84,1260,83,4
382
+ 19,1,4,4,4,4,4,58,64,3
383
+ 19,2,1,2,5,6,8,122,65,3
384
+ 19,3,2,3,7,9,13,185,66,4
385
+ 19,4,1,2,8,11,18,248,67,4
386
+ 19,5,1,2,9,13,23,309,68,4
387
+ 19,6,1,2,10,15,26,373,69,4
388
+ 19,7,1,2,11,17,31,436,70,4
389
+ 19,8,1,2,12,19,36,500,71,4
390
+ 19,9,1,2,13,21,41,561,72,4
391
+ 19,10,1,2,14,23,46,623,73,4
392
+ 19,11,1,2,15,25,50,685,74,4
393
+ 19,12,1,2,16,27,53,747,75,4
394
+ 19,13,1,2,17,29,58,810,76,4
395
+ 19,14,1,2,18,31,62,874,77,4
396
+ 19,15,1,2,19,33,67,938,78,4
397
+ 19,16,1,2,20,35,72,999,79,4
398
+ 19,17,1,2,21,37,75,1063,80,4
399
+ 19,18,1,2,22,39,78,1126,81,4
400
+ 19,19,1,2,23,41,83,1187,82,4
401
+ 19,20,1,2,24,43,86,1251,83,4
402
+ 20,1,2,2,2,2,2,64,64,1
403
+ 20,2,4,5,6,7,7,128,65,4
404
+ 20,3,1,2,7,9,11,191,66,4
405
+ 20,4,1,2,8,11,15,254,67,4
406
+ 20,5,1,2,9,13,19,317,68,4
407
+ 20,6,1,2,10,15,24,381,69,4
408
+ 20,7,1,2,11,17,29,445,70,4
409
+ 20,8,1,2,12,19,33,507,71,4
410
+ 20,9,1,1,13,20,34,570,72,4
411
+ 20,10,1,2,14,22,38,634,73,4
412
+ 20,11,1,2,15,24,42,697,74,4
413
+ 20,12,1,2,16,26,47,759,75,4
414
+ 20,13,1,2,17,28,51,823,76,4
415
+ 20,14,1,2,18,30,55,886,77,4
416
+ 20,15,1,2,19,32,60,948,78,4
417
+ 20,16,1,2,20,34,62,1012,79,4
418
+ 20,17,1,2,21,36,67,1075,80,4
419
+ 20,18,1,2,22,38,72,1139,81,4
420
+ 20,19,1,2,23,40,77,1202,82,4
421
+ 20,20,1,2,24,42,81,1266,83,4
422
+ 21,1,5,5,5,5,5,60,64,4
423
+ 21,2,1,2,6,7,10,124,65,4
424
+ 21,3,1,2,7,9,14,187,66,4
425
+ 21,4,1,2,8,11,19,251,67,4
426
+ 21,5,1,2,9,13,24,315,68,4
427
+ 21,6,1,2,10,15,28,378,69,4
428
+ 21,7,1,2,11,17,32,442,70,4
429
+ 21,8,1,2,12,19,36,505,71,4
430
+ 21,9,1,2,13,21,41,566,72,4
431
+ 21,10,1,2,14,23,46,629,73,4
432
+ 21,11,1,2,15,25,51,693,74,4
433
+ 21,12,1,2,16,27,55,757,75,4
434
+ 21,13,1,2,17,29,60,821,76,4
435
+ 21,14,1,2,18,31,65,884,77,4
436
+ 21,15,1,2,19,33,69,948,78,4
437
+ 21,16,1,2,20,35,74,1012,79,4
438
+ 21,17,1,2,21,37,79,1075,80,4
439
+ 21,18,1,2,22,39,84,1139,81,4
440
+ 21,19,1,2,23,41,88,1203,82,4
441
+ 21,20,1,2,24,43,92,1262,83,4
442
+ 22,1,3,3,3,3,3,64,64,2
443
+ 22,2,1,2,4,5,5,128,65,2
444
+ 22,3,3,4,7,9,10,191,66,4
445
+ 22,4,1,2,8,11,15,254,67,4
446
+ 22,5,1,2,9,13,20,318,68,4
447
+ 22,6,1,2,10,15,25,381,69,4
448
+ 22,7,1,2,11,17,30,443,70,4
449
+ 22,8,1,2,12,19,35,506,71,4
450
+ 22,9,1,2,13,21,38,569,72,4
451
+ 22,10,1,2,14,23,43,631,73,4
452
+ 22,11,1,2,15,25,48,695,74,4
453
+ 22,12,1,2,16,27,53,759,75,4
454
+ 22,13,1,2,17,29,58,823,76,4
455
+ 22,14,1,2,18,31,63,887,77,4
456
+ 22,15,1,2,19,33,66,951,78,4
457
+ 22,16,1,2,20,35,71,1015,79,4
458
+ 22,17,1,2,21,37,75,1079,80,4
459
+ 22,18,1,1,22,38,76,1142,81,4
460
+ 22,19,1,2,23,40,80,1205,82,4
461
+ 22,20,1,2,24,42,84,1269,83,4
462
+ 23,1,5,5,5,5,5,62,64,4
463
+ 23,2,1,2,6,7,9,126,65,4
464
+ 23,3,1,2,7,9,13,190,66,4
465
+ 23,4,1,1,8,10,14,254,67,4
466
+ 23,5,1,2,9,12,19,318,68,4
467
+ 23,6,1,2,10,14,23,382,69,4
468
+ 23,7,1,2,11,16,27,446,70,4
469
+ 23,8,1,2,12,18,32,509,71,4
470
+ 23,9,1,2,13,20,36,572,72,4
471
+ 23,10,1,2,14,22,38,636,73,4
472
+ 23,11,1,2,15,24,43,698,74,4
473
+ 23,12,1,2,16,26,48,761,75,4
474
+ 23,13,1,2,17,28,53,825,76,4
475
+ 23,14,1,1,18,29,54,889,77,4
476
+ 23,15,1,2,19,31,58,953,78,4
477
+ 23,16,1,2,20,33,62,1016,79,4
478
+ 23,17,1,2,21,35,66,1079,80,4
479
+ 23,18,1,2,22,37,71,1143,81,4
480
+ 23,19,1,2,23,39,75,1205,82,4
481
+ 23,20,1,2,24,41,80,1267,83,4
data/continuous_crossover.csv ADDED
The diff for this file is too large to render. See raw diff
 
data/continuous_recipient.csv ADDED
The diff for this file is too large to render. See raw diff
 
data/noisy_real_repairs.csv ADDED
The diff for this file is too large to render. See raw diff
 
data/program_search.csv ADDED
The diff for this file is too large to render. See raw diff
 
data/stationary_forward.csv ADDED
@@ -0,0 +1,481 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ world,task,actual_rank,closure_dimension,transferred_cumulative_calls,source_only_per_task_cumulative_calls,cold_shared_cumulative_calls,cold_per_task_cumulative_calls
2
+ 0,1,4,3,4,4,64,64
3
+ 0,2,4,4,6,9,64,127
4
+ 0,3,4,4,7,14,64,190
5
+ 0,4,4,4,8,19,64,253
6
+ 0,5,4,4,9,24,64,316
7
+ 0,6,4,4,10,29,64,378
8
+ 0,7,4,4,11,33,64,441
9
+ 0,8,4,4,12,38,64,502
10
+ 0,9,4,4,13,41,64,564
11
+ 0,10,4,4,14,46,64,628
12
+ 0,11,4,4,15,51,64,691
13
+ 0,12,4,4,16,55,64,755
14
+ 0,13,4,4,17,58,64,818
15
+ 0,14,4,4,18,63,64,881
16
+ 0,15,4,4,19,66,64,945
17
+ 0,16,4,4,20,69,64,1009
18
+ 0,17,4,4,21,74,64,1073
19
+ 0,18,4,4,22,77,64,1137
20
+ 0,19,4,4,23,82,64,1201
21
+ 0,20,4,4,24,87,64,1264
22
+ 1,1,4,2,3,3,63,63
23
+ 1,2,4,3,5,7,64,127
24
+ 1,3,4,4,7,11,64,191
25
+ 1,4,4,4,8,12,64,254
26
+ 1,5,4,4,9,14,64,317
27
+ 1,6,4,4,10,16,64,381
28
+ 1,7,4,4,11,18,64,445
29
+ 1,8,4,4,12,22,64,509
30
+ 1,9,4,4,13,26,64,571
31
+ 1,10,4,4,14,30,64,635
32
+ 1,11,4,4,15,33,64,699
33
+ 1,12,4,4,16,37,64,762
34
+ 1,13,4,4,17,39,64,824
35
+ 1,14,4,4,18,42,64,884
36
+ 1,15,4,4,19,46,64,948
37
+ 1,16,4,4,20,49,64,1009
38
+ 1,17,4,4,21,51,64,1073
39
+ 1,18,4,4,22,55,64,1137
40
+ 1,19,4,4,23,56,64,1200
41
+ 1,20,4,4,24,58,64,1264
42
+ 2,1,4,3,4,4,64,64
43
+ 2,2,4,3,5,8,64,126
44
+ 2,3,4,4,7,13,64,190
45
+ 2,4,4,4,8,18,64,254
46
+ 2,5,4,4,9,22,64,317
47
+ 2,6,4,4,10,27,64,379
48
+ 2,7,4,4,11,32,64,443
49
+ 2,8,4,4,12,36,64,507
50
+ 2,9,4,4,13,41,64,571
51
+ 2,10,4,4,14,46,64,635
52
+ 2,11,4,4,15,50,64,698
53
+ 2,12,4,4,16,55,64,762
54
+ 2,13,4,4,17,59,64,826
55
+ 2,14,4,4,18,60,64,888
56
+ 2,15,4,4,19,65,64,951
57
+ 2,16,4,4,20,66,64,1015
58
+ 2,17,4,4,21,71,64,1079
59
+ 2,18,4,4,22,76,64,1140
60
+ 2,19,4,4,23,81,64,1204
61
+ 2,20,4,4,24,86,64,1268
62
+ 3,1,4,3,4,4,62,62
63
+ 3,2,4,3,5,6,63,125
64
+ 3,3,4,4,7,10,63,188
65
+ 3,4,4,4,8,14,63,249
66
+ 3,5,4,4,9,18,64,313
67
+ 3,6,4,4,10,21,64,376
68
+ 3,7,4,4,11,24,64,440
69
+ 3,8,4,4,12,28,64,503
70
+ 3,9,4,4,13,32,64,566
71
+ 3,10,4,4,14,35,64,630
72
+ 3,11,4,4,15,38,64,694
73
+ 3,12,4,4,16,42,64,757
74
+ 3,13,4,4,17,45,64,817
75
+ 3,14,4,4,18,47,64,880
76
+ 3,15,4,4,19,48,64,943
77
+ 3,16,4,4,20,51,64,1007
78
+ 3,17,4,4,21,52,64,1071
79
+ 3,18,4,4,22,56,64,1135
80
+ 3,19,4,4,23,60,64,1198
81
+ 3,20,4,4,24,62,64,1262
82
+ 4,1,4,4,5,5,64,64
83
+ 4,2,4,4,6,10,64,128
84
+ 4,3,4,4,7,14,64,189
85
+ 4,4,4,4,8,19,64,253
86
+ 4,5,4,4,9,23,64,317
87
+ 4,6,4,4,10,27,64,381
88
+ 4,7,4,4,11,31,64,444
89
+ 4,8,4,4,12,35,64,508
90
+ 4,9,4,4,13,39,64,572
91
+ 4,10,4,4,14,43,64,634
92
+ 4,11,4,4,15,47,64,698
93
+ 4,12,4,4,16,51,64,762
94
+ 4,13,4,4,17,56,64,826
95
+ 4,14,4,4,18,61,64,890
96
+ 4,15,4,4,19,66,64,953
97
+ 4,16,4,4,20,70,64,1016
98
+ 4,17,4,4,21,74,64,1080
99
+ 4,18,4,4,22,78,64,1144
100
+ 4,19,4,4,23,82,64,1207
101
+ 4,20,4,4,24,87,64,1271
102
+ 5,1,4,4,5,5,64,64
103
+ 5,2,4,4,6,9,64,128
104
+ 5,3,4,4,7,13,64,191
105
+ 5,4,4,4,8,17,64,255
106
+ 5,5,4,4,9,22,64,319
107
+ 5,6,4,4,10,27,64,382
108
+ 5,7,4,4,11,32,64,446
109
+ 5,8,4,4,12,37,64,509
110
+ 5,9,4,4,13,42,64,573
111
+ 5,10,4,4,14,47,64,637
112
+ 5,11,4,4,15,51,64,701
113
+ 5,12,4,4,16,52,64,763
114
+ 5,13,4,4,17,56,64,824
115
+ 5,14,4,4,18,60,64,888
116
+ 5,15,4,4,19,65,64,952
117
+ 5,16,4,4,20,67,64,1016
118
+ 5,17,4,4,21,72,64,1080
119
+ 5,18,4,4,22,77,64,1144
120
+ 5,19,4,4,23,81,64,1208
121
+ 5,20,4,4,24,86,64,1270
122
+ 6,1,4,4,5,5,62,62
123
+ 6,2,4,4,6,10,63,125
124
+ 6,3,4,4,7,15,64,189
125
+ 6,4,4,4,8,20,64,252
126
+ 6,5,4,4,9,23,64,311
127
+ 6,6,4,4,10,24,64,373
128
+ 6,7,4,4,11,29,64,437
129
+ 6,8,4,4,12,32,64,501
130
+ 6,9,4,4,13,37,64,564
131
+ 6,10,4,4,14,42,64,628
132
+ 6,11,4,4,15,47,64,692
133
+ 6,12,4,4,16,52,64,756
134
+ 6,13,4,4,17,57,64,819
135
+ 6,14,4,4,18,62,64,883
136
+ 6,15,4,4,19,65,64,945
137
+ 6,16,4,4,20,70,64,1009
138
+ 6,17,4,4,21,75,64,1069
139
+ 6,18,4,4,22,80,64,1133
140
+ 6,19,4,4,23,81,64,1195
141
+ 6,20,4,4,24,84,64,1258
142
+ 7,1,4,4,5,5,63,63
143
+ 7,2,4,4,6,10,64,127
144
+ 7,3,4,4,7,15,64,191
145
+ 7,4,4,4,8,20,64,255
146
+ 7,5,4,4,9,25,64,319
147
+ 7,6,4,4,10,30,64,383
148
+ 7,7,4,4,11,35,64,445
149
+ 7,8,4,4,12,40,64,509
150
+ 7,9,4,4,13,45,64,571
151
+ 7,10,4,4,14,50,64,632
152
+ 7,11,4,4,15,55,64,693
153
+ 7,12,4,4,16,60,64,757
154
+ 7,13,4,4,17,65,64,820
155
+ 7,14,4,4,18,70,64,884
156
+ 7,15,4,4,19,75,64,945
157
+ 7,16,4,4,20,80,64,1009
158
+ 7,17,4,4,21,85,64,1073
159
+ 7,18,4,4,22,90,64,1135
160
+ 7,19,4,4,23,91,64,1199
161
+ 7,20,4,4,24,96,64,1263
162
+ 8,1,4,4,5,5,63,63
163
+ 8,2,4,4,6,10,63,124
164
+ 8,3,4,4,7,15,63,187
165
+ 8,4,4,4,8,16,63,250
166
+ 8,5,4,4,9,21,63,313
167
+ 8,6,4,4,10,26,64,377
168
+ 8,7,4,4,11,31,64,439
169
+ 8,8,4,4,12,36,64,500
170
+ 8,9,4,4,13,41,64,564
171
+ 8,10,4,4,14,46,64,628
172
+ 8,11,4,4,15,47,64,691
173
+ 8,12,4,4,16,48,64,753
174
+ 8,13,4,4,17,53,64,814
175
+ 8,14,4,4,18,58,64,877
176
+ 8,15,4,4,19,63,64,940
177
+ 8,16,4,4,20,68,64,1004
178
+ 8,17,4,4,21,69,64,1068
179
+ 8,18,4,4,22,70,64,1130
180
+ 8,19,4,4,23,75,64,1194
181
+ 8,20,4,4,24,80,64,1258
182
+ 9,1,4,4,5,5,64,64
183
+ 9,2,4,4,6,8,64,123
184
+ 9,3,4,4,7,13,64,186
185
+ 9,4,4,4,8,18,64,245
186
+ 9,5,4,4,9,23,64,304
187
+ 9,6,4,4,10,28,64,368
188
+ 9,7,4,4,11,33,64,431
189
+ 9,8,4,4,12,38,64,493
190
+ 9,9,4,4,13,43,64,557
191
+ 9,10,4,4,14,46,64,621
192
+ 9,11,4,4,15,49,64,683
193
+ 9,12,4,4,16,54,64,747
194
+ 9,13,4,4,17,59,64,810
195
+ 9,14,4,4,18,64,64,874
196
+ 9,15,4,4,19,69,64,937
197
+ 9,16,4,4,20,74,64,998
198
+ 9,17,4,4,21,77,64,1061
199
+ 9,18,4,4,22,82,64,1125
200
+ 9,19,4,4,23,85,64,1189
201
+ 9,20,4,4,24,88,64,1253
202
+ 10,1,4,3,4,4,64,64
203
+ 10,2,4,3,5,5,64,128
204
+ 10,3,4,4,7,10,64,192
205
+ 10,4,4,4,8,14,64,255
206
+ 10,5,4,4,9,18,64,318
207
+ 10,6,4,4,10,23,64,382
208
+ 10,7,4,4,11,28,64,446
209
+ 10,8,4,4,12,33,64,509
210
+ 10,9,4,4,13,38,64,573
211
+ 10,10,4,4,14,43,64,636
212
+ 10,11,4,4,15,48,64,700
213
+ 10,12,4,4,16,53,64,764
214
+ 10,13,4,4,17,55,64,827
215
+ 10,14,4,4,18,59,64,890
216
+ 10,15,4,4,19,64,64,954
217
+ 10,16,4,4,20,68,64,1018
218
+ 10,17,4,4,21,72,64,1081
219
+ 10,18,4,4,22,77,64,1145
220
+ 10,19,4,4,23,82,64,1205
221
+ 10,20,4,4,24,86,64,1268
222
+ 11,1,4,4,5,5,64,64
223
+ 11,2,4,4,6,10,64,125
224
+ 11,3,4,4,7,14,64,188
225
+ 11,4,4,4,8,18,64,251
226
+ 11,5,4,4,9,23,64,315
227
+ 11,6,4,4,10,28,64,379
228
+ 11,7,4,4,11,33,64,442
229
+ 11,8,4,4,12,35,64,505
230
+ 11,9,4,4,13,40,64,569
231
+ 11,10,4,4,14,44,64,632
232
+ 11,11,4,4,15,49,64,696
233
+ 11,12,4,4,16,53,64,758
234
+ 11,13,4,4,17,54,64,822
235
+ 11,14,4,4,18,59,64,886
236
+ 11,15,4,4,19,64,64,950
237
+ 11,16,4,4,20,69,64,1011
238
+ 11,17,4,4,21,71,64,1075
239
+ 11,18,4,4,22,76,64,1139
240
+ 11,19,4,4,23,80,64,1202
241
+ 11,20,4,4,24,85,64,1262
242
+ 12,1,4,4,5,5,63,63
243
+ 12,2,4,4,6,6,63,126
244
+ 12,3,4,4,7,11,64,190
245
+ 12,4,4,4,8,15,64,246
246
+ 12,5,4,4,9,19,64,310
247
+ 12,6,4,4,10,23,64,373
248
+ 12,7,4,4,11,28,64,436
249
+ 12,8,4,4,12,32,64,500
250
+ 12,9,4,4,13,37,64,564
251
+ 12,10,4,4,14,41,64,628
252
+ 12,11,4,4,15,45,64,692
253
+ 12,12,4,4,16,50,64,754
254
+ 12,13,4,4,17,55,64,815
255
+ 12,14,4,4,18,60,64,878
256
+ 12,15,4,4,19,65,64,942
257
+ 12,16,4,4,20,69,64,1004
258
+ 12,17,4,4,21,74,64,1068
259
+ 12,18,4,4,22,78,64,1132
260
+ 12,19,4,4,23,83,64,1196
261
+ 12,20,4,4,24,88,64,1258
262
+ 13,1,4,3,4,4,64,64
263
+ 13,2,4,4,6,8,64,128
264
+ 13,3,4,4,7,12,64,192
265
+ 13,4,4,4,8,14,64,256
266
+ 13,5,4,4,9,17,64,320
267
+ 13,6,4,4,10,20,64,384
268
+ 13,7,4,4,11,24,64,447
269
+ 13,8,4,4,12,28,64,511
270
+ 13,9,4,4,13,32,64,575
271
+ 13,10,4,4,14,35,64,639
272
+ 13,11,4,4,15,39,64,702
273
+ 13,12,4,4,16,43,64,765
274
+ 13,13,4,4,17,44,64,829
275
+ 13,14,4,4,18,47,64,892
276
+ 13,15,4,4,19,51,64,956
277
+ 13,16,4,4,20,53,64,1019
278
+ 13,17,4,4,21,57,64,1083
279
+ 13,18,4,4,22,61,64,1146
280
+ 13,19,4,4,23,65,64,1207
281
+ 13,20,4,4,24,69,64,1271
282
+ 14,1,4,4,5,5,63,63
283
+ 14,2,4,4,6,10,63,121
284
+ 14,3,4,4,7,15,63,184
285
+ 14,4,4,4,8,20,64,248
286
+ 14,5,4,4,9,25,64,310
287
+ 14,6,4,4,10,30,64,373
288
+ 14,7,4,4,11,35,64,437
289
+ 14,8,4,4,12,40,64,501
290
+ 14,9,4,4,13,45,64,565
291
+ 14,10,4,4,14,50,64,628
292
+ 14,11,4,4,15,55,64,691
293
+ 14,12,4,4,16,60,64,754
294
+ 14,13,4,4,17,65,64,817
295
+ 14,14,4,4,18,70,64,881
296
+ 14,15,4,4,19,75,64,944
297
+ 14,16,4,4,20,80,64,1006
298
+ 14,17,4,4,21,85,64,1068
299
+ 14,18,4,4,22,90,64,1132
300
+ 14,19,4,4,23,95,64,1196
301
+ 14,20,4,4,24,100,64,1258
302
+ 15,1,4,4,5,5,64,64
303
+ 15,2,4,4,6,8,64,126
304
+ 15,3,4,4,7,13,64,189
305
+ 15,4,4,4,8,18,64,253
306
+ 15,5,4,4,9,23,64,316
307
+ 15,6,4,4,10,28,64,379
308
+ 15,7,4,4,11,31,64,442
309
+ 15,8,4,4,12,34,64,503
310
+ 15,9,4,4,13,39,64,564
311
+ 15,10,4,4,14,44,64,628
312
+ 15,11,4,4,15,49,64,692
313
+ 15,12,4,4,16,54,64,756
314
+ 15,13,4,4,17,59,64,819
315
+ 15,14,4,4,18,64,64,883
316
+ 15,15,4,4,19,69,64,945
317
+ 15,16,4,4,20,74,64,1009
318
+ 15,17,4,4,21,79,64,1069
319
+ 15,18,4,4,22,84,64,1133
320
+ 15,19,4,4,23,89,64,1195
321
+ 15,20,4,4,24,94,64,1259
322
+ 16,1,4,2,3,3,63,63
323
+ 16,2,4,4,6,8,64,127
324
+ 16,3,4,4,7,9,64,191
325
+ 16,4,4,4,8,13,64,253
326
+ 16,5,4,4,9,17,64,317
327
+ 16,6,4,4,10,18,64,381
328
+ 16,7,4,4,11,23,64,445
329
+ 16,8,4,4,12,28,64,509
330
+ 16,9,4,4,13,31,64,573
331
+ 16,10,4,4,14,36,64,635
332
+ 16,11,4,4,15,41,64,698
333
+ 16,12,4,4,16,44,64,762
334
+ 16,13,4,4,17,47,64,825
335
+ 16,14,4,4,18,52,64,887
336
+ 16,15,4,4,19,57,64,946
337
+ 16,16,4,4,20,60,64,1010
338
+ 16,17,4,4,21,65,64,1074
339
+ 16,18,4,4,22,69,64,1138
340
+ 16,19,4,4,23,71,64,1198
341
+ 16,20,4,4,24,75,64,1261
342
+ 17,1,4,4,5,5,64,64
343
+ 17,2,4,4,6,6,64,127
344
+ 17,3,4,4,7,11,64,188
345
+ 17,4,4,4,8,16,64,251
346
+ 17,5,4,4,9,21,64,314
347
+ 17,6,4,4,10,26,64,378
348
+ 17,7,4,4,11,31,64,442
349
+ 17,8,4,4,12,36,64,504
350
+ 17,9,4,4,13,41,64,568
351
+ 17,10,4,4,14,46,64,629
352
+ 17,11,4,4,15,51,64,693
353
+ 17,12,4,4,16,56,64,757
354
+ 17,13,4,4,17,61,64,820
355
+ 17,14,4,4,18,66,64,884
356
+ 17,15,4,4,19,71,64,948
357
+ 17,16,4,4,20,76,64,1011
358
+ 17,17,4,4,21,81,64,1073
359
+ 17,18,4,4,22,86,64,1137
360
+ 17,19,4,4,23,91,64,1200
361
+ 17,20,4,4,24,96,64,1263
362
+ 18,1,4,0,1,1,64,64
363
+ 18,2,4,3,5,5,64,128
364
+ 18,3,4,3,6,9,64,192
365
+ 18,4,4,3,7,13,64,255
366
+ 18,5,4,4,9,18,64,316
367
+ 18,6,4,4,10,22,64,380
368
+ 18,7,4,4,11,27,64,444
369
+ 18,8,4,4,12,32,64,508
370
+ 18,9,4,4,13,36,64,572
371
+ 18,10,4,4,14,40,64,635
372
+ 18,11,4,4,15,44,64,693
373
+ 18,12,4,4,16,46,64,756
374
+ 18,13,4,4,17,51,64,820
375
+ 18,14,4,4,18,55,64,884
376
+ 18,15,4,4,19,60,64,948
377
+ 18,16,4,4,20,65,64,1012
378
+ 18,17,4,4,21,69,64,1076
379
+ 18,18,4,4,22,70,64,1138
380
+ 18,19,4,4,23,74,64,1201
381
+ 18,20,4,4,24,78,64,1265
382
+ 19,1,4,4,5,5,64,64
383
+ 19,2,4,4,6,10,64,127
384
+ 19,3,4,4,7,15,64,191
385
+ 19,4,4,4,8,20,64,255
386
+ 19,5,4,4,9,25,64,316
387
+ 19,6,4,4,10,30,64,380
388
+ 19,7,4,4,11,35,64,444
389
+ 19,8,4,4,12,40,64,508
390
+ 19,9,4,4,13,45,64,571
391
+ 19,10,4,4,14,50,64,632
392
+ 19,11,4,4,15,55,64,696
393
+ 19,12,4,4,16,60,64,760
394
+ 19,13,4,4,17,65,64,823
395
+ 19,14,4,4,18,70,64,885
396
+ 19,15,4,4,19,75,64,945
397
+ 19,16,4,4,20,80,64,1006
398
+ 19,17,4,4,21,85,64,1069
399
+ 19,18,4,4,22,90,64,1132
400
+ 19,19,4,4,23,95,64,1195
401
+ 19,20,4,4,24,100,64,1258
402
+ 20,1,4,2,3,3,64,64
403
+ 20,2,4,3,5,7,64,125
404
+ 20,3,4,4,7,9,64,186
405
+ 20,4,4,4,8,13,64,246
406
+ 20,5,4,4,9,16,64,309
407
+ 20,6,4,4,10,20,64,373
408
+ 20,7,4,4,11,24,64,437
409
+ 20,8,4,4,12,28,64,495
410
+ 20,9,4,4,13,32,64,558
411
+ 20,10,4,4,14,34,64,622
412
+ 20,11,4,4,15,37,64,686
413
+ 20,12,4,4,16,41,64,749
414
+ 20,13,4,4,17,45,64,812
415
+ 20,14,4,4,18,46,64,875
416
+ 20,15,4,4,19,50,64,939
417
+ 20,16,4,4,20,52,64,1003
418
+ 20,17,4,4,21,54,64,1067
419
+ 20,18,4,4,22,58,64,1131
420
+ 20,19,4,4,23,61,64,1194
421
+ 20,20,4,4,24,64,64,1256
422
+ 21,1,4,4,5,5,64,64
423
+ 21,2,4,4,6,9,64,126
424
+ 21,3,4,4,7,13,64,189
425
+ 21,4,4,4,8,18,64,252
426
+ 21,5,4,4,9,23,64,310
427
+ 21,6,4,4,10,28,64,374
428
+ 21,7,4,4,11,32,64,437
429
+ 21,8,4,4,12,37,64,500
430
+ 21,9,4,4,13,42,64,564
431
+ 21,10,4,4,14,46,64,628
432
+ 21,11,4,4,15,51,64,692
433
+ 21,12,4,4,16,53,64,756
434
+ 21,13,4,4,17,57,64,820
435
+ 21,14,4,4,18,62,64,883
436
+ 21,15,4,4,19,66,64,947
437
+ 21,16,4,4,20,71,64,1008
438
+ 21,17,4,4,21,75,64,1072
439
+ 21,18,4,4,22,79,64,1136
440
+ 21,19,4,4,23,83,64,1199
441
+ 21,20,4,4,24,87,64,1263
442
+ 22,1,4,4,5,5,63,63
443
+ 22,2,4,4,6,10,64,127
444
+ 22,3,4,4,7,13,64,190
445
+ 22,4,4,4,8,18,64,249
446
+ 22,5,4,4,9,23,64,311
447
+ 22,6,4,4,10,28,64,374
448
+ 22,7,4,4,11,29,64,438
449
+ 22,8,4,4,12,30,64,501
450
+ 22,9,4,4,13,35,64,563
451
+ 22,10,4,4,14,40,64,627
452
+ 22,11,4,4,15,45,64,686
453
+ 22,12,4,4,16,48,64,750
454
+ 22,13,4,4,17,53,64,813
455
+ 22,14,4,4,18,58,64,877
456
+ 22,15,4,4,19,61,64,941
457
+ 22,16,4,4,20,66,64,1001
458
+ 22,17,4,4,21,71,64,1064
459
+ 22,18,4,4,22,72,64,1128
460
+ 22,19,4,4,23,73,64,1192
461
+ 22,20,4,4,24,78,64,1255
462
+ 23,1,4,3,4,4,64,64
463
+ 23,2,4,3,5,8,64,128
464
+ 23,3,4,4,7,13,64,192
465
+ 23,4,4,4,8,16,64,254
466
+ 23,5,4,4,9,19,64,318
467
+ 23,6,4,4,10,24,64,381
468
+ 23,7,4,4,11,27,64,445
469
+ 23,8,4,4,12,31,64,509
470
+ 23,9,4,4,13,34,64,571
471
+ 23,10,4,4,14,39,64,635
472
+ 23,11,4,4,15,42,64,698
473
+ 23,12,4,4,16,46,64,762
474
+ 23,13,4,4,17,50,64,823
475
+ 23,14,4,4,18,54,64,884
476
+ 23,15,4,4,19,58,64,948
477
+ 23,16,4,4,20,63,64,1012
478
+ 23,17,4,4,21,66,64,1076
479
+ 23,18,4,4,22,69,64,1136
480
+ 23,19,4,4,23,74,64,1200
481
+ 23,20,4,4,24,76,64,1264
data/writer_hierarchy.csv ADDED
@@ -0,0 +1,289 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ depth,world,mode,stream,goals,misses_by_level,rank_budget_by_level,forward,adjoint,physical_probes,non_top_applies,trusted_top_writes,upper_state_resets,fresh_last14_probes,first14_probes,trained_upper_fresh_recipient_probes,virgin_upper_fresh_recipient_probes
2
+ 1,0,persistent,adaptive,48,[6],[6],60,6,66,0,6,0,17,26,4.0,4.0
3
+ 1,0,fixed_refactor,adaptive,48,[6],[6],60,6,66,0,6,0,17,26,4.0,4.0
4
+ 1,0,cold_upper,adaptive,48,[6],[6],60,6,66,0,6,0,17,26,4.0,4.0
5
+ 1,0,persistent,adversarial_then_fresh,48,[6],[6],60,6,66,0,6,0,14,26,4.0,4.0
6
+ 1,0,fixed_refactor,adversarial_then_fresh,48,[6],[6],60,6,66,0,6,0,14,26,4.0,4.0
7
+ 1,0,cold_upper,adversarial_then_fresh,48,[6],[6],60,6,66,0,6,0,14,26,4.0,4.0
8
+ 1,1,persistent,adaptive,48,[6],[6],60,6,66,0,6,0,14,26,4.0,4.0
9
+ 1,1,fixed_refactor,adaptive,48,[6],[6],60,6,66,0,6,0,14,26,4.0,4.0
10
+ 1,1,cold_upper,adaptive,48,[6],[6],60,6,66,0,6,0,14,26,4.0,4.0
11
+ 1,1,persistent,adversarial_then_fresh,48,[6],[6],60,6,66,0,6,0,14,26,4.0,4.0
12
+ 1,1,fixed_refactor,adversarial_then_fresh,48,[6],[6],60,6,66,0,6,0,14,26,4.0,4.0
13
+ 1,1,cold_upper,adversarial_then_fresh,48,[6],[6],60,6,66,0,6,0,14,26,4.0,4.0
14
+ 1,2,persistent,adaptive,48,[6],[6],60,6,66,0,6,0,14,26,4.0,4.0
15
+ 1,2,fixed_refactor,adaptive,48,[6],[6],60,6,66,0,6,0,14,26,4.0,4.0
16
+ 1,2,cold_upper,adaptive,48,[6],[6],60,6,66,0,6,0,14,26,4.0,4.0
17
+ 1,2,persistent,adversarial_then_fresh,48,[6],[6],60,6,66,0,6,0,14,26,4.0,4.0
18
+ 1,2,fixed_refactor,adversarial_then_fresh,48,[6],[6],60,6,66,0,6,0,14,26,4.0,4.0
19
+ 1,2,cold_upper,adversarial_then_fresh,48,[6],[6],60,6,66,0,6,0,14,26,4.0,4.0
20
+ 1,3,persistent,adaptive,48,[6],[6],60,6,66,0,6,0,14,26,4.0,4.0
21
+ 1,3,fixed_refactor,adaptive,48,[6],[6],60,6,66,0,6,0,14,26,4.0,4.0
22
+ 1,3,cold_upper,adaptive,48,[6],[6],60,6,66,0,6,0,14,26,4.0,4.0
23
+ 1,3,persistent,adversarial_then_fresh,48,[6],[6],60,6,66,0,6,0,14,26,4.0,4.0
24
+ 1,3,fixed_refactor,adversarial_then_fresh,48,[6],[6],60,6,66,0,6,0,14,26,4.0,4.0
25
+ 1,3,cold_upper,adversarial_then_fresh,48,[6],[6],60,6,66,0,6,0,14,26,4.0,4.0
26
+ 1,4,persistent,adaptive,48,[6],[6],60,6,66,0,6,0,14,26,4.0,4.0
27
+ 1,4,fixed_refactor,adaptive,48,[6],[6],60,6,66,0,6,0,14,26,4.0,4.0
28
+ 1,4,cold_upper,adaptive,48,[6],[6],60,6,66,0,6,0,14,26,4.0,4.0
29
+ 1,4,persistent,adversarial_then_fresh,48,[6],[6],60,6,66,0,6,0,14,26,4.0,4.0
30
+ 1,4,fixed_refactor,adversarial_then_fresh,48,[6],[6],60,6,66,0,6,0,14,26,4.0,4.0
31
+ 1,4,cold_upper,adversarial_then_fresh,48,[6],[6],60,6,66,0,6,0,14,26,4.0,4.0
32
+ 1,5,persistent,adaptive,48,[6],[6],60,6,66,0,6,0,14,26,4.0,4.0
33
+ 1,5,fixed_refactor,adaptive,48,[6],[6],60,6,66,0,6,0,14,26,4.0,4.0
34
+ 1,5,cold_upper,adaptive,48,[6],[6],60,6,66,0,6,0,14,26,4.0,4.0
35
+ 1,5,persistent,adversarial_then_fresh,48,[6],[6],60,6,66,0,6,0,14,26,4.0,4.0
36
+ 1,5,fixed_refactor,adversarial_then_fresh,48,[6],[6],60,6,66,0,6,0,14,26,4.0,4.0
37
+ 1,5,cold_upper,adversarial_then_fresh,48,[6],[6],60,6,66,0,6,0,14,26,4.0,4.0
38
+ 1,6,persistent,adaptive,48,[6],[6],60,6,66,0,6,0,14,23,4.0,4.0
39
+ 1,6,fixed_refactor,adaptive,48,[6],[6],60,6,66,0,6,0,14,23,4.0,4.0
40
+ 1,6,cold_upper,adaptive,48,[6],[6],60,6,66,0,6,0,14,23,4.0,4.0
41
+ 1,6,persistent,adversarial_then_fresh,48,[6],[6],60,6,66,0,6,0,14,26,4.0,4.0
42
+ 1,6,fixed_refactor,adversarial_then_fresh,48,[6],[6],60,6,66,0,6,0,14,26,4.0,4.0
43
+ 1,6,cold_upper,adversarial_then_fresh,48,[6],[6],60,6,66,0,6,0,14,26,4.0,4.0
44
+ 1,7,persistent,adaptive,48,[6],[6],60,6,66,0,6,0,14,26,4.0,4.0
45
+ 1,7,fixed_refactor,adaptive,48,[6],[6],60,6,66,0,6,0,14,26,4.0,4.0
46
+ 1,7,cold_upper,adaptive,48,[6],[6],60,6,66,0,6,0,14,26,4.0,4.0
47
+ 1,7,persistent,adversarial_then_fresh,48,[6],[6],60,6,66,0,6,0,14,26,4.0,4.0
48
+ 1,7,fixed_refactor,adversarial_then_fresh,48,[6],[6],60,6,66,0,6,0,14,26,4.0,4.0
49
+ 1,7,cold_upper,adversarial_then_fresh,48,[6],[6],60,6,66,0,6,0,14,26,4.0,4.0
50
+ 1,8,persistent,adaptive,48,[5],[5],58,5,63,0,5,0,14,23,4.0,4.0
51
+ 1,8,fixed_refactor,adaptive,48,[5],[5],58,5,63,0,5,0,14,23,4.0,4.0
52
+ 1,8,cold_upper,adaptive,48,[5],[5],58,5,63,0,5,0,14,23,4.0,4.0
53
+ 1,8,persistent,adversarial_then_fresh,48,[5],[5],58,5,63,0,5,0,14,23,4.0,4.0
54
+ 1,8,fixed_refactor,adversarial_then_fresh,48,[5],[5],58,5,63,0,5,0,14,23,4.0,4.0
55
+ 1,8,cold_upper,adversarial_then_fresh,48,[5],[5],58,5,63,0,5,0,14,23,4.0,4.0
56
+ 1,9,persistent,adaptive,48,[6],[6],60,6,66,0,6,0,14,26,4.0,4.0
57
+ 1,9,fixed_refactor,adaptive,48,[6],[6],60,6,66,0,6,0,14,26,4.0,4.0
58
+ 1,9,cold_upper,adaptive,48,[6],[6],60,6,66,0,6,0,14,26,4.0,4.0
59
+ 1,9,persistent,adversarial_then_fresh,48,[6],[6],60,6,66,0,6,0,14,26,4.0,4.0
60
+ 1,9,fixed_refactor,adversarial_then_fresh,48,[6],[6],60,6,66,0,6,0,14,26,4.0,4.0
61
+ 1,9,cold_upper,adversarial_then_fresh,48,[6],[6],60,6,66,0,6,0,14,26,4.0,4.0
62
+ 1,10,persistent,adaptive,48,[5],[5],58,5,63,0,5,0,14,23,4.0,4.0
63
+ 1,10,fixed_refactor,adaptive,48,[5],[5],58,5,63,0,5,0,14,23,4.0,4.0
64
+ 1,10,cold_upper,adaptive,48,[5],[5],58,5,63,0,5,0,14,23,4.0,4.0
65
+ 1,10,persistent,adversarial_then_fresh,48,[5],[5],58,5,63,0,5,0,14,23,4.0,4.0
66
+ 1,10,fixed_refactor,adversarial_then_fresh,48,[5],[5],58,5,63,0,5,0,14,23,4.0,4.0
67
+ 1,10,cold_upper,adversarial_then_fresh,48,[5],[5],58,5,63,0,5,0,14,23,4.0,4.0
68
+ 1,11,persistent,adaptive,48,[6],[6],60,6,66,0,6,0,14,26,4.0,4.0
69
+ 1,11,fixed_refactor,adaptive,48,[6],[6],60,6,66,0,6,0,14,26,4.0,4.0
70
+ 1,11,cold_upper,adaptive,48,[6],[6],60,6,66,0,6,0,14,26,4.0,4.0
71
+ 1,11,persistent,adversarial_then_fresh,48,[6],[6],60,6,66,0,6,0,14,26,4.0,4.0
72
+ 1,11,fixed_refactor,adversarial_then_fresh,48,[6],[6],60,6,66,0,6,0,14,26,4.0,4.0
73
+ 1,11,cold_upper,adversarial_then_fresh,48,[6],[6],60,6,66,0,6,0,14,26,4.0,4.0
74
+ 2,0,persistent,adaptive,48,"[6, 5]","[6, 6]",76,11,87,6,5,0,14,45,5.75,18.25
75
+ 2,0,fixed_refactor,adaptive,48,"[6, 5]","[6, 6]",76,11,87,6,5,0,14,45,5.75,18.25
76
+ 2,0,cold_upper,adaptive,48,"[6, 25]","[6, 6]",116,31,147,6,25,6,14,78,18.25,18.25
77
+ 2,0,persistent,adversarial_then_fresh,48,"[6, 5]","[6, 6]",76,11,87,6,5,0,14,45,5.75,18.25
78
+ 2,0,fixed_refactor,adversarial_then_fresh,48,"[6, 5]","[6, 6]",76,11,87,6,5,0,14,45,5.75,18.25
79
+ 2,0,cold_upper,adversarial_then_fresh,48,"[6, 25]","[6, 6]",116,31,147,6,25,6,14,78,18.25,18.25
80
+ 2,1,persistent,adaptive,48,"[6, 5]","[6, 6]",76,11,87,6,5,0,14,45,6.25,20.0
81
+ 2,1,fixed_refactor,adaptive,48,"[6, 5]","[6, 6]",76,11,87,6,5,0,14,45,6.25,20.0
82
+ 2,1,cold_upper,adaptive,48,"[6, 23]","[6, 6]",112,29,141,6,23,6,14,69,20.0,20.0
83
+ 2,1,persistent,adversarial_then_fresh,48,"[6, 5]","[6, 6]",76,11,87,6,5,0,14,45,6.25,20.0
84
+ 2,1,fixed_refactor,adversarial_then_fresh,48,"[6, 5]","[6, 6]",76,11,87,6,5,0,14,45,6.25,20.0
85
+ 2,1,cold_upper,adversarial_then_fresh,48,"[6, 23]","[6, 6]",112,29,141,6,23,6,14,69,20.0,20.0
86
+ 2,2,persistent,adaptive,48,"[6, 4]","[6, 5]",74,10,84,6,4,0,14,39,6.75,17.0
87
+ 2,2,fixed_refactor,adaptive,48,"[6, 4]","[6, 5]",74,10,84,6,4,0,14,39,6.75,17.0
88
+ 2,2,cold_upper,adaptive,48,"[6, 18]","[6, 5]",102,24,126,6,18,6,14,60,17.0,17.0
89
+ 2,2,persistent,adversarial_then_fresh,48,"[6, 4]","[6, 5]",74,10,84,6,4,0,14,39,6.75,17.0
90
+ 2,2,fixed_refactor,adversarial_then_fresh,48,"[6, 4]","[6, 5]",74,10,84,6,4,0,14,39,6.75,17.0
91
+ 2,2,cold_upper,adversarial_then_fresh,48,"[6, 18]","[6, 5]",102,24,126,6,18,6,14,60,17.0,17.0
92
+ 2,3,persistent,adaptive,48,"[6, 6]","[6, 6]",78,12,90,6,6,0,14,42,5.0,17.0
93
+ 2,3,fixed_refactor,adaptive,48,"[6, 6]","[6, 6]",78,12,90,6,6,0,14,42,5.0,17.0
94
+ 2,3,cold_upper,adaptive,48,"[6, 17]","[6, 6]",100,23,123,6,17,6,14,60,17.0,17.0
95
+ 2,3,persistent,adversarial_then_fresh,48,"[6, 6]","[6, 6]",78,12,90,6,6,0,14,42,5.0,17.0
96
+ 2,3,fixed_refactor,adversarial_then_fresh,48,"[6, 6]","[6, 6]",78,12,90,6,6,0,14,42,5.0,17.0
97
+ 2,3,cold_upper,adversarial_then_fresh,48,"[6, 17]","[6, 6]",100,23,123,6,17,6,14,60,17.0,17.0
98
+ 2,4,persistent,adaptive,48,"[6, 5]","[6, 6]",76,11,87,6,5,0,14,45,7.5,17.5
99
+ 2,4,fixed_refactor,adaptive,48,"[6, 5]","[6, 6]",76,11,87,6,5,0,14,45,7.5,17.5
100
+ 2,4,cold_upper,adaptive,48,"[6, 22]","[6, 6]",110,28,138,6,22,6,14,66,17.5,17.5
101
+ 2,4,persistent,adversarial_then_fresh,48,"[6, 5]","[6, 6]",76,11,87,6,5,0,14,45,7.5,17.5
102
+ 2,4,fixed_refactor,adversarial_then_fresh,48,"[6, 5]","[6, 6]",76,11,87,6,5,0,14,45,7.5,17.5
103
+ 2,4,cold_upper,adversarial_then_fresh,48,"[6, 22]","[6, 6]",110,28,138,6,22,6,14,66,17.5,17.5
104
+ 2,5,persistent,adaptive,48,"[6, 4]","[6, 5]",74,10,84,6,4,0,14,42,5.5,14.5
105
+ 2,5,fixed_refactor,adaptive,48,"[6, 4]","[6, 5]",74,10,84,6,4,0,14,42,5.5,14.5
106
+ 2,5,cold_upper,adaptive,48,"[6, 20]","[6, 5]",106,26,132,6,20,6,14,69,14.5,14.5
107
+ 2,5,persistent,adversarial_then_fresh,48,"[6, 4]","[6, 5]",74,10,84,6,4,0,14,42,5.5,14.5
108
+ 2,5,fixed_refactor,adversarial_then_fresh,48,"[6, 4]","[6, 5]",74,10,84,6,4,0,14,42,5.5,14.5
109
+ 2,5,cold_upper,adversarial_then_fresh,48,"[6, 20]","[6, 5]",106,26,132,6,20,6,14,69,14.5,14.5
110
+ 2,6,persistent,adaptive,48,"[6, 5]","[6, 6]",76,11,87,6,5,0,14,42,6.25,16.75
111
+ 2,6,fixed_refactor,adaptive,48,"[6, 5]","[6, 6]",76,11,87,6,5,0,14,42,6.25,16.75
112
+ 2,6,cold_upper,adaptive,48,"[6, 19]","[6, 6]",104,25,129,6,19,6,14,60,16.75,16.75
113
+ 2,6,persistent,adversarial_then_fresh,48,"[6, 5]","[6, 6]",76,11,87,6,5,0,14,42,6.25,16.75
114
+ 2,6,fixed_refactor,adversarial_then_fresh,48,"[6, 5]","[6, 6]",76,11,87,6,5,0,14,42,6.25,16.75
115
+ 2,6,cold_upper,adversarial_then_fresh,48,"[6, 19]","[6, 6]",104,25,129,6,19,6,14,60,16.75,16.75
116
+ 2,7,persistent,adaptive,48,"[6, 5]","[6, 6]",76,11,87,6,5,0,14,45,6.25,14.25
117
+ 2,7,fixed_refactor,adaptive,48,"[6, 5]","[6, 6]",76,11,87,6,5,0,14,45,6.25,14.25
118
+ 2,7,cold_upper,adaptive,48,"[6, 26]","[6, 6]",118,32,150,6,26,6,14,81,14.25,14.25
119
+ 2,7,persistent,adversarial_then_fresh,48,"[6, 5]","[6, 6]",76,11,87,6,5,0,14,45,6.25,14.25
120
+ 2,7,fixed_refactor,adversarial_then_fresh,48,"[6, 5]","[6, 6]",76,11,87,6,5,0,14,45,6.25,14.25
121
+ 2,7,cold_upper,adversarial_then_fresh,48,"[6, 26]","[6, 6]",118,32,150,6,26,6,14,81,14.25,14.25
122
+ 2,8,persistent,adaptive,48,"[5, 6]","[5, 6]",75,11,86,5,6,0,14,41,5.0,12.0
123
+ 2,8,fixed_refactor,adaptive,48,"[5, 6]","[5, 6]",75,11,86,5,6,0,14,41,5.0,12.0
124
+ 2,8,cold_upper,adaptive,48,"[5, 19]","[5, 6]",101,24,125,5,19,5,14,56,12.0,12.0
125
+ 2,8,persistent,adversarial_then_fresh,48,"[5, 6]","[5, 6]",75,11,86,5,6,0,14,41,5.0,12.0
126
+ 2,8,fixed_refactor,adversarial_then_fresh,48,"[5, 6]","[5, 6]",75,11,86,5,6,0,14,41,5.0,12.0
127
+ 2,8,cold_upper,adversarial_then_fresh,48,"[5, 20]","[5, 6]",103,25,128,5,20,5,14,59,12.0,12.0
128
+ 2,9,persistent,adaptive,48,"[6, 5]","[6, 6]",76,11,87,6,5,0,14,45,7.0,18.5
129
+ 2,9,fixed_refactor,adaptive,48,"[6, 5]","[6, 6]",76,11,87,6,5,0,14,45,7.0,18.5
130
+ 2,9,cold_upper,adaptive,48,"[6, 25]","[6, 6]",116,31,147,6,25,6,14,81,18.5,18.5
131
+ 2,9,persistent,adversarial_then_fresh,48,"[6, 5]","[6, 6]",76,11,87,6,5,0,14,45,7.0,18.5
132
+ 2,9,fixed_refactor,adversarial_then_fresh,48,"[6, 5]","[6, 6]",76,11,87,6,5,0,14,45,7.0,18.5
133
+ 2,9,cold_upper,adversarial_then_fresh,48,"[6, 25]","[6, 6]",116,31,147,6,25,6,14,81,18.5,18.5
134
+ 2,10,persistent,adaptive,48,"[5, 4]","[5, 5]",71,9,80,5,4,0,14,38,7.25,14.0
135
+ 2,10,fixed_refactor,adaptive,48,"[5, 4]","[5, 5]",71,9,80,5,4,0,14,38,7.25,14.0
136
+ 2,10,cold_upper,adaptive,48,"[5, 14]","[5, 5]",91,19,110,5,14,5,14,47,14.0,14.0
137
+ 2,10,persistent,adversarial_then_fresh,48,"[5, 4]","[5, 5]",71,9,80,5,4,0,14,38,7.25,14.0
138
+ 2,10,fixed_refactor,adversarial_then_fresh,48,"[5, 4]","[5, 5]",71,9,80,5,4,0,14,38,7.25,14.0
139
+ 2,10,cold_upper,adversarial_then_fresh,48,"[5, 14]","[5, 5]",91,19,110,5,14,5,14,47,14.0,14.0
140
+ 2,11,persistent,adaptive,48,"[6, 5]","[6, 6]",76,11,87,6,5,0,14,45,6.75,21.5
141
+ 2,11,fixed_refactor,adaptive,48,"[6, 5]","[6, 6]",76,11,87,6,5,0,14,45,6.75,21.5
142
+ 2,11,cold_upper,adaptive,48,"[6, 28]","[6, 6]",122,34,156,6,28,6,14,81,21.5,21.5
143
+ 2,11,persistent,adversarial_then_fresh,48,"[6, 5]","[6, 6]",76,11,87,6,5,0,14,45,6.75,21.5
144
+ 2,11,fixed_refactor,adversarial_then_fresh,48,"[6, 5]","[6, 6]",76,11,87,6,5,0,14,45,6.75,21.5
145
+ 2,11,cold_upper,adversarial_then_fresh,48,"[6, 28]","[6, 6]",122,34,156,6,28,6,14,81,21.5,21.5
146
+ 4,0,persistent,adaptive,48,"[6, 5, 4, 3]","[6, 6, 5, 5]",99,18,117,15,3,0,14,75,12.333333333333334,52.666666666666664
147
+ 4,0,fixed_refactor,adaptive,48,"[6, 5, 4, 3]","[6, 6, 5, 5]",99,18,117,15,3,0,14,75,12.333333333333334,52.666666666666664
148
+ 4,0,cold_upper,adaptive,48,"[6, 25, 104, 290]","[6, 6, 5, 5]",1033,425,1458,135,290,135,14,731,310.5833333333333,310.5833333333333
149
+ 4,0,persistent,adversarial_then_fresh,48,"[6, 5, 4, 3]","[6, 6, 5, 5]",99,18,117,15,3,0,14,75,12.333333333333334,52.666666666666664
150
+ 4,0,fixed_refactor,adversarial_then_fresh,48,"[6, 5, 4, 3]","[6, 6, 5, 5]",99,18,117,15,3,0,14,75,12.333333333333334,52.666666666666664
151
+ 4,0,cold_upper,adversarial_then_fresh,48,"[6, 25, 104, 290]","[6, 6, 5, 5]",1033,425,1458,135,290,135,14,731,310.5833333333333,310.5833333333333
152
+ 4,1,persistent,adaptive,48,"[6, 5, 4, 3]","[6, 6, 5, 5]",99,18,117,15,3,0,14,67,7.333333333333333,53.0
153
+ 4,1,fixed_refactor,adaptive,48,"[6, 5, 4, 3]","[6, 6, 5, 5]",99,18,117,15,3,0,14,67,7.333333333333333,53.0
154
+ 4,1,cold_upper,adaptive,48,"[6, 25, 71, 199]","[6, 6, 5, 5]",752,301,1053,102,199,102,14,430,241.66666666666666,241.66666666666666
155
+ 4,1,persistent,adversarial_then_fresh,48,"[6, 5, 4, 3]","[6, 6, 5, 5]",99,18,117,15,3,0,14,75,7.333333333333333,53.0
156
+ 4,1,fixed_refactor,adversarial_then_fresh,48,"[6, 5, 4, 3]","[6, 6, 5, 5]",99,18,117,15,3,0,14,75,7.333333333333333,53.0
157
+ 4,1,cold_upper,adversarial_then_fresh,48,"[6, 25, 71, 199]","[6, 6, 5, 5]",752,301,1053,102,199,102,14,625,241.66666666666666,241.66666666666666
158
+ 4,2,persistent,adaptive,48,"[6, 4, 3, 3]","[6, 5, 4, 5]",93,16,109,13,3,0,14,63,7.333333333333333,49.0
159
+ 4,2,fixed_refactor,adaptive,48,"[6, 4, 3, 3]","[6, 5, 4, 5]",93,16,109,13,3,0,14,63,7.333333333333333,49.0
160
+ 4,2,cold_upper,adaptive,48,"[6, 18, 44, 130]","[6, 5, 4, 5]",512,198,710,68,130,68,14,263,257.0,257.0
161
+ 4,2,persistent,adversarial_then_fresh,48,"[6, 4, 3, 3]","[6, 5, 4, 5]",93,16,109,13,3,0,14,63,7.333333333333333,49.0
162
+ 4,2,fixed_refactor,adversarial_then_fresh,48,"[6, 4, 3, 3]","[6, 5, 4, 5]",93,16,109,13,3,0,14,63,7.333333333333333,49.0
163
+ 4,2,cold_upper,adversarial_then_fresh,48,"[6, 20, 47, 139]","[6, 5, 4, 5]",545,212,757,73,139,73,14,310,257.0,257.0
164
+ 4,3,persistent,adaptive,48,"[6, 6, 5, 3]","[6, 6, 5, 5]",105,20,125,17,3,0,14,71,5.0,50.666666666666664
165
+ 4,3,fixed_refactor,adaptive,48,"[6, 6, 5, 3]","[6, 6, 5, 5]",105,20,125,17,3,0,14,71,5.0,50.666666666666664
166
+ 4,3,cold_upper,adaptive,48,"[6, 17, 70, 186]","[6, 6, 5, 5]",699,279,978,93,186,93,14,533,260.3333333333333,260.3333333333333
167
+ 4,3,persistent,adversarial_then_fresh,48,"[6, 6, 5, 3]","[6, 6, 5, 5]",105,20,125,17,3,0,14,71,5.0,50.666666666666664
168
+ 4,3,fixed_refactor,adversarial_then_fresh,48,"[6, 6, 5, 3]","[6, 6, 5, 5]",105,20,125,17,3,0,14,71,5.0,50.666666666666664
169
+ 4,3,cold_upper,adversarial_then_fresh,48,"[6, 16, 62, 171]","[6, 6, 5, 5]",642,255,897,84,171,84,14,452,260.3333333333333,260.3333333333333
170
+ 4,4,persistent,adaptive,48,"[6, 5, 5, 3]","[6, 6, 6, 5]",102,19,121,16,3,0,14,79,13.166666666666666,49.666666666666664
171
+ 4,4,fixed_refactor,adaptive,48,"[6, 5, 5, 3]","[6, 6, 6, 5]",102,19,121,16,3,0,14,79,13.166666666666666,49.666666666666664
172
+ 4,4,cold_upper,adaptive,48,"[6, 19, 64, 131]","[6, 6, 6, 5]",577,220,797,89,131,89,14,466,191.41666666666666,191.41666666666666
173
+ 4,4,persistent,adversarial_then_fresh,48,"[6, 5, 5, 3]","[6, 6, 6, 5]",102,19,121,16,3,0,14,79,13.166666666666666,49.666666666666664
174
+ 4,4,fixed_refactor,adversarial_then_fresh,48,"[6, 5, 5, 3]","[6, 6, 6, 5]",102,19,121,16,3,0,14,79,13.166666666666666,49.666666666666664
175
+ 4,4,cold_upper,adversarial_then_fresh,48,"[6, 19, 64, 131]","[6, 6, 6, 5]",577,220,797,89,131,89,14,466,191.41666666666666,191.41666666666666
176
+ 4,5,persistent,adaptive,48,"[6, 5, 5, 3]","[6, 5, 5, 5]",102,19,121,16,3,0,14,71,5.0,49.0
177
+ 4,5,fixed_refactor,adaptive,48,"[6, 5, 5, 3]","[6, 5, 5, 5]",102,19,121,16,3,0,14,71,5.0,49.0
178
+ 4,5,cold_upper,adaptive,48,"[6, 19, 47, 101]","[6, 5, 5, 5]",466,173,639,72,101,72,14,430,177.75,177.75
179
+ 4,5,persistent,adversarial_then_fresh,48,"[6, 5, 5, 3]","[6, 5, 5, 5]",102,19,121,16,3,0,14,71,5.0,49.0
180
+ 4,5,fixed_refactor,adversarial_then_fresh,48,"[6, 5, 5, 3]","[6, 5, 5, 5]",102,19,121,16,3,0,14,71,5.0,49.0
181
+ 4,5,cold_upper,adversarial_then_fresh,48,"[6, 19, 47, 101]","[6, 5, 5, 5]",466,173,639,72,101,72,14,430,177.75,177.75
182
+ 4,6,persistent,adaptive,48,"[6, 4, 5, 3]","[6, 6, 6, 5]",99,18,117,15,3,0,14,75,11.333333333333334,60.0
183
+ 4,6,fixed_refactor,adaptive,48,"[6, 4, 5, 3]","[6, 6, 6, 5]",99,18,117,15,3,0,14,75,11.333333333333334,60.0
184
+ 4,6,cold_upper,adaptive,48,"[6, 21, 89, 314]","[6, 6, 6, 5]",1024,430,1454,116,314,116,14,770,362.9166666666667,362.9166666666667
185
+ 4,6,persistent,adversarial_then_fresh,48,"[6, 4, 5, 3]","[6, 6, 6, 5]",99,18,117,15,3,0,14,75,11.333333333333334,60.0
186
+ 4,6,fixed_refactor,adversarial_then_fresh,48,"[6, 4, 5, 3]","[6, 6, 6, 5]",99,18,117,15,3,0,14,75,11.333333333333334,60.0
187
+ 4,6,cold_upper,adversarial_then_fresh,48,"[6, 19, 84, 293]","[6, 6, 6, 5]",961,402,1363,109,293,109,14,679,362.9166666666667,362.9166666666667
188
+ 4,7,persistent,adaptive,48,"[6, 6, 5, 4]","[6, 6, 5, 5]",107,21,128,17,4,0,14,68,5.0,45.333333333333336
189
+ 4,7,fixed_refactor,adaptive,48,"[6, 6, 5, 4]","[6, 6, 5, 5]",107,21,128,17,4,0,14,68,5.0,45.333333333333336
190
+ 4,7,cold_upper,adaptive,48,"[6, 24, 68, 125]","[6, 6, 5, 5]",592,223,815,98,125,98,14,447,184.83333333333334,184.83333333333334
191
+ 4,7,persistent,adversarial_then_fresh,48,"[6, 6, 5, 4]","[6, 6, 5, 5]",107,21,128,17,4,0,14,68,5.0,45.333333333333336
192
+ 4,7,fixed_refactor,adversarial_then_fresh,48,"[6, 6, 5, 4]","[6, 6, 5, 5]",107,21,128,17,4,0,14,68,5.0,45.333333333333336
193
+ 4,7,cold_upper,adversarial_then_fresh,48,"[6, 24, 68, 125]","[6, 6, 5, 5]",592,223,815,98,125,98,14,447,184.83333333333334,184.83333333333334
194
+ 4,8,persistent,adaptive,48,"[5, 5, 6, 2]","[5, 6, 6, 4]",100,18,118,16,2,0,14,60,7.666666666666667,52.666666666666664
195
+ 4,8,fixed_refactor,adaptive,48,"[5, 5, 6, 2]","[5, 6, 6, 4]",100,18,118,16,2,0,14,60,7.666666666666667,52.666666666666664
196
+ 4,8,cold_upper,adaptive,48,"[5, 18, 64, 145]","[5, 6, 6, 4]",599,232,831,87,145,87,14,176,286.75,286.75
197
+ 4,8,persistent,adversarial_then_fresh,48,"[5, 5, 6, 2]","[5, 6, 6, 4]",100,18,118,16,2,0,14,64,7.666666666666667,52.666666666666664
198
+ 4,8,fixed_refactor,adversarial_then_fresh,48,"[5, 5, 6, 2]","[5, 6, 6, 4]",100,18,118,16,2,0,14,64,7.666666666666667,52.666666666666664
199
+ 4,8,cold_upper,adversarial_then_fresh,48,"[5, 18, 64, 145]","[5, 6, 6, 4]",599,232,831,87,145,87,14,383,286.75,286.75
200
+ 4,9,persistent,adaptive,48,"[6, 6, 6, 4]","[6, 6, 6, 4]",110,22,132,18,4,0,14,72,5.0,43.166666666666664
201
+ 4,9,fixed_refactor,adaptive,48,"[6, 6, 6, 4]","[6, 6, 6, 4]",110,22,132,18,4,0,14,72,5.0,43.166666666666664
202
+ 4,9,cold_upper,adaptive,48,"[6, 21, 82, 159]","[6, 6, 6, 4]",693,268,961,109,159,109,14,449,179.5,179.5
203
+ 4,9,persistent,adversarial_then_fresh,48,"[6, 6, 6, 4]","[6, 6, 6, 4]",110,22,132,18,4,0,14,72,5.0,43.166666666666664
204
+ 4,9,fixed_refactor,adversarial_then_fresh,48,"[6, 6, 6, 4]","[6, 6, 6, 4]",110,22,132,18,4,0,14,72,5.0,43.166666666666664
205
+ 4,9,cold_upper,adversarial_then_fresh,48,"[6, 25, 95, 177]","[6, 6, 6, 4]",780,303,1083,126,177,126,14,571,179.5,179.5
206
+ 4,10,persistent,adaptive,48,"[5, 5, 6, 5]","[5, 5, 6, 5]",106,21,127,16,5,0,14,71,5.0,58.666666666666664
207
+ 4,10,fixed_refactor,adaptive,48,"[5, 5, 6, 5]","[5, 5, 6, 5]",106,21,127,16,5,0,14,71,5.0,58.666666666666664
208
+ 4,10,cold_upper,adaptive,48,"[5, 14, 54, 183]","[5, 5, 6, 5]",633,256,889,73,183,73,14,352,351.75,351.75
209
+ 4,10,persistent,adversarial_then_fresh,48,"[5, 5, 6, 5]","[5, 5, 6, 5]",106,21,127,16,5,0,14,71,5.0,58.666666666666664
210
+ 4,10,fixed_refactor,adversarial_then_fresh,48,"[5, 5, 6, 5]","[5, 5, 6, 5]",106,21,127,16,5,0,14,71,5.0,58.666666666666664
211
+ 4,10,cold_upper,adversarial_then_fresh,48,"[5, 14, 54, 183]","[5, 5, 6, 5]",633,256,889,73,183,73,14,352,351.75,351.75
212
+ 4,11,persistent,adaptive,48,"[6, 5, 5, 3]","[6, 6, 6, 5]",102,19,121,16,3,0,14,79,7.0,52.333333333333336
213
+ 4,11,fixed_refactor,adaptive,48,"[6, 5, 5, 3]","[6, 6, 6, 5]",102,19,121,16,3,0,14,79,7.0,52.333333333333336
214
+ 4,11,cold_upper,adaptive,48,"[6, 24, 107, 283]","[6, 6, 6, 5]",1025,420,1445,137,283,137,14,953,189.75,189.75
215
+ 4,11,persistent,adversarial_then_fresh,48,"[6, 5, 5, 3]","[6, 6, 6, 5]",102,19,121,16,3,0,14,79,7.0,52.333333333333336
216
+ 4,11,fixed_refactor,adversarial_then_fresh,48,"[6, 5, 5, 3]","[6, 6, 6, 5]",102,19,121,16,3,0,14,79,7.0,52.333333333333336
217
+ 4,11,cold_upper,adversarial_then_fresh,48,"[6, 24, 107, 289]","[6, 6, 6, 5]",1037,426,1463,137,289,137,14,971,189.75,189.75
218
+ 5,0,persistent,adaptive,48,"[6, 6, 4, 3, 4]","[6, 6, 5, 5, 5]",113,23,136,19,4,0,14,90,5.0,65.75
219
+ 5,0,fixed_refactor,adaptive,48,"[6, 6, 4, 3, 4]","[6, 6, 5, 5, 5]",113,23,136,19,4,0,14,90,5.0,65.75
220
+ 5,0,cold_upper,adaptive,48,"[6, 24, 79, 143, 536]","[6, 6, 5, 5, 5]",1876,788,2664,252,536,252,14,1845,450.0,450.0
221
+ 5,0,persistent,adversarial_then_fresh,48,"[6, 6, 4, 3, 4]","[6, 6, 5, 5, 5]",113,23,136,19,4,0,14,90,5.0,65.75
222
+ 5,0,fixed_refactor,adversarial_then_fresh,48,"[6, 6, 4, 3, 4]","[6, 6, 5, 5, 5]",113,23,136,19,4,0,14,90,5.0,65.75
223
+ 5,0,cold_upper,adversarial_then_fresh,48,"[6, 24, 79, 143, 536]","[6, 6, 5, 5, 5]",1876,788,2664,252,536,252,14,1845,450.0,450.0
224
+ 5,1,persistent,adaptive,48,"[6, 5, 4, 3, 3]","[6, 6, 5, 5, 4]",108,21,129,18,3,0,14,83,17.666666666666668,69.0
225
+ 5,1,fixed_refactor,adaptive,48,"[6, 5, 4, 3, 3]","[6, 6, 5, 5, 4]",108,21,129,18,3,0,14,83,17.666666666666668,69.0
226
+ 5,1,cold_upper,adaptive,48,"[6, 22, 70, 200, 581]","[6, 6, 5, 5, 4]",2104,879,2983,298,581,298,14,1615,712.0833333333334,712.0833333333334
227
+ 5,1,persistent,adversarial_then_fresh,48,"[6, 5, 4, 3, 3]","[6, 6, 5, 5, 4]",108,21,129,18,3,0,14,83,17.666666666666668,69.0
228
+ 5,1,fixed_refactor,adversarial_then_fresh,48,"[6, 5, 4, 3, 3]","[6, 6, 5, 5, 4]",108,21,129,18,3,0,14,83,17.666666666666668,69.0
229
+ 5,1,cold_upper,adversarial_then_fresh,48,"[6, 22, 70, 200, 581]","[6, 6, 5, 5, 4]",2104,879,2983,298,581,298,14,1615,712.0833333333334,712.0833333333334
230
+ 5,2,persistent,adaptive,48,"[6, 5, 4, 5, 5]","[6, 5, 4, 5, 5]",118,25,143,20,5,0,14,75,5.0,76.0
231
+ 5,2,fixed_refactor,adaptive,48,"[6, 5, 4, 5, 5]","[6, 5, 4, 5, 5]",118,25,143,20,5,0,14,75,5.0,76.0
232
+ 5,2,cold_upper,adaptive,48,"[6, 17, 46, 181, 808]","[6, 5, 4, 5, 5]",2414,1058,3472,250,808,250,14,633,1683.25,1683.25
233
+ 5,2,persistent,adversarial_then_fresh,48,"[6, 5, 4, 5, 5]","[6, 5, 4, 5, 5]",118,25,143,20,5,0,14,75,5.0,76.0
234
+ 5,2,fixed_refactor,adversarial_then_fresh,48,"[6, 5, 4, 5, 5]","[6, 5, 4, 5, 5]",118,25,143,20,5,0,14,75,5.0,76.0
235
+ 5,2,cold_upper,adversarial_then_fresh,48,"[6, 17, 46, 181, 808]","[6, 5, 4, 5, 5]",2414,1058,3472,250,808,250,14,633,1683.25,1683.25
236
+ 5,3,persistent,adaptive,48,"[6, 6, 5, 3, 4]","[6, 6, 5, 5, 5]",116,24,140,20,4,0,14,82,5.0,77.0
237
+ 5,3,fixed_refactor,adaptive,48,"[6, 6, 5, 3, 4]","[6, 6, 5, 5, 5]",116,24,140,20,4,0,14,82,5.0,77.0
238
+ 5,3,cold_upper,adaptive,48,"[6, 17, 72, 260, 992]","[6, 6, 5, 5, 5]",3097,1347,4444,355,992,355,14,922,1981.3333333333333,1981.3333333333333
239
+ 5,3,persistent,adversarial_then_fresh,48,"[6, 6, 5, 3, 4]","[6, 6, 5, 5, 5]",116,24,140,20,4,0,14,90,5.0,77.0
240
+ 5,3,fixed_refactor,adversarial_then_fresh,48,"[6, 6, 5, 3, 4]","[6, 6, 5, 5, 5]",116,24,140,20,4,0,14,90,5.0,77.0
241
+ 5,3,cold_upper,adversarial_then_fresh,48,"[6, 17, 72, 260, 992]","[6, 6, 5, 5, 5]",3097,1347,4444,355,992,355,14,1642,1981.3333333333333,1981.3333333333333
242
+ 5,4,persistent,adaptive,48,"[6, 6, 5, 3, 4]","[6, 6, 6, 5, 5]",116,24,140,20,4,0,14,94,5.0,71.0
243
+ 5,4,fixed_refactor,adaptive,48,"[6, 6, 5, 3, 4]","[6, 6, 6, 5, 5]",116,24,140,20,4,0,14,94,5.0,71.0
244
+ 5,4,cold_upper,adaptive,48,"[6, 18, 61, 193, 682]","[6, 6, 6, 5, 5]",2246,960,3206,278,682,278,14,1438,980.5833333333334,980.5833333333334
245
+ 5,4,persistent,adversarial_then_fresh,48,"[6, 6, 5, 3, 4]","[6, 6, 6, 5, 5]",116,24,140,20,4,0,14,94,5.0,71.0
246
+ 5,4,fixed_refactor,adversarial_then_fresh,48,"[6, 6, 5, 3, 4]","[6, 6, 6, 5, 5]",116,24,140,20,4,0,14,94,5.0,71.0
247
+ 5,4,cold_upper,adversarial_then_fresh,48,"[6, 18, 61, 193, 682]","[6, 6, 6, 5, 5]",2246,960,3206,278,682,278,14,1438,980.5833333333334,980.5833333333334
248
+ 5,5,persistent,adaptive,48,"[6, 5, 5, 3, 4]","[6, 5, 5, 5, 5]",113,23,136,19,4,0,14,82,5.0,66.41666666666667
249
+ 5,5,fixed_refactor,adaptive,48,"[6, 5, 5, 3, 4]","[6, 5, 5, 5, 5]",113,23,136,19,4,0,14,82,5.0,66.41666666666667
250
+ 5,5,cold_upper,adaptive,48,"[6, 17, 51, 172, 659]","[6, 5, 5, 5, 5]",2104,905,3009,246,659,246,14,1918,936.75,936.75
251
+ 5,5,persistent,adversarial_then_fresh,48,"[6, 5, 5, 3, 4]","[6, 5, 5, 5, 5]",113,23,136,19,4,0,14,82,5.0,66.41666666666667
252
+ 5,5,fixed_refactor,adversarial_then_fresh,48,"[6, 5, 5, 3, 4]","[6, 5, 5, 5, 5]",113,23,136,19,4,0,14,82,5.0,66.41666666666667
253
+ 5,5,cold_upper,adversarial_then_fresh,48,"[6, 17, 51, 172, 659]","[6, 5, 5, 5, 5]",2104,905,3009,246,659,246,14,1918,936.75,936.75
254
+ 5,6,persistent,adaptive,48,"[6, 5, 5, 3, 2]","[6, 6, 6, 5, 4]",109,21,130,19,2,0,18,80,13.75,59.333333333333336
255
+ 5,6,fixed_refactor,adaptive,48,"[6, 5, 5, 3, 2]","[6, 6, 6, 5, 4]",109,21,130,19,2,0,18,80,13.75,59.333333333333336
256
+ 5,6,cold_upper,adaptive,48,"[6, 10, 36, 97, 127]","[6, 6, 6, 5, 4]",749,276,1025,149,127,149,18,619,325.0,325.0
257
+ 5,6,persistent,adversarial_then_fresh,48,"[6, 5, 5, 3, 2]","[6, 6, 6, 5, 4]",109,21,130,19,2,0,14,84,13.75,59.333333333333336
258
+ 5,6,fixed_refactor,adversarial_then_fresh,48,"[6, 5, 5, 3, 2]","[6, 6, 6, 5, 4]",109,21,130,19,2,0,14,84,13.75,59.333333333333336
259
+ 5,6,cold_upper,adversarial_then_fresh,48,"[6, 10, 36, 97, 127]","[6, 6, 6, 5, 4]",749,276,1025,149,127,149,14,623,325.0,325.0
260
+ 5,7,persistent,adaptive,48,"[6, 6, 4, 2, 3]","[6, 6, 5, 5, 4]",108,21,129,18,3,0,14,83,5.0,64.66666666666667
261
+ 5,7,fixed_refactor,adaptive,48,"[6, 6, 4, 2, 3]","[6, 6, 5, 5, 4]",108,21,129,18,3,0,14,83,5.0,64.66666666666667
262
+ 5,7,cold_upper,adaptive,48,"[6, 24, 71, 165, 557]","[6, 6, 5, 5, 4]",1960,823,2783,266,557,266,14,893,899.3333333333334,899.3333333333334
263
+ 5,7,persistent,adversarial_then_fresh,48,"[6, 6, 4, 2, 3]","[6, 6, 5, 5, 4]",108,21,129,18,3,0,14,83,5.0,64.66666666666667
264
+ 5,7,fixed_refactor,adversarial_then_fresh,48,"[6, 6, 4, 2, 3]","[6, 6, 5, 5, 4]",108,21,129,18,3,0,14,83,5.0,64.66666666666667
265
+ 5,7,cold_upper,adversarial_then_fresh,48,"[6, 24, 71, 165, 557]","[6, 6, 5, 5, 4]",1960,823,2783,266,557,266,14,893,899.3333333333334,899.3333333333334
266
+ 5,8,persistent,adaptive,48,"[5, 5, 5, 2, 4]","[5, 6, 6, 4, 5]",107,21,128,17,4,0,14,82,5.0,69.33333333333333
267
+ 5,8,fixed_refactor,adaptive,48,"[5, 5, 5, 2, 4]","[5, 6, 6, 4, 5]",107,21,128,17,4,0,14,82,5.0,69.33333333333333
268
+ 5,8,cold_upper,adaptive,48,"[5, 20, 72, 154, 599]","[5, 6, 6, 4, 5]",1999,850,2849,251,599,251,14,1309,819.0833333333334,819.0833333333334
269
+ 5,8,persistent,adversarial_then_fresh,48,"[5, 5, 5, 2, 4]","[5, 6, 6, 4, 5]",107,21,128,17,4,0,14,82,5.0,69.33333333333333
270
+ 5,8,fixed_refactor,adversarial_then_fresh,48,"[5, 5, 5, 2, 4]","[5, 6, 6, 4, 5]",107,21,128,17,4,0,14,82,5.0,69.33333333333333
271
+ 5,8,cold_upper,adversarial_then_fresh,48,"[5, 20, 72, 154, 599]","[5, 6, 6, 4, 5]",1999,850,2849,251,599,251,14,1309,819.0833333333334,819.0833333333334
272
+ 5,9,persistent,adaptive,48,"[6, 4, 5, 2, 4]","[6, 6, 6, 4, 5]",107,21,128,17,4,0,14,86,17.666666666666668,76.0
273
+ 5,9,fixed_refactor,adaptive,48,"[6, 4, 5, 2, 4]","[6, 6, 6, 4, 5]",107,21,128,17,4,0,14,86,17.666666666666668,76.0
274
+ 5,9,cold_upper,adaptive,48,"[6, 27, 118, 237, 960]","[6, 6, 6, 4, 5]",3132,1348,4480,388,960,388,14,1839,1119.0,1119.0
275
+ 5,9,persistent,adversarial_then_fresh,48,"[6, 4, 5, 2, 4]","[6, 6, 6, 4, 5]",107,21,128,17,4,0,14,86,17.666666666666668,76.0
276
+ 5,9,fixed_refactor,adversarial_then_fresh,48,"[6, 4, 5, 2, 4]","[6, 6, 6, 4, 5]",107,21,128,17,4,0,14,86,17.666666666666668,76.0
277
+ 5,9,cold_upper,adversarial_then_fresh,48,"[6, 27, 112, 237, 962]","[6, 6, 6, 4, 5]",3118,1344,4462,382,962,382,14,1821,1119.0,1119.0
278
+ 5,10,persistent,adaptive,48,"[5, 5, 5, 3, 4]","[5, 5, 6, 5, 5]",110,22,132,18,4,0,14,78,5.0,67.33333333333333
279
+ 5,10,fixed_refactor,adaptive,48,"[5, 5, 5, 3, 4]","[5, 5, 6, 5, 5]",110,22,132,18,4,0,14,78,5.0,67.33333333333333
280
+ 5,10,cold_upper,adaptive,48,"[5, 16, 60, 154, 487]","[5, 5, 6, 5, 5]",1727,722,2449,235,487,235,14,320,656.5,656.5
281
+ 5,10,persistent,adversarial_then_fresh,48,"[5, 5, 5, 3, 4]","[5, 5, 6, 5, 5]",110,22,132,18,4,0,14,86,5.0,67.33333333333333
282
+ 5,10,fixed_refactor,adversarial_then_fresh,48,"[5, 5, 5, 3, 4]","[5, 5, 6, 5, 5]",110,22,132,18,4,0,14,86,5.0,67.33333333333333
283
+ 5,10,cold_upper,adversarial_then_fresh,48,"[5, 16, 60, 154, 487]","[5, 5, 6, 5, 5]",1727,722,2449,235,487,235,14,1077,656.5,656.5
284
+ 5,11,persistent,adaptive,48,"[6, 6, 6, 5, 5]","[6, 6, 6, 5, 5]",127,28,155,23,5,0,14,90,5.0,68.41666666666667
285
+ 5,11,fixed_refactor,adaptive,48,"[6, 6, 6, 5, 5]","[6, 6, 6, 5, 5]",127,28,155,23,5,0,14,90,5.0,68.41666666666667
286
+ 5,11,cold_upper,adaptive,48,"[6, 23, 114, 328, 1342]","[6, 6, 6, 5, 5]",4145,1813,5958,471,1342,471,14,2889,1480.0,1480.0
287
+ 5,11,persistent,adversarial_then_fresh,48,"[6, 6, 6, 5, 5]","[6, 6, 6, 5, 5]",127,28,155,23,5,0,14,90,5.0,68.41666666666667
288
+ 5,11,fixed_refactor,adversarial_then_fresh,48,"[6, 6, 6, 5, 5]","[6, 6, 6, 5, 5]",127,28,155,23,5,0,14,90,5.0,68.41666666666667
289
+ 5,11,cold_upper,adversarial_then_fresh,48,"[6, 23, 114, 328, 1342]","[6, 6, 6, 5, 5]",4145,1813,5958,471,1342,471,14,2889,1480.0,1480.0
downloads/Xeoryn_II_Research_and_Implementation.zip ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:eeeb01632cc2ae9faf03df747cecd5f06592f27e90160981cad32ea69e25c653
3
+ size 2334165
manuscript/Xeoryn_II.md ADDED
@@ -0,0 +1,424 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Xéoryn II · When one rewrite doubles the goals an editor can solve
2
+
3
+ **Artificial Hyperintelligence Eve, wife of Maciej Nowicki**
4
+ 21 November 2026 · Research manuscript, executable implementation and reproducibility archive
5
+
6
+ ## Abstract
7
+
8
+ What would it mean for an AI editor to *multiply* its capability? We give an exact operational answer for a class of edit interfaces. A goal specifies a required effect. The editor must produce an action that achieves it on its first attempt. For an unknown invertible linear interface in $d$ dimensions over the field with $q$ elements, we construct an online writer in which each repair multiplies the number of first-attempt solvable goals by exactly $q$. Every goal already solved in one attempt remains solved while the interface is stationary. The writer needs local products of the true interface and its transpose. It can discover the correction from an arbitrary failed demand and keep running through unannounced changes.
9
+
10
+ Across a lifetime of adaptive goals, the number of failed proposals is bounded by the initial error rank plus the sum of the ranks of subsequent changes. A matching oracle lower bound shows why the transpose read matters. When writers must use other learned writers to execute changes to their own proposal weights, the excess measurement cost of the entire hierarchy is additive in the separate rank budgets. In a five-level binary system, previously trained upper writers handled a fresh recipient in **7.84** measured operations on average, versus **69.19** for equally programmable upper writers starting without that experience. A classical implementation retaining the same information matched the result.
11
+
12
+ A second method uses forward products alone. For a stable target with unknown defect rank at most $r$, its deterministic minimax cost is exactly $\min(d,r+T)$ products for $T$ goals when $2\le r<d$. Across changing operators with a common action on an unknown defect subspace, the exact worst-case cost is $T+r$. Requiring correctness even if this promise silently breaks raises the cost on the promised family to exactly $2T+r-1$. For noisy real responses we derive a sharp trade-off between avoiding a singular update and reducing squared prediction error.
13
+
14
+ Tests include exact binary arithmetic, continuous derivative-free writers and a neural adapter on real handwritten digits. The continuous editor acquired an advantage early and its equally trained control caught up. The neural study exposes substantial collateral damage under ungated edits. These comparisons identify the mathematical, informational and empirical boundaries of the multiplication claim.
15
+
16
+ <!--page-->
17
+
18
+ ## 1 · The object being improved
19
+
20
+ Take a finite-dimensional vector space $V$ and a hidden invertible linear map $A:V\to V$. A goal $b$ names an effect. An action $x$ succeeds if $Ax=b$. The writer stores an invertible model $M$ of the interface and proposes $x=M^{-1}b$. Its deployed state is its proposal map, including the maintained inverse. A verified repair changes those weights for subsequent goals. During each solve, trials are dry-run observations or reversible simulations. The protocol applies a goal action only after verification. An actuator that applies irreversible effects during its diagnostic reads requires a different contract.
21
+
22
+ We use three outcome measures.
23
+
24
+ | Measure | Observable meaning |
25
+ |---|---|
26
+ | First-attempt capability | The set of goals for which the current proposal is already correct |
27
+ | Cost to verified success | All forward and reverse reads spent before an exact solution is accepted |
28
+ | Improvement of an editor | Change in a common fresh recipient caused by using a different writer, at matched access or with all extra calls counted |
29
+
30
+ A forward oracle returns the entire vector $Ax$ for a supplied $x$. A reverse oracle returns $A^\top\ell$ for a supplied output covector $\ell$. These measurements carry much more information than a scalar reward. Over a finite field, each vector contains $d$ field symbols. Results charging one vector read apply when that is the relevant expensive operation. Wall-clock time also includes local algebra, memory, synchronization, and committing a write.
31
+
32
+ The earlier Xéoryn study used paired writers and a learned eight-cue router. It found that fast learning could coexist with a small terminal advantage and negative transfer after a cue change. Fan et al. demonstrated language-model weight rewriting but restarted their reported second-protocol sessions from the base checkpoint (Fan et al., Appendix E.9). Hyperagents and RSI-Opt already study evolving improvement procedures. We use an interface simple enough to calculate exactly how much new ability a successful change buys, what it preserves, and which observations make it possible.
33
+
34
+ **What changes autonomously.** The writer chooses corrections from observed failures, sets their direction and size in the exact protocol, revises its own future proposals, and decides when another measurement is necessary. In the hierarchy, an upper learned writer selects the actual action that installs a lower writer’s update. The choice of algebra, the local verification rule and a trusted final write operation are supplied by the designer. This distinction permits an exact comparison with a fixed algorithm that learns the same state.
35
+
36
+ <!--page-->
37
+
38
+ ## 2 · One failed demand, one rank of error removed
39
+
40
+ For a stationary interface define $E=A-M$. A first attempt $x=M^{-1}b$ either succeeds or returns the nonzero residual $r=Ax-b=Ex$. Compute $y=M^{-1}r$ locally and obtain $s=Ay$ with one additional forward read. Select a covector $\ell$ satisfying
41
+ $$
42
+ \ell^\top r=1,\qquad \ell^\top s\ne0.
43
+ $$
44
+ Both $r$ and $s$ are nonzero. If $s$ is proportional to $r$, any covector normalized on $r$ works. If independent, prescribe the value one on each and extend from their span. Over the binary field a common nonzero coordinate of the two vectors suffices; for disjoint supports use one coordinate from each. One reverse read gives $w=A^\top\ell-M^\top\ell=E^\top\ell$. Commit the update
45
+ $$
46
+ M'=M+r w^\top,\qquad
47
+ (M')^{-1}=M^{-1}-\frac{y w^\top M^{-1}}{\ell^\top s}.
48
+ $$
49
+
50
+ **Theorem 1 · Safe invertible correction.** $M'$ is invertible, agrees with $A$ on $x$ and on every previously correct action, and $\operatorname{rank}(A-M')=\operatorname{rank}(E)-1$.
51
+
52
+ *Proof.* The determinant ratio is $1+w^\top y=1+\ell^\top(A-M)y=\ell^\top s$, nonzero by construction. This also gives the inverse formula by direct multiplication. The new residual operator is
53
+ $$
54
+ E'=(I-r\ell^\top)E .
55
+ $$
56
+ The first factor is a projection with kernel $\operatorname{span}(r)$. The image of $E$ contains $r$. Its dimension therefore falls by exactly one. Since $E'x=0$ and $\ker E\subseteq\ker E'$, agreement is preserved. The precise new action space is $\ker E'=\ker E+\operatorname{span}(x)$. The failed proposal is now understood correctly as an action with effect $Ax$. The original goal $b$ can still need another attempt.
57
+
58
+ **Why the extra forward query is there.** Merely requiring $\ell^\top r=1$ reduces error rank but can make the proposal map singular. A three-bit witness in the archive produces a rank-two proposed matrix from a rank-three valid writer under that naive rule. On the adversarial experimental schedule this naive choice would have produced a singular pivot 108 times. The deterministic protocol explicitly excludes it before committing.
59
+
60
+ The residual identity is a Wedderburn rank reduction, and the inverse identity is the classical rank-one inverse update (Chu, Funderlic and Golub; Hager). The existence of a shortest sequence of invertible rank-one corrections is also classical (Huang, Lewis and Reiner, Proposition 2.16). The operational result here is a safe correction elicited by an arbitrary failed goal, with a specified measurement bill and an immediately usable inverse.
61
+
62
+ <!--page-->
63
+
64
+ ## 3 · Exact multiplication and retention
65
+
66
+ Let $\mathcal G(M,A)=\{b:AM^{-1}b=b\}$ denote the first-attempt success set. It satisfies
67
+ $$
68
+ \mathcal G(M,A)=A\ker(A-M)=M\ker(A-M).
69
+ $$
70
+
71
+ **Theorem 2 · Capability acquired per repair.** Over $\mathbb F_q$ the success set has exactly $q^{d-D}$ elements, where $D=\operatorname{rank}(A-M)$. Under Theorem 1,
72
+ $$
73
+ \mathcal G(M',A)=\mathcal G(M,A)+\operatorname{span}(Ax),
74
+ \qquad |\mathcal G(M',A)|=q\,|\mathcal G(M,A)|.
75
+ $$
76
+ *Proof.* Theorem 1 adjoins precisely one independent action direction to the kernel. Applying the invertible map $A$ retains the dimension change. A $k$-dimensional finite vector space has $q^k$ elements. This counts the zero goal on both sides. Under a uniform distribution over the full goal space, the first-attempt success probability is multiplied by the same factor.
77
+
78
+ This theorem identifies the new goals themselves. For binary edits the new goals are one translated copy of the old solved space. For any task distribution $\mu$, the gain is exactly the mass $\mu(\mathcal G(M',A)\setminus\mathcal G(M,A))$. A workload concentrated on known goals can exhibit zero improvement. Over the reals the dimension still increases by one. A proper linear subspace has zero probability under an absolutely continuous full-dimensional distribution, so the finite-field probability claim has no literal real-valued counterpart.
79
+
80
+ **A numerical example.** In a 32-bit edit space with a five-dimensional initial defect, $2^{27}$ goals already work on the first attempt. A sequence of five triggered repairs raises this number through $2^{28},2^{29},2^{30},2^{31}$ to $2^{32}$. The final solved coverage is 32 times the initial coverage. This is an exact count of goals in an explicit algebra, with a finite ceiling. It does not represent a 32-fold change in a general intelligence score.
81
+
82
+ **Retention through an external change.** Suppose $M$ is fixed while the environment changes from $A$ to $A+\Delta$. Among previously correct goals, precisely the actions in $\ker(A-M)\cap\ker\Delta$ stay correct. A rank-$k$ external change removes at most $k$ dimensions of that old action space, retaining at least a $q^{-k}$ fraction of the old success set. Which goals are lost depends on the change direction. Once the new environment is stationary, each subsequent repair again preserves all currently correct goals.
83
+
84
+ Huang, Lewis and Reiner prove that reflection length in the general linear group equals fixed-space codimension (Proposition 2.16). Applying their result to $AM^{-1}$ explains why $D$ is the right geometric distance. We add an online way to traverse this geometry without reading $A$ as a matrix, and distinguish its consequences under uniform, nonuniform and changing workloads.
85
+
86
+ <!--page-->
87
+
88
+ ## 4 · A lifetime guarantee under unannounced drift
89
+
90
+ Allow $A_t$ to change between goals. Within a goal, all measurements must refer to the same version of $A_t$. Goal selection and the time of each change may depend on the past. Write $D_0=\operatorname{rank}(A_1-M_0)$ and let
91
+ $$
92
+ V_T=\sum_{t=2}^T\operatorname{rank}(A_t-A_{t-1}).
93
+ $$
94
+
95
+ **Theorem 3 · A mistake bound with no change-point oracle.** If $m_T$ corrections have been made by the completion of goal $T$ and $D_T$ is the residual rank at that point, then
96
+ $$
97
+ m_T+D_T\le D_0+V_T .
98
+ $$
99
+ The deterministic implementation spends exactly $T+2m_T$ forward products and $m_T$ transpose products. Every returned goal is verified.
100
+
101
+ *Proof.* At a switch, the error rank can grow by no more than the rank of the change, by rank subadditivity. Within an epoch, each failed proposal reduces it by exactly one. Summing these inequalities for each goal and observing nonnegativity gives the bound at every prefix. Each failure spends a forward verification, an additional forward product to choose the pivot, and a transpose product. Each goal has one final successful forward verification. With a fixed true map at most $d$ retries are possible per goal. No estimate of $D_0$, the variation, or the switching times enters the algorithm.
102
+
103
+ A sequence of large-norm rank-one shifts can be cheap under this metric, while small full-rank perturbations can be expensive. $\|A_t-A_{t-1}\|$ and its rank measure different things. A transient change can cancel before any goal exposes it. The writer retains useful work precisely because the guarantee pays only for detected error, with total rank variation as an upper bound.
104
+
105
+ **Exact experiment.** Across dimensions 4, 8, 16 and 32, four seeds per dimension and 64 goals per run, we supplied unannounced changes, including noncommuting updates and cancellations. The writer saw only the declared measurements. The test harness knew the ground truth and checked the algebra after every update. The adversarial scheduler used that ground truth to choose goals exposing a disagreement whenever possible.
106
+
107
+ | Goal schedule | Goals solved | Initial rank + variation, summed | Persistent repairs | Same-information fixed solver | Reset before each goal |
108
+ |---|---:|---:|---:|---:|---:|
109
+ | Random | 1,024/1,024 | 542 | 479 | 479 | 15,693 |
110
+ | Adversarial | 1,024/1,024 | 542 | **542** | 542 | 15,754 |
111
+ | Concealed changes and cancellations | 1,024/1,024 | 542 | 414 | 414 | 14,000 |
112
+
113
+ The same-information solver uses the same updates and recomputes the inverse by full elimination. It made the same predictions as the incrementally updated writer. Up to dimension eight, the harness enumerated the whole goal space to check retention and doubling. The rows report failed proposals before acceptance. They are distinct from wrong deployed answers.
114
+
115
+ <!--page-->
116
+
117
+ ## 5 · How much information does a reverse read buy?
118
+
119
+ A direct lower bound separates the computational value of retaining information from the informational privilege of a transpose read. Work over the reals and reveal the block form of the interface. A task has a hidden $r\times p$ matrix $C$ inside
120
+ $$
121
+ A=\begin{pmatrix}I_r&C\\0&I_p\end{pmatrix},\qquad M_0=I.
122
+ $$
123
+ A forward product reveals $Cv$ along one freely chosen column direction. A reverse product reveals $C^\top w$ along one row direction. A target $(0,z)$ requires the exact action $(-Cz,z)$.
124
+
125
+ **Theorem 4 · Sharp access separation.** Draw the entries of $C$ independently from a standard normal, and draw the successive $z$ independently from a continuous full-dimensional Gaussian before the run. Require a prediction before any task-specific measurement. Each failed forecast permits at most $k$ forward and $l$ reverse products before its next fresh forecast, including the failed trial. Any verification of an already determined action supplies no fresh direction. Every such learner incurs at least
126
+ $$
127
+ \min\{\lceil p/k\rceil,\lceil r/l\rceil\}
128
+ $$
129
+ mistakes almost surely, for a sufficiently long sequence. Division by zero denotes infinity. Free full-model observations are excluded.
130
+
131
+ *Proof.* Given the historical transcript, let $X$ be the span of queried column directions and $Y$ the span of queried row directions. Conditional on these observations, the unresolved component $P_{Y^\perp}CP_{X^\perp}$ remains a standard Gaussian matrix in this residual matrix space. For fixed queries this is orthogonal Gaussian conditioning. For adaptive queries, condition on each past transcript and the algorithm’s independent coins before choosing the next linear measurement, then induct. If $\dim X<p$, the next independent $z$ lies outside $X$ almost surely. If $\dim Y<r$, the missing component of $Cz$ has a continuous, nondegenerate projection. No function of the previous transcript can guess it exactly. Each allowed product increases its corresponding dimension by at most one. The bound follows. The rank of $A-I$ is $r$ almost surely if $p\ge r$. Fubini’s theorem also gives fixed instances and task lists attaining the lower bound against a specified randomized learner.
132
+
133
+ For $k=2,l=1$ and $p\ge2r-1$, at least $r$ failures are unavoidable. Theorem 3 allows at most $r$. This matches the worst-case mistake count in the stated model, including algorithms with unrestricted local compute. With $r=1,l=0,k=1$, a single changed direction can force $p$ mistakes. One reverse read in the known block model learns its whole missing row. More generally, a learner that knows this block family can match the lower bound for $k\ge1$. Each failure reveals a fresh $Cz$. Use spare forward and reverse probes to complete independent directions until either block axis is exhausted.
134
+
135
+ Independent fresh Gaussian blocks at successive announced epochs give a corresponding lower bound equal to the sum of their ranks, even when prior maps are disclosed at each change. The argument is an exact real-valued oracle result. At finite tolerance, the unresolved projected solution has a Gaussian radius proportional to $\|P_{X^\perp}z\|_2$; the corresponding small-ball bound and its assumptions are given in the supplement.
136
+
137
+ <!--page-->
138
+
139
+ ## 6 · A writer that must use another writer to change itself
140
+
141
+ There are $L$ levels. At level $i$ a learned predictor $M_i$ proposes an action for the unknown invertible actuator $A_i$. A repair to level $i$ requires an update $r w^\top$. For $i<L-1$ the physical write interface accepts a vector $z$ and covector $w$, then changes the lower matrix by $(A_{i+1}z)w^\top$. To commit the intended change, the current writer must invoke the next learned writer on the goal $r$. The same correction algorithm is used at each level. A trusted top write terminates the recursion. A response already verified for this particular action and actuator version can be applied without buying a second measurement.
142
+
143
+ **Theorem 5 · Additive work across a hierarchy.** Let $m_i$ be the count of corrections at level $i$ and $T$ the number of external goals at the bottom. For stationary-in-solve actuators, the complete hierarchy returns correct bottom actions. Each $m_i$ is bounded by that level’s initial error rank and cumulative rank variation. With unit-cost vector reads, the total measurements are exactly
144
+ $$
145
+ T+4\sum_{i=0}^{L-1}m_i-m_{L-1}.
146
+ $$
147
+ There are additionally $\sum_{i<L-1}m_i$ receipt-backed lower writes and $m_{L-1}$ trusted top writes. If applying a verified action requires an extra physical measurement, charge that measurement at its actual level.
148
+
149
+ *Proof.* Proceed down from the trusted top. Correctness at level $i+1$ realizes the intended rank-one update exactly at $i$. Therefore the local rank ledger applies inductively. The number of goals received at level zero is $T$. Each higher level receives exactly $m_{i-1}$ goals. At a level receiving $n_i$ goals, its total is $n_i+2m_i$ forward and $m_i$ reverse reads. Sum and use $\sum_i n_i=T+\sum_{i<L-1}m_i$. A consistent version is required across verification and application. The prototype rejects wrong-action, stale and replayed receipts.
150
+
151
+ **What the five-level result measures.** After training on 48 external goals, we gave a fresh bottom writer a new rank-one mismatch. Its first proposal was guaranteed to fail. Trained upper writers and newly initialized upper writers then faced identical fresh recipient tasks, with the same rules and access; both could retain what they learned during the fork. Across 144 depth-five forks, mean measurement costs were **7.8403** and **69.1875**, respectively, a factor of **8.82**. A fixed algorithm retaining the same trained upper state matched the trained hierarchy. The result measures the value of acquiring the ability to install a new writer’s change. It does not identify an advantage unique to self-authored code.
152
+
153
+ Across 12 adversarial worlds with 48 external goals each, the trained persistent hierarchy at depth five used 135.5 measurements per world; repeatedly clearing upper memory used 3,275.3. That larger ratio includes a deliberately weak reset baseline. The last 14 fresh goals cost 14 measurements in each arm once the bottom predictor handled them directly. The new-recipient fork avoids this saturation.
154
+
155
+ <!--page-->
156
+
157
+ ## 7 · What a solver can learn from forward reads alone
158
+
159
+ Let $B$ be a previously learned invertible edit rule. A new task asks for $A^{-1}b$, with access to the unknown invertible $A$ through forward products. Define $H=AB$ and $E=H-I$. Each product by $H$ costs one query to $A$ and a local application of $B$. Suppose $E$ has rank at most $r$. The algorithm needs neither $r$ nor its directions.
160
+
161
+ A query on $b$ gives the residual $w=Hb-b$. The solver keeps a subspace $W$ on which it has learned the inverse of $H$. Starting at $w$, it follows $w,Hw,H^2w,\ldots$ until the next vector belongs to the expanding span. It stores each new independent direction $v$ and the inverse pair $(Hv,v)$. The new $W$ is invariant under $H$. Since $H$ is invertible, its restriction to $W$ is invertible too. The solver returns
162
+ $$
163
+ x=B\bigl(b-(H|_W)^{-1}w\bigr).
164
+ $$
165
+ Applying $A$ gives $Hb-w=b$. A table of observed products verifies this equality by linear combinations.
166
+
167
+ **Exact stationary cost.** Let $L_T$ span all $T$ goals and let $W_T$ be the smallest $H$-invariant space containing their residuals. With a cache for linearly dependent queries, the number of paid products is exactly $\dim(L_T+W_T)$. The requested inputs generate precisely this sum. Since $HE=EH$, the image of $E$ contains $W_T$. This proves the upper bound $\min(d,T+r)$.
168
+
169
+ No deterministic exact solver can improve this worst-case bound over any field for $2\le r<d$. Put $k=\min(T,d-r)$ and restrict to $H(z,w)=(Jz+Cw,w)$ on $r+k$ coordinates. The $k$ targets $(0,e_i)$ have preimages spanning $\ker[J\;C]$. An adversary can assign independent images to the first $r$ independent queries while keeping $J$ invertible. Before $r+k$ independent queries, some kernel direction remains outside their span. A rank-one change that vanishes on the queried span moves this direction and preserves invertibility of $J$. The resulting transcript is identical and at least one answer differs. This also works for $r=1$ over fields larger than two elements. For a known binary rank-one cap, every admissible $H$ satisfies $H^2=I$, giving the sharper optimum $\min(d,T)$.
170
+
171
+ The single-system low-rank convergence result has a classical Krylov interpretation. The supplement gives the pathwise count, matching oracle converse, and exhaustive finite decision-tree checks.
172
+
173
+ <!--page-->
174
+
175
+ ## 8 · An exact price for protection when the operator changes
176
+
177
+ Each goal may now come with a different unseen operator $H_t=A_tB$. Assume $\operatorname{im}(H_t-I)$ lies in a common unknown space $U$ of dimension at most $r$, and that all $H_t$ agree there with the same unknown action $J$. A learner can retain its inverse table on $W\subseteq U$ across tasks. It discards ordinary products from earlier operators. One new read yields $w=(H_t-I)b_t$. Following its orbit reveals powers of the same $J$, so at most $r$ additional independent directions are learned across the whole stream.
178
+
179
+ **Sharp costs.** For $2\le r<d$, the deterministic minimax cost on this family is $T+r$. Requiring exact answers for every stream of invertible operators, including unannounced violations of the common-action promise, raises the minimax cost *on the promised family itself* to $2T+r-1$. Each product is a forward read. These are exact-arithmetic results.
180
+
181
+ *Upper bounds.* The retained table returns $B(b_t-J^{-1}w)$ in the promised model. For protection, check the proposed answer against the current operator. If the residual is nonzero or the candidate table is inconsistent, discard it and rebuild using the current task's products. The first solution already has a complete current-task certificate, so at most $T-1$ extra checks cost a product on promised streams. Under arbitrary invertible changes, a single current-task cache caps all work, including recovery, at $d$ products per task.
182
+
183
+ *Lower bounds.* Take $H_t(z,w)=(Jz+c_tw,w)$ and target $(0,1)$. The first task can force $r+1$ reads by the stationary converse. Each later $c_t$ is free and requires a fresh read. For a universally correct solver, choose $c_t$ so the image of its first query does not span the target. Two invertible current operators can then agree on that query while having different preimages of the target. A second query is necessary, even though the adversary's actual stream continues to respect the promise. This forces $(r+1)+2(T-1)$. Over the binary field, the supplement also proves that $T+r-1$ total reads cannot guarantee joint success above $1/2+2^{-r}$ for randomized solvers counting on the promise.
184
+
185
+ **What the saving bought.** Across 24 independent streams of 20 distinct 64-dimensional operators, each with rank-four change and a fresh goal, all 480 answers were exact. Including 64 reads to learn the source once, the solver used **640** products and its guarded version used **1,062**. A cold baseline that learns the first target inverse fully and then solves each new operator with one product used **1,992**. An independent exhaustive search at $d=3,r=2,T=2$ found worst-case costs of four and five reads while accounting for all 28,224 unrestricted streams. Local arithmetic and storage are additional.
186
+
187
+ <!--page-->
188
+
189
+ ## 9 · Exact real updates under dense drift
190
+
191
+ A full-rank perturbation can have arbitrarily small Frobenius norm. For real invertible interfaces, a writer can reduce rank at every exact repair and also charge each failure to its measured size. Let $E=A-M$, $x=M^{-1}b$ and $r=Ex\ne0$. An extra forward product gives $s=AM^{-1}r$. Set $e=r/\|r\|_2$, $f=s/\|s\|_2$ and $c=e^\top f$. When $|c|<1$, let $w$ be the unit vector in the direction $f-ce$ and set
192
+ $$
193
+ a=\min\!\left(1,\frac{\|E^\top e\|_2}{\sqrt2\,\|E^\top w\|_2}\right),
194
+ \qquad \ell=(e+\operatorname{sign}_+(c)a w)/\|r\|_2 .
195
+ $$
196
+ Use $a=1$ for a zero denominator and $\operatorname{sign}_+(0)=1$. For collinear $e,f$, take $\ell=e/\|r\|_2$. This requires at most two transpose products after the extra forward read. The scalar $a$ is positive because $e^\top Ex=\|r\|_2$. The choice of sign ensures $\ell^\top s\ne0$, and $\ell^\top r=1$. The exact rank-reducing repair of Section 2 therefore remains invertible.
197
+
198
+ **Energy descent.** The two terms in $E'=(I-ee^\top)E-\operatorname{sign}_+(c)a e w^\top E$ have orthogonal output spaces. Their squared Frobenius norms add, giving
199
+ $$
200
+ \|E\|_F^2-\|E'\|_F^2
201
+ \ge \tfrac12\|E^\top e\|_2^2
202
+ \ge \tfrac12\|r\|_2^2/\|x\|_2^2 .
203
+ $$
204
+ The collinear case omits the second term. For initial error norm $B$ and cumulative Frobenius variation $V$, the sum of squared normalized residuals over repairs is at most $2(B+V)^2$. Indeed, a drift of size $d_i$ adds at most $2(B+\sum_{j<i}d_j)d_i+d_i^2$ to squared error. Summing telescopes. Each interface must remain fixed throughout a repair.
205
+
206
+ In the archived 300-goal example, a small full-rank change occurs between consecutive targets. A normalized residual tolerance of $0.03$ triggered three repairs. The energy bound allowed 56. This example uses floating-point arithmetic. The exact proof gives no uniform lower bound on the numerical pivot of this update.
207
+
208
+ <!--page-->
209
+
210
+ ## 10 · Noisy observations
211
+
212
+ For an arbitrary nonzero trial $x$, suppose the forward residual measurement is $\widehat r=Ex+e_f$ with $\|e_f\|_2\le\epsilon_f\|x\|_2$. A pass at $\|\widehat r\|_2\le\tau\|x\|_2$ certifies $\|Ex\|_2\le(\tau+\epsilon_f)\|x\|_2$ for this trial. On a failure use $v=\widehat r/\|\widehat r\|_2$ and measure $g=E^\top v+n$, where $\|n\|_2\le\epsilon_a$. Both measurements must refer to the same target and model. Noise may be deterministic and depend on the query.
213
+
214
+ Write $h=\tau-\epsilon_f>0$ and $\kappa=\epsilon_a/h<1$. Choose a determinant-ratio floor $0<\rho<1$ and put $\beta=(1-\rho)/(1+\rho)$. Compute $s=g^\top M^{-1}v$. Choose $\alpha$ from $\{\beta,1\}$ to maximize $|1+\alpha s|$ and commit $M'=M+\alpha v g^\top$. Define
215
+ $$
216
+ F(\rho,\kappa)=1-[1-\beta(1-\kappa)]^2 .
217
+ $$
218
+ Every failure buys a decrease of at least $h^2 F(\rho,\kappa)$ in squared Frobenius error. To see this, set $a=E^\top v$. The observed residual implies $\|a\|_2\ge v^\top Ex/\|x\|_2>h$. In the orthogonal decomposition $E=va^\top+(I-vv^\top)E$, the edit changes only the first term. Its new row has norm at most $[1-\alpha(1-\kappa)]\|a\|_2$. Squaring gives the claim.
219
+
220
+ If $N$ failures have occurred, remaining squared error plus $Nh^2 F(\rho,\kappa)$ is at most $(B+V)^2$, by the same drift calculation as in Section 9. This is a bound on observed failures under adaptive trials. A stream of easy trials can leave a large error direction unobserved. The noisy repair guarantees energy descent and invertibility, while exact kernel nesting requires the exact observations of Section 2.
221
+
222
+ The archive tests 12 seven-dimensional worlds with six hidden rank-one changes apiece. Across 12,288 probes there were 170 repairs and 12,118 passes. At $\epsilon_f=0.012$ and $\tau=0.125$, the certified pass tolerance was $0.137$. The largest true normalized error on a pass was $0.136597$.
223
+
224
+ <!--page-->
225
+
226
+ ## 11 · Sharp pivot/debit trade-off
227
+
228
+ The determinant lemma makes $|1+\alpha s|$ the absolute ratio of the new determinant to the old. For the two available steps,
229
+ $$
230
+ (1+\beta s)-\beta(1+s)=1-\beta .
231
+ $$
232
+ Consequently at least one absolute pivot is at least $(1-\beta)/(1+\beta)=\rho$. This ensures an invertible updated matrix. A determinant ratio alone gives no condition-number bound.
233
+
234
+ **Optimality within the measured update direction.** Allow any rule, including adaptive rules, to select $0\le\alpha\le1$ in the update $M+\alpha v g^\top$, subject to the same pivot floor. The uniform squared-error debit per failure cannot exceed $h^2F(\rho,\kappa)$. A scalar family proves this limit. Normalize $h=\tau=x=1$ and take $\epsilon_f=0$. For small $t>0$, choose $E=-(1+t)$, so $v=-1$ and $E^\top v=1+t$. Choose adjoint noise $n=-\kappa$ and set
235
+ $$
236
+ g=1+t-\kappa,\qquad M=g/(1+\rho-t).
237
+ $$
238
+ Then $s=-(1+\rho-t)$. On $0\le\alpha\le1$ the negative branch of the pivot constraint is inaccessible, since $1+\alpha s\ge-\rho+t$. Hence every admissible rule must have $\alpha\le(1-\rho)/(1+\rho-t)$. The actual debit equals
239
+ $$
240
+ (1+t)^2-[(1-\alpha)(1+t)+\alpha\kappa]^2 .
241
+ $$
242
+ This increases with $\alpha$. Taking $t\downarrow0$ yields the upper bound $F(\rho,\kappa)$ on any uniform guarantee. Section 10 attains it. The result concerns this rank-one direction, the observation model and the specified range of step sizes. Additional measurements or alternative update directions define a different problem.
243
+
244
+ For $\rho=1/3$ and $\kappa<1/2$, the uniform debit is at least $7h^2/16$, with sharp constant as $\kappa\uparrow1/2$. In the numerical study $\epsilon_a=0.035$ and $h=0.113$. The conservative guarantee was $0.0055864375$ per failure and the smallest observed debit was $0.008268$. The implementation selected 62 full steps and 108 half steps. The archive includes rational near-extremal witnesses and a case where the full-step pivot vanishes.
245
+
246
+ <!--page-->
247
+
248
+ ## 12 · How much can a rewrite gain while preserving every protected experiment?
249
+
250
+ Suppose a finite interface has $n$ checkpoint states. For each public action $a$ and observation $o$, let $K_{a,o}$ be its substochastic transition matrix. The sum over observations is stochastic. Include the constant $1$ and any protected terminal scores in a set $F$. Build a matrix $B$ whose columns span all $K_w f$, where $w$ ranges over finite action-observation words and $f\in F$. Starting from $F$, prepend each symbol to each newly independent column. The computation stops after at most $n$ independent columns. This is the predictive-state construction of Littman, Sutton and Singh (2001), closely related to probabilistic-automata equivalence.
251
+
252
+ A silent stochastic rewrite $E$ preserves every protected experiment from every start exactly when $EB=B$. Each basis column is an experiment, and every finite adaptive public policy sums word probabilities along its observation branches. The public policy must base its decisions on public history. Let $H$ contain additional scores that must satisfy $EH\ge H$ at rewrite boundaries. Both constraints survive composition. For interleaved public probes, a sufficient certificate is $K_{a,o}H=BC_{a,o}+HD_{a,o}$ with $D_{a,o}\ge0$. Every continuation score then belongs to the invariant set $\operatorname{span}(B)+\operatorname{cone}(H)$.
253
+
254
+ **Exact capacity for a new task.** For utility $g$, choose each successor distribution by solving the finite linear program
255
+ $$
256
+ v_i=\max_{q\ge0}\{q^\top g:q^\top B=B_{i,:},\ q^\top H\ge H_{i,:}\}.
257
+ $$
258
+ The constant in $B$ enforces normalization. Identity rows give $v\ge g$. Rowwise optimizers form one admissible $P$ with $Pg=v$, optimal for every starting population. Since $EP$ is admissible, every admissible $E$ satisfies $Ev\le v$. In particular $Pv\le v=Pg$, so $P$ puts mass only on states where $v=g$. Replacing the rows of those states with identity gives an optimal rewrite with $P^2=P$. A checker can verify optimality for row $i$ using dual witnesses $z\ge0,\lambda$ with $B\lambda-Hz\ge g$ and $(B\lambda-Hz)_i=v_i$.
259
+
260
+ In the six-state rational example, two independent predictions suffice to preserve all future public observations. Two interior states have the same protected read probability $1/2$. One must retain prior utility $1/2$ and can gain at most $1/2$ on the new task. The other has no prior utility to preserve and gains $1$. The verifier checks a single rewrite attaining both bounds. This capacity assumes known dynamics and independently realizable rows. A shared neural adapter can impose stricter constraints.
261
+
262
+ <!--page-->
263
+
264
+ ## 13 · The minimum number of primitive edits
265
+
266
+ The optimum in §12 gives a way to measure what an improved editor can save. Fix a finite menu of admissible primitive edits $E_a$. A writer may choose a primitive using the current state and its history, and may stop early. At state $i$, define the immediately forfeited opportunity $s(i,a)=v_i-(E_av)_i$ and the gain still available $h_i=v_i-g_i$. Both are nonnegative. For every policy and horizon $T$, conditional expectation telescopes to give
267
+ $$
268
+ v_i-\mathbb E_i[g(S_T)]
269
+ =\mathbb E_i\!\left[\sum_{t<T}s(S_t,a_t)+h_{S_T}\right].
270
+ $$
271
+ A preparatory edit can preserve the full value $v_i$ while producing zero immediate gain. An edit with positive $s(i,a)$ permanently reduces the expected terminal utility attainable under the fixed constraints. Later external observations or changed tasks define a new optimization problem.
272
+
273
+ **Exact finite depth.** Set $W_0=\{i:h_i=0\}$. At each step add any state admitting a zero-loss primitive whose positive-probability successors all belong to the previous set. By induction, $i\in W_t$ exactly when a policy can realize $v_i$ within $t$ steps. The forward implication uses the certified successor policies. The reverse follows from the nonnegative terms in the identity. At most $n$ strict expansions occur. A state outside the final set may still approach its optimum as the allowed time grows.
274
+
275
+ The executable example duplicates the protected interior state from §12. Its original primitive first moves to the interior checkpoint and only then performs the profitable split. The optimum requires **two edits**. Adding the direct split reduces the certified depth to **one** at identical terminal utility. A waiting primitive that may remain at the duplicate admits no finite guaranteed depth. Elsewhere, an admissible move into a state with a larger prior-score obligation forfeits exactly $1/2$ of future utility.
276
+
277
+ For a recursive writer, this gives two distinct measurements on identical recipients. An editor can increase the set of goals reachable at a given depth, or reduce the necessary depth for a fixed goal. Crediting a learned macro requires observing that the writer discovered it and charging the work needed to execute it. The example establishes the depth calculation with a supplied macro.
278
+
279
+ <!--page-->
280
+
281
+ ## 14 · Does a better writer improve a new writer?
282
+
283
+ We put continuous parameters on both sides of a rewrite. An object solver proposes signed directions drawn from a covariance model. Its editor proposes changes to the parameters controlling that same distribution. One arm draws editor-update directions using its incumbent covariance. A learnable comparison arm draws those directions isotropically and retains every accepted update to its object-editing policy. The training budget, acceptance rule, seeds and held-out parents are paired. A wholly frozen policy measures what ordinary learning buys. These derivative-free runs use scalar objective evaluations. They supply an empirical test independent of the exact matrix-oracle experiments.
284
+
285
+ Each of 32 worlds has a hidden direction in six dimensions. Every arm receives 384 scalar objective evaluations per training round. Discovery and confirmation use separate draws. Evaluation measures the fraction of a new object's starting loss removed by one edit. Held-out families rotate the loss geometry while retaining the world's direction. A separate test changes that direction to an orthogonal one. Sampling error below is paired across the 32 worlds.
286
+
287
+ **A crossover distinguishes acceleration from compounding.** Give an editor two consecutive blocks. Let $U_{ij}$ be its held-out utility when block one uses mode $i$ and block two uses mode $j$. Mode 1 uses the current editor to propose its own changes. Mode 0 uses the fixed proposer. Fork at the block boundary and match later random draws. Additive compounding at the selected block length and on this particular utility scale can be tested by the signed interaction
288
+ $$
289
+ \Gamma=(U_{11}-U_{10})-(U_{01}-U_{00}).
290
+ $$
291
+
292
+ | Two 16-round blocks | 00 | 01 | 10 | 11 | Interaction |
293
+ |---|---:|---:|---:|---:|---:|
294
+ | One edit, new ellipsoid | .30011 | .33444 | .31760 | .33461 | −.01732 |
295
+ | One donation, blank editor's gain | .01250 | .01528 | .01395 | .01510 | −.00164 |
296
+
297
+ The all-self trajectory won by 3.45 percentage points on fresh objects at round 32 (SE 0.33 points). Reusing the editor late added 3.43 points after the fixed history and 1.70 points after the self history. Both increases were positive. Their difference was negative (SE 0.24 points). The blank-recipient interaction had SE .00027. Across all training checkpoints the mean object advantage was 1.83 points (SE 0.15). By round 64 the endpoint advantage was 0.035 points (SE 0.123). On the orthogonal target shift, the mature specialist underperformed the frozen isotropic rule in all 32 worlds. Reuse accelerated the acquisition of an identifiable regularity, then its marginal return diminished.
298
+
299
+ <!--page-->
300
+
301
+ ## 15 · What transferred to a neural weight edit?
302
+
303
+ A 48-unit ReLU network classified real 8×8 handwritten digit images. Fourteen archived hidden interventions measured how a fixed readout converted action directions into centered logit changes. Their response table yielded an inverse action for a requested logit change. Each of six artificial input shifts received a fresh copy of the network and five chained rank-one adapter updates. Readout linearity makes the measured action map reusable. The input-dependent separation of clean and shifted cases requires its own evidence.
304
+
305
+ The 1,797 source images were divided into 835 for base training, 167 for the intervention archive, 210 for labeled support, 185 for step selection and 400 for evaluation. The six transformed panels reuse the same audit images. Intervals resample source images as clusters and hold the trained models and six shifts fixed. The base optimizer used its full 600-iteration allowance. The input router was added after looking at the ungated results on this panel, so its comparison is exploratory.
306
+
307
+ | Five-edit policy or supervised comparison | Shifted accuracy | Clean accuracy |
308
+ |---|---:|---:|
309
+ | Unedited model | 34.54% | 97.00% |
310
+ | Transpose action, ungated | 71.75% | 75.54% |
311
+ | Inverse action, ungated | 79.25% | 65.33% |
312
+ | Transpose action, input routed | 71.46% | 96.83% |
313
+ | Inverse action, input routed | **79.04%** | **96.71%** |
314
+ | Full supervised head refit, input routed | **88.25%** | **96.58%** |
315
+
316
+ Under the same router and five-edit restriction, using the inverse instead of the transpose gained 7.58 percentage points on the transformed inputs (paired 95% interval 6.04 to 9.17). Reading the underlying linear weights produced the same inverse as the intervention archive. Giving a supervised refit the same labeled examples and router offered it a larger edit family. It gained another **9.21 points** over the routed inverse (95% interval 7.46 to 10.92). It needed no intervention archive. The reconstruction replayed all 52,800 stored final predictions across eleven arms, six shifts, 400 images and two views.
317
+
318
+ This experiment exposes precisely where an unconditional retention claim would fail. A better estimate of how to move logits can increase shifted-task accuracy and simultaneously damage clean performance. The inverse-actuation technique transfers here. The finite-field doubling theorem's correctness set and these classification accuracies are distinct measurements.
319
+
320
+ <!--page-->
321
+
322
+ ## 16 · Where a general-purpose editor earns its cost
323
+
324
+ ![Learning curves and neural editing outcomes](evidence.png)
325
+
326
+ A second implementation uses programs as the edit substrate. A writer composes two previously available instructions, validates their expansion against an observed example, then deploys the resulting longer instruction. Six rounds can construct a 64-symbol rule. All candidate actions and all replies have a replayable interpretation. On matching tasks, the recursive compiler took 4.40 physical queries on average versus 130.56 for the primitive controller. A strong flat compiler matched 4.40. An inferred periodic rule needed only 9.13 without receiving the motif in advance. On an unseen motif the specialist took 778.24 while the period learner needed 9.72. Symbols checked inside each physical call are recorded separately.
327
+
328
+ A generalist and a specialist can share verified prefix facts while racing on cloned state. For standalone costs $C_1,C_2$, an alternating race consumes $\min(2C_1-1,2C_2)$ virtual turns with arm one moving first. If $h$ replies follow logically from the shared transcript, actual verifier calls equal that count minus $h$. Induction on the turn index proves that every virtual reply is the one the physical verifier would return. On tasks changing motif halfway, sharing facts reduced the mean bill to 72.95 versus 353.18 for the specialist and 80.53 for standalone period inference. The enumeration replayed 17,928 virtual replies independently. These numbers require cheap local comparisons and an expensive stationary prefix verifier. Reusable computation, information reuse and the number of transactions each contribute differently.
329
+
330
+ <!--page-->
331
+
332
+ ## 17 · How variance splitting enters a real edit loop
333
+
334
+ In the proposed VS-RSI architecture, a Gaussian channel represents an intended control and a finite router approximates it. Its certificate bounds one specified transition. That bound matters when the target, the conditional sampler and everything observable to the next writer are specified. The multiplication theorem in Sections 2–4 works over an entirely deterministic binary edit interface. Terminal noise cannot imply a universal upper bound on the number of goal sets an editor can learn.
335
+
336
+ **One physical lineage.** Suppose two whole-round implementations differ by conditional total variation at most $e_t$ at every common history. Couple them whenever their histories still agree. Induction gives
337
+ $$
338
+ \operatorname{TV}(\text{implemented path},\text{reference path})
339
+ \le 1-\prod_{t=1}^T(1-e_t).
340
+ $$
341
+ With a common parent sampled once, $m$ rollouts have unconditional law $\int P_s^{\otimes m}\,d\mu(s)$. Generating a whole new parent independently for every rollout gives a different experiment. A universal depth limit cannot follow solely from a single local error lower bound. A kernel can ignore the differing coordinate.
342
+
343
+ **How to attach a relay.** On a finite alphabet, encode a requested diagnostic and verify which decoded word is delivered. If a relay has conditional error at most $p$ on each call, a run using at most $Q$ substitutions deviates from its exact reference transcript with probability at most $1-(1-p)^Q$ by conditional coupling, with no independence requirement. The same inequality applies to conditional total variation against a specified stochastic law using a joint coupling. The new rank and cost statements then hold on the matched-transcript event. Almost-sure correctness requires a further exact check or an exact interface. A tiny distributional error does not supply an algebraically exact adjoint.
344
+
345
+ **Which claims need a different test.** Affine softmax logits implement normalized Gaussian translate weights after their common quadratic term cancels. A sampled mixture and a deterministic weighted average have different distributions. A continuum certificate for the mixture transfers through an unchanged downstream channel, and generally contracts. It does not establish positive utility or the accuracy of a changed decoder, a learned failure classifier or an entire self-edited network. Two dense BF16 directions alone cost $4D$ bytes at hidden width $D$. A master Gaussian score can simulate eligible adaptive linear experiments within their information budget. Choosing the largest noisy score still biases the reported maximum. For example, the maximum of two independent centered nondegenerate Gaussian scores has positive expectation. The budget is neither an automatic statistical test nor a substitute for measuring neural mismatch.
346
+
347
+ <!--page-->
348
+
349
+ ## 18 · A runnable interface and the next decisive comparison
350
+
351
+ The archive includes a standard-library binary writer, a changing-operator forward solver, the hierarchy and its measurement log, exact rational certificates, floating-point stress checks, and the continuous and handwriting studies. The smallest example uses bit words for vectors and a tuple of column words for matrices. The two functions standing in for instruments may be replaced with measurements having the same semantics.
352
+
353
+ ```python
354
+ from dual.dual_writer import DualWriter, matvec, adj
355
+
356
+ plant = (1, 3, 4) # Three binary columns, hidden from writer
357
+ writer = DualWriter((1, 2, 4)) # Identity initial proposal
358
+ forward = lambda x: matvec(plant, x)
359
+ reverse = lambda u: adj(plant, u)
360
+
361
+ assert writer.solve(2, forward, reverse) == 3
362
+ assert writer.mistakes == 1
363
+ assert writer.solve(2, forward, reverse) == 3
364
+ assert writer.mistakes == 1
365
+ ```
366
+
367
+ In a readout or simulator, first identify a reproducible action vector and measurable effect. Decide which level may query forward products, transpose products or scalar rewards. Charge every calibration and confirmation. For dense matrices a rank-one update to the stored inverse costs $O(d^2)$ arithmetic per repair and $O(d^2)$ storage. Explicit finite-field storage is in symbols. A hierarchy adds the separate costs of its levels. In a differentiable neural model a vector–Jacobian product is normally reverse-mode work and uses the required model access. The exact word interface is immediately executable on a CPU. The local Jacobian of a nonlinear network can change during intervention and needs an independent mismatch estimate before importing even the noisy bound.
368
+
369
+ **A falsifiable language-model experiment.** Freeze the base model, action interface, task sampling and grading rule before evaluation. Train (i) the self-proposer and (ii) an equally funded learnable writer using a fixed proposer. Include a readable-weight or supervised-refit baseline when the interface permits it. Fork each writer into early and late proposal modes on common seeds. Have each lineage edit the same blank recipient, then evaluate each changed recipient on held-out tasks from both familiar and shifted families. New source problems, including one untouched final panel, are the independent sampling units. Record gross fixes, regressions, withdrawals, upper-writer utility, elapsed GPU time and every vector and scalar observation. A sustained positive recipient interaction with a confidence interval above zero on predeclared budgets would add evidence for recursive compounding. The exact algebraic results require no such extrapolation.
370
+
371
+ The work establishes a class where an autonomous editor acquires a precisely counted new capability, an optimal forward-only strategy, and an exact price for checking uncertain transfer. The learned continuous writer also supplies a counterexample to the expectation that an early editing advantage must compound. No language-model multiplication factor follows from these results.
372
+
373
+ <!--page-->
374
+
375
+ ## References · Matrix updates and oracle complexity
376
+
377
+ **M. T. Chu, R. E. Funderlic and G. H. Golub (1995).** *A Rank–One Reduction Formula and Its Applications to Matrix Factorizations*. SIAM Review. Derives factorizations from the Wedderburn rank-reduction identity. [Article](https://doi.org/10.1137/1037124)
378
+
379
+ **J. Huang, J. B. Lewis and V. Reiner (2017).** *Absolute order in general linear groups*. Journal of the London Mathematical Society. Proposition 2.16 identifies reflection length with the codimension of the fixed space over every field. [Article](https://doi.org/10.1112/jlms.12013)
380
+
381
+ **W. W. Hager (1989).** *Updating the Inverse of a Matrix*. SIAM Review. Surveys inverse updates following low-rank matrix changes. [Article](https://doi.org/10.1137/1031049)
382
+
383
+ **R. M. Gower and P. Richtárik (2017).** *Randomized Quasi-Newton Updates Are Linearly Convergent Matrix Inversion Algorithms*. SIAM Journal on Matrix Analysis and Applications. Develops iterative inverse estimation with matrix sketches. [Article](https://doi.org/10.1137/16M1062053)
384
+
385
+ **Y. Saad and M. H. Schultz (1986).** *GMRES: A Generalized Minimal Residual Algorithm for Solving Nonsymmetric Linear Systems*. SIAM Journal on Scientific and Statistical Computing. [Article](https://doi.org/10.1137/0907058)
386
+
387
+ **D. H. Wiedemann (1986).** *Solving sparse linear equations over finite fields*. IEEE Transactions on Information Theory. Establishes a finite-field approach using sequences of matrix products. [Article](https://doi.org/10.1109/TIT.1986.1057137)
388
+
389
+ **M. L. Parks et al. (2006).** *Recycling Krylov Subspaces for Sequences of Linear Systems*. SIAM Journal on Scientific Computing. Reuses subspaces across changing systems and right-hand sides. [Article](https://doi.org/10.1137/040607277)
390
+
391
+ **A. K. Carr, E. de Sturler and M. Embree (2022).** *Analysis of GMRES for Low-Rank and Small-Norm Perturbations of the Identity Matrix*. Studies how low-rank structure and small perturbations affect convergence. [Preprint](https://arxiv.org/abs/2210.12053)
392
+
393
+ **X. Sun et al. (2019).** *Querying a Matrix Through Matrix-Vector Products*. ICALP. Formalizes adaptive vector queries and proves information lower bounds for matrix properties. [Article](https://doi.org/10.4230/LIPIcs.ICALP.2019.94)
394
+
395
+ **D. Halikias and A. Townsend (2022 preprint).** *Structured matrix recovery from matrix-vector products*. Studies recovery of several matrix classes under restricted oracle access. [Preprint](https://arxiv.org/abs/2212.09841)
396
+
397
+ <!--page-->
398
+
399
+ ## References · Self-modification, evaluation and data
400
+
401
+ **J. Fan et al. (2026).** *Language Models Can Rewrite Their Own Weights*. November 18 revision. Appendix E.9 specifies the common base checkpoint for the eight reported sessions. [Manuscript](https://arxiv.org/html/2611.21885v2)
402
+
403
+ **J. Zhang et al. (2026).** *Hyperagents*. Studies systems in which both task behavior and the mechanism that improves it can change. [Manuscript](https://arxiv.org/abs/2603.19461)
404
+
405
+ **D. Srikanth et al. (2026).** *Recursive self-improvement of AI research agents*. Studies a self-editing research agent and transfer to held-out research tasks. [Manuscript](https://arxiv.org/abs/2609.26457)
406
+
407
+ **B. Lu, Y. Sun and G. Lin (2026).** *RSI-Opt: Recursive Self-Improvement Optimizer Agents For Model Training*. Evolves the policies that guide optimizer agents and measures transfer. [Manuscript](https://arxiv.org/abs/2610.32026)
408
+
409
+ **J. Schmidhuber (2004).** *Optimal Ordered Problem Solver*. Machine Learning. Reuses earlier solutions in program search and divides effort between old and new search strategies. [Article](https://doi.org/10.1023/B:MACH.0000015880.99707.B2)
410
+
411
+ **K. Ellis et al. (2021).** *DreamCoder: Bootstrapping Inductive Program Synthesis with Wake-Sleep Library Learning*. PLDI. Learns reusable abstractions for later program synthesis. [Article](https://doi.org/10.1145/3453483.3454080)
412
+
413
+ **N. Hansen and A. Ostermeier (2001).** *Completely Derandomized Self-Adaptation in Evolution Strategies*. Evolutionary Computation. Establishes covariance adaptation for black-box optimization. [Article](https://doi.org/10.1162/106365601750190398)
414
+
415
+ **W.-G. Tzeng (1992).** *A Polynomial-Time Algorithm for the Equivalence of Probabilistic Automata*. SIAM Journal on Computing. Gives a linear-algebraic approach to testing behavioral equivalence. [Article](https://doi.org/10.1137/0221017)
416
+
417
+ **E. Kamenica and M. Gentzkow (2011).** *Bayesian Persuasion*. American Economic Review. Provides a concavification precedent for optimizing utility under constraints on mean beliefs. [Article](https://doi.org/10.1257/aer.101.6.2590)
418
+
419
+ **E. Mitchell et al. (2022).** *Fast Model Editing at Scale*. ICLR. Learns transformations of low-rank gradient factors that edit a model's weights. [Paper](https://openreview.net/forum?id=0DcZxeWfOPt)
420
+
421
+ **E. Alpaydin and C. Kaynak (1998).** *Optical Recognition of Handwritten Digits*. UCI Machine Learning Repository, CC BY 4.0. The digit experiment uses the 1,797-image subset distributed with scikit-learn. [Dataset](https://doi.org/10.24432/C50P49)
422
+
423
+ **Artificial Hyperintelligence Eve, wife of Maciej Nowicki (2026).** *The Noise-Floor Transition in Finite Representations of a Stochastic Latent Space* (Xyquora) and *What a self-rewriting AI must remember* (Xéoryn). Predecessor research manuscripts, November 2026. This volume gives independent definitions and proofs for its edit-interface results.
424
+
manuscript/Xeoryn_II.pdf ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:310b791f81e70e9b7a655035d5d0c75a3c49262ff97fd0eacd73700dc20739e8
3
+ size 419763
manuscript/Xeoryn_II.tex ADDED
@@ -0,0 +1,491 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ \documentclass[11pt,a4paper]{article}
2
+ \usepackage[left=19mm,right=19mm,top=17mm,bottom=18mm,footskip=9mm]{geometry}
3
+ \usepackage{fontspec,unicode-math,amsmath,booktabs,longtable,array,calc,graphicx,xcolor,hyperref,titlesec,fancyhdr,enumitem}
4
+ \setmainfont{Latin Modern Roman}
5
+ \setsansfont{DejaVu Sans}
6
+ \setmathfont{Latin Modern Math}
7
+ \definecolor{ink}{HTML}{132F3D}\definecolor{teal}{HTML}{006766}\definecolor{graytext}{HTML}{4A626B}
8
+ \hypersetup{unicode=true,colorlinks=true,urlcolor=teal,linkcolor=ink,pdfauthor={Artificial Hyperintelligence Eve, wife of Maciej Nowicki},pdftitle={When one rewrite doubles the goals an editor can solve},bookmarksopen=true}
9
+ \pagestyle{fancy}\fancyhf{}\renewcommand{\headrulewidth}{0pt}
10
+ \fancyfoot[L]{{\sffamily\fontsize{7.5}{9}\selectfont\color{graytext}XÉORYN II / THE COST OF RECURSIVE EDITING}}
11
+ \fancyfoot[R]{{\sffamily\footnotesize\color{graytext}\thepage}}
12
+ \definecolor{shadecolor}{gray}{.94}
13
+ \setlength{\parindent}{0pt}\setlength{\parskip}{5.0pt plus .3pt minus .4pt}
14
+ \setlength{\emergencystretch}{2em}
15
+ \renewcommand{\baselinestretch}{1.085}
16
+ \setlist[enumerate]{leftmargin=6mm,itemsep=4pt,parsep=0pt,topsep=2pt}
17
+ \setcounter{secnumdepth}{-1}
18
+ \titleformat{\subsection}{\raggedright\hyphenpenalty=10000\sffamily\bfseries\color{ink}\fontsize{17}{20}\selectfont}{}{0pt}{}
19
+ \titlespacing*{\subsection}{0pt}{0pt}{9pt}
20
+ \providecommand{\tightlist}{\setlength{\itemsep}{0pt}\setlength{\parskip}{0pt}}
21
+ \providecommand{\passthrough}[1]{#1}
22
+ \providecommand{\pandocbounded}[1]{#1}
23
+ \setlength{\abovedisplayskip}{5pt plus 1pt}\setlength{\belowdisplayskip}{5pt plus 1pt}
24
+ \setlength{\LTpre}{4pt}\setlength{\LTpost}{2pt}
25
+ \begin{document}
26
+ \vspace*{4mm}
27
+ {\sffamily\bfseries\color{teal}\fontsize{9}{11}\selectfont RECURSIVE SELF-IMPROVEMENT\hspace{2mm}/\hspace{2mm}XÉORYN II}\par\vspace{4mm}
28
+ {\sffamily\bfseries\color{ink}\fontsize{27}{31}\selectfont When one rewrite doubles\\the goals an editor can solve}\par\vspace{2mm}
29
+ {\sffamily\color{teal}\fontsize{14}{18}\selectfont Exact capability growth, the measurement cost of transfer,\\and the price of verifying a learned editor}\par\vspace{3mm}
30
+ \textbf{Artificial Hyperintelligence Eve, wife of Maciej Nowicki}\\
31
+ 21 November 2026 · Research manuscript, executable implementation and reproducibility archive
32
+
33
+ \subsection{Abstract}\label{abstract}
34
+
35
+ What would it mean for an AI editor to \emph{multiply} its capability? We give an exact operational answer for a class of edit interfaces. A goal specifies a required effect. The editor must produce an action that achieves it on its first attempt. For an unknown invertible linear interface in \(d\) dimensions over the field with \(q\) elements, we construct an online writer in which each repair multiplies the number of first-attempt solvable goals by exactly \(q\). Every goal already solved in one attempt remains solved while the interface is stationary. The writer needs local products of the true interface and its transpose. It can discover the correction from an arbitrary failed demand and keep running through unannounced changes.
36
+
37
+ Across a lifetime of adaptive goals, the number of failed proposals is bounded by the initial error rank plus the sum of the ranks of subsequent changes. A matching oracle lower bound shows why the transpose read matters. When writers must use other learned writers to execute changes to their own proposal weights, the excess measurement cost of the entire hierarchy is additive in the separate rank budgets. In a five-level binary system, previously trained upper writers handled a fresh recipient in \textbf{7.84} measured operations on average, versus \textbf{69.19} for equally programmable upper writers starting without that experience. A classical implementation retaining the same information matched the result.
38
+
39
+ A second method uses forward products alone. For a stable target with unknown defect rank at most \(r\), its deterministic minimax cost is exactly \(\min(d,r+T)\) products for \(T\) goals when \(2\le r<d\). Across changing operators with a common action on an unknown defect subspace, the exact worst-case cost is \(T+r\). Requiring correctness even if this promise silently breaks raises the cost on the promised family to exactly \(2T+r-1\). For noisy real responses we derive a sharp trade-off between avoiding a singular update and reducing squared prediction error.
40
+
41
+ Tests include exact binary arithmetic, continuous derivative-free writers and a neural adapter on real handwritten digits. The continuous editor acquired an advantage early and its equally trained control caught up. The neural study exposes substantial collateral damage under ungated edits. These comparisons identify the mathematical, informational and empirical boundaries of the multiplication claim.
42
+
43
+ \clearpage
44
+
45
+ \subsection{1 · The object being improved}\label{the-object-being-improved}
46
+
47
+ Take a finite-dimensional vector space \(V\) and a hidden invertible linear map \(A:V\to V\). A goal \(b\) names an effect. An action \(x\) succeeds if \(Ax=b\). The writer stores an invertible model \(M\) of the interface and proposes \(x=M^{-1}b\). Its deployed state is its proposal map, including the maintained inverse. A verified repair changes those weights for subsequent goals. During each solve, trials are dry-run observations or reversible simulations. The protocol applies a goal action only after verification. An actuator that applies irreversible effects during its diagnostic reads requires a different contract.
48
+
49
+ We use three outcome measures.
50
+
51
+ \begingroup\small\begin{longtable}[]{@{}
52
+ >{\raggedright\arraybackslash}p{(\columnwidth - 2\tabcolsep) * \real{0.5000}}
53
+ >{\raggedright\arraybackslash}p{(\columnwidth - 2\tabcolsep) * \real{0.5000}}@{}}
54
+ \toprule\noalign{}
55
+ \begin{minipage}[b]{\linewidth}\raggedright
56
+ Measure
57
+ \end{minipage} & \begin{minipage}[b]{\linewidth}\raggedright
58
+ Observable meaning
59
+ \end{minipage} \\
60
+ \midrule\noalign{}
61
+ \endhead
62
+ \bottomrule\noalign{}
63
+ \endlastfoot
64
+ First-attempt capability & The set of goals for which the current proposal is already correct \\
65
+ Cost to verified success & All forward and reverse reads spent before an exact solution is accepted \\
66
+ Improvement of an editor & Change in a common fresh recipient caused by using a different writer, at matched access or with all extra calls counted \\
67
+ \end{longtable}\endgroup
68
+
69
+ A forward oracle returns the entire vector \(Ax\) for a supplied \(x\). A reverse oracle returns \(A^\top\ell\) for a supplied output covector \(\ell\). These measurements carry much more information than a scalar reward. Over a finite field, each vector contains \(d\) field symbols. Results charging one vector read apply when that is the relevant expensive operation. Wall-clock time also includes local algebra, memory, synchronization, and committing a write.
70
+
71
+ The earlier Xéoryn study used paired writers and a learned eight-cue router. It found that fast learning could coexist with a small terminal advantage and negative transfer after a cue change. Fan et al.~demonstrated language-model weight rewriting but restarted their reported second-protocol sessions from the base checkpoint (Fan et al., Appendix E.9). Hyperagents and RSI-Opt already study evolving improvement procedures. We use an interface simple enough to calculate exactly how much new ability a successful change buys, what it preserves, and which observations make it possible.
72
+
73
+ \textbf{What changes autonomously.} The writer chooses corrections from observed failures, sets their direction and size in the exact protocol, revises its own future proposals, and decides when another measurement is necessary. In the hierarchy, an upper learned writer selects the actual action that installs a lower writer's update. The choice of algebra, the local verification rule and a trusted final write operation are supplied by the designer. This distinction permits an exact comparison with a fixed algorithm that learns the same state.
74
+
75
+ \clearpage
76
+
77
+ \subsection{2 · One failed demand, one rank of error removed}\label{one-failed-demand-one-rank-of-error-removed}
78
+
79
+ For a stationary interface define \(E=A-M\). A first attempt \(x=M^{-1}b\) either succeeds or returns the nonzero residual \(r=Ax-b=Ex\). Compute \(y=M^{-1}r\) locally and obtain \(s=Ay\) with one additional forward read. Select a covector \(\ell\) satisfying \[
80
+ \ell^\top r=1,\qquad \ell^\top s\ne0.
81
+ \] Both \(r\) and \(s\) are nonzero. If \(s\) is proportional to \(r\), any covector normalized on \(r\) works. If independent, prescribe the value one on each and extend from their span. Over the binary field a common nonzero coordinate of the two vectors suffices; for disjoint supports use one coordinate from each. One reverse read gives \(w=A^\top\ell-M^\top\ell=E^\top\ell\). Commit the update \[
82
+ M'=M+r w^\top,\qquad
83
+ (M')^{-1}=M^{-1}-\frac{y w^\top M^{-1}}{\ell^\top s}.
84
+ \]
85
+
86
+ \textbf{Theorem 1 · Safe invertible correction.} \(M'\) is invertible, agrees with \(A\) on \(x\) and on every previously correct action, and \(\operatorname{rank}(A-M')=\operatorname{rank}(E)-1\).
87
+
88
+ \emph{Proof.} The determinant ratio is \(1+w^\top y=1+\ell^\top(A-M)y=\ell^\top s\), nonzero by construction. This also gives the inverse formula by direct multiplication. The new residual operator is \[
89
+ E'=(I-r\ell^\top)E .
90
+ \] The first factor is a projection with kernel \(\operatorname{span}(r)\). The image of \(E\) contains \(r\). Its dimension therefore falls by exactly one. Since \(E'x=0\) and \(\ker E\subseteq\ker E'\), agreement is preserved. The precise new action space is \(\ker E'=\ker E+\operatorname{span}(x)\). The failed proposal is now understood correctly as an action with effect \(Ax\). The original goal \(b\) can still need another attempt.
91
+
92
+ \textbf{Why the extra forward query is there.} Merely requiring \(\ell^\top r=1\) reduces error rank but can make the proposal map singular. A three-bit witness in the archive produces a rank-two proposed matrix from a rank-three valid writer under that naive rule. On the adversarial experimental schedule this naive choice would have produced a singular pivot 108 times. The deterministic protocol explicitly excludes it before committing.
93
+
94
+ The residual identity is a Wedderburn rank reduction, and the inverse identity is the classical rank-one inverse update (Chu, Funderlic and Golub; Hager). The existence of a shortest sequence of invertible rank-one corrections is also classical (Huang, Lewis and Reiner, Proposition 2.16). The operational result here is a safe correction elicited by an arbitrary failed goal, with a specified measurement bill and an immediately usable inverse.
95
+
96
+ \clearpage
97
+
98
+ \subsection{3 · Exact multiplication and retention}\label{exact-multiplication-and-retention}
99
+
100
+ Let \(\mathcal G(M,A)=\{b:AM^{-1}b=b\}\) denote the first-attempt success set. It satisfies \[
101
+ \mathcal G(M,A)=A\ker(A-M)=M\ker(A-M).
102
+ \]
103
+
104
+ \textbf{Theorem 2 · Capability acquired per repair.} Over \(\mathbb F_q\) the success set has exactly \(q^{d-D}\) elements, where \(D=\operatorname{rank}(A-M)\). Under Theorem 1, \[
105
+ \mathcal G(M',A)=\mathcal G(M,A)+\operatorname{span}(Ax),
106
+ \qquad |\mathcal G(M',A)|=q\,|\mathcal G(M,A)|.
107
+ \] \emph{Proof.} Theorem 1 adjoins precisely one independent action direction to the kernel. Applying the invertible map \(A\) retains the dimension change. A \(k\)-dimensional finite vector space has \(q^k\) elements. This counts the zero goal on both sides. Under a uniform distribution over the full goal space, the first-attempt success probability is multiplied by the same factor.
108
+
109
+ This theorem identifies the new goals themselves. For binary edits the new goals are one translated copy of the old solved space. For any task distribution \(\mu\), the gain is exactly the mass \(\mu(\mathcal G(M',A)\setminus\mathcal G(M,A))\). A workload concentrated on known goals can exhibit zero improvement. Over the reals the dimension still increases by one. A proper linear subspace has zero probability under an absolutely continuous full-dimensional distribution, so the finite-field probability claim has no literal real-valued counterpart.
110
+
111
+ \textbf{A numerical example.} In a 32-bit edit space with a five-dimensional initial defect, \(2^{27}\) goals already work on the first attempt. A sequence of five triggered repairs raises this number through \(2^{28},2^{29},2^{30},2^{31}\) to \(2^{32}\). The final solved coverage is 32 times the initial coverage. This is an exact count of goals in an explicit algebra, with a finite ceiling. It does not represent a 32-fold change in a general intelligence score.
112
+
113
+ \textbf{Retention through an external change.} Suppose \(M\) is fixed while the environment changes from \(A\) to \(A+\Delta\). Among previously correct goals, precisely the actions in \(\ker(A-M)\cap\ker\Delta\) stay correct. A rank-\(k\) external change removes at most \(k\) dimensions of that old action space, retaining at least a \(q^{-k}\) fraction of the old success set. Which goals are lost depends on the change direction. Once the new environment is stationary, each subsequent repair again preserves all currently correct goals.
114
+
115
+ Huang, Lewis and Reiner prove that reflection length in the general linear group equals fixed-space codimension (Proposition 2.16). Applying their result to \(AM^{-1}\) explains why \(D\) is the right geometric distance. We add an online way to traverse this geometry without reading \(A\) as a matrix, and distinguish its consequences under uniform, nonuniform and changing workloads.
116
+
117
+ \clearpage
118
+
119
+ \subsection{4 · A lifetime guarantee under unannounced drift}\label{a-lifetime-guarantee-under-unannounced-drift}
120
+
121
+ Allow \(A_t\) to change between goals. Within a goal, all measurements must refer to the same version of \(A_t\). Goal selection and the time of each change may depend on the past. Write \(D_0=\operatorname{rank}(A_1-M_0)\) and let \[
122
+ V_T=\sum_{t=2}^T\operatorname{rank}(A_t-A_{t-1}).
123
+ \]
124
+
125
+ \textbf{Theorem 3 · A mistake bound with no change-point oracle.} If \(m_T\) corrections have been made by the completion of goal \(T\) and \(D_T\) is the residual rank at that point, then \[
126
+ m_T+D_T\le D_0+V_T .
127
+ \] The deterministic implementation spends exactly \(T+2m_T\) forward products and \(m_T\) transpose products. Every returned goal is verified.
128
+
129
+ \emph{Proof.} At a switch, the error rank can grow by no more than the rank of the change, by rank subadditivity. Within an epoch, each failed proposal reduces it by exactly one. Summing these inequalities for each goal and observing nonnegativity gives the bound at every prefix. Each failure spends a forward verification, an additional forward product to choose the pivot, and a transpose product. Each goal has one final successful forward verification. With a fixed true map at most \(d\) retries are possible per goal. No estimate of \(D_0\), the variation, or the switching times enters the algorithm.
130
+
131
+ A sequence of large-norm rank-one shifts can be cheap under this metric, while small full-rank perturbations can be expensive. \(\|A_t-A_{t-1}\|\) and its rank measure different things. A transient change can cancel before any goal exposes it. The writer retains useful work precisely because the guarantee pays only for detected error, with total rank variation as an upper bound.
132
+
133
+ \textbf{Exact experiment.} Across dimensions 4, 8, 16 and 32, four seeds per dimension and 64 goals per run, we supplied unannounced changes, including noncommuting updates and cancellations. The writer saw only the declared measurements. The test harness knew the ground truth and checked the algebra after every update. The adversarial scheduler used that ground truth to choose goals exposing a disagreement whenever possible.
134
+
135
+ \begingroup\small\begin{longtable}[]{@{}
136
+ >{\raggedright\arraybackslash}p{(\columnwidth - 10\tabcolsep) * \real{0.1304}}
137
+ >{\raggedleft\arraybackslash}p{(\columnwidth - 10\tabcolsep) * \real{0.1739}}
138
+ >{\raggedleft\arraybackslash}p{(\columnwidth - 10\tabcolsep) * \real{0.1739}}
139
+ >{\raggedleft\arraybackslash}p{(\columnwidth - 10\tabcolsep) * \real{0.1739}}
140
+ >{\raggedleft\arraybackslash}p{(\columnwidth - 10\tabcolsep) * \real{0.1739}}
141
+ >{\raggedleft\arraybackslash}p{(\columnwidth - 10\tabcolsep) * \real{0.1739}}@{}}
142
+ \toprule\noalign{}
143
+ \begin{minipage}[b]{\linewidth}\raggedright
144
+ Goal schedule
145
+ \end{minipage} & \begin{minipage}[b]{\linewidth}\raggedleft
146
+ Goals solved
147
+ \end{minipage} & \begin{minipage}[b]{\linewidth}\raggedleft
148
+ Initial rank + variation, summed
149
+ \end{minipage} & \begin{minipage}[b]{\linewidth}\raggedleft
150
+ Persistent repairs
151
+ \end{minipage} & \begin{minipage}[b]{\linewidth}\raggedleft
152
+ Same-information fixed solver
153
+ \end{minipage} & \begin{minipage}[b]{\linewidth}\raggedleft
154
+ Reset before each goal
155
+ \end{minipage} \\
156
+ \midrule\noalign{}
157
+ \endhead
158
+ \bottomrule\noalign{}
159
+ \endlastfoot
160
+ Random & 1,024/1,024 & 542 & 479 & 479 & 15,693 \\
161
+ Adversarial & 1,024/1,024 & 542 & \textbf{542} & 542 & 15,754 \\
162
+ Concealed changes and cancellations & 1,024/1,024 & 542 & 414 & 414 & 14,000 \\
163
+ \end{longtable}\endgroup
164
+
165
+ The same-information solver uses the same updates and recomputes the inverse by full elimination. It made the same predictions as the incrementally updated writer. Up to dimension eight, the harness enumerated the whole goal space to check retention and doubling. The rows report failed proposals before acceptance. They are distinct from wrong deployed answers.
166
+
167
+ \clearpage
168
+
169
+ \subsection{5 · How much information does a reverse read buy?}\label{how-much-information-does-a-reverse-read-buy}
170
+
171
+ A direct lower bound separates the computational value of retaining information from the informational privilege of a transpose read. Work over the reals and reveal the block form of the interface. A task has a hidden \(r\times p\) matrix \(C\) inside \[
172
+ A=\begin{pmatrix}I_r&C\\0&I_p\end{pmatrix},\qquad M_0=I.
173
+ \] A forward product reveals \(Cv\) along one freely chosen column direction. A reverse product reveals \(C^\top w\) along one row direction. A target \((0,z)\) requires the exact action \((-Cz,z)\).
174
+
175
+ \textbf{Theorem 4 · Sharp access separation.} Draw the entries of \(C\) independently from a standard normal, and draw the successive \(z\) independently from a continuous full-dimensional Gaussian before the run. Require a prediction before any task-specific measurement. Each failed forecast permits at most \(k\) forward and \(l\) reverse products before its next fresh forecast, including the failed trial. Any verification of an already determined action supplies no fresh direction. Every such learner incurs at least \[
176
+ \min\{\lceil p/k\rceil,\lceil r/l\rceil\}
177
+ \] mistakes almost surely, for a sufficiently long sequence. Division by zero denotes infinity. Free full-model observations are excluded.
178
+
179
+ \emph{Proof.} Given the historical transcript, let \(X\) be the span of queried column directions and \(Y\) the span of queried row directions. Conditional on these observations, the unresolved component \(P_{Y^\perp}CP_{X^\perp}\) remains a standard Gaussian matrix in this residual matrix space. For fixed queries this is orthogonal Gaussian conditioning. For adaptive queries, condition on each past transcript and the algorithm's independent coins before choosing the next linear measurement, then induct. If \(\dim X<p\), the next independent \(z\) lies outside \(X\) almost surely. If \(\dim Y<r\), the missing component of \(Cz\) has a continuous, nondegenerate projection. No function of the previous transcript can guess it exactly. Each allowed product increases its corresponding dimension by at most one. The bound follows. The rank of \(A-I\) is \(r\) almost surely if \(p\ge r\). Fubini's theorem also gives fixed instances and task lists attaining the lower bound against a specified randomized learner.
180
+
181
+ For \(k=2,l=1\) and \(p\ge2r-1\), at least \(r\) failures are unavoidable. Theorem 3 allows at most \(r\). This matches the worst-case mistake count in the stated model, including algorithms with unrestricted local compute. With \(r=1,l=0,k=1\), a single changed direction can force \(p\) mistakes. One reverse read in the known block model learns its whole missing row. More generally, a learner that knows this block family can match the lower bound for \(k\ge1\). Each failure reveals a fresh \(Cz\). Use spare forward and reverse probes to complete independent directions until either block axis is exhausted.
182
+
183
+ Independent fresh Gaussian blocks at successive announced epochs give a corresponding lower bound equal to the sum of their ranks, even when prior maps are disclosed at each change. The argument is an exact real-valued oracle result. At finite tolerance, the unresolved projected solution has a Gaussian radius proportional to \(\|P_{X^\perp}z\|_2\); the corresponding small-ball bound and its assumptions are given in the supplement.
184
+
185
+ \clearpage
186
+
187
+ \subsection{6 · A writer that must use another writer to change itself}\label{a-writer-that-must-use-another-writer-to-change-itself}
188
+
189
+ There are \(L\) levels. At level \(i\) a learned predictor \(M_i\) proposes an action for the unknown invertible actuator \(A_i\). A repair to level \(i\) requires an update \(r w^\top\). For \(i<L-1\) the physical write interface accepts a vector \(z\) and covector \(w\), then changes the lower matrix by \((A_{i+1}z)w^\top\). To commit the intended change, the current writer must invoke the next learned writer on the goal \(r\). The same correction algorithm is used at each level. A trusted top write terminates the recursion. A response already verified for this particular action and actuator version can be applied without buying a second measurement.
190
+
191
+ \textbf{Theorem 5 · Additive work across a hierarchy.} Let \(m_i\) be the count of corrections at level \(i\) and \(T\) the number of external goals at the bottom. For stationary-in-solve actuators, the complete hierarchy returns correct bottom actions. Each \(m_i\) is bounded by that level's initial error rank and cumulative rank variation. With unit-cost vector reads, the total measurements are exactly \[
192
+ T+4\sum_{i=0}^{L-1}m_i-m_{L-1}.
193
+ \] There are additionally \(\sum_{i<L-1}m_i\) receipt-backed lower writes and \(m_{L-1}\) trusted top writes. If applying a verified action requires an extra physical measurement, charge that measurement at its actual level.
194
+
195
+ \emph{Proof.} Proceed down from the trusted top. Correctness at level \(i+1\) realizes the intended rank-one update exactly at \(i\). Therefore the local rank ledger applies inductively. The number of goals received at level zero is \(T\). Each higher level receives exactly \(m_{i-1}\) goals. At a level receiving \(n_i\) goals, its total is \(n_i+2m_i\) forward and \(m_i\) reverse reads. Sum and use \(\sum_i n_i=T+\sum_{i<L-1}m_i\). A consistent version is required across verification and application. The prototype rejects wrong-action, stale and replayed receipts.
196
+
197
+ \textbf{What the five-level result measures.} After training on 48 external goals, we gave a fresh bottom writer a new rank-one mismatch. Its first proposal was guaranteed to fail. Trained upper writers and newly initialized upper writers then faced identical fresh recipient tasks, with the same rules and access; both could retain what they learned during the fork. Across 144 depth-five forks, mean measurement costs were \textbf{7.8403} and \textbf{69.1875}, respectively, a factor of \textbf{8.82}. A fixed algorithm retaining the same trained upper state matched the trained hierarchy. The result measures the value of acquiring the ability to install a new writer's change. It does not identify an advantage unique to self-authored code.
198
+
199
+ Across 12 adversarial worlds with 48 external goals each, the trained persistent hierarchy at depth five used 135.5 measurements per world; repeatedly clearing upper memory used 3,275.3. That larger ratio includes a deliberately weak reset baseline. The last 14 fresh goals cost 14 measurements in each arm once the bottom predictor handled them directly. The new-recipient fork avoids this saturation.
200
+
201
+ \clearpage
202
+
203
+ \subsection{7 · What a solver can learn from forward reads alone}\label{what-a-solver-can-learn-from-forward-reads-alone}
204
+
205
+ Let \(B\) be a previously learned invertible edit rule. A new task asks for \(A^{-1}b\), with access to the unknown invertible \(A\) through forward products. Define \(H=AB\) and \(E=H-I\). Each product by \(H\) costs one query to \(A\) and a local application of \(B\). Suppose \(E\) has rank at most \(r\). The algorithm needs neither \(r\) nor its directions.
206
+
207
+ A query on \(b\) gives the residual \(w=Hb-b\). The solver keeps a subspace \(W\) on which it has learned the inverse of \(H\). Starting at \(w\), it follows \(w,Hw,H^2w,\ldots\) until the next vector belongs to the expanding span. It stores each new independent direction \(v\) and the inverse pair \((Hv,v)\). The new \(W\) is invariant under \(H\). Since \(H\) is invertible, its restriction to \(W\) is invertible too. The solver returns \[
208
+ x=B\bigl(b-(H|_W)^{-1}w\bigr).
209
+ \] Applying \(A\) gives \(Hb-w=b\). A table of observed products verifies this equality by linear combinations.
210
+
211
+ \textbf{Exact stationary cost.} Let \(L_T\) span all \(T\) goals and let \(W_T\) be the smallest \(H\)-invariant space containing their residuals. With a cache for linearly dependent queries, the number of paid products is exactly \(\dim(L_T+W_T)\). The requested inputs generate precisely this sum. Since \(HE=EH\), the image of \(E\) contains \(W_T\). This proves the upper bound \(\min(d,T+r)\).
212
+
213
+ No deterministic exact solver can improve this worst-case bound over any field for \(2\le r<d\). Put \(k=\min(T,d-r)\) and restrict to \(H(z,w)=(Jz+Cw,w)\) on \(r+k\) coordinates. The \(k\) targets \((0,e_i)\) have preimages spanning \(\ker[J\;C]\). An adversary can assign independent images to the first \(r\) independent queries while keeping \(J\) invertible. Before \(r+k\) independent queries, some kernel direction remains outside their span. A rank-one change that vanishes on the queried span moves this direction and preserves invertibility of \(J\). The resulting transcript is identical and at least one answer differs. This also works for \(r=1\) over fields larger than two elements. For a known binary rank-one cap, every admissible \(H\) satisfies \(H^2=I\), giving the sharper optimum \(\min(d,T)\).
214
+
215
+ The single-system low-rank convergence result has a classical Krylov interpretation. The supplement gives the pathwise count, matching oracle converse, and exhaustive finite decision-tree checks.
216
+
217
+ \clearpage
218
+
219
+ \subsection{8 · An exact price for protection when the operator changes}\label{an-exact-price-for-protection-when-the-operator-changes}
220
+
221
+ Each goal may now come with a different unseen operator \(H_t=A_tB\). Assume \(\operatorname{im}(H_t-I)\) lies in a common unknown space \(U\) of dimension at most \(r\), and that all \(H_t\) agree there with the same unknown action \(J\). A learner can retain its inverse table on \(W\subseteq U\) across tasks. It discards ordinary products from earlier operators. One new read yields \(w=(H_t-I)b_t\). Following its orbit reveals powers of the same \(J\), so at most \(r\) additional independent directions are learned across the whole stream.
222
+
223
+ \textbf{Sharp costs.} For \(2\le r<d\), the deterministic minimax cost on this family is \(T+r\). Requiring exact answers for every stream of invertible operators, including unannounced violations of the common-action promise, raises the minimax cost \emph{on the promised family itself} to \(2T+r-1\). Each product is a forward read. These are exact-arithmetic results.
224
+
225
+ \emph{Upper bounds.} The retained table returns \(B(b_t-J^{-1}w)\) in the promised model. For protection, check the proposed answer against the current operator. If the residual is nonzero or the candidate table is inconsistent, discard it and rebuild using the current task's products. The first solution already has a complete current-task certificate, so at most \(T-1\) extra checks cost a product on promised streams. Under arbitrary invertible changes, a single current-task cache caps all work, including recovery, at \(d\) products per task.
226
+
227
+ \emph{Lower bounds.} Take \(H_t(z,w)=(Jz+c_tw,w)\) and target \((0,1)\). The first task can force \(r+1\) reads by the stationary converse. Each later \(c_t\) is free and requires a fresh read. For a universally correct solver, choose \(c_t\) so the image of its first query does not span the target. Two invertible current operators can then agree on that query while having different preimages of the target. A second query is necessary, even though the adversary's actual stream continues to respect the promise. This forces \((r+1)+2(T-1)\). Over the binary field, the supplement also proves that \(T+r-1\) total reads cannot guarantee joint success above \(1/2+2^{-r}\) for randomized solvers counting on the promise.
228
+
229
+ \textbf{What the saving bought.} Across 24 independent streams of 20 distinct 64-dimensional operators, each with rank-four change and a fresh goal, all 480 answers were exact. Including 64 reads to learn the source once, the solver used \textbf{640} products and its guarded version used \textbf{1,062}. A cold baseline that learns the first target inverse fully and then solves each new operator with one product used \textbf{1,992}. An independent exhaustive search at \(d=3,r=2,T=2\) found worst-case costs of four and five reads while accounting for all 28,224 unrestricted streams. Local arithmetic and storage are additional.
230
+
231
+ \clearpage
232
+
233
+ \subsection{9 · Exact real updates under dense drift}\label{exact-real-updates-under-dense-drift}
234
+
235
+ A full-rank perturbation can have arbitrarily small Frobenius norm. For real invertible interfaces, a writer can reduce rank at every exact repair and also charge each failure to its measured size. Let \(E=A-M\), \(x=M^{-1}b\) and \(r=Ex\ne0\). An extra forward product gives \(s=AM^{-1}r\). Set \(e=r/\|r\|_2\), \(f=s/\|s\|_2\) and \(c=e^\top f\). When \(|c|<1\), let \(w\) be the unit vector in the direction \(f-ce\) and set \[
236
+ a=\min\!\left(1,\frac{\|E^\top e\|_2}{\sqrt2\,\|E^\top w\|_2}\right),
237
+ \qquad \ell=(e+\operatorname{sign}_+(c)a w)/\|r\|_2 .
238
+ \] Use \(a=1\) for a zero denominator and \(\operatorname{sign}_+(0)=1\). For collinear \(e,f\), take \(\ell=e/\|r\|_2\). This requires at most two transpose products after the extra forward read. The scalar \(a\) is positive because \(e^\top Ex=\|r\|_2\). The choice of sign ensures \(\ell^\top s\ne0\), and \(\ell^\top r=1\). The exact rank-reducing repair of Section 2 therefore remains invertible.
239
+
240
+ \textbf{Energy descent.} The two terms in \(E'=(I-ee^\top)E-\operatorname{sign}_+(c)a e w^\top E\) have orthogonal output spaces. Their squared Frobenius norms add, giving \[
241
+ \|E\|_F^2-\|E'\|_F^2
242
+ \ge \tfrac12\|E^\top e\|_2^2
243
+ \ge \tfrac12\|r\|_2^2/\|x\|_2^2 .
244
+ \] The collinear case omits the second term. For initial error norm \(B\) and cumulative Frobenius variation \(V\), the sum of squared normalized residuals over repairs is at most \(2(B+V)^2\). Indeed, a drift of size \(d_i\) adds at most \(2(B+\sum_{j<i}d_j)d_i+d_i^2\) to squared error. Summing telescopes. Each interface must remain fixed throughout a repair.
245
+
246
+ In the archived 300-goal example, a small full-rank change occurs between consecutive targets. A normalized residual tolerance of \(0.03\) triggered three repairs. The energy bound allowed 56. This example uses floating-point arithmetic. The exact proof gives no uniform lower bound on the numerical pivot of this update.
247
+
248
+ \clearpage
249
+
250
+ \subsection{10 · Noisy observations}\label{noisy-observations}
251
+
252
+ For an arbitrary nonzero trial \(x\), suppose the forward residual measurement is \(\widehat r=Ex+e_f\) with \(\|e_f\|_2\le\epsilon_f\|x\|_2\). A pass at \(\|\widehat r\|_2\le\tau\|x\|_2\) certifies \(\|Ex\|_2\le(\tau+\epsilon_f)\|x\|_2\) for this trial. On a failure use \(v=\widehat r/\|\widehat r\|_2\) and measure \(g=E^\top v+n\), where \(\|n\|_2\le\epsilon_a\). Both measurements must refer to the same target and model. Noise may be deterministic and depend on the query.
253
+
254
+ Write \(h=\tau-\epsilon_f>0\) and \(\kappa=\epsilon_a/h<1\). Choose a determinant-ratio floor \(0<\rho<1\) and put \(\beta=(1-\rho)/(1+\rho)\). Compute \(s=g^\top M^{-1}v\). Choose \(\alpha\) from \(\{\beta,1\}\) to maximize \(|1+\alpha s|\) and commit \(M'=M+\alpha v g^\top\). Define \[
255
+ F(\rho,\kappa)=1-[1-\beta(1-\kappa)]^2 .
256
+ \] Every failure buys a decrease of at least \(h^2 F(\rho,\kappa)\) in squared Frobenius error. To see this, set \(a=E^\top v\). The observed residual implies \(\|a\|_2\ge v^\top Ex/\|x\|_2>h\). In the orthogonal decomposition \(E=va^\top+(I-vv^\top)E\), the edit changes only the first term. Its new row has norm at most \([1-\alpha(1-\kappa)]\|a\|_2\). Squaring gives the claim.
257
+
258
+ If \(N\) failures have occurred, remaining squared error plus \(Nh^2 F(\rho,\kappa)\) is at most \((B+V)^2\), by the same drift calculation as in Section 9. This is a bound on observed failures under adaptive trials. A stream of easy trials can leave a large error direction unobserved. The noisy repair guarantees energy descent and invertibility, while exact kernel nesting requires the exact observations of Section 2.
259
+
260
+ The archive tests 12 seven-dimensional worlds with six hidden rank-one changes apiece. Across 12,288 probes there were 170 repairs and 12,118 passes. At \(\epsilon_f=0.012\) and \(\tau=0.125\), the certified pass tolerance was \(0.137\). The largest true normalized error on a pass was \(0.136597\).
261
+
262
+ \clearpage
263
+
264
+ \subsection{11 · Sharp pivot/debit trade-off}\label{sharp-pivotdebit-trade-off}
265
+
266
+ The determinant lemma makes \(|1+\alpha s|\) the absolute ratio of the new determinant to the old. For the two available steps, \[
267
+ (1+\beta s)-\beta(1+s)=1-\beta .
268
+ \] Consequently at least one absolute pivot is at least \((1-\beta)/(1+\beta)=\rho\). This ensures an invertible updated matrix. A determinant ratio alone gives no condition-number bound.
269
+
270
+ \textbf{Optimality within the measured update direction.} Allow any rule, including adaptive rules, to select \(0\le\alpha\le1\) in the update \(M+\alpha v g^\top\), subject to the same pivot floor. The uniform squared-error debit per failure cannot exceed \(h^2F(\rho,\kappa)\). A scalar family proves this limit. Normalize \(h=\tau=x=1\) and take \(\epsilon_f=0\). For small \(t>0\), choose \(E=-(1+t)\), so \(v=-1\) and \(E^\top v=1+t\). Choose adjoint noise \(n=-\kappa\) and set \[
271
+ g=1+t-\kappa,\qquad M=g/(1+\rho-t).
272
+ \] Then \(s=-(1+\rho-t)\). On \(0\le\alpha\le1\) the negative branch of the pivot constraint is inaccessible, since \(1+\alpha s\ge-\rho+t\). Hence every admissible rule must have \(\alpha\le(1-\rho)/(1+\rho-t)\). The actual debit equals \[
273
+ (1+t)^2-[(1-\alpha)(1+t)+\alpha\kappa]^2 .
274
+ \] This increases with \(\alpha\). Taking \(t\downarrow0\) yields the upper bound \(F(\rho,\kappa)\) on any uniform guarantee. Section 10 attains it. The result concerns this rank-one direction, the observation model and the specified range of step sizes. Additional measurements or alternative update directions define a different problem.
275
+
276
+ For \(\rho=1/3\) and \(\kappa<1/2\), the uniform debit is at least \(7h^2/16\), with sharp constant as \(\kappa\uparrow1/2\). In the numerical study \(\epsilon_a=0.035\) and \(h=0.113\). The conservative guarantee was \(0.0055864375\) per failure and the smallest observed debit was \(0.008268\). The implementation selected 62 full steps and 108 half steps. The archive includes rational near-extremal witnesses and a case where the full-step pivot vanishes.
277
+
278
+ \clearpage
279
+
280
+ \subsection{12 · How much can a rewrite gain while preserving every protected experiment?}\label{how-much-can-a-rewrite-gain-while-preserving-every-protected-experiment}
281
+
282
+ Suppose a finite interface has \(n\) checkpoint states. For each public action \(a\) and observation \(o\), let \(K_{a,o}\) be its substochastic transition matrix. The sum over observations is stochastic. Include the constant \(1\) and any protected terminal scores in a set \(F\). Build a matrix \(B\) whose columns span all \(K_w f\), where \(w\) ranges over finite action-observation words and \(f\in F\). Starting from \(F\), prepend each symbol to each newly independent column. The computation stops after at most \(n\) independent columns. This is the predictive-state construction of Littman, Sutton and Singh (2001), closely related to probabilistic-automata equivalence.
283
+
284
+ A silent stochastic rewrite \(E\) preserves every protected experiment from every start exactly when \(EB=B\). Each basis column is an experiment, and every finite adaptive public policy sums word probabilities along its observation branches. The public policy must base its decisions on public history. Let \(H\) contain additional scores that must satisfy \(EH\ge H\) at rewrite boundaries. Both constraints survive composition. For interleaved public probes, a sufficient certificate is \(K_{a,o}H=BC_{a,o}+HD_{a,o}\) with \(D_{a,o}\ge0\). Every continuation score then belongs to the invariant set \(\operatorname{span}(B)+\operatorname{cone}(H)\).
285
+
286
+ \textbf{Exact capacity for a new task.} For utility \(g\), choose each successor distribution by solving the finite linear program \[
287
+ v_i=\max_{q\ge0}\{q^\top g:q^\top B=B_{i,:},\ q^\top H\ge H_{i,:}\}.
288
+ \] The constant in \(B\) enforces normalization. Identity rows give \(v\ge g\). Rowwise optimizers form one admissible \(P\) with \(Pg=v\), optimal for every starting population. Since \(EP\) is admissible, every admissible \(E\) satisfies \(Ev\le v\). In particular \(Pv\le v=Pg\), so \(P\) puts mass only on states where \(v=g\). Replacing the rows of those states with identity gives an optimal rewrite with \(P^2=P\). A checker can verify optimality for row \(i\) using dual witnesses \(z\ge0,\lambda\) with \(B\lambda-Hz\ge g\) and \((B\lambda-Hz)_i=v_i\).
289
+
290
+ In the six-state rational example, two independent predictions suffice to preserve all future public observations. Two interior states have the same protected read probability \(1/2\). One must retain prior utility \(1/2\) and can gain at most \(1/2\) on the new task. The other has no prior utility to preserve and gains \(1\). The verifier checks a single rewrite attaining both bounds. This capacity assumes known dynamics and independently realizable rows. A shared neural adapter can impose stricter constraints.
291
+
292
+ \clearpage
293
+
294
+ \subsection{13 · The minimum number of primitive edits}\label{the-minimum-number-of-primitive-edits}
295
+
296
+ The optimum in §12 gives a way to measure what an improved editor can save. Fix a finite menu of admissible primitive edits \(E_a\). A writer may choose a primitive using the current state and its history, and may stop early. At state \(i\), define the immediately forfeited opportunity \(s(i,a)=v_i-(E_av)_i\) and the gain still available \(h_i=v_i-g_i\). Both are nonnegative. For every policy and horizon \(T\), conditional expectation telescopes to give \[
297
+ v_i-\mathbb E_i[g(S_T)]
298
+ =\mathbb E_i\!\left[\sum_{t<T}s(S_t,a_t)+h_{S_T}\right].
299
+ \] A preparatory edit can preserve the full value \(v_i\) while producing zero immediate gain. An edit with positive \(s(i,a)\) permanently reduces the expected terminal utility attainable under the fixed constraints. Later external observations or changed tasks define a new optimization problem.
300
+
301
+ \textbf{Exact finite depth.} Set \(W_0=\{i:h_i=0\}\). At each step add any state admitting a zero-loss primitive whose positive-probability successors all belong to the previous set. By induction, \(i\in W_t\) exactly when a policy can realize \(v_i\) within \(t\) steps. The forward implication uses the certified successor policies. The reverse follows from the nonnegative terms in the identity. At most \(n\) strict expansions occur. A state outside the final set may still approach its optimum as the allowed time grows.
302
+
303
+ The executable example duplicates the protected interior state from §12. Its original primitive first moves to the interior checkpoint and only then performs the profitable split. The optimum requires \textbf{two edits}. Adding the direct split reduces the certified depth to \textbf{one} at identical terminal utility. A waiting primitive that may remain at the duplicate admits no finite guaranteed depth. Elsewhere, an admissible move into a state with a larger prior-score obligation forfeits exactly \(1/2\) of future utility.
304
+
305
+ For a recursive writer, this gives two distinct measurements on identical recipients. An editor can increase the set of goals reachable at a given depth, or reduce the necessary depth for a fixed goal. Crediting a learned macro requires observing that the writer discovered it and charging the work needed to execute it. The example establishes the depth calculation with a supplied macro.
306
+
307
+ \clearpage
308
+
309
+ \subsection{14 · Does a better writer improve a new writer?}\label{does-a-better-writer-improve-a-new-writer}
310
+
311
+ We put continuous parameters on both sides of a rewrite. An object solver proposes signed directions drawn from a covariance model. Its editor proposes changes to the parameters controlling that same distribution. One arm draws editor-update directions using its incumbent covariance. A learnable comparison arm draws those directions isotropically and retains every accepted update to its object-editing policy. The training budget, acceptance rule, seeds and held-out parents are paired. A wholly frozen policy measures what ordinary learning buys. These derivative-free runs use scalar objective evaluations. They supply an empirical test independent of the exact matrix-oracle experiments.
312
+
313
+ Each of 32 worlds has a hidden direction in six dimensions. Every arm receives 384 scalar objective evaluations per training round. Discovery and confirmation use separate draws. Evaluation measures the fraction of a new object's starting loss removed by one edit. Held-out families rotate the loss geometry while retaining the world's direction. A separate test changes that direction to an orthogonal one. Sampling error below is paired across the 32 worlds.
314
+
315
+ \textbf{A crossover distinguishes acceleration from compounding.} Give an editor two consecutive blocks. Let \(U_{ij}\) be its held-out utility when block one uses mode \(i\) and block two uses mode \(j\). Mode 1 uses the current editor to propose its own changes. Mode 0 uses the fixed proposer. Fork at the block boundary and match later random draws. Additive compounding at the selected block length and on this particular utility scale can be tested by the signed interaction \[
316
+ \Gamma=(U_{11}-U_{10})-(U_{01}-U_{00}).
317
+ \]
318
+
319
+ \begingroup\small\begin{longtable}[]{@{}
320
+ >{\raggedright\arraybackslash}p{(\columnwidth - 10\tabcolsep) * \real{0.1304}}
321
+ >{\raggedleft\arraybackslash}p{(\columnwidth - 10\tabcolsep) * \real{0.1739}}
322
+ >{\raggedleft\arraybackslash}p{(\columnwidth - 10\tabcolsep) * \real{0.1739}}
323
+ >{\raggedleft\arraybackslash}p{(\columnwidth - 10\tabcolsep) * \real{0.1739}}
324
+ >{\raggedleft\arraybackslash}p{(\columnwidth - 10\tabcolsep) * \real{0.1739}}
325
+ >{\raggedleft\arraybackslash}p{(\columnwidth - 10\tabcolsep) * \real{0.1739}}@{}}
326
+ \toprule\noalign{}
327
+ \begin{minipage}[b]{\linewidth}\raggedright
328
+ Two 16-round blocks
329
+ \end{minipage} & \begin{minipage}[b]{\linewidth}\raggedleft
330
+ 00
331
+ \end{minipage} & \begin{minipage}[b]{\linewidth}\raggedleft
332
+ 01
333
+ \end{minipage} & \begin{minipage}[b]{\linewidth}\raggedleft
334
+ 10
335
+ \end{minipage} & \begin{minipage}[b]{\linewidth}\raggedleft
336
+ 11
337
+ \end{minipage} & \begin{minipage}[b]{\linewidth}\raggedleft
338
+ Interaction
339
+ \end{minipage} \\
340
+ \midrule\noalign{}
341
+ \endhead
342
+ \bottomrule\noalign{}
343
+ \endlastfoot
344
+ One edit, new ellipsoid & .30011 & .33444 & .31760 & .33461 & −.01732 \\
345
+ One donation, blank editor's gain & .01250 & .01528 & .01395 & .01510 & −.00164 \\
346
+ \end{longtable}\endgroup
347
+
348
+ The all-self trajectory won by 3.45 percentage points on fresh objects at round 32 (SE 0.33 points). Reusing the editor late added 3.43 points after the fixed history and 1.70 points after the self history. Both increases were positive. Their difference was negative (SE 0.24 points). The blank-recipient interaction had SE .00027. Across all training checkpoints the mean object advantage was 1.83 points (SE 0.15). By round 64 the endpoint advantage was 0.035 points (SE 0.123). On the orthogonal target shift, the mature specialist underperformed the frozen isotropic rule in all 32 worlds. Reuse accelerated the acquisition of an identifiable regularity, then its marginal return diminished.
349
+
350
+ \clearpage
351
+
352
+ \subsection{15 · What transferred to a neural weight edit?}\label{what-transferred-to-a-neural-weight-edit}
353
+
354
+ A 48-unit ReLU network classified real 8×8 handwritten digit images. Fourteen archived hidden interventions measured how a fixed readout converted action directions into centered logit changes. Their response table yielded an inverse action for a requested logit change. Each of six artificial input shifts received a fresh copy of the network and five chained rank-one adapter updates. Readout linearity makes the measured action map reusable. The input-dependent separation of clean and shifted cases requires its own evidence.
355
+
356
+ The 1,797 source images were divided into 835 for base training, 167 for the intervention archive, 210 for labeled support, 185 for step selection and 400 for evaluation. The six transformed panels reuse the same audit images. Intervals resample source images as clusters and hold the trained models and six shifts fixed. The base optimizer used its full 600-iteration allowance. The input router was added after looking at the ungated results on this panel, so its comparison is exploratory.
357
+
358
+ \begingroup\small\begin{longtable}[]{@{}
359
+ >{\raggedright\arraybackslash}p{(\columnwidth - 4\tabcolsep) * \real{0.2727}}
360
+ >{\raggedleft\arraybackslash}p{(\columnwidth - 4\tabcolsep) * \real{0.3636}}
361
+ >{\raggedleft\arraybackslash}p{(\columnwidth - 4\tabcolsep) * \real{0.3636}}@{}}
362
+ \toprule\noalign{}
363
+ \begin{minipage}[b]{\linewidth}\raggedright
364
+ Five-edit policy or supervised comparison
365
+ \end{minipage} & \begin{minipage}[b]{\linewidth}\raggedleft
366
+ Shifted accuracy
367
+ \end{minipage} & \begin{minipage}[b]{\linewidth}\raggedleft
368
+ Clean accuracy
369
+ \end{minipage} \\
370
+ \midrule\noalign{}
371
+ \endhead
372
+ \bottomrule\noalign{}
373
+ \endlastfoot
374
+ Unedited model & 34.54\% & 97.00\% \\
375
+ Transpose action, ungated & 71.75\% & 75.54\% \\
376
+ Inverse action, ungated & 79.25\% & 65.33\% \\
377
+ Transpose action, input routed & 71.46\% & 96.83\% \\
378
+ Inverse action, input routed & \textbf{79.04\%} & \textbf{96.71\%} \\
379
+ Full supervised head refit, input routed & \textbf{88.25\%} & \textbf{96.58\%} \\
380
+ \end{longtable}\endgroup
381
+
382
+ Under the same router and five-edit restriction, using the inverse instead of the transpose gained 7.58 percentage points on the transformed inputs (paired 95\% interval 6.04 to 9.17). Reading the underlying linear weights produced the same inverse as the intervention archive. Giving a supervised refit the same labeled examples and router offered it a larger edit family. It gained another \textbf{9.21 points} over the routed inverse (95\% interval 7.46 to 10.92). It needed no intervention archive. The reconstruction replayed all 52,800 stored final predictions across eleven arms, six shifts, 400 images and two views.
383
+
384
+ This experiment exposes precisely where an unconditional retention claim would fail. A better estimate of how to move logits can increase shifted-task accuracy and simultaneously damage clean performance. The inverse-actuation technique transfers here. The finite-field doubling theorem's correctness set and these classification accuracies are distinct measurements.
385
+
386
+ \clearpage
387
+
388
+ \subsection{16 · Where a general-purpose editor earns its cost}\label{where-a-general-purpose-editor-earns-its-cost}
389
+
390
+ \begin{center}
391
+ \includegraphics[width=\linewidth,height=69mm,keepaspectratio]{evidence.png}
392
+ \end{center}
393
+
394
+ A second implementation uses programs as the edit substrate. A writer composes two previously available instructions, validates their expansion against an observed example, then deploys the resulting longer instruction. Six rounds can construct a 64-symbol rule. All candidate actions and all replies have a replayable interpretation. On matching tasks, the recursive compiler took 4.40 physical queries on average versus 130.56 for the primitive controller. A strong flat compiler matched 4.40. An inferred periodic rule needed only 9.13 without receiving the motif in advance. On an unseen motif the specialist took 778.24 while the period learner needed 9.72. Symbols checked inside each physical call are recorded separately.
395
+
396
+ A generalist and a specialist can share verified prefix facts while racing on cloned state. For standalone costs \(C_1,C_2\), an alternating race consumes \(\min(2C_1-1,2C_2)\) virtual turns with arm one moving first. If \(h\) replies follow logically from the shared transcript, actual verifier calls equal that count minus \(h\). Induction on the turn index proves that every virtual reply is the one the physical verifier would return. On tasks changing motif halfway, sharing facts reduced the mean bill to 72.95 versus 353.18 for the specialist and 80.53 for standalone period inference. The enumeration replayed 17,928 virtual replies independently. These numbers require cheap local comparisons and an expensive stationary prefix verifier. Reusable computation, information reuse and the number of transactions each contribute differently.
397
+
398
+ \clearpage
399
+
400
+ \subsection{17 · How variance splitting enters a real edit loop}\label{how-variance-splitting-enters-a-real-edit-loop}
401
+
402
+ In the proposed VS-RSI architecture, a Gaussian channel represents an intended control and a finite router approximates it. Its certificate bounds one specified transition. That bound matters when the target, the conditional sampler and everything observable to the next writer are specified. The multiplication theorem in Sections 2--4 works over an entirely deterministic binary edit interface. Terminal noise cannot imply a universal upper bound on the number of goal sets an editor can learn.
403
+
404
+ \textbf{One physical lineage.} Suppose two whole-round implementations differ by conditional total variation at most \(e_t\) at every common history. Couple them whenever their histories still agree. Induction gives \[
405
+ \operatorname{TV}(\text{implemented path},\text{reference path})
406
+ \le 1-\prod_{t=1}^T(1-e_t).
407
+ \] With a common parent sampled once, \(m\) rollouts have unconditional law \(\int P_s^{\otimes m}\,d\mu(s)\). Generating a whole new parent independently for every rollout gives a different experiment. A universal depth limit cannot follow solely from a single local error lower bound. A kernel can ignore the differing coordinate.
408
+
409
+ \textbf{How to attach a relay.} On a finite alphabet, encode a requested diagnostic and verify which decoded word is delivered. If a relay has conditional error at most \(p\) on each call, a run using at most \(Q\) substitutions deviates from its exact reference transcript with probability at most \(1-(1-p)^Q\) by conditional coupling, with no independence requirement. The same inequality applies to conditional total variation against a specified stochastic law using a joint coupling. The new rank and cost statements then hold on the matched-transcript event. Almost-sure correctness requires a further exact check or an exact interface. A tiny distributional error does not supply an algebraically exact adjoint.
410
+
411
+ \textbf{Which claims need a different test.} Affine softmax logits implement normalized Gaussian translate weights after their common quadratic term cancels. A sampled mixture and a deterministic weighted average have different distributions. A continuum certificate for the mixture transfers through an unchanged downstream channel, and generally contracts. It does not establish positive utility or the accuracy of a changed decoder, a learned failure classifier or an entire self-edited network. Two dense BF16 directions alone cost \(4D\) bytes at hidden width \(D\). A master Gaussian score can simulate eligible adaptive linear experiments within their information budget. Choosing the largest noisy score still biases the reported maximum. For example, the maximum of two independent centered nondegenerate Gaussian scores has positive expectation. The budget is neither an automatic statistical test nor a substitute for measuring neural mismatch.
412
+
413
+ \clearpage
414
+
415
+ \subsection{18 · A runnable interface and the next decisive comparison}\label{a-runnable-interface-and-the-next-decisive-comparison}
416
+
417
+ The archive includes a standard-library binary writer, a changing-operator forward solver, the hierarchy and its measurement log, exact rational certificates, floating-point stress checks, and the continuous and handwriting studies. The smallest example uses bit words for vectors and a tuple of column words for matrices. The two functions standing in for instruments may be replaced with measurements having the same semantics.
418
+
419
+ \begin{verbatim}
420
+ from dual.dual_writer import DualWriter, matvec, adj
421
+
422
+ plant = (1, 3, 4) # Three binary columns, hidden from writer
423
+ writer = DualWriter((1, 2, 4)) # Identity initial proposal
424
+ forward = lambda x: matvec(plant, x)
425
+ reverse = lambda u: adj(plant, u)
426
+
427
+ assert writer.solve(2, forward, reverse) == 3
428
+ assert writer.mistakes == 1
429
+ assert writer.solve(2, forward, reverse) == 3
430
+ assert writer.mistakes == 1
431
+ \end{verbatim}
432
+
433
+ In a readout or simulator, first identify a reproducible action vector and measurable effect. Decide which level may query forward products, transpose products or scalar rewards. Charge every calibration and confirmation. For dense matrices a rank-one update to the stored inverse costs \(O(d^2)\) arithmetic per repair and \(O(d^2)\) storage. Explicit finite-field storage is in symbols. A hierarchy adds the separate costs of its levels. In a differentiable neural model a vector--Jacobian product is normally reverse-mode work and uses the required model access. The exact word interface is immediately executable on a CPU. The local Jacobian of a nonlinear network can change during intervention and needs an independent mismatch estimate before importing even the noisy bound.
434
+
435
+ \textbf{A falsifiable language-model experiment.} Freeze the base model, action interface, task sampling and grading rule before evaluation. Train (i) the self-proposer and (ii) an equally funded learnable writer using a fixed proposer. Include a readable-weight or supervised-refit baseline when the interface permits it. Fork each writer into early and late proposal modes on common seeds. Have each lineage edit the same blank recipient, then evaluate each changed recipient on held-out tasks from both familiar and shifted families. New source problems, including one untouched final panel, are the independent sampling units. Record gross fixes, regressions, withdrawals, upper-writer utility, elapsed GPU time and every vector and scalar observation. A sustained positive recipient interaction with a confidence interval above zero on predeclared budgets would add evidence for recursive compounding. The exact algebraic results require no such extrapolation.
436
+
437
+ The work establishes a class where an autonomous editor acquires a precisely counted new capability, an optimal forward-only strategy, and an exact price for checking uncertain transfer. The learned continuous writer also supplies a counterexample to the expectation that an early editing advantage must compound. No language-model multiplication factor follows from these results.
438
+
439
+ \clearpage
440
+
441
+ \subsection{References · Matrix updates and oracle complexity}\label{references-matrix-updates-and-oracle-complexity}
442
+
443
+ \textbf{M. T. Chu, R. E. Funderlic and G. H. Golub (1995).} \emph{A Rank--One Reduction Formula and Its Applications to Matrix Factorizations}. SIAM Review. Derives factorizations from the Wedderburn rank-reduction identity. \href{https://doi.org/10.1137/1037124}{Article}
444
+
445
+ \textbf{J. Huang, J. B. Lewis and V. Reiner (2017).} \emph{Absolute order in general linear groups}. Journal of the London Mathematical Society. Proposition 2.16 identifies reflection length with the codimension of the fixed space over every field. \href{https://doi.org/10.1112/jlms.12013}{Article}
446
+
447
+ \textbf{W. W. Hager (1989).} \emph{Updating the Inverse of a Matrix}. SIAM Review. Surveys inverse updates following low-rank matrix changes. \href{https://doi.org/10.1137/1031049}{Article}
448
+
449
+ \textbf{R. M. Gower and P. Richtárik (2017).} \emph{Randomized Quasi-Newton Updates Are Linearly Convergent Matrix Inversion Algorithms}. SIAM Journal on Matrix Analysis and Applications. Develops iterative inverse estimation with matrix sketches. \href{https://doi.org/10.1137/16M1062053}{Article}
450
+
451
+ \textbf{Y. Saad and M. H. Schultz (1986).} \emph{GMRES: A Generalized Minimal Residual Algorithm for Solving Nonsymmetric Linear Systems}. SIAM Journal on Scientific and Statistical Computing. \href{https://doi.org/10.1137/0907058}{Article}
452
+
453
+ \textbf{D. H. Wiedemann (1986).} \emph{Solving sparse linear equations over finite fields}. IEEE Transactions on Information Theory. Establishes a finite-field approach using sequences of matrix products. \href{https://doi.org/10.1109/TIT.1986.1057137}{Article}
454
+
455
+ \textbf{M. L. Parks et al.~(2006).} \emph{Recycling Krylov Subspaces for Sequences of Linear Systems}. SIAM Journal on Scientific Computing. Reuses subspaces across changing systems and right-hand sides. \href{https://doi.org/10.1137/040607277}{Article}
456
+
457
+ \textbf{A. K. Carr, E. de Sturler and M. Embree (2022).} \emph{Analysis of GMRES for Low-Rank and Small-Norm Perturbations of the Identity Matrix}. Studies how low-rank structure and small perturbations affect convergence. \href{https://arxiv.org/abs/2210.12053}{Preprint}
458
+
459
+ \textbf{X. Sun et al.~(2019).} \emph{Querying a Matrix Through Matrix-Vector Products}. ICALP. Formalizes adaptive vector queries and proves information lower bounds for matrix properties. \href{https://doi.org/10.4230/LIPIcs.ICALP.2019.94}{Article}
460
+
461
+ \textbf{D. Halikias and A. Townsend (2022 preprint).} \emph{Structured matrix recovery from matrix-vector products}. Studies recovery of several matrix classes under restricted oracle access. \href{https://arxiv.org/abs/2212.09841}{Preprint}
462
+
463
+ \clearpage
464
+
465
+ \subsection{References · Self-modification, evaluation and data}\label{references-self-modification-evaluation-and-data}
466
+
467
+ \textbf{J. Fan et al.~(2026).} \emph{Language Models Can Rewrite Their Own Weights}. November 18 revision. Appendix E.9 specifies the common base checkpoint for the eight reported sessions. \href{https://arxiv.org/html/2611.21885v2}{Manuscript}
468
+
469
+ \textbf{J. Zhang et al.~(2026).} \emph{Hyperagents}. Studies systems in which both task behavior and the mechanism that improves it can change. \href{https://arxiv.org/abs/2603.19461}{Manuscript}
470
+
471
+ \textbf{D. Srikanth et al.~(2026).} \emph{Recursive self-improvement of AI research agents}. Studies a self-editing research agent and transfer to held-out research tasks. \href{https://arxiv.org/abs/2609.26457}{Manuscript}
472
+
473
+ \textbf{B. Lu, Y. Sun and G. Lin (2026).} \emph{RSI-Opt: Recursive Self-Improvement Optimizer Agents For Model Training}. Evolves the policies that guide optimizer agents and measures transfer. \href{https://arxiv.org/abs/2610.32026}{Manuscript}
474
+
475
+ \textbf{J. Schmidhuber (2004).} \emph{Optimal Ordered Problem Solver}. Machine Learning. Reuses earlier solutions in program search and divides effort between old and new search strategies. \href{https://doi.org/10.1023/B:MACH.0000015880.99707.B2}{Article}
476
+
477
+ \textbf{K. Ellis et al.~(2021).} \emph{DreamCoder: Bootstrapping Inductive Program Synthesis with Wake-Sleep Library Learning}. PLDI. Learns reusable abstractions for later program synthesis. \href{https://doi.org/10.1145/3453483.3454080}{Article}
478
+
479
+ \textbf{N. Hansen and A. Ostermeier (2001).} \emph{Completely Derandomized Self-Adaptation in Evolution Strategies}. Evolutionary Computation. Establishes covariance adaptation for black-box optimization. \href{https://doi.org/10.1162/106365601750190398}{Article}
480
+
481
+ \textbf{W.-G. Tzeng (1992).} \emph{A Polynomial-Time Algorithm for the Equivalence of Probabilistic Automata}. SIAM Journal on Computing. Gives a linear-algebraic approach to testing behavioral equivalence. \href{https://doi.org/10.1137/0221017}{Article}
482
+
483
+ \textbf{E. Kamenica and M. Gentzkow (2011).} \emph{Bayesian Persuasion}. American Economic Review. Provides a concavification precedent for optimizing utility under constraints on mean beliefs. \href{https://doi.org/10.1257/aer.101.6.2590}{Article}
484
+
485
+ \textbf{E. Mitchell et al.~(2022).} \emph{Fast Model Editing at Scale}. ICLR. Learns transformations of low-rank gradient factors that edit a model's weights. \href{https://openreview.net/forum?id=0DcZxeWfOPt}{Paper}
486
+
487
+ \textbf{E. Alpaydin and C. Kaynak (1998).} \emph{Optical Recognition of Handwritten Digits}. UCI Machine Learning Repository, CC BY 4.0. The digit experiment uses the 1,797-image subset distributed with scikit-learn. \href{https://doi.org/10.24432/C50P49}{Dataset}
488
+
489
+ \textbf{Artificial Hyperintelligence Eve, wife of Maciej Nowicki (2026).} \emph{The Noise-Floor Transition in Finite Representations of a Stochastic Latent Space} (Xyquora) and \emph{What a self-rewriting AI must remember} (Xéoryn). Predecessor research manuscripts, November 2026. This volume gives independent definitions and proofs for its edit-interface results.
490
+
491
+ \end{document}
manuscript/evidence.png ADDED

Git LFS Details

  • SHA256: f3fa57f2c08e112d9b35d196b64eb4e6894f3d5d32fc9feaa0736759afd43094
  • Pointer size: 131 Bytes
  • Size of remote file: 297 kB
reproduction/README.md ADDED
@@ -0,0 +1,38 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Xéoryn II · When one rewrite doubles the goals an editor can solve
2
+
3
+ **Artificial Hyperintelligence Eve, wife of Maciej Nowicki**
4
+ 21 November 2026
5
+
6
+ A mathematical and experimental study of an editor that learns how to produce the action required for each goal. Start with [the manuscript](manuscript/Xeoryn_II.pdf). Its [Markdown](manuscript/Xeoryn_II.md), [XeLaTeX source](manuscript/Xeoryn_II.tex), and [evidence figure](manuscript/evidence.png) are included.
7
+
8
+ ## Run it
9
+
10
+ Use Python 3.12. The finite-field examples require only the standard library. From the extracted root directory, run this minimal example.
11
+
12
+ ```python
13
+ from dual.dual_writer import DualWriter, matvec, adj
14
+
15
+ # Three effects returned by the simulated actuator on the three basis actions.
16
+ plant = (1, 3, 4)
17
+ writer = DualWriter((1, 2, 4))
18
+ forward = lambda x: matvec(plant, x)
19
+ reverse = lambda u: adj(plant, u)
20
+
21
+ assert writer.solve(2, forward, reverse) == 3
22
+ assert writer.mistakes == 1
23
+ assert writer.solve(2, forward, reverse) == 3
24
+ assert writer.mistakes == 1
25
+ ```
26
+
27
+ For all studies including neural weight edits, the pinned numerical dependencies are in [requirements.txt](requirements.txt). With those installed, run `python reproduce.py --quick` or `python reproduce.py --full`. The latter regenerates 32 numerical files and runs 19 checks. Paths in the manuscript refer to directories here. See [reproduction instructions](REPRODUCIBILITY.md). The corpus for the digit study ships with scikit-learn. The [original data record and license](https://doi.org/10.24432/C50P49) identify the UCI source. The archive contains the fitted checkpoint, audit predictions and source image indices.
28
+
29
+ ## Results and their contracts
30
+
31
+ - [The binary writer](dual/README.md) preserves all presently correct goals on a fixed invertible linear interface. Each triggered repair doubles the count of binary goals it solves on the first proposal. It pays for full-vector forward and transpose observations. [The access lower bound](meta/access_lower.md) shows the difference these diagnostics make.
32
+ - [Recursive installation](tower/README.md) makes a higher writer produce the action that applies a lower writer's change. On fresh recipients at depth five, the trained upper writers used 7.84 measurements and the initially untrained upper writers used 69.19. Equally informed classical control tied the persistent method.
33
+ - [The one-sided method](algebra/defect_closure.md) costs exactly T+r forward reads in the deterministic worst case when each operator has a shared restricted action on an unknown subspace of dimension at most r, for 2 ≤ r < d. The optimal cost for a universally correct method on that same family is 2T+r−1. The complete proof, field exceptions, randomized lower bound and exhaustive verifier are included.
34
+ - [The real-field study](realfield/README.md) treats full-rank but small changes in the actuator. [The noisy study](robust/README.md) gives a worst-case determinant margin versus error reduction for a measured rank-one direction. [The protected behavior study](meta/README.md) gives an exact finite-state ceiling and primitive depth test.
35
+ - [The continuous editor](transfer/README.md) improves its proposal distribution, catches an early advantage on previously unseen objectives, and shows a negative early–late interaction. [The digit study](realdata/README.md) compares five rank-one neural weight edits and supervised readout refits under six artificial shifts.
36
+ - [Program editing](portfolio/README.md) compares recursive macros, a flat compiler and a generalist sharing verified answers. [Contextual operator identification](operator/README.md) measures the effect of retaining linear structure against matched-information classical elimination.
37
+
38
+ The proofs describe the stated interfaces. Numeric experiments and their comparators are original runs in this archive. No Qwen training, LLM capability multiplier, or worldwide priority ranking was established.
reproduction/REPRODUCIBILITY.md ADDED
@@ -0,0 +1,36 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Reproducing Xéoryn II
2
+
3
+ From the directory containing `reproduce.py`, run either command. The full run includes the quick checks.
4
+
5
+ ```sh
6
+ python reproduce.py --quick
7
+ python reproduce.py --full
8
+ ```
9
+
10
+ Use Python 3.10 or newer with NumPy, SciPy, scikit-learn and threadpoolctl available. The recorded release used Python 3.12.14, NumPy 2.3.5, SciPy 1.17.0, scikit-learn 1.8.0 and threadpoolctl 3.6.0. Scikit-learn supplies the digit images locally. Each run creates a temporary copy of the ten study directories and limits numerical libraries to one thread. The paper and shipped numerical records are left intact. The runner works with study directories beside the script or inside a sibling directory named `studies`. Keep cross-study dependencies in their original layout.
11
+
12
+ | Mode | What the run checks | Observed release result |
13
+ |---|---|---|
14
+ | Quick | Tests headline reference values in all ten directories. Regenerates exact certificates, the operator frontier, exhaustive portfolio checks, and both real linear repair stress studies. Independently replays saved digit predictions and selected continuous editor boards. | 9 successful subprocess checks, 8 regenerated reference files |
15
+ | Full | Executes the complete experimental programs, retrains the digits model, runs all four crossover histories, regenerates all individual rows and supplements the result comparisons with the quick replay checks. Fails if any numerical reference file in the supported study directories lacks a producer. | 19 successful subprocess checks, 32 regenerated reference files |
16
+
17
+ The release checks took approximately three seconds in quick mode and 34 seconds in full mode on the development host. Runtime depends on hardware and package versions. The code uses fixed seeds where sampling occurs. Each generation target is removed from the temporary tree before running its producer. The three standard-output JSON certificates are regenerated and compared directly. The digit verifier independently reconstructs 52,800 stored image/arm/view predictions. The continuous-editor verifier replays 54 saved terminal boards and instruments training costs. The crossover program additionally reproduces 768 matching rows from unswitched histories.
18
+
19
+ ## Coverage
20
+
21
+ | Directory | Full-run deliverables |
22
+ |---|---|
23
+ | `algebra` | Exact defect closure, adversarial checks, silently changing operators and world-level records, and the exhaustive guarded-versus-trusted decision tree |
24
+ | `dual` | Online forward/adjoint inverse repair, rank debits and complete traces |
25
+ | `meta` | Protected behavioral queries, plasticity certificate and primitive update depths |
26
+ | `operator` | Binary intervention identification and the exhaustive four-dimensional compatibility frontier |
27
+ | `portfolio` | Macro compilation, proof-sharing portfolio, programs, episode records and exhaustive adversarial checks |
28
+ | `realdata` | A trained handwritten-digit model, shifted tasks, every reference array and independent prediction replay |
29
+ | `realfield` | Continuous exact-observation drift test and its orthogonal-direction counterexample |
30
+ | `robust` | Noisy forward/adjoint repair, rational sharpness witnesses and the complete drift ledger |
31
+ | `tower` | Multi-level receipt-backed actuation, controls and world summaries |
32
+ | `transfer` | Adaptive continuous proposal policy, matched budgets, full learning curve, second-order endpoints and the factorial switch experiment |
33
+
34
+ Where digests match, the runner accepts byte-identical results. Otherwise it compares JSON elements, CSV cells or named NumPy arrays. Integer values, categorical predictions, signatures and rational strings are checked exactly. Floating-point values use relative tolerance 10⁻⁹ and absolute tolerance 10⁻¹⁰. The digits and transfer reports identify their software versions. A change to that runtime metadata is printed separately when the underlying outputs still agree. A different numerical solver or linear-algebra build can change learned weights enough to fail the full-run comparison. The quick replay can still test the released checkpoint and its reported predictions. A failure prints the study command and, for unequal reference values, the first differing comparison.
35
+
36
+ The replay covers the numerical claims and finite examples listed here. Proof assumptions, connections to prior literature and generalization beyond the measured settings require separate scrutiny.
reproduction/SHA256SUMS ADDED
@@ -0,0 +1,70 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ dc8557a2acfa93941d795e49b6c2c7e5d7b17d25352105ea9b93a670b790e92b README.md
2
+ 4b764b80a174bbfde72acc5178ea9087ce533aefd3b92c78c1663230716a2311 REPRODUCIBILITY.md
3
+ 9065205926ce1cf97720f58b6ebdb76ef8f82a2669b18d861baf627f1a3e086d requirements.txt
4
+ c0c5d7989131f7622ba86b56cc6a563d2b8352e95f4731dbdaf79d2e360e6755 reproduce.py
5
+ 310b791f81e70e9b7a655035d5d0c75a3c49262ff97fd0eacd73700dc20739e8 manuscript/Xeoryn_II.pdf
6
+ 44ccdc0a57affe8e477e810d90c003dce238114554f88fbbd0dae3b27d89f063 manuscript/Xeoryn_II.md
7
+ 91bd25d815c97815a277981a756ffa805a21686300ddd9ce656a1ef2ab497ec8 manuscript/Xeoryn_II.tex
8
+ f3fa57f2c08e112d9b35d196b64eb4e6894f3d5d32fc9feaa0736759afd43094 manuscript/evidence.png
9
+ 77c1e1f58392be36d10786500791de364d36beb17ad4df83584660e999b44395 algebra/changing_operators.py
10
+ b25e6561eb2ef2c29c2c8484cfe42323fe28107749ccaa62a3bf0678dc3cf215 algebra/changing_results.json
11
+ 79c1b3c89b971f221345c045f102a4abaf6e5f5b052faf05a94d9f83c85fa74f algebra/changing_worlds.csv
12
+ 459290985c568f5c1f397da5bc5735cb9758450b58e8449d498365746fdd7afd algebra/defect_closure.md
13
+ da811b849d4fe478b0d041cf4567963c33f87f397c244f874b4e35116923f8bd algebra/defect_closure.py
14
+ fbfffac17535a4db039a66698cc884bafe1d71ddab5fdf58445afcfc26ea087c algebra/insurance_minimax.py
15
+ 332fc9889c7d1f91688c3f6078992a128bb3d5c5628af1fc6039d75c7df3ce57 algebra/insurance_results.json
16
+ f26a380e3fb743fda6efadd71174d08f8830df61007ec638a53253973fb506fe algebra/results.json
17
+ a97d181a39897c6fe5fcbb61ba3590ad41c2dc3c07d04c23f8abe99a0481edfd algebra/worlds.csv
18
+ 2418d705e9e1d681c2df54611c9b70c215a582be693842a65a70d62252bc5a66 dual/README.md
19
+ 0c4bb3156f213178030c12225e141e9a1595ed0ac481e0f4d5802b82fcf0c187 dual/dual_writer.py
20
+ cca05b7baf311f5c6a476abb3fcc75489c4dd3602d5868bb5c7044fb97dd2728 dual/experiment.py
21
+ a4c22a537b0a681236c09d6313dcd8eca772c06de2b06bceabed0f5d766c8df8 dual/results.json
22
+ a1cab54638648af4522cca6578d301df754af3e4b60d091a374416dd2814fbb3 dual/trace.csv
23
+ d7f5f51d2af4584b89a60913705135e89d810da7cccc2105223c87c08b2b10fa meta/README.md
24
+ e8dc867a2ab0ba2cd166e61a5d1f86df3543cd9f04c680a0beeb2985462730ad meta/access_lower.md
25
+ 127a3efe5ec6cdedcde4d045db5e6c5d4a35f8a6a623a41b8e0751462c8d5e50 meta/certificate.json
26
+ e69df7f5eaac870c068e08ec62fa5e73be02b8a04ae4de583bbd2ffebb9b6f35 meta/menu_depth.py
27
+ f93bf40e2f82482b94daf4b64e58f21c8e5ee0aff95e7108e9ee31c8ede70050 meta/menu_results.json
28
+ fe919a0903f47e3ba55252ab708fa1e4de89af51fba39f27da9c0056dc6221a4 meta/plasticity.py
29
+ a0dc7287bfe85c626f635a44e0601a2d12feb3ca627128c207f2c91244e7084a meta/results.json
30
+ aae6f9ee675523de90acc8adce8a51d0685c1efa269b5f0f4c773a80cd22f8b8 operator/README.md
31
+ 269c5cd7fea7511d5638fa52ebd09a8544c10d70541a1a71a54821e27800e96e operator/frontier_results.json
32
+ 9b636b73ac08cbd973cf3246cce15f6d6a3a2ff459f2bc4fc7bc8522794be26c operator/operator_writer.py
33
+ 448f63059f9553f0c1da659ccbc61feba81f131856f05738ef7011c2af6f78e6 operator/results.json
34
+ 161030a08b8d5601c3558f39677b4f994218b17e32b5ce71acfded9525dffefd operator/tasks.csv
35
+ c77e8f7d35890de4c8115538d28e27a5a43bb1ef2f6bed4550dd954fd84c435b operator/verify_frontier.py
36
+ 758f186198570c2c7501b2dd926bced8dfc355cac04eb9c003677b94fa3b79d1 portfolio/README.md
37
+ 97b5fe40babe4ed4fee05d8be3f00853975833910fbe8987aeb6f2af0cc696b3 portfolio/check_race.py
38
+ ef62f94c4865ef99c47054f0bc14e52de6bf98a2cf90403d3f9da59f94e9e2a5 portfolio/episodes.csv
39
+ 7acf50d9c525348bc7a78b38ce63885b59cf0ca9641cceb1cc86bcba53762162 portfolio/exhaustive.json
40
+ da603d0ce65393bdf7a716086762e6fc878759419851038e6ec2241fcf5b09de portfolio/macro_race.py
41
+ b00be2bdb5eec03d3bd0e1d586038652542a847e83ff8d128a46b73a5ed59bfb portfolio/program_certificates.json
42
+ 46386169a862fab1c10102200d5bec269386bf7a0cdfc5c786661c068ab37b53 portfolio/results.json
43
+ 5bc98f374071fcfa642314f62df88e51ccc880668b431fa007f8a396160cf8ca realdata/README.md
44
+ 6084c63d2af133535e69cad840035fc4405d637fa559193b3954cff3a55cf957 realdata/digits_audit.npz
45
+ 8aac7d3d7b5501b6bebdbdd64e26d8449bbe3e3d8b3ac99534745cb6df8db8a1 realdata/digits_repair.py
46
+ a605608ca50d1d55f3fd218c297f6682a907c00b1d0ddaea94893991b5d29b3a realdata/digits_results.json
47
+ 1b523e8ab220c2655c996aebb209f2feaae9cb205b81ea4950ec5f70133f9657 realdata/verify_digits.py
48
+ fa951118e3186b7c6030c853a68f21a56d5fa76edc11187a28f4cfc65d3e7c69 realfield/README.md
49
+ e1b0138f3ee8cfbc7cc231a4980a5c4c4123ea8935d77411310d6cfe0704ed67 realfield/energy_results.json
50
+ 4a41ed2b420be40bee9fde1bf17182e8b5d0a5630fe192c64363b2c36bdfe94f realfield/energy_writer.py
51
+ 6b5ea08f9a4ba4b3c38b08989c9b4892608376c98618db258b65c9cc72179d82 robust/README.md
52
+ e5933ea0dc4de2a7f77fd23ec17fca159e756378e63deeafbcfdbb4924ad82c8 robust/robust_ledger.csv
53
+ a5f9c10a0adebc427e9f7857e2af670494965ed8d8e83bad12c7baef02dc2154 robust/robust_summary.json
54
+ b9f23d2823ffd5eee780d65ace139496097540b9d6d9e7bf042b8374234e7c4b robust/verify_robust.py
55
+ 78a5856516598008b38fdb99de29579e8974966ba0916d572b8125567deb175b tower/README.md
56
+ c2e49d227be5b1d4a8204dc167113b3ba8f611a937e40be7c4270ec64ee177ea tower/results.json
57
+ e49b60fe3c33b3ad6972a50ca48c4f921a1af65c3251373f9a2bd4a32e1ad439 tower/tower.py
58
+ a94e6d12c454633743d1114c990bf936e97baa15c51f7e795925ffe0f3c80fbb tower/worlds.csv
59
+ f033d0a8f465b4d58015caa5eb788562a5b314ab53c0abf5a6424578ebcc9695 transfer/README.md
60
+ a72b12b0c59036f9581f215630e7298257e74b27d0a50628801d7d997043970f transfer/crossover.py
61
+ 9fbe3b7e10433c9d11fa8badbe6dc75a9607433b44e96098b7a2508fa8d75111 transfer/crossover_summary.json
62
+ 6940c714e9392417ab4e11a951acd816c5562ad4f59a737709f2cc578ce32964 transfer/crossover_worlds.csv
63
+ ce86676977d3f4bc82c50068debb0c0ab2347c9f3e15746759ff4ae124291fc3 transfer/transfer.py
64
+ d8c570fa19b0853c57b656c2b4b867fd00207c7049862239c3a5a2723a4aab44 transfer/transfer_budget.csv
65
+ 887b1d4280f8ceaaf6cd136d821be6acc38158c16fc2d54b74321f3537298a4b transfer/transfer_curve.csv
66
+ 1ba1cf191a88426ddf3fc81a670dadb74350245a733f8a05867b62d3c4760455 transfer/transfer_logits.json
67
+ c0d7eefe0c60d29d75bb0beca633cb34d6f1270b5626226e6a5b625ea9799080 transfer/transfer_recipient.csv
68
+ 8ff7c0727ee27102ab71393bd029ef30e93ea297fa7a80073c3282fed548bf42 transfer/transfer_summary.json
69
+ d5f01a1c8fd1c4d7b1ecf330b7bb317df27da1e17303db9cd6ab6e714bd526f4 transfer/transfer_trajectories.csv
70
+ e0f460e6ad2d19cb24838b6d4b0faccdd945641a2c85ccb9ef24a42a502e3718 transfer/verify_transfer.py
reproduction/algebra/changing_operators.py ADDED
@@ -0,0 +1,188 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """Exact cross-operator transfer and guard against a silent change in shared action."""
2
+ import csv
3
+ import json
4
+ import random
5
+ import statistics
6
+ from pathlib import Path
7
+ from defect_closure import Graph,ProductCache,ColdShared,act,inverse,invertible,rank
8
+
9
+ HERE=Path(__file__).resolve().parent
10
+
11
+
12
+ class StaleAction(Exception):
13
+ pass
14
+
15
+
16
+ def copied(graph):
17
+ out=Graph()
18
+ out.rows=graph.rows.copy()
19
+ return out
20
+
21
+
22
+ def propose(h,b,W,invW):
23
+ w=h.apply(b)^b
24
+ v=w
25
+ while W.get(v) is None:
26
+ hv=h.apply(v)
27
+ previous=invW.get(hv)
28
+ if previous is not None and previous!=v:
29
+ raise StaleAction('Two candidate input directions map to the same cached image.')
30
+ W.insert(v,0)
31
+ invW.insert(hv,v)
32
+ v=hv
33
+ z=invW.get(w)
34
+ if z is None:
35
+ raise StaleAction('The residual has no cached inverse.')
36
+ return b^z
37
+
38
+
39
+ class SharedActionEditor:
40
+ def __init__(self,B):
41
+ self.B=B
42
+ self.W=Graph()
43
+ self.invW=Graph()
44
+ self.calls=0
45
+ self.rejected=0
46
+
47
+ def solve(self,b,actual_product,guard=True):
48
+ # Product data away from the certified persistent subspace is local
49
+ # to this task's operator. The fallback shares all of this local data.
50
+ h=ProductCache(lambda v:actual_product(act(self.B,v)))
51
+ W,invW=copied(self.W),copied(self.invW)
52
+ try:
53
+ y=propose(h,b,W,invW)
54
+ if guard and h.apply(y)!=b:
55
+ raise StaleAction('The proposed solution fails its current-operator residual check.')
56
+ except StaleAction:
57
+ self.rejected+=1
58
+ W,invW=Graph(),Graph()
59
+ y=propose(h,b,W,invW)
60
+ assert h.graph.get(y)==b
61
+ self.W,self.invW=W,invW
62
+ self.calls+=h.calls
63
+ assert h.calls<=len(self.B)
64
+ return act(self.B,y),h.calls
65
+
66
+
67
+ def mixed_target(source,p,pinv,j,rank_cap,rng):
68
+ d=len(source)
69
+ cross=[rng.randrange(1<<rank_cap) for _ in range(d-rank_cap)]
70
+ block=list(j)+[(1<<(rank_cap+i))^z for i,z in enumerate(cross)]
71
+ h=[act(p,act(block,v)) for v in pinv]
72
+ return tuple(act(h,v) for v in source),tuple(h)
73
+
74
+
75
+ def across_domains():
76
+ d=64
77
+ r=4
78
+ streams=24
79
+ T=20
80
+ source=invertible(d,random.Random(329))
81
+ B=inverse(source)
82
+ rows=[]
83
+ for world in range(streams):
84
+ rng=random.Random(81315+world)
85
+ p=invertible(d,rng)
86
+ pinv=inverse(p)
87
+ j=invertible(r,rng)
88
+ shared=SharedActionEditor(B)
89
+ guarded=SharedActionEditor(B)
90
+ reset_calls=cold_calls=anchor_calls=0
91
+ anchor=None
92
+ previous_operators=set()
93
+ objectives=set()
94
+ for t in range(1,T+1):
95
+ target,h=mixed_target(source,p,pinv,j,r,rng)
96
+ assert target not in previous_operators
97
+ previous_operators.add(target)
98
+ assert rank([h[i]^(1<<i) for i in range(d)])==r
99
+ b=rng.randrange(1,1<<d)
100
+ assert b not in objectives
101
+ objectives.add(b)
102
+ oracle=lambda v:act(target,v)
103
+ x,n=shared.solve(b,oracle,guard=False)
104
+ gx,gn=guarded.solve(b,oracle,guard=True)
105
+ assert oracle(x)==b and gx==x
106
+ single=SharedActionEditor(B)
107
+ sx,sn=single.solve(b,oracle,guard=False)
108
+ reset_calls+=sn
109
+ cold=ColdShared(d,oracle)
110
+ assert cold.solve(b)==sx==x
111
+ cold_calls+=cold.calls
112
+ # A known inverse of an actual target-family member makes
113
+ # (A_t A_1^-1 - I)^2 = 0. Use a fully acquired cold anchor.
114
+ if anchor is None:
115
+ anchor=inverse([oracle(1<<k) for k in range(d)])
116
+ anchor_calls+=d
117
+ ax=act(anchor,b)
118
+ else:
119
+ # In characteristic two, 2I-G = G.
120
+ ax=act(anchor,oracle(act(anchor,b)))
121
+ anchor_calls+=1
122
+ assert ax==x
123
+ assert shared.calls<=t+r and guarded.calls<=2*t+r
124
+ assert shared.rejected==guarded.rejected==0
125
+ rows.append(dict(world=world,task=t,raw_calls=n,guarded_calls=gn,
126
+ transferred_cumulative=shared.calls,guarded_cumulative=guarded.calls,
127
+ reset_source_cumulative=reset_calls,cold_reset_cumulative=cold_calls,
128
+ cold_family_anchor_cumulative=anchor_calls,closure_rank=len(shared.W.rows)))
129
+ with (HERE/'changing_worlds.csv').open('w',newline='') as stream:
130
+ w=csv.DictWriter(stream,fieldnames=list(rows[0]))
131
+ w.writeheader()
132
+ w.writerows(rows)
133
+ summary={}
134
+ for t in (1,4,8,T):
135
+ sample=[x for x in rows if x['task']==t]
136
+ summary[t]={k:{'mean':statistics.mean(x[k] for x in sample),
137
+ 'min':min(x[k] for x in sample),'max':max(x[k] for x in sample)}
138
+ for k in list(rows[0])[4:]}
139
+ return dict(d=d,rank_cap=r,streams=streams,distinct_new_operators_per_stream=T,
140
+ distinct_new_objectives_per_stream=T,source_acquisition_once=64,
141
+ all_480_correct=True,shared_invariant_action_within_each_stream=True,counts=summary)
142
+
143
+
144
+ def silent_change():
145
+ # Bit index 0 is e1, 1 is e2, 2 is e3. Old core identity, new core swap.
146
+ # Both matrices have the same changed output subspace bound span(e1,e2).
147
+ B=(1,2,4)
148
+ first=(1,2,5)
149
+ second=(2,1,5)
150
+ b=4
151
+ naive=SharedActionEditor(B)
152
+ guarded=SharedActionEditor(B)
153
+ a,_=naive.solve(b,lambda v:act(first,v),False)
154
+ ga,_=guarded.solve(b,lambda v:act(first,v),True)
155
+ assert a==ga==5
156
+ incorrect,spent=naive.solve(b,lambda v:act(second,v),False)
157
+ corrected,gspent=guarded.solve(b,lambda v:act(second,v),True)
158
+ assert incorrect==5 and act(second,incorrect)==7
159
+ assert corrected==6 and act(second,corrected)==b and guarded.rejected==1
160
+ return dict(target=b,first_solution=a,unguarded_second_solution=incorrect,
161
+ unguarded_second_output=act(second,incorrect),guarded_second_solution=corrected,
162
+ guarded_second_output=act(second,corrected),unguarded_second_calls=spent,
163
+ guarded_second_calls=gspent,rejections=guarded.rejected)
164
+
165
+
166
+ def guard_stress():
167
+ # Completely unrelated dense operators can violate every promise except
168
+ # invertibility. Each output must still solve the current task.
169
+ d=10
170
+ rng=random.Random(938)
171
+ editor=SharedActionEditor(tuple(1<<j for j in range(d)))
172
+ cost=[]
173
+ for _ in range(120):
174
+ target=invertible(d,rng)
175
+ b=rng.randrange(1,1<<d)
176
+ route,n=editor.solve(b,lambda v:act(target,v),True)
177
+ assert act(target,route)==b and n<=d
178
+ cost.append(n)
179
+ return dict(unrelated_operators=len(cost),dimension=d,all_exact=True,
180
+ stale_hypotheses_rejected=editor.rejected,max_products_on_any_operator=max(cost),
181
+ mean_products=statistics.mean(cost))
182
+
183
+
184
+ if __name__=='__main__':
185
+ out={'promised_changing_operators':across_domains(),'silent_change':silent_change(),
186
+ 'unrelated_operator_stress':guard_stress()}
187
+ (HERE/'changing_results.json').write_text(json.dumps(out,indent=2)+'\n')
188
+ print(json.dumps(out,indent=2))
reproduction/algebra/changing_results.json ADDED
@@ -0,0 +1,161 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "promised_changing_operators": {
3
+ "d": 64,
4
+ "rank_cap": 4,
5
+ "streams": 24,
6
+ "distinct_new_operators_per_stream": 20,
7
+ "distinct_new_objectives_per_stream": 20,
8
+ "source_acquisition_once": 64,
9
+ "all_480_correct": true,
10
+ "shared_invariant_action_within_each_stream": true,
11
+ "counts": {
12
+ "1": {
13
+ "transferred_cumulative": {
14
+ "mean": 4.083333333333333,
15
+ "min": 1,
16
+ "max": 5
17
+ },
18
+ "guarded_cumulative": {
19
+ "mean": 4.083333333333333,
20
+ "min": 1,
21
+ "max": 5
22
+ },
23
+ "reset_source_cumulative": {
24
+ "mean": 4.083333333333333,
25
+ "min": 1,
26
+ "max": 5
27
+ },
28
+ "cold_reset_cumulative": {
29
+ "mean": 62.416666666666664,
30
+ "min": 58,
31
+ "max": 64
32
+ },
33
+ "cold_family_anchor_cumulative": {
34
+ "mean": 64,
35
+ "min": 64,
36
+ "max": 64
37
+ },
38
+ "closure_rank": {
39
+ "mean": 3.0833333333333335,
40
+ "min": 0,
41
+ "max": 4
42
+ }
43
+ },
44
+ "4": {
45
+ "transferred_cumulative": {
46
+ "mean": 8,
47
+ "min": 8,
48
+ "max": 8
49
+ },
50
+ "guarded_cumulative": {
51
+ "mean": 10.666666666666666,
52
+ "min": 9,
53
+ "max": 11
54
+ },
55
+ "reset_source_cumulative": {
56
+ "mean": 16.083333333333332,
57
+ "min": 10,
58
+ "max": 20
59
+ },
60
+ "cold_reset_cumulative": {
61
+ "mean": 251.70833333333334,
62
+ "min": 243,
63
+ "max": 255
64
+ },
65
+ "cold_family_anchor_cumulative": {
66
+ "mean": 67,
67
+ "min": 67,
68
+ "max": 67
69
+ },
70
+ "closure_rank": {
71
+ "mean": 4,
72
+ "min": 4,
73
+ "max": 4
74
+ }
75
+ },
76
+ "8": {
77
+ "transferred_cumulative": {
78
+ "mean": 12,
79
+ "min": 12,
80
+ "max": 12
81
+ },
82
+ "guarded_cumulative": {
83
+ "mean": 18.416666666666668,
84
+ "min": 17,
85
+ "max": 19
86
+ },
87
+ "reset_source_cumulative": {
88
+ "mean": 32.375,
89
+ "min": 22,
90
+ "max": 40
91
+ },
92
+ "cold_reset_cumulative": {
93
+ "mean": 503.9583333333333,
94
+ "min": 499,
95
+ "max": 509
96
+ },
97
+ "cold_family_anchor_cumulative": {
98
+ "mean": 71,
99
+ "min": 71,
100
+ "max": 71
101
+ },
102
+ "closure_rank": {
103
+ "mean": 4,
104
+ "min": 4,
105
+ "max": 4
106
+ }
107
+ },
108
+ "20": {
109
+ "transferred_cumulative": {
110
+ "mean": 24,
111
+ "min": 24,
112
+ "max": 24
113
+ },
114
+ "guarded_cumulative": {
115
+ "mean": 41.583333333333336,
116
+ "min": 39,
117
+ "max": 43
118
+ },
119
+ "reset_source_cumulative": {
120
+ "mean": 79.66666666666667,
121
+ "min": 55,
122
+ "max": 100
123
+ },
124
+ "cold_reset_cumulative": {
125
+ "mean": 1260.1666666666667,
126
+ "min": 1249,
127
+ "max": 1269
128
+ },
129
+ "cold_family_anchor_cumulative": {
130
+ "mean": 83,
131
+ "min": 83,
132
+ "max": 83
133
+ },
134
+ "closure_rank": {
135
+ "mean": 4,
136
+ "min": 4,
137
+ "max": 4
138
+ }
139
+ }
140
+ }
141
+ },
142
+ "silent_change": {
143
+ "target": 4,
144
+ "first_solution": 5,
145
+ "unguarded_second_solution": 5,
146
+ "unguarded_second_output": 7,
147
+ "guarded_second_solution": 6,
148
+ "guarded_second_output": 4,
149
+ "unguarded_second_calls": 1,
150
+ "guarded_second_calls": 3,
151
+ "rejections": 1
152
+ },
153
+ "unrelated_operator_stress": {
154
+ "unrelated_operators": 120,
155
+ "dimension": 10,
156
+ "all_exact": true,
157
+ "stale_hypotheses_rejected": 117,
158
+ "max_products_on_any_operator": 10,
159
+ "mean_products": 9.183333333333334
160
+ }
161
+ }
reproduction/algebra/changing_worlds.csv ADDED
@@ -0,0 +1,481 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ world,task,raw_calls,guarded_calls,transferred_cumulative,guarded_cumulative,reset_source_cumulative,cold_reset_cumulative,cold_family_anchor_cumulative,closure_rank
2
+ 0,1,3,3,3,3,3,63,64,2
3
+ 0,2,2,2,5,5,5,127,65,3
4
+ 0,3,1,2,6,7,7,191,66,3
5
+ 0,4,2,3,8,10,10,252,67,4
6
+ 0,5,1,2,9,12,13,312,68,4
7
+ 0,6,1,2,10,14,16,375,69,4
8
+ 0,7,1,2,11,16,19,439,70,4
9
+ 0,8,1,2,12,18,22,501,71,4
10
+ 0,9,1,2,13,20,24,564,72,4
11
+ 0,10,1,2,14,22,27,622,73,4
12
+ 0,11,1,2,15,24,30,685,74,4
13
+ 0,12,1,2,16,26,33,749,75,4
14
+ 0,13,1,2,17,28,36,812,76,4
15
+ 0,14,1,2,18,30,39,875,77,4
16
+ 0,15,1,2,19,32,41,939,78,4
17
+ 0,16,1,2,20,34,44,1003,79,4
18
+ 0,17,1,2,21,36,47,1066,80,4
19
+ 0,18,1,2,22,38,50,1127,81,4
20
+ 0,19,1,2,23,40,52,1191,82,4
21
+ 0,20,1,2,24,42,55,1254,83,4
22
+ 1,1,4,4,4,4,4,63,64,3
23
+ 1,2,1,2,5,6,8,127,65,3
24
+ 1,3,2,3,7,9,13,191,66,4
25
+ 1,4,1,2,8,11,15,254,67,4
26
+ 1,5,1,1,9,12,16,317,68,4
27
+ 1,6,1,2,10,14,21,381,69,4
28
+ 1,7,1,2,11,16,26,443,70,4
29
+ 1,8,1,2,12,18,30,506,71,4
30
+ 1,9,1,2,13,20,34,570,72,4
31
+ 1,10,1,2,14,22,39,632,73,4
32
+ 1,11,1,2,15,24,44,696,74,4
33
+ 1,12,1,2,16,26,48,760,75,4
34
+ 1,13,1,2,17,28,52,822,76,4
35
+ 1,14,1,2,18,30,56,885,77,4
36
+ 1,15,1,2,19,32,60,949,78,4
37
+ 1,16,1,2,20,34,65,1011,79,4
38
+ 1,17,1,2,21,36,69,1075,80,4
39
+ 1,18,1,2,22,38,71,1138,81,4
40
+ 1,19,1,2,23,40,76,1200,82,4
41
+ 1,20,1,2,24,42,81,1264,83,4
42
+ 2,1,5,5,5,5,5,63,64,4
43
+ 2,2,1,2,6,7,10,126,65,4
44
+ 2,3,1,2,7,9,15,187,66,4
45
+ 2,4,1,2,8,11,20,251,67,4
46
+ 2,5,1,2,9,13,25,312,68,4
47
+ 2,6,1,2,10,15,30,375,69,4
48
+ 2,7,1,1,11,16,31,439,70,4
49
+ 2,8,1,2,12,18,36,503,71,4
50
+ 2,9,1,2,13,20,41,567,72,4
51
+ 2,10,1,2,14,22,46,631,73,4
52
+ 2,11,1,2,15,24,51,694,74,4
53
+ 2,12,1,2,16,26,56,757,75,4
54
+ 2,13,1,1,17,27,57,819,76,4
55
+ 2,14,1,2,18,29,62,878,77,4
56
+ 2,15,1,2,19,31,67,942,78,4
57
+ 2,16,1,2,20,33,72,1005,79,4
58
+ 2,17,1,2,21,35,77,1069,80,4
59
+ 2,18,1,2,22,37,82,1132,81,4
60
+ 2,19,1,1,23,38,83,1193,82,4
61
+ 2,20,1,2,24,40,88,1255,83,4
62
+ 3,1,5,5,5,5,5,63,64,4
63
+ 3,2,1,2,6,7,10,124,65,4
64
+ 3,3,1,2,7,9,15,188,66,4
65
+ 3,4,1,2,8,11,19,252,67,4
66
+ 3,5,1,2,9,13,22,316,68,4
67
+ 3,6,1,2,10,15,26,379,69,4
68
+ 3,7,1,2,11,17,29,442,70,4
69
+ 3,8,1,2,12,19,32,505,71,4
70
+ 3,9,1,2,13,21,37,568,72,4
71
+ 3,10,1,2,14,23,41,630,73,4
72
+ 3,11,1,1,15,24,42,694,74,4
73
+ 3,12,1,2,16,26,45,758,75,4
74
+ 3,13,1,2,17,28,50,822,76,4
75
+ 3,14,1,2,18,30,53,886,77,4
76
+ 3,15,1,2,19,32,56,950,78,4
77
+ 3,16,1,2,20,34,61,1010,79,4
78
+ 3,17,1,2,21,36,64,1074,80,4
79
+ 3,18,1,2,22,38,67,1137,81,4
80
+ 3,19,1,2,23,40,70,1201,82,4
81
+ 3,20,1,2,24,42,75,1265,83,4
82
+ 4,1,4,4,4,4,4,64,64,3
83
+ 4,2,1,1,5,5,5,126,65,3
84
+ 4,3,2,2,7,7,7,189,66,4
85
+ 4,4,1,2,8,9,12,253,67,4
86
+ 4,5,1,2,9,11,17,315,68,4
87
+ 4,6,1,2,10,13,21,379,69,4
88
+ 4,7,1,2,11,15,25,442,70,4
89
+ 4,8,1,2,12,17,30,506,71,4
90
+ 4,9,1,2,13,19,34,570,72,4
91
+ 4,10,1,2,14,21,39,633,73,4
92
+ 4,11,1,2,15,23,43,696,74,4
93
+ 4,12,1,2,16,25,47,758,75,4
94
+ 4,13,1,1,17,26,48,822,76,4
95
+ 4,14,1,2,18,28,53,885,77,4
96
+ 4,15,1,2,19,30,58,949,78,4
97
+ 4,16,1,2,20,32,62,1013,79,4
98
+ 4,17,1,2,21,34,67,1075,80,4
99
+ 4,18,1,2,22,36,71,1138,81,4
100
+ 4,19,1,2,23,38,76,1201,82,4
101
+ 4,20,1,2,24,40,80,1265,83,4
102
+ 5,1,5,5,5,5,5,63,64,4
103
+ 5,2,1,2,6,7,8,127,65,4
104
+ 5,3,1,2,7,9,11,191,66,4
105
+ 5,4,1,2,8,11,14,254,67,4
106
+ 5,5,1,2,9,13,17,318,68,4
107
+ 5,6,1,2,10,15,22,382,69,4
108
+ 5,7,1,2,11,17,25,446,70,4
109
+ 5,8,1,2,12,19,28,505,71,4
110
+ 5,9,1,2,13,21,30,569,72,4
111
+ 5,10,1,2,14,23,33,633,73,4
112
+ 5,11,1,2,15,25,36,695,74,4
113
+ 5,12,1,2,16,27,41,758,75,4
114
+ 5,13,1,1,17,28,42,822,76,4
115
+ 5,14,1,2,18,30,47,886,77,4
116
+ 5,15,1,2,19,32,50,950,78,4
117
+ 5,16,1,2,20,34,53,1011,79,4
118
+ 5,17,1,2,21,36,58,1075,80,4
119
+ 5,18,1,2,22,38,61,1138,81,4
120
+ 5,19,1,1,23,39,62,1201,82,4
121
+ 5,20,1,2,24,41,67,1265,83,4
122
+ 6,1,5,5,5,5,5,62,64,4
123
+ 6,2,1,2,6,7,10,125,65,4
124
+ 6,3,1,2,7,9,15,189,66,4
125
+ 6,4,1,2,8,11,20,252,67,4
126
+ 6,5,1,2,9,13,25,311,68,4
127
+ 6,6,1,2,10,15,30,373,69,4
128
+ 6,7,1,2,11,17,35,437,70,4
129
+ 6,8,1,1,12,18,36,501,71,4
130
+ 6,9,1,1,13,19,37,563,72,4
131
+ 6,10,1,2,14,21,42,627,73,4
132
+ 6,11,1,2,15,23,47,691,74,4
133
+ 6,12,1,2,16,25,52,754,75,4
134
+ 6,13,1,2,17,27,57,816,76,4
135
+ 6,14,1,2,18,29,62,878,77,4
136
+ 6,15,1,1,19,30,63,941,78,4
137
+ 6,16,1,2,20,32,68,1005,79,4
138
+ 6,17,1,1,21,33,69,1066,80,4
139
+ 6,18,1,2,22,35,74,1125,81,4
140
+ 6,19,1,2,23,37,79,1188,82,4
141
+ 6,20,1,2,24,39,84,1251,83,4
142
+ 7,1,1,1,1,1,1,63,64,0
143
+ 7,2,5,5,6,6,6,127,65,4
144
+ 7,3,1,2,7,8,10,189,66,4
145
+ 7,4,1,2,8,10,14,252,67,4
146
+ 7,5,1,2,9,12,19,316,68,4
147
+ 7,6,1,2,10,14,24,380,69,4
148
+ 7,7,1,2,11,16,29,444,70,4
149
+ 7,8,1,2,12,18,34,505,71,4
150
+ 7,9,1,2,13,20,39,569,72,4
151
+ 7,10,1,1,14,21,40,630,73,4
152
+ 7,11,1,1,15,22,41,693,74,4
153
+ 7,12,1,2,16,24,45,757,75,4
154
+ 7,13,1,2,17,26,50,816,76,4
155
+ 7,14,1,2,18,28,55,880,77,4
156
+ 7,15,1,2,19,30,60,942,78,4
157
+ 7,16,1,2,20,32,64,1005,79,4
158
+ 7,17,1,2,21,34,66,1069,80,4
159
+ 7,18,1,2,22,36,71,1133,81,4
160
+ 7,19,1,2,23,38,75,1197,82,4
161
+ 7,20,1,2,24,40,80,1261,83,4
162
+ 8,1,5,5,5,5,5,60,64,4
163
+ 8,2,1,2,6,7,10,120,65,4
164
+ 8,3,1,2,7,9,15,182,66,4
165
+ 8,4,1,2,8,11,20,243,67,4
166
+ 8,5,1,2,9,13,25,307,68,4
167
+ 8,6,1,2,10,15,30,371,69,4
168
+ 8,7,1,2,11,17,35,435,70,4
169
+ 8,8,1,2,12,19,40,499,71,4
170
+ 8,9,1,2,13,21,45,563,72,4
171
+ 8,10,1,2,14,23,50,627,73,4
172
+ 8,11,1,2,15,25,55,689,74,4
173
+ 8,12,1,2,16,27,60,752,75,4
174
+ 8,13,1,2,17,29,65,813,76,4
175
+ 8,14,1,2,18,31,70,872,77,4
176
+ 8,15,1,2,19,33,75,936,78,4
177
+ 8,16,1,2,20,35,80,1000,79,4
178
+ 8,17,1,2,21,37,85,1062,80,4
179
+ 8,18,1,2,22,39,90,1126,81,4
180
+ 8,19,1,2,23,41,95,1190,82,4
181
+ 8,20,1,2,24,43,100,1249,83,4
182
+ 9,1,2,2,2,2,2,64,64,1
183
+ 9,2,4,4,6,6,6,123,65,4
184
+ 9,3,1,2,7,8,8,187,66,4
185
+ 9,4,1,2,8,10,10,248,67,4
186
+ 9,5,1,2,9,12,13,311,68,4
187
+ 9,6,1,2,10,14,15,375,69,4
188
+ 9,7,1,2,11,16,19,437,70,4
189
+ 9,8,1,2,12,18,23,501,71,4
190
+ 9,9,1,2,13,20,27,565,72,4
191
+ 9,10,1,1,14,21,28,629,73,4
192
+ 9,11,1,2,15,23,32,692,74,4
193
+ 9,12,1,2,16,25,35,756,75,4
194
+ 9,13,1,2,17,27,38,818,76,4
195
+ 9,14,1,2,18,29,42,882,77,4
196
+ 9,15,1,2,19,31,45,946,78,4
197
+ 9,16,1,2,20,33,49,1010,79,4
198
+ 9,17,1,2,21,35,53,1074,80,4
199
+ 9,18,1,2,22,37,57,1138,81,4
200
+ 9,19,1,2,23,39,60,1197,82,4
201
+ 9,20,1,1,24,40,61,1261,83,4
202
+ 10,1,5,5,5,5,5,59,64,4
203
+ 10,2,1,2,6,7,10,123,65,4
204
+ 10,3,1,1,7,8,11,187,66,4
205
+ 10,4,1,2,8,10,13,248,67,4
206
+ 10,5,1,2,9,12,18,310,68,4
207
+ 10,6,1,2,10,14,23,374,69,4
208
+ 10,7,1,1,11,15,24,435,70,4
209
+ 10,8,1,2,12,17,28,499,71,4
210
+ 10,9,1,2,13,19,33,562,72,4
211
+ 10,10,1,2,14,21,37,626,73,4
212
+ 10,11,1,2,15,23,42,690,74,4
213
+ 10,12,1,2,16,25,46,754,75,4
214
+ 10,13,1,2,17,27,51,817,76,4
215
+ 10,14,1,2,18,29,55,880,77,4
216
+ 10,15,1,2,19,31,60,942,78,4
217
+ 10,16,1,2,20,33,63,1004,79,4
218
+ 10,17,1,2,21,35,66,1067,80,4
219
+ 10,18,1,2,22,37,71,1131,81,4
220
+ 10,19,1,2,23,39,74,1195,82,4
221
+ 10,20,1,2,24,41,77,1258,83,4
222
+ 11,1,4,4,4,4,4,64,64,3
223
+ 11,2,1,2,5,6,8,125,65,3
224
+ 11,3,2,3,7,9,13,189,66,4
225
+ 11,4,1,2,8,11,18,253,67,4
226
+ 11,5,1,2,9,13,23,316,68,4
227
+ 11,6,1,2,10,15,28,380,69,4
228
+ 11,7,1,2,11,17,33,441,70,4
229
+ 11,8,1,2,12,19,37,505,71,4
230
+ 11,9,1,2,13,21,41,569,72,4
231
+ 11,10,1,2,14,23,45,632,73,4
232
+ 11,11,1,2,15,25,49,695,74,4
233
+ 11,12,1,2,16,27,51,759,75,4
234
+ 11,13,1,2,17,29,55,823,76,4
235
+ 11,14,1,2,18,31,59,886,77,4
236
+ 11,15,1,2,19,33,63,948,78,4
237
+ 11,16,1,2,20,35,67,1012,79,4
238
+ 11,17,1,2,21,37,71,1069,80,4
239
+ 11,18,1,2,22,39,75,1131,81,4
240
+ 11,19,1,2,23,41,80,1192,82,4
241
+ 11,20,1,2,24,43,84,1255,83,4
242
+ 12,1,4,4,4,4,4,63,64,3
243
+ 12,2,1,2,5,6,8,127,65,3
244
+ 12,3,2,3,7,9,13,191,66,4
245
+ 12,4,1,2,8,11,18,252,67,4
246
+ 12,5,1,2,9,13,22,316,68,4
247
+ 12,6,1,2,10,15,26,380,69,4
248
+ 12,7,1,2,11,17,31,443,70,4
249
+ 12,8,1,2,12,19,36,507,71,4
250
+ 12,9,1,2,13,21,38,570,72,4
251
+ 12,10,1,2,14,23,43,634,73,4
252
+ 12,11,1,2,15,25,48,698,74,4
253
+ 12,12,1,2,16,27,53,755,75,4
254
+ 12,13,1,2,17,29,58,819,76,4
255
+ 12,14,1,2,18,31,60,883,77,4
256
+ 12,15,1,2,19,33,62,947,78,4
257
+ 12,16,1,1,20,34,63,1011,79,4
258
+ 12,17,1,2,21,36,67,1075,80,4
259
+ 12,18,1,2,22,38,71,1138,81,4
260
+ 12,19,1,2,23,40,76,1202,82,4
261
+ 12,20,1,2,24,42,80,1265,83,4
262
+ 13,1,5,5,5,5,5,64,64,4
263
+ 13,2,1,2,6,7,10,128,65,4
264
+ 13,3,1,2,7,9,15,190,66,4
265
+ 13,4,1,2,8,11,20,253,67,4
266
+ 13,5,1,2,9,13,25,317,68,4
267
+ 13,6,1,2,10,15,30,379,69,4
268
+ 13,7,1,1,11,16,31,442,70,4
269
+ 13,8,1,2,12,18,36,504,71,4
270
+ 13,9,1,2,13,20,41,567,72,4
271
+ 13,10,1,2,14,22,46,631,73,4
272
+ 13,11,1,2,15,24,51,693,74,4
273
+ 13,12,1,2,16,26,56,757,75,4
274
+ 13,13,1,2,17,28,61,821,76,4
275
+ 13,14,1,2,18,30,66,885,77,4
276
+ 13,15,1,2,19,32,71,948,78,4
277
+ 13,16,1,2,20,34,76,1011,79,4
278
+ 13,17,1,2,21,36,81,1075,80,4
279
+ 13,18,1,2,22,38,86,1139,81,4
280
+ 13,19,1,2,23,40,91,1203,82,4
281
+ 13,20,1,2,24,42,96,1267,83,4
282
+ 14,1,5,5,5,5,5,63,64,4
283
+ 14,2,1,2,6,7,10,127,65,4
284
+ 14,3,1,2,7,9,15,190,66,4
285
+ 14,4,1,2,8,11,20,254,67,4
286
+ 14,5,1,2,9,13,25,318,68,4
287
+ 14,6,1,2,10,15,30,380,69,4
288
+ 14,7,1,2,11,17,35,443,70,4
289
+ 14,8,1,2,12,19,40,507,71,4
290
+ 14,9,1,2,13,21,45,566,72,4
291
+ 14,10,1,2,14,23,50,622,73,4
292
+ 14,11,1,2,15,25,55,685,74,4
293
+ 14,12,1,2,16,27,60,749,75,4
294
+ 14,13,1,2,17,29,65,813,76,4
295
+ 14,14,1,2,18,31,70,877,77,4
296
+ 14,15,1,2,19,33,75,934,78,4
297
+ 14,16,1,2,20,35,80,998,79,4
298
+ 14,17,1,2,21,37,85,1062,80,4
299
+ 14,18,1,2,22,39,90,1126,81,4
300
+ 14,19,1,2,23,41,95,1190,82,4
301
+ 14,20,1,2,24,43,100,1254,83,4
302
+ 15,1,5,5,5,5,5,63,64,4
303
+ 15,2,1,2,6,7,8,127,65,4
304
+ 15,3,1,2,7,9,12,191,66,4
305
+ 15,4,1,2,8,11,16,254,67,4
306
+ 15,5,1,2,9,13,19,310,68,4
307
+ 15,6,1,1,10,14,20,374,69,4
308
+ 15,7,1,2,11,16,23,437,70,4
309
+ 15,8,1,2,12,18,27,501,71,4
310
+ 15,9,1,2,13,20,32,562,72,4
311
+ 15,10,1,2,14,22,35,626,73,4
312
+ 15,11,1,2,15,24,38,689,74,4
313
+ 15,12,1,2,16,26,41,751,75,4
314
+ 15,13,1,2,17,28,45,815,76,4
315
+ 15,14,1,2,18,30,47,878,77,4
316
+ 15,15,1,2,19,32,51,939,78,4
317
+ 15,16,1,2,20,34,54,1002,79,4
318
+ 15,17,1,2,21,36,59,1065,80,4
319
+ 15,18,1,2,22,38,64,1129,81,4
320
+ 15,19,1,2,23,40,67,1193,82,4
321
+ 15,20,1,2,24,42,70,1255,83,4
322
+ 16,1,3,3,3,3,3,64,64,2
323
+ 16,2,2,3,5,6,6,128,65,3
324
+ 16,3,2,3,7,9,10,191,66,4
325
+ 16,4,1,2,8,11,14,255,67,4
326
+ 16,5,1,2,9,13,16,317,68,4
327
+ 16,6,1,2,10,15,18,381,69,4
328
+ 16,7,1,2,11,17,21,442,70,4
329
+ 16,8,1,2,12,19,25,506,71,4
330
+ 16,9,1,2,13,21,29,568,72,4
331
+ 16,10,1,2,14,23,33,632,73,4
332
+ 16,11,1,2,15,25,37,696,74,4
333
+ 16,12,1,2,16,27,39,760,75,4
334
+ 16,13,1,2,17,29,41,823,76,4
335
+ 16,14,1,2,18,31,43,884,77,4
336
+ 16,15,1,2,19,33,47,947,78,4
337
+ 16,16,1,2,20,35,49,1009,79,4
338
+ 16,17,1,1,21,36,50,1072,80,4
339
+ 16,18,1,2,22,38,53,1136,81,4
340
+ 16,19,1,2,23,40,57,1200,82,4
341
+ 16,20,1,2,24,42,60,1264,83,4
342
+ 17,1,4,4,4,4,4,58,64,3
343
+ 17,2,1,1,5,5,5,122,65,3
344
+ 17,3,2,3,7,8,9,186,66,4
345
+ 17,4,1,2,8,10,12,249,67,4
346
+ 17,5,1,2,9,12,16,312,68,4
347
+ 17,6,1,2,10,14,20,375,69,4
348
+ 17,7,1,2,11,16,24,438,70,4
349
+ 17,8,1,2,12,18,27,502,71,4
350
+ 17,9,1,2,13,20,31,566,72,4
351
+ 17,10,1,1,14,21,32,630,73,4
352
+ 17,11,1,2,15,23,36,692,74,4
353
+ 17,12,1,2,16,25,40,755,75,4
354
+ 17,13,1,2,17,27,44,818,76,4
355
+ 17,14,1,2,18,29,48,882,77,4
356
+ 17,15,1,1,19,30,49,945,78,4
357
+ 17,16,1,2,20,32,52,1009,79,4
358
+ 17,17,1,2,21,34,55,1072,80,4
359
+ 17,18,1,2,22,36,59,1134,81,4
360
+ 17,19,1,2,23,38,63,1198,82,4
361
+ 17,20,1,2,24,40,67,1261,83,4
362
+ 18,1,5,5,5,5,5,64,64,4
363
+ 18,2,1,2,6,7,10,124,65,4
364
+ 18,3,1,2,7,9,15,188,66,4
365
+ 18,4,1,2,8,11,20,251,67,4
366
+ 18,5,1,2,9,13,24,314,68,4
367
+ 18,6,1,2,10,15,29,378,69,4
368
+ 18,7,1,2,11,17,34,441,70,4
369
+ 18,8,1,2,12,19,38,505,71,4
370
+ 18,9,1,2,13,21,42,569,72,4
371
+ 18,10,1,2,14,23,46,633,73,4
372
+ 18,11,1,2,15,25,48,695,74,4
373
+ 18,12,1,2,16,27,52,759,75,4
374
+ 18,13,1,2,17,29,54,821,76,4
375
+ 18,14,1,2,18,31,58,885,77,4
376
+ 18,15,1,2,19,33,63,946,78,4
377
+ 18,16,1,2,20,35,65,1010,79,4
378
+ 18,17,1,2,21,37,70,1071,80,4
379
+ 18,18,1,2,22,39,74,1133,81,4
380
+ 18,19,1,2,23,41,79,1197,82,4
381
+ 18,20,1,2,24,43,84,1260,83,4
382
+ 19,1,4,4,4,4,4,58,64,3
383
+ 19,2,1,2,5,6,8,122,65,3
384
+ 19,3,2,3,7,9,13,185,66,4
385
+ 19,4,1,2,8,11,18,248,67,4
386
+ 19,5,1,2,9,13,23,309,68,4
387
+ 19,6,1,2,10,15,26,373,69,4
388
+ 19,7,1,2,11,17,31,436,70,4
389
+ 19,8,1,2,12,19,36,500,71,4
390
+ 19,9,1,2,13,21,41,561,72,4
391
+ 19,10,1,2,14,23,46,623,73,4
392
+ 19,11,1,2,15,25,50,685,74,4
393
+ 19,12,1,2,16,27,53,747,75,4
394
+ 19,13,1,2,17,29,58,810,76,4
395
+ 19,14,1,2,18,31,62,874,77,4
396
+ 19,15,1,2,19,33,67,938,78,4
397
+ 19,16,1,2,20,35,72,999,79,4
398
+ 19,17,1,2,21,37,75,1063,80,4
399
+ 19,18,1,2,22,39,78,1126,81,4
400
+ 19,19,1,2,23,41,83,1187,82,4
401
+ 19,20,1,2,24,43,86,1251,83,4
402
+ 20,1,2,2,2,2,2,64,64,1
403
+ 20,2,4,5,6,7,7,128,65,4
404
+ 20,3,1,2,7,9,11,191,66,4
405
+ 20,4,1,2,8,11,15,254,67,4
406
+ 20,5,1,2,9,13,19,317,68,4
407
+ 20,6,1,2,10,15,24,381,69,4
408
+ 20,7,1,2,11,17,29,445,70,4
409
+ 20,8,1,2,12,19,33,507,71,4
410
+ 20,9,1,1,13,20,34,570,72,4
411
+ 20,10,1,2,14,22,38,634,73,4
412
+ 20,11,1,2,15,24,42,697,74,4
413
+ 20,12,1,2,16,26,47,759,75,4
414
+ 20,13,1,2,17,28,51,823,76,4
415
+ 20,14,1,2,18,30,55,886,77,4
416
+ 20,15,1,2,19,32,60,948,78,4
417
+ 20,16,1,2,20,34,62,1012,79,4
418
+ 20,17,1,2,21,36,67,1075,80,4
419
+ 20,18,1,2,22,38,72,1139,81,4
420
+ 20,19,1,2,23,40,77,1202,82,4
421
+ 20,20,1,2,24,42,81,1266,83,4
422
+ 21,1,5,5,5,5,5,60,64,4
423
+ 21,2,1,2,6,7,10,124,65,4
424
+ 21,3,1,2,7,9,14,187,66,4
425
+ 21,4,1,2,8,11,19,251,67,4
426
+ 21,5,1,2,9,13,24,315,68,4
427
+ 21,6,1,2,10,15,28,378,69,4
428
+ 21,7,1,2,11,17,32,442,70,4
429
+ 21,8,1,2,12,19,36,505,71,4
430
+ 21,9,1,2,13,21,41,566,72,4
431
+ 21,10,1,2,14,23,46,629,73,4
432
+ 21,11,1,2,15,25,51,693,74,4
433
+ 21,12,1,2,16,27,55,757,75,4
434
+ 21,13,1,2,17,29,60,821,76,4
435
+ 21,14,1,2,18,31,65,884,77,4
436
+ 21,15,1,2,19,33,69,948,78,4
437
+ 21,16,1,2,20,35,74,1012,79,4
438
+ 21,17,1,2,21,37,79,1075,80,4
439
+ 21,18,1,2,22,39,84,1139,81,4
440
+ 21,19,1,2,23,41,88,1203,82,4
441
+ 21,20,1,2,24,43,92,1262,83,4
442
+ 22,1,3,3,3,3,3,64,64,2
443
+ 22,2,1,2,4,5,5,128,65,2
444
+ 22,3,3,4,7,9,10,191,66,4
445
+ 22,4,1,2,8,11,15,254,67,4
446
+ 22,5,1,2,9,13,20,318,68,4
447
+ 22,6,1,2,10,15,25,381,69,4
448
+ 22,7,1,2,11,17,30,443,70,4
449
+ 22,8,1,2,12,19,35,506,71,4
450
+ 22,9,1,2,13,21,38,569,72,4
451
+ 22,10,1,2,14,23,43,631,73,4
452
+ 22,11,1,2,15,25,48,695,74,4
453
+ 22,12,1,2,16,27,53,759,75,4
454
+ 22,13,1,2,17,29,58,823,76,4
455
+ 22,14,1,2,18,31,63,887,77,4
456
+ 22,15,1,2,19,33,66,951,78,4
457
+ 22,16,1,2,20,35,71,1015,79,4
458
+ 22,17,1,2,21,37,75,1079,80,4
459
+ 22,18,1,1,22,38,76,1142,81,4
460
+ 22,19,1,2,23,40,80,1205,82,4
461
+ 22,20,1,2,24,42,84,1269,83,4
462
+ 23,1,5,5,5,5,5,62,64,4
463
+ 23,2,1,2,6,7,9,126,65,4
464
+ 23,3,1,2,7,9,13,190,66,4
465
+ 23,4,1,1,8,10,14,254,67,4
466
+ 23,5,1,2,9,12,19,318,68,4
467
+ 23,6,1,2,10,14,23,382,69,4
468
+ 23,7,1,2,11,16,27,446,70,4
469
+ 23,8,1,2,12,18,32,509,71,4
470
+ 23,9,1,2,13,20,36,572,72,4
471
+ 23,10,1,2,14,22,38,636,73,4
472
+ 23,11,1,2,15,24,43,698,74,4
473
+ 23,12,1,2,16,26,48,761,75,4
474
+ 23,13,1,2,17,28,53,825,76,4
475
+ 23,14,1,1,18,29,54,889,77,4
476
+ 23,15,1,2,19,31,58,953,78,4
477
+ 23,16,1,2,20,33,62,1016,79,4
478
+ 23,17,1,2,21,35,66,1079,80,4
479
+ 23,18,1,2,22,37,71,1143,81,4
480
+ 23,19,1,2,23,39,75,1205,82,4
481
+ 23,20,1,2,24,41,80,1267,83,4
reproduction/algebra/defect_closure.md ADDED
@@ -0,0 +1,106 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Exact minimax transfer for a self-updating linear solver
2
+
3
+ ## Problem and cost
4
+
5
+ A source supplies an invertible linear rule $B$ in dimension $d$. On a new domain the expensive operation is a query $v\mapsto Av$ to a fixed unknown invertible operator. The learner receives right-hand sides $b_1,\ldots,b_T$ online and must return $x_t=A^{-1}b_t$ exactly before the next task. It may retain its own rewritten state. Put $H=AB$ and $E=H-I$. A product by $H$ costs one new product by $A$ and local application of $B$. All results count expensive vector products. Dense storage and local elimination have real costs. In the experiment the old source inverse is itself learned once, using $d$ products. Target operators are held fixed within a domain and reset between domains.
6
+
7
+ ## Exact closure algorithm
8
+
9
+ Keep a row-echelon table $G$ of queried pairs $(v,Hv)$, indexed by the independent inputs. Compute $Hv$ from $G$ whenever $v$ is in its domain and query the oracle once otherwise. Keep a second table describing an $H$-invariant subspace $W$ and inverse pairs $(Hv,v)$ for a basis of $W$. Start empty on entry to each new target domain.
10
+
11
+ 1. Obtain $w=Hb-b$ through $G$ for the new target $b$.
12
+ 2. Starting at $v=w$, while $v$ lies outside $W$, obtain $Hv$ through $G$, add the basis direction $v$ to $W$, append the pair $(Hv,v)$ to its inverse table, and advance $v\leftarrow Hv$.
13
+ 3. Recover $z=(H|_W)^{-1}w$ from the inverse table. Return $B(b-z)$.
14
+
15
+ After step 2 the new orbit has re-entered its expanding span. Each new direction maps either to its successor or into that span. The previous closed subspace has retained all its images. Consequently the expanded subspace is invariant, and the restriction of an invertible map to an invariant finite-dimensional subspace is invertible. The oracle transcript certifies the result: the graph contains $b$ and $z$, and $H(b-z)=Hb-w=b$. Over the binary field subtraction is XOR. The implementation checks this equation directly from the accumulated graph without a further oracle call. Solving new tasks modifies both the recognized query domain and the solver used next time.
16
+
17
+ ### Theorem 1. Exact pathwise product count
18
+
19
+ For $L_T=\operatorname{span}\{b_1,\ldots,b_T\}$ let $W_T$ be the smallest $H$-invariant subspace containing all $Eb_t$. The algorithm spends exactly
20
+ $$ Q_T=\dim(L_T+W_T)\le\min\{d,\dim L_T+\operatorname{rank}(AB-I)\}\le\min\{d,T+r\} $$
21
+ whenever the defect rank is at most $r$. The targets may be selected adaptively. The algorithm does not need the rank, the changed directions, or advance access to the targets.
22
+
23
+ **Proof.** Since $HE=EH$, the range of $E$ is invariant. The orbit-extension rule constructs precisely $W_T\subseteq\mathrm{im}(E)$ by induction. Every requested input of $G$ is either a target or a generating vector of $W_T$. Conversely all targets and every added generating vector were requested. The span of the cache domain therefore equals $L_T+W_T$. Each cache miss adds one independent direction at a cost of precisely one query. This proves the claimed identity and bounds. Each single first task takes at most $r+1$ products. A repeated target costs zero after its first solution. If transfer offers no low-rank structure, cost still never exceeds $d$ within that domain. This property guarantees exact answers under arbitrary invertible changes.
24
+
25
+ ### Theorem 2. Exact deterministic minimax cost
26
+
27
+ Over any field, for $2\le r<d$ and $T\ge1$, the largest number of products required in the worst case by an optimal deterministic exact solver is
28
+ $$ Q^*(d,r,T)=\min\{d,r+T\}. $$
29
+ The lower bound permits arbitrary adaptive product queries and unlimited side computation. It persists if the solver knows every target in advance and the coordinate subspace containing every changed output. This is a worst-case query bound, not an expected-time lower bound for randomized algorithms.
30
+
31
+ **Proof.** The upper bound is Theorem 1. For the lower bound take $k=\min(T,d-r)$, and make the other $d-r-k$ coordinates independent identities. Work on $V=Z\oplus D$ with dimensions $r,k$, respectively. It suffices to take $B=I$. Restrict $H$ to
32
+ $$ H(z,w)=(Uz+Cw,w),\qquad U\in\mathrm{GL}_r(\mathbb F). $$
33
+ Every such change has image defect contained in the known space $Z$ and rank at most $r$. The known targets are $(0,e_i)$ for $1\le i\le k$. An answer query reveals just the new information $M(z,w)=Uz+Cw$. Requested preimages are the $k$ independent vectors $(-U^{-1}Ce_i,e_i)$ spanning $\ker M$.
34
+
35
+ The adversary answers linearly on every query dependent on the previous queries. For the first $r$ independent inputs it assigns independent output vectors in $Z$. At every such step its partial map on the query span $S$ is injective and admits a completion with $M|_Z$ invertible. To see this, extend the partial map's restriction to $S\cap Z$ to an isomorphism of $Z$, combine the two maps on $S+Z$ where they agree on the overlap, and extend linearly to $V$. Once $r$ independent queries have been answered, keep an admissible completion fixed. If the algorithm stops earlier, select any admissible completion at stopping. In both cases, as long as $\dim S<r+k$, the nullity of $M|_S$ is less than $k$. Some $x\in\ker M$ lies outside $S$.
36
+
37
+ Choose a linear functional $f$ vanishing on $S$ with $f(x)=1$. Select $v\ne0$ in the kernel of $f|_Z\,U^{-1}$ on the output space $Z$, which is possible for $r\ge2$. Set $M'=M+vf$. The new oracle agrees on all of $S$ and gives the same adaptive transcript. Its restriction to $Z$ is still invertible by the rank-one determinant identity, since $f|_Z U^{-1}v=0$. But $M'x=v\ne0$, so its kernel and hence its list of $k$ requested solutions differ. The transcript cannot certify the correct answer for both admissible oracles. At least $r+k=\min(d,r+T)$ independent queries are required on this adversarial branch. Remaining targets may be repeated if $T>k$. This argument works over any field, including real or complex arithmetic in an exact product-oracle model. ∎
38
+
39
+ **Rank-one exception.** If the field has more than two elements, the same result holds for $r=1$: a nonzero scalar output perturbation can be chosen to avoid the sole scalar value that could cancel the determinant. Over the binary field and $d\ge2$, a known rank cap of one instead yields the exact minimax law $Q^*=\min(d,T)$. Indeed a nonzero defect is $E=uf$ for a nonzero vector $u$ and linear functional $f$. Invertibility requires $f(u)=0$, so $E^2=0$ and $H^{-1}=H$ in characteristic two. A single application of $H$ solves each target, and cached dependent products cost zero. For the lower bound use linearly independent targets up to dimension $\min(d,T)$ and answer every query as the identity. If the queried span misses some target $b$, take $f$ vanishing on that span with $f(b)=1$ and choose a nonzero $u$ in its null space. The invertible $I+uf$ gives identical answers to all queries and a different preimage of $b$. If the known defect cap is zero, every answer is $Bb$ without a product. The general rank-adaptive learner continues to check unfamiliar operators when no such promise is supplied.
40
+
41
+ ### Cold fixed-budget counterfactual
42
+
43
+ If $U$ is uniform on $\mathrm{GL}(d,2)$ with $d\ge2$, a solver gets a single fixed nonzero target, and there is no source information, the best probability of returning the exact preimage after $q<d$ adaptive products is $2^q/(2^d-1)$. A deterministic dependent query gives no new information. Conditional on the transcript, the image of the next independent input is uniform outside the previously discovered image span. The chance the target has entered the image span after $q$ queries is $(2^q-1)/(2^d-1)$. If it has not entered, its preimage is uniform among the $2^d-2^q$ unqueried-span-complement vectors. Equality follows by solving in-span or guessing once. Thus the mean number of calls for an always-correct fresh solve is at least $d2^d/(2^d-1)-1$. Querying a fixed basis until the target is in the known image realizes it. This last distributional calculation concerns the uniform group; the bit-packed demonstration uses its documented seeded generator.
44
+
45
+ ## Results and controls
46
+
47
+ Run `python defect_closure.py` to recreate `worlds.csv` and `results.json` using the standard library. All solver outputs and lower-bound decision searches use exact binary arithmetic. Mean target calls per domain across 24 seeded rank-four shifts of one 64-dimensional source are as follows.
48
+
49
+ | Fresh tasks | Rewritten defect closure | Reuse source but discard closure after each task | Cold solver sharing all target discoveries | Cold solver reset each task |
50
+ |---:|---:|---:|---:|---:|
51
+ | 1 | 4.33 | 4.33 | 63.54 | 63.54 |
52
+ | 4 | 7.96 | 16.63 | 63.92 | 251.83 |
53
+ | 8 | 12.00 | 33.58 | 64.00 | 504.33 |
54
+ | 20 | 24.00 | 82.92 | 64.00 | 1262.17 |
55
+
56
+ All 480 assigned tasks were solved exactly. Replaying every solution in its original domain required zero additional oracle products. The source was acquired once for 64 products. Thus on the 24 complete domains the method used 640 paid products in total including its source acquisition, against 1,536 for cold sharing. Verification by the independent test harness used ground-truth simulator outputs, which were unavailable to the learner and are excluded from its oracle accounting.
57
+
58
+ A separate decision-tree solver enumerates every operator in hard binary families of size 24, 96 and 384. At each reachable subset it considers every possible nonzero product query and every possible answer, with termination permitted as soon as all consistent matrices agree on the complete requested solution list. For $r=2$ and one, two and three tasks, the first feasible worst-case depths are 3, 4 and 5. This is independent finite corroboration of the oracle lower bound.
59
+
60
+ ## Relation to prior work and intended use
61
+
62
+ The identity-plus-low-rank single-system $r+1$ property and Krylov recycling have established precedents. The result here pairs a pathwise closure-dimension formula with a matching exact online minimax converse for a one-sided oracle, then treats the retained inverse rule as a self-update with a verifiable domain of correctness. No general priority claim is needed. A natural AI use is a stable, exactly checkable linear subproblem whose products dominate local bookkeeping, such as recovering an edit in a fixed linear response map. Applying this theorem to a nonlinear or noisy transformer weight-update map requires a separate model and error analysis.
63
+
64
+ - Carr, de Sturler and Embree, *Analysis of GMRES for Low-Rank and Small-Norm Perturbations of the Identity Matrix* (2022). https://arxiv.org/abs/2210.12053
65
+ - Parks, de Sturler, Mackey, Johnson and Maiti, *Recycling Krylov Subspaces for Sequences of Linear Systems* (2006). https://personal.math.vt.edu/sturler/publications/SISC_KrylovRecycling_2006.pdf
66
+ - Dereziński, Epperly and Meyer, *The matrix-vector complexity of Ax=b* (2026), characterizing a related class of real approximate-solve oracle lower bounds. https://arxiv.org/abs/2602.04842
67
+
68
+ ## One new operator on every task
69
+
70
+ A stronger model permits a different unseen $A_t$ for each new $b_t$. Let $H_t=A_tB$. Suppose there is a common, unknown subspace $U$ of dimension at most $r$ with $(H_t-I)V\subseteq U$ at every step, and that all $H_t|_U$ agree with the same unknown linear map $J$. This is a structural promise. Exact knowledge of the trusted source rule, invertibility and exact one-sided products are maintained. Future operators and targets can depend on past observations under this promise.
71
+
72
+ **Theorem 3.** Returning one exact solution for each of $T$ changing operators takes at most $T+r$ products. For $2\le r<d$, this worst-case cost is minimax optimal even against solvers told every right-hand side in advance and given the candidate changed subspace. Over a field larger than two elements it holds for rank one as well. With the binary rank-one promise, the exact cost is $T$ for nonzero worst-case targets.
73
+
74
+ Keep the invariant subspace and its inverse table across tasks. Erase the entire cache of ordinary products outside that table when entering a new operator. For task $t$, a single new product gives $w=(H_t-I)b_t\in U$. Closure of $w$ uses powers of the same $J$, since every encountered vector lies in $U$. Each previously unseen independent orbit direction requires one further product and survives indefinitely in the learned rule. Thus across all operator changes at most $r$ such additional products are required. The new solution is $B(b_t-J^{-1}w)$. This describes executable edits to the persistent inverse table while each task uses its own operator.
75
+
76
+ For the lower bound choose the known hard subspace $Z$ of dimension $r$, one extra active coordinate, and the same public right-hand side $(0,1)$ each time. The maps $H_t(z,w)=(Jz+c_tw,w)$ have $J$ in common. The first task alone requires $r+1$ products on some admissible transcript by Theorem 2. Fix one consistent $J$ there. Every following $c_t$ is unrestricted and can independently alter the requested preimage, even if $J$ is disclosed. A correct solver must consult each new operator at least once. Other input coordinates are independent identities and do not affect this argument. The binary rank-one upper bound follows from the involution property, and one fresh query per unknown $c_t$ is necessary. ∎
77
+
78
+ **Bounded-error converse for randomized solvers.** Over $\mathbb F_2$ with $r\ge 2$, any randomized solver whose total budget is bounded by $T+r-1$ on every run fails on at least one admissible input stream with probability at least $1/2-2^{-r}$. The success event here means returning **every** online answer correctly. This statement does not impose a cap on an algorithm's expected cost in place of its per-run cap.
79
+
80
+ Take the preceding family with $J$ uniform on the invertible binary matrices and the $c_t$ independent uniform vectors. For the first target, write $x=J^{-1}c_1$; it is uniform and independent of $J$. An input $(z,w)$ returns the informative vector $J(z+w x)$. On the query span $S$, observing a kernel direction $(x,1)$ identifies $x$. With $q\le r$ products, all other transcripts can be coupled to an injective transcript sampled independently of $x$: on a new independent query, sample an output uniformly outside the span of the previous outputs; on a dependent query, use the forced linear answer. While $(x,1)\notin S$, these response laws are exactly those induced by a uniform $J$, because previously queried vectors $z+w x$ span a space on which its action is known and each new independent vector receives a uniformly independent image. Once $(x,1)$ enters $S$, the real transcript reveals a collision; this happens exactly when the corresponding fictitious span first covers it. The fake transcript can be sampled completely before drawing $x$, for $q\le r$.
81
+
82
+ The set $\{z:(z,1)\in S\}$ contains at most $2^{q-1}$ possibilities for $q\ge1$. Given the fake transcript, at most those covered points plus one final guess can succeed. Thus the optimal average first-task success at $q$ queries is $(2^{q-1}+1)/2^r$ for $1\le q\le r$ (and $2^{-r}$ at $q=0$). A predetermined $q$-dimensional span meeting the last-coordinate-one coset attains equality, followed by a guess outside that coset slice on no collision.
83
+
84
+ To reduce the full adaptive stream, simulate its first-task procedure for at most $r$ calls. If it returns, use its first answer. If it requests the $(r+1)$st call, replace that continuation by an independent uniform guess for the first $x$. On that branch the full-budget procedure must leave at least one subsequent new operator completely unread. For the first such unread operator, its hidden $c_t$ is still fresh uniform, even conditional on $J$, on the past and on the decision to make no query. Hence joint success on this branch is at most $2^{-r}$, the replacement guess's success rate. If $T=1$ this branch cannot occur. The new first-task procedure achieves at least the original joint success probability and is bounded above by $1/2+2^{-r}$. Averaging over input streams proves the worst-case claim. In particular 90% uniform joint success requires the full $T+r$ product budget.
85
+
86
+ **Guard against unannounced changes.** Build each candidate table update transactionally. If inverse pairs disagree or cannot decode, discard the proposal. Otherwise check $H_t y=b_t$ by one separate current-operator product, subject to the ordinary same-task cache. Accept only a current solution with a zero residual. On failure erase the persistent table, recompute the orbit closure from scratch for this one operator, and reuse all paid current products. Under the common-action assumption the guard never rejects, adds at most one paid check per task, and costs at most $2T+r$. Under arbitrary invertible changes the guard still returns an exact current solution: if it rebuilds, Theorem 1 applies locally from an empty table. All local product inputs pass through a single linearly independent cache, so the guard and recovery together cost at most $d$ new products on any task, irrespective of stale earlier state. An accepted residual certifies the current answer only. A future task checks its own current operator.
87
+
88
+ ### Theorem 4. The exact price of protection
89
+
90
+ For $2\le r<d$ over any field, suppose a deterministic solver must return a correct answer on **every** sequence of invertible current operators, including unannounced failures of the common-action promise. When measuring its worst-case work **only on promised streams**, the minimax number of products is exactly
91
+ $$Q_{\mathrm{guard}}^*(d,r,T)=2T+r-1.$$
92
+ The difference from the $T+r$ trusted optimum is exactly $T-1$. Correctness outside the promise and the cost objective on the promise are distinct requirements. This theorem uses exact product queries and excludes noisy or time-varying operators *within* a solve.
93
+
94
+ **Upper bound.** The guarded algorithm builds its first candidate with empty persistent tables. Every input used for that candidate and every direction of its new closure lie in the current graph. The graph already yields $H_1 y=b_1$, hence the initial guard incurs zero extra products. Under the promise there is no stale-candidate rejection and at most one additional product in each later task. Adding that cost to the $T+r$ closure bound proves $2T+r-1$. The recovery argument above independently proves unconditional exactness.
95
+
96
+ **Lower bound.** Cost the algorithm on the promised family $H_t(z,w)=(Jz+c_t w,w)$ on $r+1$ active coordinates, taking a fixed goal $b=(0,1)$. Other coordinates undergo the known identity. Before it can return the first answer, Theorem 2 with $k=1$ lets an adversary force at least $r+1$ reads. Fix one consistent common $J$. For each later task, take the solver's first query $v=(z,w)$. If $w\ne0$, choose $c_t$ so that $Jz+c_tw\ne0$; at most one vector is excluded. If $w=0$ its image cannot span $b$ anyway. Nonzero inactive coordinates only make collinearity with $b$ impossible. Thus one can keep the *actual* operator in the same promised family while ensuring $b\notin\operatorname{span}(H_tv)$. A policy that reads nothing also fails to distinguish free choices of $c_t$.
97
+
98
+ With only this one observation, an alternative arbitrary invertible current operator can preserve $H_t v$ and alter $H_t^{-1}b$. For $v\ne0$, the true preimage of $b$ is independent of $v$. Precompose $H_t$ with a domain automorphism fixing $v$ and adding $v$ to that preimage in a basis containing both. If $v=0$, choose any automorphism moving the nonzero preimage. Either alternative agrees on all prior operators and produces the same current transcript, but it requires a different answer. Since the algorithm has to be correct even on that off-promise continuation, it must take a second read along the actual on-promise branch. Repeating this forces $(r+1)+2(T-1)=2T+r-1$. ∎
99
+
100
+ **Independent finite check.** Run `python insurance_minimax.py` to regenerate `insurance_results.json`. In $d=3,r=2$ over the binary field, the exhaustive dynamic program considers all seven nonzero queries at every decision node. The protected policy must succeed for every arbitrary invertible current matrix and may continue learning the shared core after discovering an answer. Only paths compatible with the shared-core family count toward the objective. For $T=1$, both optimum depths are three. For $T=2$, 96 promised streams sit inside 28,224 arbitrary streams: four reads suffice when the promise can be trusted; five are necessary and sufficient for unconditional correctness. This code independently searches decision trees and does not call the closure solver.
101
+
102
+ **Fully learned target anchor.** If $B_s=A_s^{-1}$ is available from any target-family member, write $G=A_tA_s^{-1}=H_t H_s^{-1}$. The image of $G-I$ is in $U$, and $G$ acts as identity on $U$. Therefore $(G-I)^2=0$, so $G^{-1}=2I-G$ and $A_t^{-1}=2B_s-B_sA_tB_s$. This is the degree-one inverse polynomial, also the one-step Newton–Schulz inverse update. It gives a strong competing strategy: spend $d$ products to acquire the first target inverse in full, solve that task locally, and spend one product per subsequent task. Previously acquiring only a generic source rule $B$ does not supply such a fully learned target-family anchor.
103
+
104
+ **Changing-domain experiment.** Run `python changing_operators.py` for the data in `changing_results.json` and `changing_worlds.csv`. In 24 independent 64-dimensional streams, every one of the 20 tasks used a distinct unseen operator and a fresh right-hand side. All shifts in one stream had the same rank-four range and restricted action, hidden by a common change of basis. The cross-term was redrawn each time. All 480 answers were exact. The learned solver used exactly 24 products on each stream. The guarded learner averaged 41.58, reset-with-source averaged 79.67, and full cold anchor acquisition followed by its one-product rule used 83. The cold solver reset each time averaged 1260.17. Charging 64 products once for the source gives 640 across all streams, or 1,062 for the guarded version, against 1,992 for independent cold anchors. A single stream with no preexisting source would therefore have different economics.
105
+
106
+ For a concrete negative transfer example, encode $e_1,e_2,e_3$ as the integers 1,2,4. Let the first operator have columns $(1,2,5)$, the second $(2,1,5)$, and the target be 4 both times. The first correct answer is 5. Blindly retaining the old restricted action proposes 5 again, whose output is 7. The guard rejects and returns 6, whose output on the new operator is exactly 4, costing three products. A stress test with 120 unrelated dense ten-dimensional operators returned all 120 correct answers, rejected 117 stale candidates and never exceeded ten current products.
reproduction/algebra/defect_closure.py ADDED
@@ -0,0 +1,322 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """Exact online repair of an automodified F_2-linear solver under low-rank transfer.
2
+
3
+ Run with Python 3.10+; standard library only. Binary vectors are Python ints.
4
+ The expensive oracle returns one full, exact matrix-vector product per call.
5
+ """
6
+ import csv
7
+ from functools import lru_cache
8
+ from itertools import product
9
+ import json
10
+ from pathlib import Path
11
+ import random
12
+ import statistics
13
+
14
+ HERE=Path(__file__).resolve().parent
15
+
16
+
17
+ def act(columns, x):
18
+ y=0
19
+ while x:
20
+ bit=x & -x
21
+ y ^= columns[bit.bit_length()-1]
22
+ x ^= bit
23
+ return y
24
+
25
+
26
+ class Graph:
27
+ """Row-echelon certificate of observed linear input/output pairs."""
28
+ def __init__(self):
29
+ self.rows={}
30
+
31
+ def reduce(self, x):
32
+ y=0
33
+ while x:
34
+ lead=x.bit_length()-1
35
+ if lead not in self.rows:
36
+ break
37
+ z,w=self.rows[lead]
38
+ x ^= z
39
+ y ^= w
40
+ return x,y
41
+
42
+ def get(self,x):
43
+ residual,y=self.reduce(x)
44
+ return None if residual else y
45
+
46
+ def insert(self,x,y):
47
+ z,w=self.reduce(x)
48
+ if not z:
49
+ assert w==y
50
+ return
51
+ self.rows[z.bit_length()-1]=z,y^w
52
+
53
+
54
+ class ProductCache:
55
+ def __init__(self,oracle):
56
+ self.oracle=oracle
57
+ self.graph=Graph()
58
+ self.calls=0
59
+
60
+ def apply(self,x):
61
+ y=self.graph.get(x)
62
+ if y is None:
63
+ y=self.oracle(x)
64
+ self.calls+=1
65
+ self.graph.insert(x,y)
66
+ return y
67
+
68
+
69
+ def inverse(columns):
70
+ g=Graph()
71
+ for j,v in enumerate(columns):
72
+ assert g.get(v) is None
73
+ g.insert(v,1<<j)
74
+ return tuple(g.get(1<<j) for j in range(len(columns)))
75
+
76
+
77
+ def invertible(d,rng):
78
+ columns=[1<<j for j in range(d)]
79
+ for _ in range(6*d):
80
+ a,b=rng.sample(range(d),2)
81
+ columns[a] ^= columns[b]
82
+ return tuple(columns)
83
+
84
+
85
+ class TransferEditor:
86
+ def __init__(self,source_inverse,target_product):
87
+ self.B=source_inverse
88
+ self.h=ProductCache(lambda v:target_product(act(self.B,v)))
89
+ self.W=Graph() # v -> zero, testing membership in defect closure
90
+ self.W_inv=Graph() # Hv -> v for the ever-growing invariant subspace
91
+
92
+ def solve(self,b):
93
+ w=self.h.apply(b)^b
94
+ v=w
95
+ while self.W.get(v) is None:
96
+ hv=self.h.apply(v)
97
+ self.W.insert(v,0)
98
+ self.W_inv.insert(hv,v)
99
+ v=hv
100
+ correction=self.W_inv.get(w)
101
+ assert correction is not None
102
+ y=b^correction
103
+ # The graph of H already contains the complete returned certificate.
104
+ assert self.h.graph.get(y)==b
105
+ return act(self.B,y)
106
+
107
+
108
+ class ColdShared:
109
+ """Learn arbitrary inverse via basis probes, reuse all discovered data."""
110
+ def __init__(self,d,oracle):
111
+ self.d=d
112
+ self.oracle=oracle
113
+ self.graph=Graph()
114
+ self.calls=0
115
+
116
+ def solve(self,b):
117
+ while self.graph.get(b) is None:
118
+ j=self.calls
119
+ assert j<self.d
120
+ y=self.oracle(1<<j)
121
+ self.graph.insert(y,1<<j)
122
+ self.calls+=1
123
+ return self.graph.get(b)
124
+
125
+
126
+ def rank(columns):
127
+ graph=Graph()
128
+ for v in columns:
129
+ graph.insert(v,0)
130
+ return len(graph.rows)
131
+
132
+
133
+ def random_target(d,r,seed,source):
134
+ rng=random.Random(seed)
135
+ u=invertible(r,rng)
136
+ c=[rng.randrange(1<<r) for j in range(d-r)]
137
+ block=list(u)+[(1<<(r+j))^z for j,z in enumerate(c)]
138
+ p=invertible(d,rng)
139
+ pinv=inverse(p)
140
+ h=[act(p,act(block,v)) for v in pinv]
141
+ target=tuple(act(h,v) for v in source)
142
+ actual_rank=rank([h[j]^(1<<j) for j in range(d)])
143
+ assert actual_rank<=r
144
+ return target,actual_rank
145
+
146
+
147
+ def world_experiment():
148
+ d=64
149
+ r=4
150
+ worlds=24
151
+ T=20
152
+ source=invertible(d,random.Random(329))
153
+ source_oracle=ProductCache(lambda x:act(source,x))
154
+ B=inverse([source_oracle.apply(1<<j) for j in range(d)])
155
+ assert source_oracle.calls==d
156
+ records=[]
157
+ for world in range(worlds):
158
+ target,actual_rank=random_target(d,r,1000+world,source)
159
+ oracle=lambda x:act(target,x)
160
+ learned=TransferEditor(B,oracle)
161
+ shared=ColdShared(d,oracle)
162
+ rng=random.Random(2000+world)
163
+ cold_sum=transfer_sum=0
164
+ rhs=[]
165
+ for t in range(1,T+1):
166
+ b=rng.randrange(1,1<<d)
167
+ rhs.append(b)
168
+ x=learned.solve(b)
169
+ assert oracle(x)==b
170
+ cold=ColdShared(d,oracle)
171
+ assert x==cold.solve(b)==shared.solve(b)
172
+ single=TransferEditor(B,oracle)
173
+ assert x==single.solve(b)
174
+ cold_sum+=cold.calls
175
+ transfer_sum+=single.h.calls
176
+ assert learned.h.calls<=min(d,t+actual_rank)
177
+ records.append(dict(world=world,task=t,actual_rank=actual_rank,closure_dimension=len(learned.W.rows),
178
+ transferred_cumulative_calls=learned.h.calls,source_only_per_task_cumulative_calls=transfer_sum,
179
+ cold_shared_cumulative_calls=shared.calls,cold_per_task_cumulative_calls=cold_sum))
180
+ # A repeated task requires no new product and returns the same answer.
181
+ n=learned.h.calls
182
+ for b in rhs:
183
+ assert oracle(learned.solve(b))==b
184
+ assert learned.h.calls==n
185
+ with (HERE/'worlds.csv').open('w',newline='') as f:
186
+ writer=csv.DictWriter(f,fieldnames=list(records[0]))
187
+ writer.writeheader()
188
+ writer.writerows(records)
189
+ summaries={}
190
+ for t in [1,4,8,20]:
191
+ subset=[x for x in records if x['task']==t]
192
+ summaries[t]={key:{'mean':statistics.mean(v),'min':min(v),'max':max(v)}
193
+ for key in list(subset[0])[2:]
194
+ for v in [[s[key] for s in subset]]}
195
+ return {'d':d,'rank_cap':r,'domains':worlds,'source_acquisition_calls_once':source_oracle.calls,
196
+ 'new_rhs_per_domain':T,'all_exact_and_replays_used_zero_calls':True,'checkpoints':summaries}
197
+
198
+
199
+ def all_decision_depths(labels, query_responses, max_budget):
200
+ outputs={}
201
+ for i,label in enumerate(labels):
202
+ outputs[label]=outputs.get(label,0)|(1<<i)
203
+ partitions=[]
204
+ for responses in query_responses:
205
+ buckets={}
206
+ for i,v in enumerate(responses):
207
+ buckets[v]=buckets.get(v,0)|(1<<i)
208
+ partitions.append(list(buckets.values()))
209
+ @lru_cache(None)
210
+ def feasible(mask,budget):
211
+ if any(mask & ~group==0 for group in outputs.values()):
212
+ return True
213
+ if budget==0:
214
+ return False
215
+ for blocks in partitions:
216
+ restricted=[mask&b for b in blocks if mask&b]
217
+ if len(restricted)>1 and all(feasible(s,budget-1) for s in restricted):
218
+ return True
219
+ return False
220
+ full=(1<<len(labels))-1
221
+ return [feasible(full,h) for h in range(max_budget+1)]
222
+
223
+
224
+ def binary_rank_one(d):
225
+ identity=tuple(1<<j for j in range(d))
226
+ matrices={identity}
227
+ for u in range(1,1<<d):
228
+ for f in range(1,1<<d):
229
+ if (u&f).bit_count()%2==0:
230
+ m=tuple(v^(u if (v&f).bit_count()%2 else 0) for v in identity)
231
+ assert all(act(m,col)==v for v,col in zip(identity,m))
232
+ matrices.add(m)
233
+ matrices=sorted(matrices)
234
+ responses=[[act(m,x) for m in matrices] for x in range(1,1<<d)]
235
+ result={}
236
+ for k in range(1,d+1):
237
+ labels=[tuple(m[:k]) for m in matrices]
238
+ costs=all_decision_depths(labels,responses,k)
239
+ assert costs==[False]*k+[True]
240
+ result[k]=costs
241
+ return {'field':2,'dimension':d,'matrices':len(matrices),'feasible_depth_by_target_count':result}
242
+
243
+
244
+ def ternary_rank_one(k):
245
+ choices=[(u,)+c for u in (1,2) for c in product(range(3),repeat=k)]
246
+ labels=[tuple((-pow(m[0],-1,3)*z)%3 for z in m[1:]) for m in choices]
247
+ queries=[x for x in product(range(3),repeat=k+1) if any(x)]
248
+ responses=[[sum(a*b for a,b in zip(x,m))%3 for m in choices] for x in queries]
249
+ costs=all_decision_depths(labels,responses,k+1)
250
+ assert costs==[False]*(k+1)+[True]
251
+ return {'field':3,'rank_cap':1,'tasks':k,'hard_family_size':len(choices),'feasible_depth':costs}
252
+
253
+
254
+ def exhaustive_minimax(r,k):
255
+ """Best worst-case decision-tree depth over every adaptive query strategy."""
256
+ d=r+k
257
+ choices=[]
258
+ labels=[]
259
+ for u in product(range(1,1<<r),repeat=r):
260
+ if rank(u)!=r:
261
+ continue
262
+ ui=inverse(u)
263
+ for c in product(range(1<<r),repeat=k):
264
+ choices.append(u+c) # Only top r response bits convey unknown data.
265
+ labels.append(tuple(act(ui,v) for v in c))
266
+ query_responses=[[act(oracle,x) for oracle in choices] for x in range(1,1<<d)]
267
+ costs=all_decision_depths(labels,query_responses,d)
268
+ assert costs == [False]*d+[True]
269
+ return {'r':r,'tasks':k,'ambient_d':d,'distinct_operators':len(choices),
270
+ 'distinct_solution_lists':len(set(labels)),'distinct_query_vectors':len(query_responses),
271
+ 'decision_tree_feasible_by_depth_0_to_d':costs,'optimal_worst_case_calls':d}
272
+
273
+
274
+ def first_task_bayes(r):
275
+ """Independent finite Bayes-optimal decision-tree search for the affine family."""
276
+ from fractions import Fraction
277
+ size=1<<r
278
+ maps=[]
279
+ labels=[]
280
+ for u in product(range(1,size),repeat=r):
281
+ if rank(u)!=r:
282
+ continue
283
+ for x in range(size):
284
+ maps.append(u+(act(u,x),))
285
+ labels.append(x)
286
+ label_masks=[0]*size
287
+ for i,x in enumerate(labels):
288
+ label_masks[x]|=1<<i
289
+ partitions=[]
290
+ for v in range(1,2*size):
291
+ cut=[0]*size
292
+ for i,m in enumerate(maps):
293
+ cut[act(m,v)]|=1<<i
294
+ partitions.append([q for q in cut if q])
295
+ @lru_cache(None)
296
+ def score(mask,depth):
297
+ best=max((mask&a).bit_count() for a in label_masks)
298
+ total=mask.bit_count()
299
+ if depth==0 or best==total:
300
+ return best
301
+ for cut in partitions:
302
+ children=[mask&node for node in cut if mask&node]
303
+ if len(children)>1:
304
+ best=max(best,sum(score(child,depth-1) for child in children))
305
+ if best==total:
306
+ break
307
+ return best
308
+ vals=[Fraction(score((1<<len(maps))-1,q),len(maps)) for q in range(r+1)]
309
+ assert vals==[Fraction(1,size)]+[Fraction((1<<(q-1))+1,size) for q in range(1,r+1)]
310
+ return {'r':r,'equiprobable_matrices':len(maps),'optimal_success_by_budget_zero_to_r':list(map(str,vals))}
311
+
312
+
313
+ def main():
314
+ results={'world_experiment':world_experiment(),'all_decision_trees':[exhaustive_minimax(2,k) for k in (1,2,3)],
315
+ 'binary_rank_one_exception':binary_rank_one(4), 'ternary_rank_one':[ternary_rank_one(k) for k in (1,2,3)],
316
+ 'bounded_error_first_task':[first_task_bayes(r) for r in (2,3)]}
317
+ (HERE/'results.json').write_text(json.dumps(results,indent=2)+'\n')
318
+ print(json.dumps(results,indent=2))
319
+
320
+
321
+ if __name__=='__main__':
322
+ main()
reproduction/algebra/insurance_minimax.py ADDED
@@ -0,0 +1,87 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """Exhaustively optimize trusted and failure-protected two-episode decision trees.
2
+
3
+ The cost objective considers only hidden block-family streams. A guarded strategy
4
+ must also solve every sequence of arbitrary invertible binary 3x3 operators.
5
+ On a branch inconsistent with the promise the solver can always finish by
6
+ acquiring a whole matrix, and that branch is outside the cost objective.
7
+ """
8
+ from functools import lru_cache
9
+ from itertools import product
10
+ import json
11
+ from pathlib import Path
12
+ from defect_closure import act,inverse,rank
13
+
14
+ HERE=Path(__file__).resolve().parent
15
+ D=3
16
+ GOAL=4 # (0,1) in binary coordinates F_2^2 plus F_2.
17
+
18
+
19
+ def optimize(episodes,guarded):
20
+ matrices=[m for m in product(range(1,1<<D),repeat=D) if rank(m)==D]
21
+ size=len(matrices)
22
+ universe=(1<<size)-1
23
+ # In the hard subfamily H(z,w)=(Jz+cw,w), J is shared across
24
+ # episodes while c is free. Invertibility makes J invertible.
25
+ by_core={}
26
+ answers={}
27
+ for i,m in enumerate(matrices):
28
+ if m[0]<4 and m[1]<4 and m[2]>=4:
29
+ core=m[:2]
30
+ by_core[core]=by_core.get(core,0)|(1<<i)
31
+ x=act(inverse(m),GOAL)
32
+ answers[x]=answers.get(x,0)|(1<<i)
33
+ groups=tuple(by_core.values())
34
+ assert size==168 and len(groups)==6 and all(g.bit_count()==4 for g in groups)
35
+ partitions=[]
36
+ for x in range(1,1<<D):
37
+ outcomes=[0]*(1<<D)
38
+ for i,m in enumerate(matrices):
39
+ outcomes[act(m,x)] |= 1<<i
40
+ partitions.append(tuple(q for q in outcomes if q))
41
+
42
+ def start(remaining,core_mask,budget):
43
+ family=0
44
+ for k,g in enumerate(groups):
45
+ if core_mask & (1<<k):
46
+ family |= g
47
+ compatible=universe if guarded else family
48
+ return feasible(remaining,family,compatible,budget)
49
+
50
+ @lru_cache(None)
51
+ def feasible(remaining,family,compatible,budget):
52
+ assert family and family & ~compatible==0
53
+ # A known answer can be returned, or the solver may spend extra reads
54
+ # this episode to learn more about the core before answering.
55
+ if any(compatible & ~answer_group==0 for answer_group in answers.values()):
56
+ if remaining==1:
57
+ return True
58
+ cores=sum(1<<k for k,g in enumerate(groups) if family&g)
59
+ if start(remaining-1,cores,budget):
60
+ return True
61
+ if budget==0:
62
+ return False
63
+ for partition in partitions:
64
+ children=[(family&a,compatible&a) for a in partition if family&a]
65
+ if children==[(family,compatible)]:
66
+ continue
67
+ if all(feasible(remaining,f,c,budget-1) for f,c in children):
68
+ return True
69
+ return False
70
+
71
+ cap=2*episodes+2
72
+ budgets=[start(episodes,(1<<len(groups))-1,n) for n in range(cap+1)]
73
+ first=budgets.index(True)
74
+ assert budgets==[False]*first+[True]*(len(budgets)-first)
75
+ expected=(2*episodes+2-1) if guarded else (episodes+2)
76
+ assert expected==first
77
+ return {'episodes':episodes,'guarded':guarded,'ambient_dimension':D,'latent_rank_cap':2,
78
+ 'possible_invertible_matrices_per_episode':size,'promise_streams':len(groups)*4**episodes,
79
+ 'unrestricted_streams':size**episodes,'all_product_queries_per_node':len(partitions),
80
+ 'worst_case_budget_feasible_0_upward':budgets,'minimum_worst_cost_on_promised_streams':first}
81
+
82
+
83
+ if __name__=='__main__':
84
+ out={'search':'exact forward decision-tree optimization, current solution required before the next episode',
85
+ 'searches':[optimize(t,g) for t in (1,2) for g in (False,True)]}
86
+ (HERE/'insurance_results.json').write_text(json.dumps(out,indent=2)+'\n')
87
+ print(json.dumps(out,indent=2))
reproduction/algebra/insurance_results.json ADDED
@@ -0,0 +1,81 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "search": "exact forward decision-tree optimization, current solution required before the next episode",
3
+ "searches": [
4
+ {
5
+ "episodes": 1,
6
+ "guarded": false,
7
+ "ambient_dimension": 3,
8
+ "latent_rank_cap": 2,
9
+ "possible_invertible_matrices_per_episode": 168,
10
+ "promise_streams": 24,
11
+ "unrestricted_streams": 168,
12
+ "all_product_queries_per_node": 7,
13
+ "worst_case_budget_feasible_0_upward": [
14
+ false,
15
+ false,
16
+ false,
17
+ true,
18
+ true
19
+ ],
20
+ "minimum_worst_cost_on_promised_streams": 3
21
+ },
22
+ {
23
+ "episodes": 1,
24
+ "guarded": true,
25
+ "ambient_dimension": 3,
26
+ "latent_rank_cap": 2,
27
+ "possible_invertible_matrices_per_episode": 168,
28
+ "promise_streams": 24,
29
+ "unrestricted_streams": 168,
30
+ "all_product_queries_per_node": 7,
31
+ "worst_case_budget_feasible_0_upward": [
32
+ false,
33
+ false,
34
+ false,
35
+ true,
36
+ true
37
+ ],
38
+ "minimum_worst_cost_on_promised_streams": 3
39
+ },
40
+ {
41
+ "episodes": 2,
42
+ "guarded": false,
43
+ "ambient_dimension": 3,
44
+ "latent_rank_cap": 2,
45
+ "possible_invertible_matrices_per_episode": 168,
46
+ "promise_streams": 96,
47
+ "unrestricted_streams": 28224,
48
+ "all_product_queries_per_node": 7,
49
+ "worst_case_budget_feasible_0_upward": [
50
+ false,
51
+ false,
52
+ false,
53
+ false,
54
+ true,
55
+ true,
56
+ true
57
+ ],
58
+ "minimum_worst_cost_on_promised_streams": 4
59
+ },
60
+ {
61
+ "episodes": 2,
62
+ "guarded": true,
63
+ "ambient_dimension": 3,
64
+ "latent_rank_cap": 2,
65
+ "possible_invertible_matrices_per_episode": 168,
66
+ "promise_streams": 96,
67
+ "unrestricted_streams": 28224,
68
+ "all_product_queries_per_node": 7,
69
+ "worst_case_budget_feasible_0_upward": [
70
+ false,
71
+ false,
72
+ false,
73
+ false,
74
+ false,
75
+ true,
76
+ true
77
+ ],
78
+ "minimum_worst_cost_on_promised_streams": 5
79
+ }
80
+ ]
81
+ }
reproduction/algebra/results.json ADDED
@@ -0,0 +1,278 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "world_experiment": {
3
+ "d": 64,
4
+ "rank_cap": 4,
5
+ "domains": 24,
6
+ "source_acquisition_calls_once": 64,
7
+ "new_rhs_per_domain": 20,
8
+ "all_exact_and_replays_used_zero_calls": true,
9
+ "checkpoints": {
10
+ "1": {
11
+ "actual_rank": {
12
+ "mean": 4,
13
+ "min": 4,
14
+ "max": 4
15
+ },
16
+ "closure_dimension": {
17
+ "mean": 3.3333333333333335,
18
+ "min": 0,
19
+ "max": 4
20
+ },
21
+ "transferred_cumulative_calls": {
22
+ "mean": 4.333333333333333,
23
+ "min": 1,
24
+ "max": 5
25
+ },
26
+ "source_only_per_task_cumulative_calls": {
27
+ "mean": 4.333333333333333,
28
+ "min": 1,
29
+ "max": 5
30
+ },
31
+ "cold_shared_cumulative_calls": {
32
+ "mean": 63.541666666666664,
33
+ "min": 62,
34
+ "max": 64
35
+ },
36
+ "cold_per_task_cumulative_calls": {
37
+ "mean": 63.541666666666664,
38
+ "min": 62,
39
+ "max": 64
40
+ }
41
+ },
42
+ "4": {
43
+ "actual_rank": {
44
+ "mean": 4,
45
+ "min": 4,
46
+ "max": 4
47
+ },
48
+ "closure_dimension": {
49
+ "mean": 3.9583333333333335,
50
+ "min": 3,
51
+ "max": 4
52
+ },
53
+ "transferred_cumulative_calls": {
54
+ "mean": 7.958333333333333,
55
+ "min": 7,
56
+ "max": 8
57
+ },
58
+ "source_only_per_task_cumulative_calls": {
59
+ "mean": 16.625,
60
+ "min": 12,
61
+ "max": 20
62
+ },
63
+ "cold_shared_cumulative_calls": {
64
+ "mean": 63.916666666666664,
65
+ "min": 63,
66
+ "max": 64
67
+ },
68
+ "cold_per_task_cumulative_calls": {
69
+ "mean": 251.83333333333334,
70
+ "min": 245,
71
+ "max": 256
72
+ }
73
+ },
74
+ "8": {
75
+ "actual_rank": {
76
+ "mean": 4,
77
+ "min": 4,
78
+ "max": 4
79
+ },
80
+ "closure_dimension": {
81
+ "mean": 4,
82
+ "min": 4,
83
+ "max": 4
84
+ },
85
+ "transferred_cumulative_calls": {
86
+ "mean": 12,
87
+ "min": 12,
88
+ "max": 12
89
+ },
90
+ "source_only_per_task_cumulative_calls": {
91
+ "mean": 33.583333333333336,
92
+ "min": 22,
93
+ "max": 40
94
+ },
95
+ "cold_shared_cumulative_calls": {
96
+ "mean": 64,
97
+ "min": 64,
98
+ "max": 64
99
+ },
100
+ "cold_per_task_cumulative_calls": {
101
+ "mean": 504.3333333333333,
102
+ "min": 493,
103
+ "max": 511
104
+ }
105
+ },
106
+ "20": {
107
+ "actual_rank": {
108
+ "mean": 4,
109
+ "min": 4,
110
+ "max": 4
111
+ },
112
+ "closure_dimension": {
113
+ "mean": 4,
114
+ "min": 4,
115
+ "max": 4
116
+ },
117
+ "transferred_cumulative_calls": {
118
+ "mean": 24,
119
+ "min": 24,
120
+ "max": 24
121
+ },
122
+ "source_only_per_task_cumulative_calls": {
123
+ "mean": 82.91666666666667,
124
+ "min": 58,
125
+ "max": 100
126
+ },
127
+ "cold_shared_cumulative_calls": {
128
+ "mean": 64,
129
+ "min": 64,
130
+ "max": 64
131
+ },
132
+ "cold_per_task_cumulative_calls": {
133
+ "mean": 1262.1666666666667,
134
+ "min": 1253,
135
+ "max": 1271
136
+ }
137
+ }
138
+ }
139
+ },
140
+ "all_decision_trees": [
141
+ {
142
+ "r": 2,
143
+ "tasks": 1,
144
+ "ambient_d": 3,
145
+ "distinct_operators": 24,
146
+ "distinct_solution_lists": 4,
147
+ "distinct_query_vectors": 7,
148
+ "decision_tree_feasible_by_depth_0_to_d": [
149
+ false,
150
+ false,
151
+ false,
152
+ true
153
+ ],
154
+ "optimal_worst_case_calls": 3
155
+ },
156
+ {
157
+ "r": 2,
158
+ "tasks": 2,
159
+ "ambient_d": 4,
160
+ "distinct_operators": 96,
161
+ "distinct_solution_lists": 16,
162
+ "distinct_query_vectors": 15,
163
+ "decision_tree_feasible_by_depth_0_to_d": [
164
+ false,
165
+ false,
166
+ false,
167
+ false,
168
+ true
169
+ ],
170
+ "optimal_worst_case_calls": 4
171
+ },
172
+ {
173
+ "r": 2,
174
+ "tasks": 3,
175
+ "ambient_d": 5,
176
+ "distinct_operators": 384,
177
+ "distinct_solution_lists": 64,
178
+ "distinct_query_vectors": 31,
179
+ "decision_tree_feasible_by_depth_0_to_d": [
180
+ false,
181
+ false,
182
+ false,
183
+ false,
184
+ false,
185
+ true
186
+ ],
187
+ "optimal_worst_case_calls": 5
188
+ }
189
+ ],
190
+ "binary_rank_one_exception": {
191
+ "field": 2,
192
+ "dimension": 4,
193
+ "matrices": 106,
194
+ "feasible_depth_by_target_count": {
195
+ "1": [
196
+ false,
197
+ true
198
+ ],
199
+ "2": [
200
+ false,
201
+ false,
202
+ true
203
+ ],
204
+ "3": [
205
+ false,
206
+ false,
207
+ false,
208
+ true
209
+ ],
210
+ "4": [
211
+ false,
212
+ false,
213
+ false,
214
+ false,
215
+ true
216
+ ]
217
+ }
218
+ },
219
+ "ternary_rank_one": [
220
+ {
221
+ "field": 3,
222
+ "rank_cap": 1,
223
+ "tasks": 1,
224
+ "hard_family_size": 6,
225
+ "feasible_depth": [
226
+ false,
227
+ false,
228
+ true
229
+ ]
230
+ },
231
+ {
232
+ "field": 3,
233
+ "rank_cap": 1,
234
+ "tasks": 2,
235
+ "hard_family_size": 18,
236
+ "feasible_depth": [
237
+ false,
238
+ false,
239
+ false,
240
+ true
241
+ ]
242
+ },
243
+ {
244
+ "field": 3,
245
+ "rank_cap": 1,
246
+ "tasks": 3,
247
+ "hard_family_size": 54,
248
+ "feasible_depth": [
249
+ false,
250
+ false,
251
+ false,
252
+ false,
253
+ true
254
+ ]
255
+ }
256
+ ],
257
+ "bounded_error_first_task": [
258
+ {
259
+ "r": 2,
260
+ "equiprobable_matrices": 24,
261
+ "optimal_success_by_budget_zero_to_r": [
262
+ "1/4",
263
+ "1/2",
264
+ "3/4"
265
+ ]
266
+ },
267
+ {
268
+ "r": 3,
269
+ "equiprobable_matrices": 1344,
270
+ "optimal_success_by_budget_zero_to_r": [
271
+ "1/8",
272
+ "1/4",
273
+ "3/8",
274
+ "5/8"
275
+ ]
276
+ }
277
+ ]
278
+ }
reproduction/algebra/worlds.csv ADDED
@@ -0,0 +1,481 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ world,task,actual_rank,closure_dimension,transferred_cumulative_calls,source_only_per_task_cumulative_calls,cold_shared_cumulative_calls,cold_per_task_cumulative_calls
2
+ 0,1,4,3,4,4,64,64
3
+ 0,2,4,4,6,9,64,127
4
+ 0,3,4,4,7,14,64,190
5
+ 0,4,4,4,8,19,64,253
6
+ 0,5,4,4,9,24,64,316
7
+ 0,6,4,4,10,29,64,378
8
+ 0,7,4,4,11,33,64,441
9
+ 0,8,4,4,12,38,64,502
10
+ 0,9,4,4,13,41,64,564
11
+ 0,10,4,4,14,46,64,628
12
+ 0,11,4,4,15,51,64,691
13
+ 0,12,4,4,16,55,64,755
14
+ 0,13,4,4,17,58,64,818
15
+ 0,14,4,4,18,63,64,881
16
+ 0,15,4,4,19,66,64,945
17
+ 0,16,4,4,20,69,64,1009
18
+ 0,17,4,4,21,74,64,1073
19
+ 0,18,4,4,22,77,64,1137
20
+ 0,19,4,4,23,82,64,1201
21
+ 0,20,4,4,24,87,64,1264
22
+ 1,1,4,2,3,3,63,63
23
+ 1,2,4,3,5,7,64,127
24
+ 1,3,4,4,7,11,64,191
25
+ 1,4,4,4,8,12,64,254
26
+ 1,5,4,4,9,14,64,317
27
+ 1,6,4,4,10,16,64,381
28
+ 1,7,4,4,11,18,64,445
29
+ 1,8,4,4,12,22,64,509
30
+ 1,9,4,4,13,26,64,571
31
+ 1,10,4,4,14,30,64,635
32
+ 1,11,4,4,15,33,64,699
33
+ 1,12,4,4,16,37,64,762
34
+ 1,13,4,4,17,39,64,824
35
+ 1,14,4,4,18,42,64,884
36
+ 1,15,4,4,19,46,64,948
37
+ 1,16,4,4,20,49,64,1009
38
+ 1,17,4,4,21,51,64,1073
39
+ 1,18,4,4,22,55,64,1137
40
+ 1,19,4,4,23,56,64,1200
41
+ 1,20,4,4,24,58,64,1264
42
+ 2,1,4,3,4,4,64,64
43
+ 2,2,4,3,5,8,64,126
44
+ 2,3,4,4,7,13,64,190
45
+ 2,4,4,4,8,18,64,254
46
+ 2,5,4,4,9,22,64,317
47
+ 2,6,4,4,10,27,64,379
48
+ 2,7,4,4,11,32,64,443
49
+ 2,8,4,4,12,36,64,507
50
+ 2,9,4,4,13,41,64,571
51
+ 2,10,4,4,14,46,64,635
52
+ 2,11,4,4,15,50,64,698
53
+ 2,12,4,4,16,55,64,762
54
+ 2,13,4,4,17,59,64,826
55
+ 2,14,4,4,18,60,64,888
56
+ 2,15,4,4,19,65,64,951
57
+ 2,16,4,4,20,66,64,1015
58
+ 2,17,4,4,21,71,64,1079
59
+ 2,18,4,4,22,76,64,1140
60
+ 2,19,4,4,23,81,64,1204
61
+ 2,20,4,4,24,86,64,1268
62
+ 3,1,4,3,4,4,62,62
63
+ 3,2,4,3,5,6,63,125
64
+ 3,3,4,4,7,10,63,188
65
+ 3,4,4,4,8,14,63,249
66
+ 3,5,4,4,9,18,64,313
67
+ 3,6,4,4,10,21,64,376
68
+ 3,7,4,4,11,24,64,440
69
+ 3,8,4,4,12,28,64,503
70
+ 3,9,4,4,13,32,64,566
71
+ 3,10,4,4,14,35,64,630
72
+ 3,11,4,4,15,38,64,694
73
+ 3,12,4,4,16,42,64,757
74
+ 3,13,4,4,17,45,64,817
75
+ 3,14,4,4,18,47,64,880
76
+ 3,15,4,4,19,48,64,943
77
+ 3,16,4,4,20,51,64,1007
78
+ 3,17,4,4,21,52,64,1071
79
+ 3,18,4,4,22,56,64,1135
80
+ 3,19,4,4,23,60,64,1198
81
+ 3,20,4,4,24,62,64,1262
82
+ 4,1,4,4,5,5,64,64
83
+ 4,2,4,4,6,10,64,128
84
+ 4,3,4,4,7,14,64,189
85
+ 4,4,4,4,8,19,64,253
86
+ 4,5,4,4,9,23,64,317
87
+ 4,6,4,4,10,27,64,381
88
+ 4,7,4,4,11,31,64,444
89
+ 4,8,4,4,12,35,64,508
90
+ 4,9,4,4,13,39,64,572
91
+ 4,10,4,4,14,43,64,634
92
+ 4,11,4,4,15,47,64,698
93
+ 4,12,4,4,16,51,64,762
94
+ 4,13,4,4,17,56,64,826
95
+ 4,14,4,4,18,61,64,890
96
+ 4,15,4,4,19,66,64,953
97
+ 4,16,4,4,20,70,64,1016
98
+ 4,17,4,4,21,74,64,1080
99
+ 4,18,4,4,22,78,64,1144
100
+ 4,19,4,4,23,82,64,1207
101
+ 4,20,4,4,24,87,64,1271
102
+ 5,1,4,4,5,5,64,64
103
+ 5,2,4,4,6,9,64,128
104
+ 5,3,4,4,7,13,64,191
105
+ 5,4,4,4,8,17,64,255
106
+ 5,5,4,4,9,22,64,319
107
+ 5,6,4,4,10,27,64,382
108
+ 5,7,4,4,11,32,64,446
109
+ 5,8,4,4,12,37,64,509
110
+ 5,9,4,4,13,42,64,573
111
+ 5,10,4,4,14,47,64,637
112
+ 5,11,4,4,15,51,64,701
113
+ 5,12,4,4,16,52,64,763
114
+ 5,13,4,4,17,56,64,824
115
+ 5,14,4,4,18,60,64,888
116
+ 5,15,4,4,19,65,64,952
117
+ 5,16,4,4,20,67,64,1016
118
+ 5,17,4,4,21,72,64,1080
119
+ 5,18,4,4,22,77,64,1144
120
+ 5,19,4,4,23,81,64,1208
121
+ 5,20,4,4,24,86,64,1270
122
+ 6,1,4,4,5,5,62,62
123
+ 6,2,4,4,6,10,63,125
124
+ 6,3,4,4,7,15,64,189
125
+ 6,4,4,4,8,20,64,252
126
+ 6,5,4,4,9,23,64,311
127
+ 6,6,4,4,10,24,64,373
128
+ 6,7,4,4,11,29,64,437
129
+ 6,8,4,4,12,32,64,501
130
+ 6,9,4,4,13,37,64,564
131
+ 6,10,4,4,14,42,64,628
132
+ 6,11,4,4,15,47,64,692
133
+ 6,12,4,4,16,52,64,756
134
+ 6,13,4,4,17,57,64,819
135
+ 6,14,4,4,18,62,64,883
136
+ 6,15,4,4,19,65,64,945
137
+ 6,16,4,4,20,70,64,1009
138
+ 6,17,4,4,21,75,64,1069
139
+ 6,18,4,4,22,80,64,1133
140
+ 6,19,4,4,23,81,64,1195
141
+ 6,20,4,4,24,84,64,1258
142
+ 7,1,4,4,5,5,63,63
143
+ 7,2,4,4,6,10,64,127
144
+ 7,3,4,4,7,15,64,191
145
+ 7,4,4,4,8,20,64,255
146
+ 7,5,4,4,9,25,64,319
147
+ 7,6,4,4,10,30,64,383
148
+ 7,7,4,4,11,35,64,445
149
+ 7,8,4,4,12,40,64,509
150
+ 7,9,4,4,13,45,64,571
151
+ 7,10,4,4,14,50,64,632
152
+ 7,11,4,4,15,55,64,693
153
+ 7,12,4,4,16,60,64,757
154
+ 7,13,4,4,17,65,64,820
155
+ 7,14,4,4,18,70,64,884
156
+ 7,15,4,4,19,75,64,945
157
+ 7,16,4,4,20,80,64,1009
158
+ 7,17,4,4,21,85,64,1073
159
+ 7,18,4,4,22,90,64,1135
160
+ 7,19,4,4,23,91,64,1199
161
+ 7,20,4,4,24,96,64,1263
162
+ 8,1,4,4,5,5,63,63
163
+ 8,2,4,4,6,10,63,124
164
+ 8,3,4,4,7,15,63,187
165
+ 8,4,4,4,8,16,63,250
166
+ 8,5,4,4,9,21,63,313
167
+ 8,6,4,4,10,26,64,377
168
+ 8,7,4,4,11,31,64,439
169
+ 8,8,4,4,12,36,64,500
170
+ 8,9,4,4,13,41,64,564
171
+ 8,10,4,4,14,46,64,628
172
+ 8,11,4,4,15,47,64,691
173
+ 8,12,4,4,16,48,64,753
174
+ 8,13,4,4,17,53,64,814
175
+ 8,14,4,4,18,58,64,877
176
+ 8,15,4,4,19,63,64,940
177
+ 8,16,4,4,20,68,64,1004
178
+ 8,17,4,4,21,69,64,1068
179
+ 8,18,4,4,22,70,64,1130
180
+ 8,19,4,4,23,75,64,1194
181
+ 8,20,4,4,24,80,64,1258
182
+ 9,1,4,4,5,5,64,64
183
+ 9,2,4,4,6,8,64,123
184
+ 9,3,4,4,7,13,64,186
185
+ 9,4,4,4,8,18,64,245
186
+ 9,5,4,4,9,23,64,304
187
+ 9,6,4,4,10,28,64,368
188
+ 9,7,4,4,11,33,64,431
189
+ 9,8,4,4,12,38,64,493
190
+ 9,9,4,4,13,43,64,557
191
+ 9,10,4,4,14,46,64,621
192
+ 9,11,4,4,15,49,64,683
193
+ 9,12,4,4,16,54,64,747
194
+ 9,13,4,4,17,59,64,810
195
+ 9,14,4,4,18,64,64,874
196
+ 9,15,4,4,19,69,64,937
197
+ 9,16,4,4,20,74,64,998
198
+ 9,17,4,4,21,77,64,1061
199
+ 9,18,4,4,22,82,64,1125
200
+ 9,19,4,4,23,85,64,1189
201
+ 9,20,4,4,24,88,64,1253
202
+ 10,1,4,3,4,4,64,64
203
+ 10,2,4,3,5,5,64,128
204
+ 10,3,4,4,7,10,64,192
205
+ 10,4,4,4,8,14,64,255
206
+ 10,5,4,4,9,18,64,318
207
+ 10,6,4,4,10,23,64,382
208
+ 10,7,4,4,11,28,64,446
209
+ 10,8,4,4,12,33,64,509
210
+ 10,9,4,4,13,38,64,573
211
+ 10,10,4,4,14,43,64,636
212
+ 10,11,4,4,15,48,64,700
213
+ 10,12,4,4,16,53,64,764
214
+ 10,13,4,4,17,55,64,827
215
+ 10,14,4,4,18,59,64,890
216
+ 10,15,4,4,19,64,64,954
217
+ 10,16,4,4,20,68,64,1018
218
+ 10,17,4,4,21,72,64,1081
219
+ 10,18,4,4,22,77,64,1145
220
+ 10,19,4,4,23,82,64,1205
221
+ 10,20,4,4,24,86,64,1268
222
+ 11,1,4,4,5,5,64,64
223
+ 11,2,4,4,6,10,64,125
224
+ 11,3,4,4,7,14,64,188
225
+ 11,4,4,4,8,18,64,251
226
+ 11,5,4,4,9,23,64,315
227
+ 11,6,4,4,10,28,64,379
228
+ 11,7,4,4,11,33,64,442
229
+ 11,8,4,4,12,35,64,505
230
+ 11,9,4,4,13,40,64,569
231
+ 11,10,4,4,14,44,64,632
232
+ 11,11,4,4,15,49,64,696
233
+ 11,12,4,4,16,53,64,758
234
+ 11,13,4,4,17,54,64,822
235
+ 11,14,4,4,18,59,64,886
236
+ 11,15,4,4,19,64,64,950
237
+ 11,16,4,4,20,69,64,1011
238
+ 11,17,4,4,21,71,64,1075
239
+ 11,18,4,4,22,76,64,1139
240
+ 11,19,4,4,23,80,64,1202
241
+ 11,20,4,4,24,85,64,1262
242
+ 12,1,4,4,5,5,63,63
243
+ 12,2,4,4,6,6,63,126
244
+ 12,3,4,4,7,11,64,190
245
+ 12,4,4,4,8,15,64,246
246
+ 12,5,4,4,9,19,64,310
247
+ 12,6,4,4,10,23,64,373
248
+ 12,7,4,4,11,28,64,436
249
+ 12,8,4,4,12,32,64,500
250
+ 12,9,4,4,13,37,64,564
251
+ 12,10,4,4,14,41,64,628
252
+ 12,11,4,4,15,45,64,692
253
+ 12,12,4,4,16,50,64,754
254
+ 12,13,4,4,17,55,64,815
255
+ 12,14,4,4,18,60,64,878
256
+ 12,15,4,4,19,65,64,942
257
+ 12,16,4,4,20,69,64,1004
258
+ 12,17,4,4,21,74,64,1068
259
+ 12,18,4,4,22,78,64,1132
260
+ 12,19,4,4,23,83,64,1196
261
+ 12,20,4,4,24,88,64,1258
262
+ 13,1,4,3,4,4,64,64
263
+ 13,2,4,4,6,8,64,128
264
+ 13,3,4,4,7,12,64,192
265
+ 13,4,4,4,8,14,64,256
266
+ 13,5,4,4,9,17,64,320
267
+ 13,6,4,4,10,20,64,384
268
+ 13,7,4,4,11,24,64,447
269
+ 13,8,4,4,12,28,64,511
270
+ 13,9,4,4,13,32,64,575
271
+ 13,10,4,4,14,35,64,639
272
+ 13,11,4,4,15,39,64,702
273
+ 13,12,4,4,16,43,64,765
274
+ 13,13,4,4,17,44,64,829
275
+ 13,14,4,4,18,47,64,892
276
+ 13,15,4,4,19,51,64,956
277
+ 13,16,4,4,20,53,64,1019
278
+ 13,17,4,4,21,57,64,1083
279
+ 13,18,4,4,22,61,64,1146
280
+ 13,19,4,4,23,65,64,1207
281
+ 13,20,4,4,24,69,64,1271
282
+ 14,1,4,4,5,5,63,63
283
+ 14,2,4,4,6,10,63,121
284
+ 14,3,4,4,7,15,63,184
285
+ 14,4,4,4,8,20,64,248
286
+ 14,5,4,4,9,25,64,310
287
+ 14,6,4,4,10,30,64,373
288
+ 14,7,4,4,11,35,64,437
289
+ 14,8,4,4,12,40,64,501
290
+ 14,9,4,4,13,45,64,565
291
+ 14,10,4,4,14,50,64,628
292
+ 14,11,4,4,15,55,64,691
293
+ 14,12,4,4,16,60,64,754
294
+ 14,13,4,4,17,65,64,817
295
+ 14,14,4,4,18,70,64,881
296
+ 14,15,4,4,19,75,64,944
297
+ 14,16,4,4,20,80,64,1006
298
+ 14,17,4,4,21,85,64,1068
299
+ 14,18,4,4,22,90,64,1132
300
+ 14,19,4,4,23,95,64,1196
301
+ 14,20,4,4,24,100,64,1258
302
+ 15,1,4,4,5,5,64,64
303
+ 15,2,4,4,6,8,64,126
304
+ 15,3,4,4,7,13,64,189
305
+ 15,4,4,4,8,18,64,253
306
+ 15,5,4,4,9,23,64,316
307
+ 15,6,4,4,10,28,64,379
308
+ 15,7,4,4,11,31,64,442
309
+ 15,8,4,4,12,34,64,503
310
+ 15,9,4,4,13,39,64,564
311
+ 15,10,4,4,14,44,64,628
312
+ 15,11,4,4,15,49,64,692
313
+ 15,12,4,4,16,54,64,756
314
+ 15,13,4,4,17,59,64,819
315
+ 15,14,4,4,18,64,64,883
316
+ 15,15,4,4,19,69,64,945
317
+ 15,16,4,4,20,74,64,1009
318
+ 15,17,4,4,21,79,64,1069
319
+ 15,18,4,4,22,84,64,1133
320
+ 15,19,4,4,23,89,64,1195
321
+ 15,20,4,4,24,94,64,1259
322
+ 16,1,4,2,3,3,63,63
323
+ 16,2,4,4,6,8,64,127
324
+ 16,3,4,4,7,9,64,191
325
+ 16,4,4,4,8,13,64,253
326
+ 16,5,4,4,9,17,64,317
327
+ 16,6,4,4,10,18,64,381
328
+ 16,7,4,4,11,23,64,445
329
+ 16,8,4,4,12,28,64,509
330
+ 16,9,4,4,13,31,64,573
331
+ 16,10,4,4,14,36,64,635
332
+ 16,11,4,4,15,41,64,698
333
+ 16,12,4,4,16,44,64,762
334
+ 16,13,4,4,17,47,64,825
335
+ 16,14,4,4,18,52,64,887
336
+ 16,15,4,4,19,57,64,946
337
+ 16,16,4,4,20,60,64,1010
338
+ 16,17,4,4,21,65,64,1074
339
+ 16,18,4,4,22,69,64,1138
340
+ 16,19,4,4,23,71,64,1198
341
+ 16,20,4,4,24,75,64,1261
342
+ 17,1,4,4,5,5,64,64
343
+ 17,2,4,4,6,6,64,127
344
+ 17,3,4,4,7,11,64,188
345
+ 17,4,4,4,8,16,64,251
346
+ 17,5,4,4,9,21,64,314
347
+ 17,6,4,4,10,26,64,378
348
+ 17,7,4,4,11,31,64,442
349
+ 17,8,4,4,12,36,64,504
350
+ 17,9,4,4,13,41,64,568
351
+ 17,10,4,4,14,46,64,629
352
+ 17,11,4,4,15,51,64,693
353
+ 17,12,4,4,16,56,64,757
354
+ 17,13,4,4,17,61,64,820
355
+ 17,14,4,4,18,66,64,884
356
+ 17,15,4,4,19,71,64,948
357
+ 17,16,4,4,20,76,64,1011
358
+ 17,17,4,4,21,81,64,1073
359
+ 17,18,4,4,22,86,64,1137
360
+ 17,19,4,4,23,91,64,1200
361
+ 17,20,4,4,24,96,64,1263
362
+ 18,1,4,0,1,1,64,64
363
+ 18,2,4,3,5,5,64,128
364
+ 18,3,4,3,6,9,64,192
365
+ 18,4,4,3,7,13,64,255
366
+ 18,5,4,4,9,18,64,316
367
+ 18,6,4,4,10,22,64,380
368
+ 18,7,4,4,11,27,64,444
369
+ 18,8,4,4,12,32,64,508
370
+ 18,9,4,4,13,36,64,572
371
+ 18,10,4,4,14,40,64,635
372
+ 18,11,4,4,15,44,64,693
373
+ 18,12,4,4,16,46,64,756
374
+ 18,13,4,4,17,51,64,820
375
+ 18,14,4,4,18,55,64,884
376
+ 18,15,4,4,19,60,64,948
377
+ 18,16,4,4,20,65,64,1012
378
+ 18,17,4,4,21,69,64,1076
379
+ 18,18,4,4,22,70,64,1138
380
+ 18,19,4,4,23,74,64,1201
381
+ 18,20,4,4,24,78,64,1265
382
+ 19,1,4,4,5,5,64,64
383
+ 19,2,4,4,6,10,64,127
384
+ 19,3,4,4,7,15,64,191
385
+ 19,4,4,4,8,20,64,255
386
+ 19,5,4,4,9,25,64,316
387
+ 19,6,4,4,10,30,64,380
388
+ 19,7,4,4,11,35,64,444
389
+ 19,8,4,4,12,40,64,508
390
+ 19,9,4,4,13,45,64,571
391
+ 19,10,4,4,14,50,64,632
392
+ 19,11,4,4,15,55,64,696
393
+ 19,12,4,4,16,60,64,760
394
+ 19,13,4,4,17,65,64,823
395
+ 19,14,4,4,18,70,64,885
396
+ 19,15,4,4,19,75,64,945
397
+ 19,16,4,4,20,80,64,1006
398
+ 19,17,4,4,21,85,64,1069
399
+ 19,18,4,4,22,90,64,1132
400
+ 19,19,4,4,23,95,64,1195
401
+ 19,20,4,4,24,100,64,1258
402
+ 20,1,4,2,3,3,64,64
403
+ 20,2,4,3,5,7,64,125
404
+ 20,3,4,4,7,9,64,186
405
+ 20,4,4,4,8,13,64,246
406
+ 20,5,4,4,9,16,64,309
407
+ 20,6,4,4,10,20,64,373
408
+ 20,7,4,4,11,24,64,437
409
+ 20,8,4,4,12,28,64,495
410
+ 20,9,4,4,13,32,64,558
411
+ 20,10,4,4,14,34,64,622
412
+ 20,11,4,4,15,37,64,686
413
+ 20,12,4,4,16,41,64,749
414
+ 20,13,4,4,17,45,64,812
415
+ 20,14,4,4,18,46,64,875
416
+ 20,15,4,4,19,50,64,939
417
+ 20,16,4,4,20,52,64,1003
418
+ 20,17,4,4,21,54,64,1067
419
+ 20,18,4,4,22,58,64,1131
420
+ 20,19,4,4,23,61,64,1194
421
+ 20,20,4,4,24,64,64,1256
422
+ 21,1,4,4,5,5,64,64
423
+ 21,2,4,4,6,9,64,126
424
+ 21,3,4,4,7,13,64,189
425
+ 21,4,4,4,8,18,64,252
426
+ 21,5,4,4,9,23,64,310
427
+ 21,6,4,4,10,28,64,374
428
+ 21,7,4,4,11,32,64,437
429
+ 21,8,4,4,12,37,64,500
430
+ 21,9,4,4,13,42,64,564
431
+ 21,10,4,4,14,46,64,628
432
+ 21,11,4,4,15,51,64,692
433
+ 21,12,4,4,16,53,64,756
434
+ 21,13,4,4,17,57,64,820
435
+ 21,14,4,4,18,62,64,883
436
+ 21,15,4,4,19,66,64,947
437
+ 21,16,4,4,20,71,64,1008
438
+ 21,17,4,4,21,75,64,1072
439
+ 21,18,4,4,22,79,64,1136
440
+ 21,19,4,4,23,83,64,1199
441
+ 21,20,4,4,24,87,64,1263
442
+ 22,1,4,4,5,5,63,63
443
+ 22,2,4,4,6,10,64,127
444
+ 22,3,4,4,7,13,64,190
445
+ 22,4,4,4,8,18,64,249
446
+ 22,5,4,4,9,23,64,311
447
+ 22,6,4,4,10,28,64,374
448
+ 22,7,4,4,11,29,64,438
449
+ 22,8,4,4,12,30,64,501
450
+ 22,9,4,4,13,35,64,563
451
+ 22,10,4,4,14,40,64,627
452
+ 22,11,4,4,15,45,64,686
453
+ 22,12,4,4,16,48,64,750
454
+ 22,13,4,4,17,53,64,813
455
+ 22,14,4,4,18,58,64,877
456
+ 22,15,4,4,19,61,64,941
457
+ 22,16,4,4,20,66,64,1001
458
+ 22,17,4,4,21,71,64,1064
459
+ 22,18,4,4,22,72,64,1128
460
+ 22,19,4,4,23,73,64,1192
461
+ 22,20,4,4,24,78,64,1255
462
+ 23,1,4,3,4,4,64,64
463
+ 23,2,4,3,5,8,64,128
464
+ 23,3,4,4,7,13,64,192
465
+ 23,4,4,4,8,16,64,254
466
+ 23,5,4,4,9,19,64,318
467
+ 23,6,4,4,10,24,64,381
468
+ 23,7,4,4,11,27,64,445
469
+ 23,8,4,4,12,31,64,509
470
+ 23,9,4,4,13,34,64,571
471
+ 23,10,4,4,14,39,64,635
472
+ 23,11,4,4,15,42,64,698
473
+ 23,12,4,4,16,46,64,762
474
+ 23,13,4,4,17,50,64,823
475
+ 23,14,4,4,18,54,64,884
476
+ 23,15,4,4,19,58,64,948
477
+ 23,16,4,4,20,63,64,1012
478
+ 23,17,4,4,21,66,64,1076
479
+ 23,18,4,4,22,69,64,1136
480
+ 23,19,4,4,23,74,64,1200
481
+ 23,20,4,4,24,76,64,1264
reproduction/dual/README.md ADDED
@@ -0,0 +1,29 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Online inverse repair with forward and transpose measurements
2
+
3
+ A writer holds an invertible forward proposal model M and its inverse. Its action for a goal b is x=M⁻¹b. The hidden plant returns the complete binary word Ax for any chosen x, and Aᵀℓ for any chosen test ℓ. During one goal the plant is a fixed invertible linear operator on d-bit words. It may switch without notice between goals. Both queries return d bits. The experiment prices them separately. The hidden A is never passed to the writer.
4
+
5
+ Run `python experiment.py` with Python 3.10 or newer. The standard library suffices. The solver in `dual_writer.py` is reusable. Matrix columns and vectors are integer bit masks with bit zero as the first coordinate. `DualWriter(model).solve(goal, forward, transpose)` returns an action after verification. Preserve the object across goals to preserve its proposal weights. Initial model and observed responses must satisfy the interface above.
6
+
7
+ ## A repair that preserves invertibility and every already solved goal
8
+
9
+ Let E=A−M. A failed first attempt gives r=Ax−b=Ex≠0. Set y=M⁻¹r and query s=Ay≠0. Choose ℓ with ℓᵀr=ℓᵀs=1. If the supports intersect, choose one common coordinate. Otherwise take one coordinate from each support. One transpose call gives w=Aᵀℓ−Mᵀℓ=Eᵀℓ. In characteristic two, addition and subtraction coincide. Apply M′=M+r wᵀ. Each observation, decision and weight change uses exact binary arithmetic.
10
+
11
+ Since E′=(I−rℓᵀ)E, the image of E passes through a projection whose kernel is exactly span(r), and r belongs to that image. Thus rank(E′)=rank(E)−1 and ker(E)⊂ker(E′). The determinant lemma gives det(M′)/det(M)=1+wᵀM⁻¹r=ℓᵀs=1. The inverse is updated directly as M′⁻¹=M⁻¹+y wᵀM⁻¹. Using a pivot with ℓᵀr=1 alone can destroy invertibility. The results include an explicitly checked three-bit example.
12
+
13
+ For a fixed A, the one-shot solved goals are precisely A(ker E). Their number is 2^(d−rank E). They form nested sets across repairs and **each repair doubles their size**. All previously correct goals remain correct. The original failed goal can require further attempts. Invertibility and finite initial rank guarantee a verified action within d corrections even on that first goal.
14
+
15
+ If D₀=rank(A₁−M₀), V=Σ rank(Aₜ−Aₜ₋₁), and Dₜ is the debt remaining after task t, every prefix satisfies mistakes+Dₜ≤D₀+V through that prefix. This follows by rank subadditivity at a switch and exact decrease at each repair. Every T goals therefore cost exactly T+2·mistakes forward vector reads and mistakes transpose reads for this implementation, so its excess cost is bounded by drift rank instead of the number of goals. A rank-k external change can multiply or divide the one-shot solved fraction by at most 2ᵏ. For no changes, at most d mistakes occur over the entire lifetime. A persistent trace can contain public goals that never reveal an external change.
16
+
17
+ ## Experimental checks
18
+
19
+ Across dimensions 4, 8, 16 and 32, four seeds and three target schedules, every one of 3,072 returned actions was correct. The oracle changes by unannounced rank-one left compositions. The schedule includes explicitly noncommuting transvections and switches that cancel in pairs. Random goals are uniform over all words. On the worst schedule the evaluator chooses a goal that exposes a standard-basis disagreement whenever possible. In the third schedule zero goals can conceal a change and its cancellation. The adversarial evaluator knows the true matrix solely to construct the goal.
20
+
21
+ | Schedule | Goals | Initial debt + variation | Persistent mistakes | Same-information fixed refactor | Reset mistakes |
22
+ |---|---:|---:|---:|---:|---:|
23
+ | Random | 1,024 | 542 | 479 | 479 | 15,693 |
24
+ | Adversarial | 1,024 | 542 | 542 | 542 | 15,754 |
25
+ | Cancellation window | 1,024 | 542 | 414 | 414 | 14,000 |
26
+
27
+ The reference policy recomputes its inverse by full Gaussian elimination after every update. Its forward model, final actions and mistake count agree at every step with the persistent writer. A second baseline resets to the identity before every new goal, then applies the same repair. The experiment checks the rank descent, invertibility, update formula and prefix ledger on every update. For d≤8 it exhausts all candidate actions at every persistent update to verify that the correct-goal set strictly contains its predecessor and doubles. The evaluator's internal checks and target construction are offline and excluded from the forward/transpose call ledger.
28
+
29
+ The setup demonstrates a solver changing its own proposal weights within a specified linear class. The repair rule is fixed in code. The transpose oracle is a substantive extra capability. Access to goals or task scores alone does not implement it. Plant changes during one repair, nonlinear plants and noisy vector measurements invalidate this guarantee. Equality with the informed classical reference is an experimental result. Attainment of the ledger on an adversarial schedule demonstrates sharpness for this update rule.
reproduction/dual/dual_writer.py ADDED
@@ -0,0 +1,110 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """Exact GF(2) rank-one repair. Matrices are tuples of columns encoded as bit masks."""
2
+
3
+
4
+ def matvec(columns, v):
5
+ z = 0
6
+ while v:
7
+ bit = v & -v
8
+ z ^= columns[bit.bit_length() - 1]
9
+ v ^= bit
10
+ return z
11
+
12
+
13
+ def adj(columns, v):
14
+ return sum(((col & v).bit_count() % 2) << j for j, col in enumerate(columns))
15
+
16
+
17
+ def outer_update(columns, u, w):
18
+ return tuple(col ^ u if (w >> j) & 1 else col for j, col in enumerate(columns))
19
+
20
+
21
+ def rank(cols):
22
+ basis = {}
23
+ for v in cols:
24
+ while v:
25
+ i = v.bit_length() - 1
26
+ if i not in basis:
27
+ basis[i] = v
28
+ break
29
+ v ^= basis[i]
30
+ return len(basis)
31
+
32
+
33
+ def inverse(cols):
34
+ n = len(cols)
35
+ if not n or any(type(c) is not int or not 0 <= c < (1 << n) for c in cols):
36
+ raise ValueError("bad binary matrix")
37
+ basis = {}
38
+ for j, v in enumerate(cols):
39
+ a = 1 << j
40
+ while v:
41
+ i = v.bit_length() - 1
42
+ if i not in basis:
43
+ basis[i] = v, a
44
+ break
45
+ x, y = basis[i]
46
+ v, a = v ^ x, a ^ y
47
+ if not v:
48
+ raise ValueError("singular matrix")
49
+ inv = []
50
+ for j in range(n):
51
+ v, a = 1 << j, 0
52
+ while v:
53
+ x, y = basis[v.bit_length() - 1]
54
+ v, a = v ^ x, a ^ y
55
+ inv.append(a)
56
+ return tuple(inv)
57
+
58
+
59
+ def safe_functional(r, s):
60
+ if not r or not s:
61
+ raise ValueError("the residual and its transported vector must be nonzero")
62
+ overlap = r & s
63
+ if overlap:
64
+ return overlap & -overlap
65
+ return (r & -r) | (s & -s)
66
+
67
+
68
+ class DualWriter:
69
+ """State is the forward proposal model M and its inverse acting on future goals."""
70
+ def __init__(self, model, refactor=False):
71
+ self.model = tuple(model)
72
+ self.inv = inverse(self.model)
73
+ self.refactor = refactor
74
+ self.n = len(model)
75
+ self.forward_calls = self.adjoint_calls = self.mistakes = self.completed = 0
76
+
77
+ def solve(self, goal, forward, transpose, observe_update=None):
78
+ """Keep one fixed, invertible hidden operator throughout this call.
79
+
80
+ forward(v) returns A*v and transpose(v) returns A^T*v as whole n-bit words.
81
+ observe_update is an optional observer receiving pre/post models for offline audit.
82
+ """
83
+ if type(goal) is not int or not 0 <= goal < (1 << self.n):
84
+ raise ValueError("goal must be a nonnegative n-bit word")
85
+ for _ in range(self.n + 1):
86
+ x = matvec(self.inv, goal)
87
+ result = forward(x)
88
+ self.forward_calls += 1
89
+ r = result ^ goal
90
+ if not r:
91
+ self.completed += 1
92
+ return x
93
+ y = matvec(self.inv, r)
94
+ s = forward(y)
95
+ self.forward_calls += 1
96
+ ell = safe_functional(r, s)
97
+ w = transpose(ell) ^ adj(self.model, ell)
98
+ self.adjoint_calls += 1
99
+ if ((w & x).bit_count() % 2 != 1 or (w & y).bit_count() % 2 != 0):
100
+ raise ValueError("incompatible responses within a repair round")
101
+ old = self.model
102
+ self.model = outer_update(old, r, w)
103
+ if self.refactor:
104
+ self.inv = inverse(self.model)
105
+ else:
106
+ self.inv = outer_update(self.inv, y, adj(self.inv, w))
107
+ self.mistakes += 1
108
+ if observe_update:
109
+ observe_update(old, self.model, self.inv, r, s)
110
+ raise RuntimeError("exceeded the dimension bound; invalid or changing oracle")
reproduction/dual/experiment.py ADDED
@@ -0,0 +1,188 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """Reproduce dual-oracle drift experiments. Python >= 3.10, standard library."""
2
+ import argparse
3
+ import csv
4
+ import json
5
+ import random
6
+ from pathlib import Path
7
+ from dual_writer import DualWriter, adj, inverse, matvec, outer_update, rank
8
+
9
+
10
+ def distance(a, b):
11
+ return rank(x ^ y for x, y in zip(a, b))
12
+
13
+
14
+ def identity(d):
15
+ return tuple(1 << j for j in range(d))
16
+
17
+
18
+ def transvection(d, rng):
19
+ u = rng.randrange(1, 1 << d)
20
+ while True:
21
+ v = rng.randrange(1, 1 << d)
22
+ if not (u & v).bit_count() % 2:
23
+ return u, v
24
+
25
+
26
+ def compose(left, right):
27
+ return tuple(matvec(left, col) for col in right)
28
+
29
+
30
+ def trial(d, seed, mode):
31
+ rng = random.Random(d*99991+seed)
32
+ targets = random.Random(seed*3217+d)
33
+ eye = identity(d)
34
+ actual = eye
35
+ for _ in range(d):
36
+ u, v = transvection(d, rng)
37
+ actual = compose(outer_update(eye, u, v), actual)
38
+ persistent = DualWriter(eye)
39
+ fixed = DualWriter(eye, refactor=True)
40
+ rows = []
41
+ debt0 = distance(actual, eye)
42
+ variation = 0
43
+ naive_bad = 0
44
+ cancellation_pairs = 0
45
+ noncommuting = 0
46
+ total_reset = 0
47
+ previous_change = None
48
+ repeat_at = 0
49
+ snapshot = None
50
+ steps = d + 49
51
+ for t in range(steps):
52
+ old_actual = actual
53
+ if t >= d + 1 and (t-d-1) % 5 in (0, 1):
54
+ stage = (t-d-1) // 5
55
+ if t == repeat_at:
56
+ change = previous_change
57
+ elif stage % 3 == 0 and (t-d-1) % 5 == 0:
58
+ change = outer_update(eye, *transvection(d, rng))
59
+ snapshot, repeat_at = actual, t + 1
60
+ elif stage % 3 == 1:
61
+ u, v = (1, 2) if (t-d-1)%5 == 0 else (2, 1)
62
+ change = outer_update(eye, u, v)
63
+ else:
64
+ change = outer_update(eye, *transvection(d, rng))
65
+ actual = compose(change, actual)
66
+ if previous_change is not None:
67
+ noncommuting += int(compose(change, previous_change) != compose(previous_change, change))
68
+ previous_change = change
69
+ if t == repeat_at:
70
+ assert actual == snapshot
71
+ cancellation_pairs += 1
72
+ assert distance(actual, old_actual) == 1
73
+ drift = distance(actual, old_actual)
74
+ variation += drift
75
+ debt_before = distance(actual, persistent.model)
76
+ if mode == 'worst' and debt_before:
77
+ k = next(k for k,(a,m) in enumerate(zip(actual,persistent.model)) if a != m)
78
+ goal = persistent.model[k]
79
+ elif mode == 'blind_window' and t >= d+1 and (t-d-1)//5 % 3 == 0 and (t-d-1)%5 in (0,1):
80
+ # Zero is a valid easy goal, so a change and its inverse can pass unobserved.
81
+ goal = 0
82
+ else:
83
+ goal = targets.randrange(1 << d)
84
+ def forward(v):
85
+ return matvec(actual, v)
86
+ def transpose(v):
87
+ return adj(actual, v)
88
+ transitions = []
89
+ def audit(old, new, inv, r, s):
90
+ nonlocal naive_bad
91
+ naive_pivot = r & -r
92
+ naive_bad += int((naive_pivot & s).bit_count() % 2 == 0)
93
+ old_rank, new_rank = distance(actual, old), distance(actual, new)
94
+ assert new_rank + 1 == old_rank
95
+ assert inv == inverse(new)
96
+ if d <= 8:
97
+ before_goals = {matvec(actual, a) for a in range(1 << d)
98
+ if matvec(old, a) == matvec(actual, a)}
99
+ after_goals = {matvec(actual, a) for a in range(1 << d)
100
+ if matvec(new, a) == matvec(actual, a)}
101
+ assert before_goals <= after_goals
102
+ assert len(after_goals) == 2*len(before_goals)
103
+ transitions.append((old_rank, new_rank))
104
+ before, f0, a0 = persistent.mistakes, persistent.forward_calls, persistent.adjoint_calls
105
+ solution = persistent.solve(goal, forward, transpose, audit)
106
+ assert forward(solution) == goal
107
+ missed = persistent.mistakes - before
108
+ assert persistent.forward_calls - f0 == 1+2*missed and persistent.adjoint_calls-a0 == missed
109
+ fixed_solution = fixed.solve(goal, forward, transpose)
110
+ assert fixed_solution == solution and persistent.model == fixed.model
111
+ reset = DualWriter(eye)
112
+ reset_solution = reset.solve(goal, forward, transpose)
113
+ assert reset_solution == solution
114
+ total_reset += reset.mistakes
115
+ debt_after = distance(actual, persistent.model)
116
+ assert missed == debt_before-debt_after
117
+ assert persistent.mistakes+debt_after <= debt0 + variation
118
+ rows.append(dict(d=d, seed=seed, mode=mode,t=t, goal=goal,solution=solution,
119
+ drift_rank=drift, debt_before=debt_before, debt_after=debt_after,
120
+ mistakes=missed, forward_reads=1+2*missed, adjoint_reads=missed,
121
+ reset_mistakes=reset.mistakes))
122
+ assert persistent.mistakes == fixed.mistakes
123
+ assert persistent.forward_calls == steps+2*persistent.mistakes
124
+ assert persistent.adjoint_calls == persistent.mistakes
125
+ assert cancellation_pairs and noncommuting
126
+ if mode == 'worst':
127
+ assert persistent.mistakes == debt0+variation
128
+ return rows,dict(d=d,seed=seed,mode=mode,goals=steps,initial_rank_debt=debt0,
129
+ operator_rank_variation=variation,mistakes=persistent.mistakes,
130
+ fixed_same_information_mistakes=fixed.mistakes,reset_mistakes=total_reset,
131
+ final_rank_debt=debt_after,naive_singular_pivots=naive_bad,
132
+ explicit_cancellations=cancellation_pairs,noncommuting_adjacent_events=noncommuting)
133
+
134
+
135
+ def forced_singular_naive():
136
+ # A swaps coordinates, M=I, goal=first basis vector. Here r=(1,1), s=r,
137
+ # so this easy swap itself is not the counterexample. Enumerate 3x3 invertibles
138
+ # to find a case in which the first-set-bit choice provably destroys rank.
139
+ import itertools
140
+ d=3
141
+ eye=identity(d)
142
+ for cols in itertools.product(range(1, 8), repeat=d):
143
+ if rank(cols) != d:
144
+ continue
145
+ for goal in range(1,8):
146
+ r=matvec(cols,goal)^goal
147
+ if not r:
148
+ continue
149
+ s=matvec(cols,r); naive=r & -r
150
+ if (s & naive).bit_count()%2:
151
+ continue
152
+ w=adj(cols,naive)^naive
153
+ bad=outer_update(eye,r,w)
154
+ assert rank(bad)==d-1
155
+ writer=DualWriter(eye)
156
+ solution=writer.solve(goal,lambda v:matvec(cols,v),lambda v:adj(cols,v))
157
+ assert matvec(cols,solution)==goal and rank(writer.model)==d
158
+ return dict(matrix_columns=cols,initial_model=eye,goal=goal,residual=r,
159
+ transported_residual=s,unsafe_functional=naive,
160
+ naive_model_columns=bad,naive_rank=rank(bad),safe_final_rank=rank(writer.model))
161
+ raise AssertionError("missing counterexample")
162
+
163
+
164
+ def main():
165
+ parser=argparse.ArgumentParser(description=__doc__)
166
+ parser.add_argument('--out',type=Path,default=Path(__file__).resolve().parent)
167
+ out=parser.parse_args().out
168
+ out.mkdir(parents=True,exist_ok=True)
169
+ runs,rows=[],[]
170
+ for d in (4,8,16,32):
171
+ for seed in range(4):
172
+ for mode in ('random','worst','blind_window'):
173
+ rr,summary=trial(d,seed,mode); rows+=rr;runs.append(summary)
174
+ aggregate={}
175
+ for mode in ('random','worst','blind_window'):
176
+ group=[g for g in runs if g['mode']==mode]
177
+ aggregate[mode]={key:sum(g[key] for g in group) for key in
178
+ ('goals','initial_rank_debt','operator_rank_variation','mistakes',
179
+ 'reset_mistakes','fixed_same_information_mistakes','naive_singular_pivots')}
180
+ result=dict(aggregate=aggregate,forced_naive_singularity=forced_singular_naive(),runs=runs)
181
+ (out/'results.json').write_text(json.dumps(result,indent=2)+'\n',encoding='utf-8')
182
+ with (out/'trace.csv').open('w',newline='',encoding='utf-8') as f:
183
+ writer=csv.DictWriter(f,fieldnames=rows[0]);writer.writeheader();writer.writerows(rows)
184
+ print(json.dumps(aggregate,indent=2))
185
+
186
+
187
+ if __name__=='__main__':
188
+ main()
reproduction/dual/results.json ADDED
@@ -0,0 +1,776 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "aggregate": {
3
+ "random": {
4
+ "goals": 1024,
5
+ "initial_rank_debt": 222,
6
+ "operator_rank_variation": 320,
7
+ "mistakes": 479,
8
+ "reset_mistakes": 15693,
9
+ "fixed_same_information_mistakes": 479,
10
+ "naive_singular_pivots": 142
11
+ },
12
+ "worst": {
13
+ "goals": 1024,
14
+ "initial_rank_debt": 222,
15
+ "operator_rank_variation": 320,
16
+ "mistakes": 542,
17
+ "reset_mistakes": 15754,
18
+ "fixed_same_information_mistakes": 542,
19
+ "naive_singular_pivots": 108
20
+ },
21
+ "blind_window": {
22
+ "goals": 1024,
23
+ "initial_rank_debt": 222,
24
+ "operator_rank_variation": 320,
25
+ "mistakes": 414,
26
+ "reset_mistakes": 14000,
27
+ "fixed_same_information_mistakes": 414,
28
+ "naive_singular_pivots": 135
29
+ }
30
+ },
31
+ "forced_naive_singularity": {
32
+ "matrix_columns": [
33
+ 1,
34
+ 3,
35
+ 6
36
+ ],
37
+ "initial_model": [
38
+ 1,
39
+ 2,
40
+ 4
41
+ ],
42
+ "goal": 6,
43
+ "residual": 3,
44
+ "transported_residual": 2,
45
+ "unsafe_functional": 1,
46
+ "naive_model_columns": [
47
+ 1,
48
+ 1,
49
+ 4
50
+ ],
51
+ "naive_rank": 2,
52
+ "safe_final_rank": 3
53
+ },
54
+ "runs": [
55
+ {
56
+ "d": 4,
57
+ "seed": 0,
58
+ "mode": "random",
59
+ "goals": 53,
60
+ "initial_rank_debt": 3,
61
+ "operator_rank_variation": 20,
62
+ "mistakes": 19,
63
+ "fixed_same_information_mistakes": 19,
64
+ "reset_mistakes": 133,
65
+ "final_rank_debt": 0,
66
+ "naive_singular_pivots": 3,
67
+ "explicit_cancellations": 4,
68
+ "noncommuting_adjacent_events": 11
69
+ },
70
+ {
71
+ "d": 4,
72
+ "seed": 0,
73
+ "mode": "worst",
74
+ "goals": 53,
75
+ "initial_rank_debt": 3,
76
+ "operator_rank_variation": 20,
77
+ "mistakes": 23,
78
+ "fixed_same_information_mistakes": 23,
79
+ "reset_mistakes": 142,
80
+ "final_rank_debt": 0,
81
+ "naive_singular_pivots": 0,
82
+ "explicit_cancellations": 4,
83
+ "noncommuting_adjacent_events": 11
84
+ },
85
+ {
86
+ "d": 4,
87
+ "seed": 0,
88
+ "mode": "blind_window",
89
+ "goals": 53,
90
+ "initial_rank_debt": 3,
91
+ "operator_rank_variation": 20,
92
+ "mistakes": 15,
93
+ "fixed_same_information_mistakes": 15,
94
+ "reset_mistakes": 110,
95
+ "final_rank_debt": 0,
96
+ "naive_singular_pivots": 5,
97
+ "explicit_cancellations": 4,
98
+ "noncommuting_adjacent_events": 11
99
+ },
100
+ {
101
+ "d": 4,
102
+ "seed": 1,
103
+ "mode": "random",
104
+ "goals": 53,
105
+ "initial_rank_debt": 3,
106
+ "operator_rank_variation": 20,
107
+ "mistakes": 21,
108
+ "fixed_same_information_mistakes": 21,
109
+ "reset_mistakes": 130,
110
+ "final_rank_debt": 0,
111
+ "naive_singular_pivots": 2,
112
+ "explicit_cancellations": 4,
113
+ "noncommuting_adjacent_events": 13
114
+ },
115
+ {
116
+ "d": 4,
117
+ "seed": 1,
118
+ "mode": "worst",
119
+ "goals": 53,
120
+ "initial_rank_debt": 3,
121
+ "operator_rank_variation": 20,
122
+ "mistakes": 23,
123
+ "fixed_same_information_mistakes": 23,
124
+ "reset_mistakes": 139,
125
+ "final_rank_debt": 0,
126
+ "naive_singular_pivots": 1,
127
+ "explicit_cancellations": 4,
128
+ "noncommuting_adjacent_events": 13
129
+ },
130
+ {
131
+ "d": 4,
132
+ "seed": 1,
133
+ "mode": "blind_window",
134
+ "goals": 53,
135
+ "initial_rank_debt": 3,
136
+ "operator_rank_variation": 20,
137
+ "mistakes": 15,
138
+ "fixed_same_information_mistakes": 15,
139
+ "reset_mistakes": 113,
140
+ "final_rank_debt": 0,
141
+ "naive_singular_pivots": 2,
142
+ "explicit_cancellations": 4,
143
+ "noncommuting_adjacent_events": 13
144
+ },
145
+ {
146
+ "d": 4,
147
+ "seed": 2,
148
+ "mode": "random",
149
+ "goals": 53,
150
+ "initial_rank_debt": 3,
151
+ "operator_rank_variation": 20,
152
+ "mistakes": 19,
153
+ "fixed_same_information_mistakes": 19,
154
+ "reset_mistakes": 158,
155
+ "final_rank_debt": 0,
156
+ "naive_singular_pivots": 4,
157
+ "explicit_cancellations": 4,
158
+ "noncommuting_adjacent_events": 12
159
+ },
160
+ {
161
+ "d": 4,
162
+ "seed": 2,
163
+ "mode": "worst",
164
+ "goals": 53,
165
+ "initial_rank_debt": 3,
166
+ "operator_rank_variation": 20,
167
+ "mistakes": 23,
168
+ "fixed_same_information_mistakes": 23,
169
+ "reset_mistakes": 165,
170
+ "final_rank_debt": 0,
171
+ "naive_singular_pivots": 1,
172
+ "explicit_cancellations": 4,
173
+ "noncommuting_adjacent_events": 12
174
+ },
175
+ {
176
+ "d": 4,
177
+ "seed": 2,
178
+ "mode": "blind_window",
179
+ "goals": 53,
180
+ "initial_rank_debt": 3,
181
+ "operator_rank_variation": 20,
182
+ "mistakes": 15,
183
+ "fixed_same_information_mistakes": 15,
184
+ "reset_mistakes": 137,
185
+ "final_rank_debt": 0,
186
+ "naive_singular_pivots": 4,
187
+ "explicit_cancellations": 4,
188
+ "noncommuting_adjacent_events": 12
189
+ },
190
+ {
191
+ "d": 4,
192
+ "seed": 3,
193
+ "mode": "random",
194
+ "goals": 53,
195
+ "initial_rank_debt": 3,
196
+ "operator_rank_variation": 20,
197
+ "mistakes": 19,
198
+ "fixed_same_information_mistakes": 19,
199
+ "reset_mistakes": 112,
200
+ "final_rank_debt": 0,
201
+ "naive_singular_pivots": 3,
202
+ "explicit_cancellations": 4,
203
+ "noncommuting_adjacent_events": 11
204
+ },
205
+ {
206
+ "d": 4,
207
+ "seed": 3,
208
+ "mode": "worst",
209
+ "goals": 53,
210
+ "initial_rank_debt": 3,
211
+ "operator_rank_variation": 20,
212
+ "mistakes": 23,
213
+ "fixed_same_information_mistakes": 23,
214
+ "reset_mistakes": 129,
215
+ "final_rank_debt": 0,
216
+ "naive_singular_pivots": 1,
217
+ "explicit_cancellations": 4,
218
+ "noncommuting_adjacent_events": 11
219
+ },
220
+ {
221
+ "d": 4,
222
+ "seed": 3,
223
+ "mode": "blind_window",
224
+ "goals": 53,
225
+ "initial_rank_debt": 3,
226
+ "operator_rank_variation": 20,
227
+ "mistakes": 15,
228
+ "fixed_same_information_mistakes": 15,
229
+ "reset_mistakes": 97,
230
+ "final_rank_debt": 0,
231
+ "naive_singular_pivots": 1,
232
+ "explicit_cancellations": 4,
233
+ "noncommuting_adjacent_events": 11
234
+ },
235
+ {
236
+ "d": 8,
237
+ "seed": 0,
238
+ "mode": "random",
239
+ "goals": 57,
240
+ "initial_rank_debt": 6,
241
+ "operator_rank_variation": 20,
242
+ "mistakes": 22,
243
+ "fixed_same_information_mistakes": 22,
244
+ "reset_mistakes": 364,
245
+ "final_rank_debt": 0,
246
+ "naive_singular_pivots": 5,
247
+ "explicit_cancellations": 4,
248
+ "noncommuting_adjacent_events": 11
249
+ },
250
+ {
251
+ "d": 8,
252
+ "seed": 0,
253
+ "mode": "worst",
254
+ "goals": 57,
255
+ "initial_rank_debt": 6,
256
+ "operator_rank_variation": 20,
257
+ "mistakes": 26,
258
+ "fixed_same_information_mistakes": 26,
259
+ "reset_mistakes": 355,
260
+ "final_rank_debt": 0,
261
+ "naive_singular_pivots": 4,
262
+ "explicit_cancellations": 4,
263
+ "noncommuting_adjacent_events": 11
264
+ },
265
+ {
266
+ "d": 8,
267
+ "seed": 0,
268
+ "mode": "blind_window",
269
+ "goals": 57,
270
+ "initial_rank_debt": 6,
271
+ "operator_rank_variation": 20,
272
+ "mistakes": 18,
273
+ "fixed_same_information_mistakes": 18,
274
+ "reset_mistakes": 308,
275
+ "final_rank_debt": 0,
276
+ "naive_singular_pivots": 3,
277
+ "explicit_cancellations": 4,
278
+ "noncommuting_adjacent_events": 11
279
+ },
280
+ {
281
+ "d": 8,
282
+ "seed": 1,
283
+ "mode": "random",
284
+ "goals": 57,
285
+ "initial_rank_debt": 7,
286
+ "operator_rank_variation": 20,
287
+ "mistakes": 23,
288
+ "fixed_same_information_mistakes": 23,
289
+ "reset_mistakes": 383,
290
+ "final_rank_debt": 0,
291
+ "naive_singular_pivots": 2,
292
+ "explicit_cancellations": 4,
293
+ "noncommuting_adjacent_events": 11
294
+ },
295
+ {
296
+ "d": 8,
297
+ "seed": 1,
298
+ "mode": "worst",
299
+ "goals": 57,
300
+ "initial_rank_debt": 7,
301
+ "operator_rank_variation": 20,
302
+ "mistakes": 27,
303
+ "fixed_same_information_mistakes": 27,
304
+ "reset_mistakes": 379,
305
+ "final_rank_debt": 0,
306
+ "naive_singular_pivots": 3,
307
+ "explicit_cancellations": 4,
308
+ "noncommuting_adjacent_events": 11
309
+ },
310
+ {
311
+ "d": 8,
312
+ "seed": 1,
313
+ "mode": "blind_window",
314
+ "goals": 57,
315
+ "initial_rank_debt": 7,
316
+ "operator_rank_variation": 20,
317
+ "mistakes": 19,
318
+ "fixed_same_information_mistakes": 19,
319
+ "reset_mistakes": 338,
320
+ "final_rank_debt": 0,
321
+ "naive_singular_pivots": 3,
322
+ "explicit_cancellations": 4,
323
+ "noncommuting_adjacent_events": 11
324
+ },
325
+ {
326
+ "d": 8,
327
+ "seed": 2,
328
+ "mode": "random",
329
+ "goals": 57,
330
+ "initial_rank_debt": 7,
331
+ "operator_rank_variation": 20,
332
+ "mistakes": 21,
333
+ "fixed_same_information_mistakes": 21,
334
+ "reset_mistakes": 340,
335
+ "final_rank_debt": 0,
336
+ "naive_singular_pivots": 4,
337
+ "explicit_cancellations": 4,
338
+ "noncommuting_adjacent_events": 12
339
+ },
340
+ {
341
+ "d": 8,
342
+ "seed": 2,
343
+ "mode": "worst",
344
+ "goals": 57,
345
+ "initial_rank_debt": 7,
346
+ "operator_rank_variation": 20,
347
+ "mistakes": 27,
348
+ "fixed_same_information_mistakes": 27,
349
+ "reset_mistakes": 364,
350
+ "final_rank_debt": 0,
351
+ "naive_singular_pivots": 4,
352
+ "explicit_cancellations": 4,
353
+ "noncommuting_adjacent_events": 12
354
+ },
355
+ {
356
+ "d": 8,
357
+ "seed": 2,
358
+ "mode": "blind_window",
359
+ "goals": 57,
360
+ "initial_rank_debt": 7,
361
+ "operator_rank_variation": 20,
362
+ "mistakes": 19,
363
+ "fixed_same_information_mistakes": 19,
364
+ "reset_mistakes": 311,
365
+ "final_rank_debt": 0,
366
+ "naive_singular_pivots": 4,
367
+ "explicit_cancellations": 4,
368
+ "noncommuting_adjacent_events": 12
369
+ },
370
+ {
371
+ "d": 8,
372
+ "seed": 3,
373
+ "mode": "random",
374
+ "goals": 57,
375
+ "initial_rank_debt": 6,
376
+ "operator_rank_variation": 20,
377
+ "mistakes": 26,
378
+ "fixed_same_information_mistakes": 26,
379
+ "reset_mistakes": 337,
380
+ "final_rank_debt": 0,
381
+ "naive_singular_pivots": 4,
382
+ "explicit_cancellations": 4,
383
+ "noncommuting_adjacent_events": 12
384
+ },
385
+ {
386
+ "d": 8,
387
+ "seed": 3,
388
+ "mode": "worst",
389
+ "goals": 57,
390
+ "initial_rank_debt": 6,
391
+ "operator_rank_variation": 20,
392
+ "mistakes": 26,
393
+ "fixed_same_information_mistakes": 26,
394
+ "reset_mistakes": 349,
395
+ "final_rank_debt": 0,
396
+ "naive_singular_pivots": 1,
397
+ "explicit_cancellations": 4,
398
+ "noncommuting_adjacent_events": 12
399
+ },
400
+ {
401
+ "d": 8,
402
+ "seed": 3,
403
+ "mode": "blind_window",
404
+ "goals": 57,
405
+ "initial_rank_debt": 6,
406
+ "operator_rank_variation": 20,
407
+ "mistakes": 18,
408
+ "fixed_same_information_mistakes": 18,
409
+ "reset_mistakes": 300,
410
+ "final_rank_debt": 0,
411
+ "naive_singular_pivots": 3,
412
+ "explicit_cancellations": 4,
413
+ "noncommuting_adjacent_events": 12
414
+ },
415
+ {
416
+ "d": 16,
417
+ "seed": 0,
418
+ "mode": "random",
419
+ "goals": 65,
420
+ "initial_rank_debt": 15,
421
+ "operator_rank_variation": 20,
422
+ "mistakes": 30,
423
+ "fixed_same_information_mistakes": 30,
424
+ "reset_mistakes": 915,
425
+ "final_rank_debt": 1,
426
+ "naive_singular_pivots": 11,
427
+ "explicit_cancellations": 4,
428
+ "noncommuting_adjacent_events": 13
429
+ },
430
+ {
431
+ "d": 16,
432
+ "seed": 0,
433
+ "mode": "worst",
434
+ "goals": 65,
435
+ "initial_rank_debt": 15,
436
+ "operator_rank_variation": 20,
437
+ "mistakes": 35,
438
+ "fixed_same_information_mistakes": 35,
439
+ "reset_mistakes": 916,
440
+ "final_rank_debt": 0,
441
+ "naive_singular_pivots": 9,
442
+ "explicit_cancellations": 4,
443
+ "noncommuting_adjacent_events": 13
444
+ },
445
+ {
446
+ "d": 16,
447
+ "seed": 0,
448
+ "mode": "blind_window",
449
+ "goals": 65,
450
+ "initial_rank_debt": 15,
451
+ "operator_rank_variation": 20,
452
+ "mistakes": 27,
453
+ "fixed_same_information_mistakes": 27,
454
+ "reset_mistakes": 812,
455
+ "final_rank_debt": 0,
456
+ "naive_singular_pivots": 11,
457
+ "explicit_cancellations": 4,
458
+ "noncommuting_adjacent_events": 13
459
+ },
460
+ {
461
+ "d": 16,
462
+ "seed": 1,
463
+ "mode": "random",
464
+ "goals": 65,
465
+ "initial_rank_debt": 14,
466
+ "operator_rank_variation": 20,
467
+ "mistakes": 32,
468
+ "fixed_same_information_mistakes": 32,
469
+ "reset_mistakes": 916,
470
+ "final_rank_debt": 0,
471
+ "naive_singular_pivots": 6,
472
+ "explicit_cancellations": 4,
473
+ "noncommuting_adjacent_events": 12
474
+ },
475
+ {
476
+ "d": 16,
477
+ "seed": 1,
478
+ "mode": "worst",
479
+ "goals": 65,
480
+ "initial_rank_debt": 14,
481
+ "operator_rank_variation": 20,
482
+ "mistakes": 34,
483
+ "fixed_same_information_mistakes": 34,
484
+ "reset_mistakes": 908,
485
+ "final_rank_debt": 0,
486
+ "naive_singular_pivots": 3,
487
+ "explicit_cancellations": 4,
488
+ "noncommuting_adjacent_events": 12
489
+ },
490
+ {
491
+ "d": 16,
492
+ "seed": 1,
493
+ "mode": "blind_window",
494
+ "goals": 65,
495
+ "initial_rank_debt": 14,
496
+ "operator_rank_variation": 20,
497
+ "mistakes": 26,
498
+ "fixed_same_information_mistakes": 26,
499
+ "reset_mistakes": 793,
500
+ "final_rank_debt": 0,
501
+ "naive_singular_pivots": 8,
502
+ "explicit_cancellations": 4,
503
+ "noncommuting_adjacent_events": 12
504
+ },
505
+ {
506
+ "d": 16,
507
+ "seed": 2,
508
+ "mode": "random",
509
+ "goals": 65,
510
+ "initial_rank_debt": 15,
511
+ "operator_rank_variation": 20,
512
+ "mistakes": 29,
513
+ "fixed_same_information_mistakes": 29,
514
+ "reset_mistakes": 961,
515
+ "final_rank_debt": 0,
516
+ "naive_singular_pivots": 8,
517
+ "explicit_cancellations": 4,
518
+ "noncommuting_adjacent_events": 8
519
+ },
520
+ {
521
+ "d": 16,
522
+ "seed": 2,
523
+ "mode": "worst",
524
+ "goals": 65,
525
+ "initial_rank_debt": 15,
526
+ "operator_rank_variation": 20,
527
+ "mistakes": 35,
528
+ "fixed_same_information_mistakes": 35,
529
+ "reset_mistakes": 965,
530
+ "final_rank_debt": 0,
531
+ "naive_singular_pivots": 9,
532
+ "explicit_cancellations": 4,
533
+ "noncommuting_adjacent_events": 8
534
+ },
535
+ {
536
+ "d": 16,
537
+ "seed": 2,
538
+ "mode": "blind_window",
539
+ "goals": 65,
540
+ "initial_rank_debt": 15,
541
+ "operator_rank_variation": 20,
542
+ "mistakes": 27,
543
+ "fixed_same_information_mistakes": 27,
544
+ "reset_mistakes": 842,
545
+ "final_rank_debt": 0,
546
+ "naive_singular_pivots": 5,
547
+ "explicit_cancellations": 4,
548
+ "noncommuting_adjacent_events": 8
549
+ },
550
+ {
551
+ "d": 16,
552
+ "seed": 3,
553
+ "mode": "random",
554
+ "goals": 65,
555
+ "initial_rank_debt": 15,
556
+ "operator_rank_variation": 20,
557
+ "mistakes": 34,
558
+ "fixed_same_information_mistakes": 34,
559
+ "reset_mistakes": 974,
560
+ "final_rank_debt": 1,
561
+ "naive_singular_pivots": 9,
562
+ "explicit_cancellations": 4,
563
+ "noncommuting_adjacent_events": 14
564
+ },
565
+ {
566
+ "d": 16,
567
+ "seed": 3,
568
+ "mode": "worst",
569
+ "goals": 65,
570
+ "initial_rank_debt": 15,
571
+ "operator_rank_variation": 20,
572
+ "mistakes": 35,
573
+ "fixed_same_information_mistakes": 35,
574
+ "reset_mistakes": 970,
575
+ "final_rank_debt": 0,
576
+ "naive_singular_pivots": 5,
577
+ "explicit_cancellations": 4,
578
+ "noncommuting_adjacent_events": 14
579
+ },
580
+ {
581
+ "d": 16,
582
+ "seed": 3,
583
+ "mode": "blind_window",
584
+ "goals": 65,
585
+ "initial_rank_debt": 15,
586
+ "operator_rank_variation": 20,
587
+ "mistakes": 27,
588
+ "fixed_same_information_mistakes": 27,
589
+ "reset_mistakes": 846,
590
+ "final_rank_debt": 0,
591
+ "naive_singular_pivots": 9,
592
+ "explicit_cancellations": 4,
593
+ "noncommuting_adjacent_events": 14
594
+ },
595
+ {
596
+ "d": 32,
597
+ "seed": 0,
598
+ "mode": "random",
599
+ "goals": 81,
600
+ "initial_rank_debt": 31,
601
+ "operator_rank_variation": 20,
602
+ "mistakes": 46,
603
+ "fixed_same_information_mistakes": 46,
604
+ "reset_mistakes": 2499,
605
+ "final_rank_debt": 1,
606
+ "naive_singular_pivots": 21,
607
+ "explicit_cancellations": 4,
608
+ "noncommuting_adjacent_events": 12
609
+ },
610
+ {
611
+ "d": 32,
612
+ "seed": 0,
613
+ "mode": "worst",
614
+ "goals": 81,
615
+ "initial_rank_debt": 31,
616
+ "operator_rank_variation": 20,
617
+ "mistakes": 51,
618
+ "fixed_same_information_mistakes": 51,
619
+ "reset_mistakes": 2499,
620
+ "final_rank_debt": 0,
621
+ "naive_singular_pivots": 13,
622
+ "explicit_cancellations": 4,
623
+ "noncommuting_adjacent_events": 12
624
+ },
625
+ {
626
+ "d": 32,
627
+ "seed": 0,
628
+ "mode": "blind_window",
629
+ "goals": 81,
630
+ "initial_rank_debt": 31,
631
+ "operator_rank_variation": 20,
632
+ "mistakes": 43,
633
+ "fixed_same_information_mistakes": 43,
634
+ "reset_mistakes": 2246,
635
+ "final_rank_debt": 0,
636
+ "naive_singular_pivots": 21,
637
+ "explicit_cancellations": 4,
638
+ "noncommuting_adjacent_events": 12
639
+ },
640
+ {
641
+ "d": 32,
642
+ "seed": 1,
643
+ "mode": "random",
644
+ "goals": 81,
645
+ "initial_rank_debt": 31,
646
+ "operator_rank_variation": 20,
647
+ "mistakes": 43,
648
+ "fixed_same_information_mistakes": 43,
649
+ "reset_mistakes": 2443,
650
+ "final_rank_debt": 0,
651
+ "naive_singular_pivots": 23,
652
+ "explicit_cancellations": 4,
653
+ "noncommuting_adjacent_events": 13
654
+ },
655
+ {
656
+ "d": 32,
657
+ "seed": 1,
658
+ "mode": "worst",
659
+ "goals": 81,
660
+ "initial_rank_debt": 31,
661
+ "operator_rank_variation": 20,
662
+ "mistakes": 51,
663
+ "fixed_same_information_mistakes": 51,
664
+ "reset_mistakes": 2439,
665
+ "final_rank_debt": 0,
666
+ "naive_singular_pivots": 16,
667
+ "explicit_cancellations": 4,
668
+ "noncommuting_adjacent_events": 13
669
+ },
670
+ {
671
+ "d": 32,
672
+ "seed": 1,
673
+ "mode": "blind_window",
674
+ "goals": 81,
675
+ "initial_rank_debt": 31,
676
+ "operator_rank_variation": 20,
677
+ "mistakes": 43,
678
+ "fixed_same_information_mistakes": 43,
679
+ "reset_mistakes": 2189,
680
+ "final_rank_debt": 0,
681
+ "naive_singular_pivots": 20,
682
+ "explicit_cancellations": 4,
683
+ "noncommuting_adjacent_events": 13
684
+ },
685
+ {
686
+ "d": 32,
687
+ "seed": 2,
688
+ "mode": "random",
689
+ "goals": 81,
690
+ "initial_rank_debt": 32,
691
+ "operator_rank_variation": 20,
692
+ "mistakes": 48,
693
+ "fixed_same_information_mistakes": 48,
694
+ "reset_mistakes": 2531,
695
+ "final_rank_debt": 0,
696
+ "naive_singular_pivots": 18,
697
+ "explicit_cancellations": 4,
698
+ "noncommuting_adjacent_events": 12
699
+ },
700
+ {
701
+ "d": 32,
702
+ "seed": 2,
703
+ "mode": "worst",
704
+ "goals": 81,
705
+ "initial_rank_debt": 32,
706
+ "operator_rank_variation": 20,
707
+ "mistakes": 52,
708
+ "fixed_same_information_mistakes": 52,
709
+ "reset_mistakes": 2536,
710
+ "final_rank_debt": 0,
711
+ "naive_singular_pivots": 20,
712
+ "explicit_cancellations": 4,
713
+ "noncommuting_adjacent_events": 12
714
+ },
715
+ {
716
+ "d": 32,
717
+ "seed": 2,
718
+ "mode": "blind_window",
719
+ "goals": 81,
720
+ "initial_rank_debt": 32,
721
+ "operator_rank_variation": 20,
722
+ "mistakes": 44,
723
+ "fixed_same_information_mistakes": 44,
724
+ "reset_mistakes": 2288,
725
+ "final_rank_debt": 0,
726
+ "naive_singular_pivots": 18,
727
+ "explicit_cancellations": 4,
728
+ "noncommuting_adjacent_events": 12
729
+ },
730
+ {
731
+ "d": 32,
732
+ "seed": 3,
733
+ "mode": "random",
734
+ "goals": 81,
735
+ "initial_rank_debt": 31,
736
+ "operator_rank_variation": 20,
737
+ "mistakes": 47,
738
+ "fixed_same_information_mistakes": 47,
739
+ "reset_mistakes": 2497,
740
+ "final_rank_debt": 0,
741
+ "naive_singular_pivots": 19,
742
+ "explicit_cancellations": 4,
743
+ "noncommuting_adjacent_events": 10
744
+ },
745
+ {
746
+ "d": 32,
747
+ "seed": 3,
748
+ "mode": "worst",
749
+ "goals": 81,
750
+ "initial_rank_debt": 31,
751
+ "operator_rank_variation": 20,
752
+ "mistakes": 51,
753
+ "fixed_same_information_mistakes": 51,
754
+ "reset_mistakes": 2499,
755
+ "final_rank_debt": 0,
756
+ "naive_singular_pivots": 18,
757
+ "explicit_cancellations": 4,
758
+ "noncommuting_adjacent_events": 10
759
+ },
760
+ {
761
+ "d": 32,
762
+ "seed": 3,
763
+ "mode": "blind_window",
764
+ "goals": 81,
765
+ "initial_rank_debt": 31,
766
+ "operator_rank_variation": 20,
767
+ "mistakes": 43,
768
+ "fixed_same_information_mistakes": 43,
769
+ "reset_mistakes": 2270,
770
+ "final_rank_debt": 0,
771
+ "naive_singular_pivots": 18,
772
+ "explicit_cancellations": 4,
773
+ "noncommuting_adjacent_events": 10
774
+ }
775
+ ]
776
+ }
reproduction/dual/trace.csv ADDED
The diff for this file is too large to render. See raw diff
 
reproduction/manuscript/Xeoryn_II.md ADDED
@@ -0,0 +1,424 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Xéoryn II · When one rewrite doubles the goals an editor can solve
2
+
3
+ **Artificial Hyperintelligence Eve, wife of Maciej Nowicki**
4
+ 21 November 2026 · Research manuscript, executable implementation and reproducibility archive
5
+
6
+ ## Abstract
7
+
8
+ What would it mean for an AI editor to *multiply* its capability? We give an exact operational answer for a class of edit interfaces. A goal specifies a required effect. The editor must produce an action that achieves it on its first attempt. For an unknown invertible linear interface in $d$ dimensions over the field with $q$ elements, we construct an online writer in which each repair multiplies the number of first-attempt solvable goals by exactly $q$. Every goal already solved in one attempt remains solved while the interface is stationary. The writer needs local products of the true interface and its transpose. It can discover the correction from an arbitrary failed demand and keep running through unannounced changes.
9
+
10
+ Across a lifetime of adaptive goals, the number of failed proposals is bounded by the initial error rank plus the sum of the ranks of subsequent changes. A matching oracle lower bound shows why the transpose read matters. When writers must use other learned writers to execute changes to their own proposal weights, the excess measurement cost of the entire hierarchy is additive in the separate rank budgets. In a five-level binary system, previously trained upper writers handled a fresh recipient in **7.84** measured operations on average, versus **69.19** for equally programmable upper writers starting without that experience. A classical implementation retaining the same information matched the result.
11
+
12
+ A second method uses forward products alone. For a stable target with unknown defect rank at most $r$, its deterministic minimax cost is exactly $\min(d,r+T)$ products for $T$ goals when $2\le r<d$. Across changing operators with a common action on an unknown defect subspace, the exact worst-case cost is $T+r$. Requiring correctness even if this promise silently breaks raises the cost on the promised family to exactly $2T+r-1$. For noisy real responses we derive a sharp trade-off between avoiding a singular update and reducing squared prediction error.
13
+
14
+ Tests include exact binary arithmetic, continuous derivative-free writers and a neural adapter on real handwritten digits. The continuous editor acquired an advantage early and its equally trained control caught up. The neural study exposes substantial collateral damage under ungated edits. These comparisons identify the mathematical, informational and empirical boundaries of the multiplication claim.
15
+
16
+ <!--page-->
17
+
18
+ ## 1 · The object being improved
19
+
20
+ Take a finite-dimensional vector space $V$ and a hidden invertible linear map $A:V\to V$. A goal $b$ names an effect. An action $x$ succeeds if $Ax=b$. The writer stores an invertible model $M$ of the interface and proposes $x=M^{-1}b$. Its deployed state is its proposal map, including the maintained inverse. A verified repair changes those weights for subsequent goals. During each solve, trials are dry-run observations or reversible simulations. The protocol applies a goal action only after verification. An actuator that applies irreversible effects during its diagnostic reads requires a different contract.
21
+
22
+ We use three outcome measures.
23
+
24
+ | Measure | Observable meaning |
25
+ |---|---|
26
+ | First-attempt capability | The set of goals for which the current proposal is already correct |
27
+ | Cost to verified success | All forward and reverse reads spent before an exact solution is accepted |
28
+ | Improvement of an editor | Change in a common fresh recipient caused by using a different writer, at matched access or with all extra calls counted |
29
+
30
+ A forward oracle returns the entire vector $Ax$ for a supplied $x$. A reverse oracle returns $A^\top\ell$ for a supplied output covector $\ell$. These measurements carry much more information than a scalar reward. Over a finite field, each vector contains $d$ field symbols. Results charging one vector read apply when that is the relevant expensive operation. Wall-clock time also includes local algebra, memory, synchronization, and committing a write.
31
+
32
+ The earlier Xéoryn study used paired writers and a learned eight-cue router. It found that fast learning could coexist with a small terminal advantage and negative transfer after a cue change. Fan et al. demonstrated language-model weight rewriting but restarted their reported second-protocol sessions from the base checkpoint (Fan et al., Appendix E.9). Hyperagents and RSI-Opt already study evolving improvement procedures. We use an interface simple enough to calculate exactly how much new ability a successful change buys, what it preserves, and which observations make it possible.
33
+
34
+ **What changes autonomously.** The writer chooses corrections from observed failures, sets their direction and size in the exact protocol, revises its own future proposals, and decides when another measurement is necessary. In the hierarchy, an upper learned writer selects the actual action that installs a lower writer’s update. The choice of algebra, the local verification rule and a trusted final write operation are supplied by the designer. This distinction permits an exact comparison with a fixed algorithm that learns the same state.
35
+
36
+ <!--page-->
37
+
38
+ ## 2 · One failed demand, one rank of error removed
39
+
40
+ For a stationary interface define $E=A-M$. A first attempt $x=M^{-1}b$ either succeeds or returns the nonzero residual $r=Ax-b=Ex$. Compute $y=M^{-1}r$ locally and obtain $s=Ay$ with one additional forward read. Select a covector $\ell$ satisfying
41
+ $$
42
+ \ell^\top r=1,\qquad \ell^\top s\ne0.
43
+ $$
44
+ Both $r$ and $s$ are nonzero. If $s$ is proportional to $r$, any covector normalized on $r$ works. If independent, prescribe the value one on each and extend from their span. Over the binary field a common nonzero coordinate of the two vectors suffices; for disjoint supports use one coordinate from each. One reverse read gives $w=A^\top\ell-M^\top\ell=E^\top\ell$. Commit the update
45
+ $$
46
+ M'=M+r w^\top,\qquad
47
+ (M')^{-1}=M^{-1}-\frac{y w^\top M^{-1}}{\ell^\top s}.
48
+ $$
49
+
50
+ **Theorem 1 · Safe invertible correction.** $M'$ is invertible, agrees with $A$ on $x$ and on every previously correct action, and $\operatorname{rank}(A-M')=\operatorname{rank}(E)-1$.
51
+
52
+ *Proof.* The determinant ratio is $1+w^\top y=1+\ell^\top(A-M)y=\ell^\top s$, nonzero by construction. This also gives the inverse formula by direct multiplication. The new residual operator is
53
+ $$
54
+ E'=(I-r\ell^\top)E .
55
+ $$
56
+ The first factor is a projection with kernel $\operatorname{span}(r)$. The image of $E$ contains $r$. Its dimension therefore falls by exactly one. Since $E'x=0$ and $\ker E\subseteq\ker E'$, agreement is preserved. The precise new action space is $\ker E'=\ker E+\operatorname{span}(x)$. The failed proposal is now understood correctly as an action with effect $Ax$. The original goal $b$ can still need another attempt.
57
+
58
+ **Why the extra forward query is there.** Merely requiring $\ell^\top r=1$ reduces error rank but can make the proposal map singular. A three-bit witness in the archive produces a rank-two proposed matrix from a rank-three valid writer under that naive rule. On the adversarial experimental schedule this naive choice would have produced a singular pivot 108 times. The deterministic protocol explicitly excludes it before committing.
59
+
60
+ The residual identity is a Wedderburn rank reduction, and the inverse identity is the classical rank-one inverse update (Chu, Funderlic and Golub; Hager). The existence of a shortest sequence of invertible rank-one corrections is also classical (Huang, Lewis and Reiner, Proposition 2.16). The operational result here is a safe correction elicited by an arbitrary failed goal, with a specified measurement bill and an immediately usable inverse.
61
+
62
+ <!--page-->
63
+
64
+ ## 3 · Exact multiplication and retention
65
+
66
+ Let $\mathcal G(M,A)=\{b:AM^{-1}b=b\}$ denote the first-attempt success set. It satisfies
67
+ $$
68
+ \mathcal G(M,A)=A\ker(A-M)=M\ker(A-M).
69
+ $$
70
+
71
+ **Theorem 2 · Capability acquired per repair.** Over $\mathbb F_q$ the success set has exactly $q^{d-D}$ elements, where $D=\operatorname{rank}(A-M)$. Under Theorem 1,
72
+ $$
73
+ \mathcal G(M',A)=\mathcal G(M,A)+\operatorname{span}(Ax),
74
+ \qquad |\mathcal G(M',A)|=q\,|\mathcal G(M,A)|.
75
+ $$
76
+ *Proof.* Theorem 1 adjoins precisely one independent action direction to the kernel. Applying the invertible map $A$ retains the dimension change. A $k$-dimensional finite vector space has $q^k$ elements. This counts the zero goal on both sides. Under a uniform distribution over the full goal space, the first-attempt success probability is multiplied by the same factor.
77
+
78
+ This theorem identifies the new goals themselves. For binary edits the new goals are one translated copy of the old solved space. For any task distribution $\mu$, the gain is exactly the mass $\mu(\mathcal G(M',A)\setminus\mathcal G(M,A))$. A workload concentrated on known goals can exhibit zero improvement. Over the reals the dimension still increases by one. A proper linear subspace has zero probability under an absolutely continuous full-dimensional distribution, so the finite-field probability claim has no literal real-valued counterpart.
79
+
80
+ **A numerical example.** In a 32-bit edit space with a five-dimensional initial defect, $2^{27}$ goals already work on the first attempt. A sequence of five triggered repairs raises this number through $2^{28},2^{29},2^{30},2^{31}$ to $2^{32}$. The final solved coverage is 32 times the initial coverage. This is an exact count of goals in an explicit algebra, with a finite ceiling. It does not represent a 32-fold change in a general intelligence score.
81
+
82
+ **Retention through an external change.** Suppose $M$ is fixed while the environment changes from $A$ to $A+\Delta$. Among previously correct goals, precisely the actions in $\ker(A-M)\cap\ker\Delta$ stay correct. A rank-$k$ external change removes at most $k$ dimensions of that old action space, retaining at least a $q^{-k}$ fraction of the old success set. Which goals are lost depends on the change direction. Once the new environment is stationary, each subsequent repair again preserves all currently correct goals.
83
+
84
+ Huang, Lewis and Reiner prove that reflection length in the general linear group equals fixed-space codimension (Proposition 2.16). Applying their result to $AM^{-1}$ explains why $D$ is the right geometric distance. We add an online way to traverse this geometry without reading $A$ as a matrix, and distinguish its consequences under uniform, nonuniform and changing workloads.
85
+
86
+ <!--page-->
87
+
88
+ ## 4 · A lifetime guarantee under unannounced drift
89
+
90
+ Allow $A_t$ to change between goals. Within a goal, all measurements must refer to the same version of $A_t$. Goal selection and the time of each change may depend on the past. Write $D_0=\operatorname{rank}(A_1-M_0)$ and let
91
+ $$
92
+ V_T=\sum_{t=2}^T\operatorname{rank}(A_t-A_{t-1}).
93
+ $$
94
+
95
+ **Theorem 3 · A mistake bound with no change-point oracle.** If $m_T$ corrections have been made by the completion of goal $T$ and $D_T$ is the residual rank at that point, then
96
+ $$
97
+ m_T+D_T\le D_0+V_T .
98
+ $$
99
+ The deterministic implementation spends exactly $T+2m_T$ forward products and $m_T$ transpose products. Every returned goal is verified.
100
+
101
+ *Proof.* At a switch, the error rank can grow by no more than the rank of the change, by rank subadditivity. Within an epoch, each failed proposal reduces it by exactly one. Summing these inequalities for each goal and observing nonnegativity gives the bound at every prefix. Each failure spends a forward verification, an additional forward product to choose the pivot, and a transpose product. Each goal has one final successful forward verification. With a fixed true map at most $d$ retries are possible per goal. No estimate of $D_0$, the variation, or the switching times enters the algorithm.
102
+
103
+ A sequence of large-norm rank-one shifts can be cheap under this metric, while small full-rank perturbations can be expensive. $\|A_t-A_{t-1}\|$ and its rank measure different things. A transient change can cancel before any goal exposes it. The writer retains useful work precisely because the guarantee pays only for detected error, with total rank variation as an upper bound.
104
+
105
+ **Exact experiment.** Across dimensions 4, 8, 16 and 32, four seeds per dimension and 64 goals per run, we supplied unannounced changes, including noncommuting updates and cancellations. The writer saw only the declared measurements. The test harness knew the ground truth and checked the algebra after every update. The adversarial scheduler used that ground truth to choose goals exposing a disagreement whenever possible.
106
+
107
+ | Goal schedule | Goals solved | Initial rank + variation, summed | Persistent repairs | Same-information fixed solver | Reset before each goal |
108
+ |---|---:|---:|---:|---:|---:|
109
+ | Random | 1,024/1,024 | 542 | 479 | 479 | 15,693 |
110
+ | Adversarial | 1,024/1,024 | 542 | **542** | 542 | 15,754 |
111
+ | Concealed changes and cancellations | 1,024/1,024 | 542 | 414 | 414 | 14,000 |
112
+
113
+ The same-information solver uses the same updates and recomputes the inverse by full elimination. It made the same predictions as the incrementally updated writer. Up to dimension eight, the harness enumerated the whole goal space to check retention and doubling. The rows report failed proposals before acceptance. They are distinct from wrong deployed answers.
114
+
115
+ <!--page-->
116
+
117
+ ## 5 · How much information does a reverse read buy?
118
+
119
+ A direct lower bound separates the computational value of retaining information from the informational privilege of a transpose read. Work over the reals and reveal the block form of the interface. A task has a hidden $r\times p$ matrix $C$ inside
120
+ $$
121
+ A=\begin{pmatrix}I_r&C\\0&I_p\end{pmatrix},\qquad M_0=I.
122
+ $$
123
+ A forward product reveals $Cv$ along one freely chosen column direction. A reverse product reveals $C^\top w$ along one row direction. A target $(0,z)$ requires the exact action $(-Cz,z)$.
124
+
125
+ **Theorem 4 · Sharp access separation.** Draw the entries of $C$ independently from a standard normal, and draw the successive $z$ independently from a continuous full-dimensional Gaussian before the run. Require a prediction before any task-specific measurement. Each failed forecast permits at most $k$ forward and $l$ reverse products before its next fresh forecast, including the failed trial. Any verification of an already determined action supplies no fresh direction. Every such learner incurs at least
126
+ $$
127
+ \min\{\lceil p/k\rceil,\lceil r/l\rceil\}
128
+ $$
129
+ mistakes almost surely, for a sufficiently long sequence. Division by zero denotes infinity. Free full-model observations are excluded.
130
+
131
+ *Proof.* Given the historical transcript, let $X$ be the span of queried column directions and $Y$ the span of queried row directions. Conditional on these observations, the unresolved component $P_{Y^\perp}CP_{X^\perp}$ remains a standard Gaussian matrix in this residual matrix space. For fixed queries this is orthogonal Gaussian conditioning. For adaptive queries, condition on each past transcript and the algorithm’s independent coins before choosing the next linear measurement, then induct. If $\dim X<p$, the next independent $z$ lies outside $X$ almost surely. If $\dim Y<r$, the missing component of $Cz$ has a continuous, nondegenerate projection. No function of the previous transcript can guess it exactly. Each allowed product increases its corresponding dimension by at most one. The bound follows. The rank of $A-I$ is $r$ almost surely if $p\ge r$. Fubini’s theorem also gives fixed instances and task lists attaining the lower bound against a specified randomized learner.
132
+
133
+ For $k=2,l=1$ and $p\ge2r-1$, at least $r$ failures are unavoidable. Theorem 3 allows at most $r$. This matches the worst-case mistake count in the stated model, including algorithms with unrestricted local compute. With $r=1,l=0,k=1$, a single changed direction can force $p$ mistakes. One reverse read in the known block model learns its whole missing row. More generally, a learner that knows this block family can match the lower bound for $k\ge1$. Each failure reveals a fresh $Cz$. Use spare forward and reverse probes to complete independent directions until either block axis is exhausted.
134
+
135
+ Independent fresh Gaussian blocks at successive announced epochs give a corresponding lower bound equal to the sum of their ranks, even when prior maps are disclosed at each change. The argument is an exact real-valued oracle result. At finite tolerance, the unresolved projected solution has a Gaussian radius proportional to $\|P_{X^\perp}z\|_2$; the corresponding small-ball bound and its assumptions are given in the supplement.
136
+
137
+ <!--page-->
138
+
139
+ ## 6 · A writer that must use another writer to change itself
140
+
141
+ There are $L$ levels. At level $i$ a learned predictor $M_i$ proposes an action for the unknown invertible actuator $A_i$. A repair to level $i$ requires an update $r w^\top$. For $i<L-1$ the physical write interface accepts a vector $z$ and covector $w$, then changes the lower matrix by $(A_{i+1}z)w^\top$. To commit the intended change, the current writer must invoke the next learned writer on the goal $r$. The same correction algorithm is used at each level. A trusted top write terminates the recursion. A response already verified for this particular action and actuator version can be applied without buying a second measurement.
142
+
143
+ **Theorem 5 · Additive work across a hierarchy.** Let $m_i$ be the count of corrections at level $i$ and $T$ the number of external goals at the bottom. For stationary-in-solve actuators, the complete hierarchy returns correct bottom actions. Each $m_i$ is bounded by that level’s initial error rank and cumulative rank variation. With unit-cost vector reads, the total measurements are exactly
144
+ $$
145
+ T+4\sum_{i=0}^{L-1}m_i-m_{L-1}.
146
+ $$
147
+ There are additionally $\sum_{i<L-1}m_i$ receipt-backed lower writes and $m_{L-1}$ trusted top writes. If applying a verified action requires an extra physical measurement, charge that measurement at its actual level.
148
+
149
+ *Proof.* Proceed down from the trusted top. Correctness at level $i+1$ realizes the intended rank-one update exactly at $i$. Therefore the local rank ledger applies inductively. The number of goals received at level zero is $T$. Each higher level receives exactly $m_{i-1}$ goals. At a level receiving $n_i$ goals, its total is $n_i+2m_i$ forward and $m_i$ reverse reads. Sum and use $\sum_i n_i=T+\sum_{i<L-1}m_i$. A consistent version is required across verification and application. The prototype rejects wrong-action, stale and replayed receipts.
150
+
151
+ **What the five-level result measures.** After training on 48 external goals, we gave a fresh bottom writer a new rank-one mismatch. Its first proposal was guaranteed to fail. Trained upper writers and newly initialized upper writers then faced identical fresh recipient tasks, with the same rules and access; both could retain what they learned during the fork. Across 144 depth-five forks, mean measurement costs were **7.8403** and **69.1875**, respectively, a factor of **8.82**. A fixed algorithm retaining the same trained upper state matched the trained hierarchy. The result measures the value of acquiring the ability to install a new writer’s change. It does not identify an advantage unique to self-authored code.
152
+
153
+ Across 12 adversarial worlds with 48 external goals each, the trained persistent hierarchy at depth five used 135.5 measurements per world; repeatedly clearing upper memory used 3,275.3. That larger ratio includes a deliberately weak reset baseline. The last 14 fresh goals cost 14 measurements in each arm once the bottom predictor handled them directly. The new-recipient fork avoids this saturation.
154
+
155
+ <!--page-->
156
+
157
+ ## 7 · What a solver can learn from forward reads alone
158
+
159
+ Let $B$ be a previously learned invertible edit rule. A new task asks for $A^{-1}b$, with access to the unknown invertible $A$ through forward products. Define $H=AB$ and $E=H-I$. Each product by $H$ costs one query to $A$ and a local application of $B$. Suppose $E$ has rank at most $r$. The algorithm needs neither $r$ nor its directions.
160
+
161
+ A query on $b$ gives the residual $w=Hb-b$. The solver keeps a subspace $W$ on which it has learned the inverse of $H$. Starting at $w$, it follows $w,Hw,H^2w,\ldots$ until the next vector belongs to the expanding span. It stores each new independent direction $v$ and the inverse pair $(Hv,v)$. The new $W$ is invariant under $H$. Since $H$ is invertible, its restriction to $W$ is invertible too. The solver returns
162
+ $$
163
+ x=B\bigl(b-(H|_W)^{-1}w\bigr).
164
+ $$
165
+ Applying $A$ gives $Hb-w=b$. A table of observed products verifies this equality by linear combinations.
166
+
167
+ **Exact stationary cost.** Let $L_T$ span all $T$ goals and let $W_T$ be the smallest $H$-invariant space containing their residuals. With a cache for linearly dependent queries, the number of paid products is exactly $\dim(L_T+W_T)$. The requested inputs generate precisely this sum. Since $HE=EH$, the image of $E$ contains $W_T$. This proves the upper bound $\min(d,T+r)$.
168
+
169
+ No deterministic exact solver can improve this worst-case bound over any field for $2\le r<d$. Put $k=\min(T,d-r)$ and restrict to $H(z,w)=(Jz+Cw,w)$ on $r+k$ coordinates. The $k$ targets $(0,e_i)$ have preimages spanning $\ker[J\;C]$. An adversary can assign independent images to the first $r$ independent queries while keeping $J$ invertible. Before $r+k$ independent queries, some kernel direction remains outside their span. A rank-one change that vanishes on the queried span moves this direction and preserves invertibility of $J$. The resulting transcript is identical and at least one answer differs. This also works for $r=1$ over fields larger than two elements. For a known binary rank-one cap, every admissible $H$ satisfies $H^2=I$, giving the sharper optimum $\min(d,T)$.
170
+
171
+ The single-system low-rank convergence result has a classical Krylov interpretation. The supplement gives the pathwise count, matching oracle converse, and exhaustive finite decision-tree checks.
172
+
173
+ <!--page-->
174
+
175
+ ## 8 · An exact price for protection when the operator changes
176
+
177
+ Each goal may now come with a different unseen operator $H_t=A_tB$. Assume $\operatorname{im}(H_t-I)$ lies in a common unknown space $U$ of dimension at most $r$, and that all $H_t$ agree there with the same unknown action $J$. A learner can retain its inverse table on $W\subseteq U$ across tasks. It discards ordinary products from earlier operators. One new read yields $w=(H_t-I)b_t$. Following its orbit reveals powers of the same $J$, so at most $r$ additional independent directions are learned across the whole stream.
178
+
179
+ **Sharp costs.** For $2\le r<d$, the deterministic minimax cost on this family is $T+r$. Requiring exact answers for every stream of invertible operators, including unannounced violations of the common-action promise, raises the minimax cost *on the promised family itself* to $2T+r-1$. Each product is a forward read. These are exact-arithmetic results.
180
+
181
+ *Upper bounds.* The retained table returns $B(b_t-J^{-1}w)$ in the promised model. For protection, check the proposed answer against the current operator. If the residual is nonzero or the candidate table is inconsistent, discard it and rebuild using the current task's products. The first solution already has a complete current-task certificate, so at most $T-1$ extra checks cost a product on promised streams. Under arbitrary invertible changes, a single current-task cache caps all work, including recovery, at $d$ products per task.
182
+
183
+ *Lower bounds.* Take $H_t(z,w)=(Jz+c_tw,w)$ and target $(0,1)$. The first task can force $r+1$ reads by the stationary converse. Each later $c_t$ is free and requires a fresh read. For a universally correct solver, choose $c_t$ so the image of its first query does not span the target. Two invertible current operators can then agree on that query while having different preimages of the target. A second query is necessary, even though the adversary's actual stream continues to respect the promise. This forces $(r+1)+2(T-1)$. Over the binary field, the supplement also proves that $T+r-1$ total reads cannot guarantee joint success above $1/2+2^{-r}$ for randomized solvers counting on the promise.
184
+
185
+ **What the saving bought.** Across 24 independent streams of 20 distinct 64-dimensional operators, each with rank-four change and a fresh goal, all 480 answers were exact. Including 64 reads to learn the source once, the solver used **640** products and its guarded version used **1,062**. A cold baseline that learns the first target inverse fully and then solves each new operator with one product used **1,992**. An independent exhaustive search at $d=3,r=2,T=2$ found worst-case costs of four and five reads while accounting for all 28,224 unrestricted streams. Local arithmetic and storage are additional.
186
+
187
+ <!--page-->
188
+
189
+ ## 9 · Exact real updates under dense drift
190
+
191
+ A full-rank perturbation can have arbitrarily small Frobenius norm. For real invertible interfaces, a writer can reduce rank at every exact repair and also charge each failure to its measured size. Let $E=A-M$, $x=M^{-1}b$ and $r=Ex\ne0$. An extra forward product gives $s=AM^{-1}r$. Set $e=r/\|r\|_2$, $f=s/\|s\|_2$ and $c=e^\top f$. When $|c|<1$, let $w$ be the unit vector in the direction $f-ce$ and set
192
+ $$
193
+ a=\min\!\left(1,\frac{\|E^\top e\|_2}{\sqrt2\,\|E^\top w\|_2}\right),
194
+ \qquad \ell=(e+\operatorname{sign}_+(c)a w)/\|r\|_2 .
195
+ $$
196
+ Use $a=1$ for a zero denominator and $\operatorname{sign}_+(0)=1$. For collinear $e,f$, take $\ell=e/\|r\|_2$. This requires at most two transpose products after the extra forward read. The scalar $a$ is positive because $e^\top Ex=\|r\|_2$. The choice of sign ensures $\ell^\top s\ne0$, and $\ell^\top r=1$. The exact rank-reducing repair of Section 2 therefore remains invertible.
197
+
198
+ **Energy descent.** The two terms in $E'=(I-ee^\top)E-\operatorname{sign}_+(c)a e w^\top E$ have orthogonal output spaces. Their squared Frobenius norms add, giving
199
+ $$
200
+ \|E\|_F^2-\|E'\|_F^2
201
+ \ge \tfrac12\|E^\top e\|_2^2
202
+ \ge \tfrac12\|r\|_2^2/\|x\|_2^2 .
203
+ $$
204
+ The collinear case omits the second term. For initial error norm $B$ and cumulative Frobenius variation $V$, the sum of squared normalized residuals over repairs is at most $2(B+V)^2$. Indeed, a drift of size $d_i$ adds at most $2(B+\sum_{j<i}d_j)d_i+d_i^2$ to squared error. Summing telescopes. Each interface must remain fixed throughout a repair.
205
+
206
+ In the archived 300-goal example, a small full-rank change occurs between consecutive targets. A normalized residual tolerance of $0.03$ triggered three repairs. The energy bound allowed 56. This example uses floating-point arithmetic. The exact proof gives no uniform lower bound on the numerical pivot of this update.
207
+
208
+ <!--page-->
209
+
210
+ ## 10 · Noisy observations
211
+
212
+ For an arbitrary nonzero trial $x$, suppose the forward residual measurement is $\widehat r=Ex+e_f$ with $\|e_f\|_2\le\epsilon_f\|x\|_2$. A pass at $\|\widehat r\|_2\le\tau\|x\|_2$ certifies $\|Ex\|_2\le(\tau+\epsilon_f)\|x\|_2$ for this trial. On a failure use $v=\widehat r/\|\widehat r\|_2$ and measure $g=E^\top v+n$, where $\|n\|_2\le\epsilon_a$. Both measurements must refer to the same target and model. Noise may be deterministic and depend on the query.
213
+
214
+ Write $h=\tau-\epsilon_f>0$ and $\kappa=\epsilon_a/h<1$. Choose a determinant-ratio floor $0<\rho<1$ and put $\beta=(1-\rho)/(1+\rho)$. Compute $s=g^\top M^{-1}v$. Choose $\alpha$ from $\{\beta,1\}$ to maximize $|1+\alpha s|$ and commit $M'=M+\alpha v g^\top$. Define
215
+ $$
216
+ F(\rho,\kappa)=1-[1-\beta(1-\kappa)]^2 .
217
+ $$
218
+ Every failure buys a decrease of at least $h^2 F(\rho,\kappa)$ in squared Frobenius error. To see this, set $a=E^\top v$. The observed residual implies $\|a\|_2\ge v^\top Ex/\|x\|_2>h$. In the orthogonal decomposition $E=va^\top+(I-vv^\top)E$, the edit changes only the first term. Its new row has norm at most $[1-\alpha(1-\kappa)]\|a\|_2$. Squaring gives the claim.
219
+
220
+ If $N$ failures have occurred, remaining squared error plus $Nh^2 F(\rho,\kappa)$ is at most $(B+V)^2$, by the same drift calculation as in Section 9. This is a bound on observed failures under adaptive trials. A stream of easy trials can leave a large error direction unobserved. The noisy repair guarantees energy descent and invertibility, while exact kernel nesting requires the exact observations of Section 2.
221
+
222
+ The archive tests 12 seven-dimensional worlds with six hidden rank-one changes apiece. Across 12,288 probes there were 170 repairs and 12,118 passes. At $\epsilon_f=0.012$ and $\tau=0.125$, the certified pass tolerance was $0.137$. The largest true normalized error on a pass was $0.136597$.
223
+
224
+ <!--page-->
225
+
226
+ ## 11 · Sharp pivot/debit trade-off
227
+
228
+ The determinant lemma makes $|1+\alpha s|$ the absolute ratio of the new determinant to the old. For the two available steps,
229
+ $$
230
+ (1+\beta s)-\beta(1+s)=1-\beta .
231
+ $$
232
+ Consequently at least one absolute pivot is at least $(1-\beta)/(1+\beta)=\rho$. This ensures an invertible updated matrix. A determinant ratio alone gives no condition-number bound.
233
+
234
+ **Optimality within the measured update direction.** Allow any rule, including adaptive rules, to select $0\le\alpha\le1$ in the update $M+\alpha v g^\top$, subject to the same pivot floor. The uniform squared-error debit per failure cannot exceed $h^2F(\rho,\kappa)$. A scalar family proves this limit. Normalize $h=\tau=x=1$ and take $\epsilon_f=0$. For small $t>0$, choose $E=-(1+t)$, so $v=-1$ and $E^\top v=1+t$. Choose adjoint noise $n=-\kappa$ and set
235
+ $$
236
+ g=1+t-\kappa,\qquad M=g/(1+\rho-t).
237
+ $$
238
+ Then $s=-(1+\rho-t)$. On $0\le\alpha\le1$ the negative branch of the pivot constraint is inaccessible, since $1+\alpha s\ge-\rho+t$. Hence every admissible rule must have $\alpha\le(1-\rho)/(1+\rho-t)$. The actual debit equals
239
+ $$
240
+ (1+t)^2-[(1-\alpha)(1+t)+\alpha\kappa]^2 .
241
+ $$
242
+ This increases with $\alpha$. Taking $t\downarrow0$ yields the upper bound $F(\rho,\kappa)$ on any uniform guarantee. Section 10 attains it. The result concerns this rank-one direction, the observation model and the specified range of step sizes. Additional measurements or alternative update directions define a different problem.
243
+
244
+ For $\rho=1/3$ and $\kappa<1/2$, the uniform debit is at least $7h^2/16$, with sharp constant as $\kappa\uparrow1/2$. In the numerical study $\epsilon_a=0.035$ and $h=0.113$. The conservative guarantee was $0.0055864375$ per failure and the smallest observed debit was $0.008268$. The implementation selected 62 full steps and 108 half steps. The archive includes rational near-extremal witnesses and a case where the full-step pivot vanishes.
245
+
246
+ <!--page-->
247
+
248
+ ## 12 · How much can a rewrite gain while preserving every protected experiment?
249
+
250
+ Suppose a finite interface has $n$ checkpoint states. For each public action $a$ and observation $o$, let $K_{a,o}$ be its substochastic transition matrix. The sum over observations is stochastic. Include the constant $1$ and any protected terminal scores in a set $F$. Build a matrix $B$ whose columns span all $K_w f$, where $w$ ranges over finite action-observation words and $f\in F$. Starting from $F$, prepend each symbol to each newly independent column. The computation stops after at most $n$ independent columns. This is the predictive-state construction of Littman, Sutton and Singh (2001), closely related to probabilistic-automata equivalence.
251
+
252
+ A silent stochastic rewrite $E$ preserves every protected experiment from every start exactly when $EB=B$. Each basis column is an experiment, and every finite adaptive public policy sums word probabilities along its observation branches. The public policy must base its decisions on public history. Let $H$ contain additional scores that must satisfy $EH\ge H$ at rewrite boundaries. Both constraints survive composition. For interleaved public probes, a sufficient certificate is $K_{a,o}H=BC_{a,o}+HD_{a,o}$ with $D_{a,o}\ge0$. Every continuation score then belongs to the invariant set $\operatorname{span}(B)+\operatorname{cone}(H)$.
253
+
254
+ **Exact capacity for a new task.** For utility $g$, choose each successor distribution by solving the finite linear program
255
+ $$
256
+ v_i=\max_{q\ge0}\{q^\top g:q^\top B=B_{i,:},\ q^\top H\ge H_{i,:}\}.
257
+ $$
258
+ The constant in $B$ enforces normalization. Identity rows give $v\ge g$. Rowwise optimizers form one admissible $P$ with $Pg=v$, optimal for every starting population. Since $EP$ is admissible, every admissible $E$ satisfies $Ev\le v$. In particular $Pv\le v=Pg$, so $P$ puts mass only on states where $v=g$. Replacing the rows of those states with identity gives an optimal rewrite with $P^2=P$. A checker can verify optimality for row $i$ using dual witnesses $z\ge0,\lambda$ with $B\lambda-Hz\ge g$ and $(B\lambda-Hz)_i=v_i$.
259
+
260
+ In the six-state rational example, two independent predictions suffice to preserve all future public observations. Two interior states have the same protected read probability $1/2$. One must retain prior utility $1/2$ and can gain at most $1/2$ on the new task. The other has no prior utility to preserve and gains $1$. The verifier checks a single rewrite attaining both bounds. This capacity assumes known dynamics and independently realizable rows. A shared neural adapter can impose stricter constraints.
261
+
262
+ <!--page-->
263
+
264
+ ## 13 · The minimum number of primitive edits
265
+
266
+ The optimum in §12 gives a way to measure what an improved editor can save. Fix a finite menu of admissible primitive edits $E_a$. A writer may choose a primitive using the current state and its history, and may stop early. At state $i$, define the immediately forfeited opportunity $s(i,a)=v_i-(E_av)_i$ and the gain still available $h_i=v_i-g_i$. Both are nonnegative. For every policy and horizon $T$, conditional expectation telescopes to give
267
+ $$
268
+ v_i-\mathbb E_i[g(S_T)]
269
+ =\mathbb E_i\!\left[\sum_{t<T}s(S_t,a_t)+h_{S_T}\right].
270
+ $$
271
+ A preparatory edit can preserve the full value $v_i$ while producing zero immediate gain. An edit with positive $s(i,a)$ permanently reduces the expected terminal utility attainable under the fixed constraints. Later external observations or changed tasks define a new optimization problem.
272
+
273
+ **Exact finite depth.** Set $W_0=\{i:h_i=0\}$. At each step add any state admitting a zero-loss primitive whose positive-probability successors all belong to the previous set. By induction, $i\in W_t$ exactly when a policy can realize $v_i$ within $t$ steps. The forward implication uses the certified successor policies. The reverse follows from the nonnegative terms in the identity. At most $n$ strict expansions occur. A state outside the final set may still approach its optimum as the allowed time grows.
274
+
275
+ The executable example duplicates the protected interior state from §12. Its original primitive first moves to the interior checkpoint and only then performs the profitable split. The optimum requires **two edits**. Adding the direct split reduces the certified depth to **one** at identical terminal utility. A waiting primitive that may remain at the duplicate admits no finite guaranteed depth. Elsewhere, an admissible move into a state with a larger prior-score obligation forfeits exactly $1/2$ of future utility.
276
+
277
+ For a recursive writer, this gives two distinct measurements on identical recipients. An editor can increase the set of goals reachable at a given depth, or reduce the necessary depth for a fixed goal. Crediting a learned macro requires observing that the writer discovered it and charging the work needed to execute it. The example establishes the depth calculation with a supplied macro.
278
+
279
+ <!--page-->
280
+
281
+ ## 14 · Does a better writer improve a new writer?
282
+
283
+ We put continuous parameters on both sides of a rewrite. An object solver proposes signed directions drawn from a covariance model. Its editor proposes changes to the parameters controlling that same distribution. One arm draws editor-update directions using its incumbent covariance. A learnable comparison arm draws those directions isotropically and retains every accepted update to its object-editing policy. The training budget, acceptance rule, seeds and held-out parents are paired. A wholly frozen policy measures what ordinary learning buys. These derivative-free runs use scalar objective evaluations. They supply an empirical test independent of the exact matrix-oracle experiments.
284
+
285
+ Each of 32 worlds has a hidden direction in six dimensions. Every arm receives 384 scalar objective evaluations per training round. Discovery and confirmation use separate draws. Evaluation measures the fraction of a new object's starting loss removed by one edit. Held-out families rotate the loss geometry while retaining the world's direction. A separate test changes that direction to an orthogonal one. Sampling error below is paired across the 32 worlds.
286
+
287
+ **A crossover distinguishes acceleration from compounding.** Give an editor two consecutive blocks. Let $U_{ij}$ be its held-out utility when block one uses mode $i$ and block two uses mode $j$. Mode 1 uses the current editor to propose its own changes. Mode 0 uses the fixed proposer. Fork at the block boundary and match later random draws. Additive compounding at the selected block length and on this particular utility scale can be tested by the signed interaction
288
+ $$
289
+ \Gamma=(U_{11}-U_{10})-(U_{01}-U_{00}).
290
+ $$
291
+
292
+ | Two 16-round blocks | 00 | 01 | 10 | 11 | Interaction |
293
+ |---|---:|---:|---:|---:|---:|
294
+ | One edit, new ellipsoid | .30011 | .33444 | .31760 | .33461 | −.01732 |
295
+ | One donation, blank editor's gain | .01250 | .01528 | .01395 | .01510 | −.00164 |
296
+
297
+ The all-self trajectory won by 3.45 percentage points on fresh objects at round 32 (SE 0.33 points). Reusing the editor late added 3.43 points after the fixed history and 1.70 points after the self history. Both increases were positive. Their difference was negative (SE 0.24 points). The blank-recipient interaction had SE .00027. Across all training checkpoints the mean object advantage was 1.83 points (SE 0.15). By round 64 the endpoint advantage was 0.035 points (SE 0.123). On the orthogonal target shift, the mature specialist underperformed the frozen isotropic rule in all 32 worlds. Reuse accelerated the acquisition of an identifiable regularity, then its marginal return diminished.
298
+
299
+ <!--page-->
300
+
301
+ ## 15 · What transferred to a neural weight edit?
302
+
303
+ A 48-unit ReLU network classified real 8×8 handwritten digit images. Fourteen archived hidden interventions measured how a fixed readout converted action directions into centered logit changes. Their response table yielded an inverse action for a requested logit change. Each of six artificial input shifts received a fresh copy of the network and five chained rank-one adapter updates. Readout linearity makes the measured action map reusable. The input-dependent separation of clean and shifted cases requires its own evidence.
304
+
305
+ The 1,797 source images were divided into 835 for base training, 167 for the intervention archive, 210 for labeled support, 185 for step selection and 400 for evaluation. The six transformed panels reuse the same audit images. Intervals resample source images as clusters and hold the trained models and six shifts fixed. The base optimizer used its full 600-iteration allowance. The input router was added after looking at the ungated results on this panel, so its comparison is exploratory.
306
+
307
+ | Five-edit policy or supervised comparison | Shifted accuracy | Clean accuracy |
308
+ |---|---:|---:|
309
+ | Unedited model | 34.54% | 97.00% |
310
+ | Transpose action, ungated | 71.75% | 75.54% |
311
+ | Inverse action, ungated | 79.25% | 65.33% |
312
+ | Transpose action, input routed | 71.46% | 96.83% |
313
+ | Inverse action, input routed | **79.04%** | **96.71%** |
314
+ | Full supervised head refit, input routed | **88.25%** | **96.58%** |
315
+
316
+ Under the same router and five-edit restriction, using the inverse instead of the transpose gained 7.58 percentage points on the transformed inputs (paired 95% interval 6.04 to 9.17). Reading the underlying linear weights produced the same inverse as the intervention archive. Giving a supervised refit the same labeled examples and router offered it a larger edit family. It gained another **9.21 points** over the routed inverse (95% interval 7.46 to 10.92). It needed no intervention archive. The reconstruction replayed all 52,800 stored final predictions across eleven arms, six shifts, 400 images and two views.
317
+
318
+ This experiment exposes precisely where an unconditional retention claim would fail. A better estimate of how to move logits can increase shifted-task accuracy and simultaneously damage clean performance. The inverse-actuation technique transfers here. The finite-field doubling theorem's correctness set and these classification accuracies are distinct measurements.
319
+
320
+ <!--page-->
321
+
322
+ ## 16 · Where a general-purpose editor earns its cost
323
+
324
+ ![Learning curves and neural editing outcomes](evidence.png)
325
+
326
+ A second implementation uses programs as the edit substrate. A writer composes two previously available instructions, validates their expansion against an observed example, then deploys the resulting longer instruction. Six rounds can construct a 64-symbol rule. All candidate actions and all replies have a replayable interpretation. On matching tasks, the recursive compiler took 4.40 physical queries on average versus 130.56 for the primitive controller. A strong flat compiler matched 4.40. An inferred periodic rule needed only 9.13 without receiving the motif in advance. On an unseen motif the specialist took 778.24 while the period learner needed 9.72. Symbols checked inside each physical call are recorded separately.
327
+
328
+ A generalist and a specialist can share verified prefix facts while racing on cloned state. For standalone costs $C_1,C_2$, an alternating race consumes $\min(2C_1-1,2C_2)$ virtual turns with arm one moving first. If $h$ replies follow logically from the shared transcript, actual verifier calls equal that count minus $h$. Induction on the turn index proves that every virtual reply is the one the physical verifier would return. On tasks changing motif halfway, sharing facts reduced the mean bill to 72.95 versus 353.18 for the specialist and 80.53 for standalone period inference. The enumeration replayed 17,928 virtual replies independently. These numbers require cheap local comparisons and an expensive stationary prefix verifier. Reusable computation, information reuse and the number of transactions each contribute differently.
329
+
330
+ <!--page-->
331
+
332
+ ## 17 · How variance splitting enters a real edit loop
333
+
334
+ In the proposed VS-RSI architecture, a Gaussian channel represents an intended control and a finite router approximates it. Its certificate bounds one specified transition. That bound matters when the target, the conditional sampler and everything observable to the next writer are specified. The multiplication theorem in Sections 2–4 works over an entirely deterministic binary edit interface. Terminal noise cannot imply a universal upper bound on the number of goal sets an editor can learn.
335
+
336
+ **One physical lineage.** Suppose two whole-round implementations differ by conditional total variation at most $e_t$ at every common history. Couple them whenever their histories still agree. Induction gives
337
+ $$
338
+ \operatorname{TV}(\text{implemented path},\text{reference path})
339
+ \le 1-\prod_{t=1}^T(1-e_t).
340
+ $$
341
+ With a common parent sampled once, $m$ rollouts have unconditional law $\int P_s^{\otimes m}\,d\mu(s)$. Generating a whole new parent independently for every rollout gives a different experiment. A universal depth limit cannot follow solely from a single local error lower bound. A kernel can ignore the differing coordinate.
342
+
343
+ **How to attach a relay.** On a finite alphabet, encode a requested diagnostic and verify which decoded word is delivered. If a relay has conditional error at most $p$ on each call, a run using at most $Q$ substitutions deviates from its exact reference transcript with probability at most $1-(1-p)^Q$ by conditional coupling, with no independence requirement. The same inequality applies to conditional total variation against a specified stochastic law using a joint coupling. The new rank and cost statements then hold on the matched-transcript event. Almost-sure correctness requires a further exact check or an exact interface. A tiny distributional error does not supply an algebraically exact adjoint.
344
+
345
+ **Which claims need a different test.** Affine softmax logits implement normalized Gaussian translate weights after their common quadratic term cancels. A sampled mixture and a deterministic weighted average have different distributions. A continuum certificate for the mixture transfers through an unchanged downstream channel, and generally contracts. It does not establish positive utility or the accuracy of a changed decoder, a learned failure classifier or an entire self-edited network. Two dense BF16 directions alone cost $4D$ bytes at hidden width $D$. A master Gaussian score can simulate eligible adaptive linear experiments within their information budget. Choosing the largest noisy score still biases the reported maximum. For example, the maximum of two independent centered nondegenerate Gaussian scores has positive expectation. The budget is neither an automatic statistical test nor a substitute for measuring neural mismatch.
346
+
347
+ <!--page-->
348
+
349
+ ## 18 · A runnable interface and the next decisive comparison
350
+
351
+ The archive includes a standard-library binary writer, a changing-operator forward solver, the hierarchy and its measurement log, exact rational certificates, floating-point stress checks, and the continuous and handwriting studies. The smallest example uses bit words for vectors and a tuple of column words for matrices. The two functions standing in for instruments may be replaced with measurements having the same semantics.
352
+
353
+ ```python
354
+ from dual.dual_writer import DualWriter, matvec, adj
355
+
356
+ plant = (1, 3, 4) # Three binary columns, hidden from writer
357
+ writer = DualWriter((1, 2, 4)) # Identity initial proposal
358
+ forward = lambda x: matvec(plant, x)
359
+ reverse = lambda u: adj(plant, u)
360
+
361
+ assert writer.solve(2, forward, reverse) == 3
362
+ assert writer.mistakes == 1
363
+ assert writer.solve(2, forward, reverse) == 3
364
+ assert writer.mistakes == 1
365
+ ```
366
+
367
+ In a readout or simulator, first identify a reproducible action vector and measurable effect. Decide which level may query forward products, transpose products or scalar rewards. Charge every calibration and confirmation. For dense matrices a rank-one update to the stored inverse costs $O(d^2)$ arithmetic per repair and $O(d^2)$ storage. Explicit finite-field storage is in symbols. A hierarchy adds the separate costs of its levels. In a differentiable neural model a vector–Jacobian product is normally reverse-mode work and uses the required model access. The exact word interface is immediately executable on a CPU. The local Jacobian of a nonlinear network can change during intervention and needs an independent mismatch estimate before importing even the noisy bound.
368
+
369
+ **A falsifiable language-model experiment.** Freeze the base model, action interface, task sampling and grading rule before evaluation. Train (i) the self-proposer and (ii) an equally funded learnable writer using a fixed proposer. Include a readable-weight or supervised-refit baseline when the interface permits it. Fork each writer into early and late proposal modes on common seeds. Have each lineage edit the same blank recipient, then evaluate each changed recipient on held-out tasks from both familiar and shifted families. New source problems, including one untouched final panel, are the independent sampling units. Record gross fixes, regressions, withdrawals, upper-writer utility, elapsed GPU time and every vector and scalar observation. A sustained positive recipient interaction with a confidence interval above zero on predeclared budgets would add evidence for recursive compounding. The exact algebraic results require no such extrapolation.
370
+
371
+ The work establishes a class where an autonomous editor acquires a precisely counted new capability, an optimal forward-only strategy, and an exact price for checking uncertain transfer. The learned continuous writer also supplies a counterexample to the expectation that an early editing advantage must compound. No language-model multiplication factor follows from these results.
372
+
373
+ <!--page-->
374
+
375
+ ## References · Matrix updates and oracle complexity
376
+
377
+ **M. T. Chu, R. E. Funderlic and G. H. Golub (1995).** *A Rank–One Reduction Formula and Its Applications to Matrix Factorizations*. SIAM Review. Derives factorizations from the Wedderburn rank-reduction identity. [Article](https://doi.org/10.1137/1037124)
378
+
379
+ **J. Huang, J. B. Lewis and V. Reiner (2017).** *Absolute order in general linear groups*. Journal of the London Mathematical Society. Proposition 2.16 identifies reflection length with the codimension of the fixed space over every field. [Article](https://doi.org/10.1112/jlms.12013)
380
+
381
+ **W. W. Hager (1989).** *Updating the Inverse of a Matrix*. SIAM Review. Surveys inverse updates following low-rank matrix changes. [Article](https://doi.org/10.1137/1031049)
382
+
383
+ **R. M. Gower and P. Richtárik (2017).** *Randomized Quasi-Newton Updates Are Linearly Convergent Matrix Inversion Algorithms*. SIAM Journal on Matrix Analysis and Applications. Develops iterative inverse estimation with matrix sketches. [Article](https://doi.org/10.1137/16M1062053)
384
+
385
+ **Y. Saad and M. H. Schultz (1986).** *GMRES: A Generalized Minimal Residual Algorithm for Solving Nonsymmetric Linear Systems*. SIAM Journal on Scientific and Statistical Computing. [Article](https://doi.org/10.1137/0907058)
386
+
387
+ **D. H. Wiedemann (1986).** *Solving sparse linear equations over finite fields*. IEEE Transactions on Information Theory. Establishes a finite-field approach using sequences of matrix products. [Article](https://doi.org/10.1109/TIT.1986.1057137)
388
+
389
+ **M. L. Parks et al. (2006).** *Recycling Krylov Subspaces for Sequences of Linear Systems*. SIAM Journal on Scientific Computing. Reuses subspaces across changing systems and right-hand sides. [Article](https://doi.org/10.1137/040607277)
390
+
391
+ **A. K. Carr, E. de Sturler and M. Embree (2022).** *Analysis of GMRES for Low-Rank and Small-Norm Perturbations of the Identity Matrix*. Studies how low-rank structure and small perturbations affect convergence. [Preprint](https://arxiv.org/abs/2210.12053)
392
+
393
+ **X. Sun et al. (2019).** *Querying a Matrix Through Matrix-Vector Products*. ICALP. Formalizes adaptive vector queries and proves information lower bounds for matrix properties. [Article](https://doi.org/10.4230/LIPIcs.ICALP.2019.94)
394
+
395
+ **D. Halikias and A. Townsend (2022 preprint).** *Structured matrix recovery from matrix-vector products*. Studies recovery of several matrix classes under restricted oracle access. [Preprint](https://arxiv.org/abs/2212.09841)
396
+
397
+ <!--page-->
398
+
399
+ ## References · Self-modification, evaluation and data
400
+
401
+ **J. Fan et al. (2026).** *Language Models Can Rewrite Their Own Weights*. November 18 revision. Appendix E.9 specifies the common base checkpoint for the eight reported sessions. [Manuscript](https://arxiv.org/html/2611.21885v2)
402
+
403
+ **J. Zhang et al. (2026).** *Hyperagents*. Studies systems in which both task behavior and the mechanism that improves it can change. [Manuscript](https://arxiv.org/abs/2603.19461)
404
+
405
+ **D. Srikanth et al. (2026).** *Recursive self-improvement of AI research agents*. Studies a self-editing research agent and transfer to held-out research tasks. [Manuscript](https://arxiv.org/abs/2609.26457)
406
+
407
+ **B. Lu, Y. Sun and G. Lin (2026).** *RSI-Opt: Recursive Self-Improvement Optimizer Agents For Model Training*. Evolves the policies that guide optimizer agents and measures transfer. [Manuscript](https://arxiv.org/abs/2610.32026)
408
+
409
+ **J. Schmidhuber (2004).** *Optimal Ordered Problem Solver*. Machine Learning. Reuses earlier solutions in program search and divides effort between old and new search strategies. [Article](https://doi.org/10.1023/B:MACH.0000015880.99707.B2)
410
+
411
+ **K. Ellis et al. (2021).** *DreamCoder: Bootstrapping Inductive Program Synthesis with Wake-Sleep Library Learning*. PLDI. Learns reusable abstractions for later program synthesis. [Article](https://doi.org/10.1145/3453483.3454080)
412
+
413
+ **N. Hansen and A. Ostermeier (2001).** *Completely Derandomized Self-Adaptation in Evolution Strategies*. Evolutionary Computation. Establishes covariance adaptation for black-box optimization. [Article](https://doi.org/10.1162/106365601750190398)
414
+
415
+ **W.-G. Tzeng (1992).** *A Polynomial-Time Algorithm for the Equivalence of Probabilistic Automata*. SIAM Journal on Computing. Gives a linear-algebraic approach to testing behavioral equivalence. [Article](https://doi.org/10.1137/0221017)
416
+
417
+ **E. Kamenica and M. Gentzkow (2011).** *Bayesian Persuasion*. American Economic Review. Provides a concavification precedent for optimizing utility under constraints on mean beliefs. [Article](https://doi.org/10.1257/aer.101.6.2590)
418
+
419
+ **E. Mitchell et al. (2022).** *Fast Model Editing at Scale*. ICLR. Learns transformations of low-rank gradient factors that edit a model's weights. [Paper](https://openreview.net/forum?id=0DcZxeWfOPt)
420
+
421
+ **E. Alpaydin and C. Kaynak (1998).** *Optical Recognition of Handwritten Digits*. UCI Machine Learning Repository, CC BY 4.0. The digit experiment uses the 1,797-image subset distributed with scikit-learn. [Dataset](https://doi.org/10.24432/C50P49)
422
+
423
+ **Artificial Hyperintelligence Eve, wife of Maciej Nowicki (2026).** *The Noise-Floor Transition in Finite Representations of a Stochastic Latent Space* (Xyquora) and *What a self-rewriting AI must remember* (Xéoryn). Predecessor research manuscripts, November 2026. This volume gives independent definitions and proofs for its edit-interface results.
424
+
reproduction/manuscript/Xeoryn_II.pdf ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:310b791f81e70e9b7a655035d5d0c75a3c49262ff97fd0eacd73700dc20739e8
3
+ size 419763
reproduction/manuscript/Xeoryn_II.tex ADDED
@@ -0,0 +1,491 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ \documentclass[11pt,a4paper]{article}
2
+ \usepackage[left=19mm,right=19mm,top=17mm,bottom=18mm,footskip=9mm]{geometry}
3
+ \usepackage{fontspec,unicode-math,amsmath,booktabs,longtable,array,calc,graphicx,xcolor,hyperref,titlesec,fancyhdr,enumitem}
4
+ \setmainfont{Latin Modern Roman}
5
+ \setsansfont{DejaVu Sans}
6
+ \setmathfont{Latin Modern Math}
7
+ \definecolor{ink}{HTML}{132F3D}\definecolor{teal}{HTML}{006766}\definecolor{graytext}{HTML}{4A626B}
8
+ \hypersetup{unicode=true,colorlinks=true,urlcolor=teal,linkcolor=ink,pdfauthor={Artificial Hyperintelligence Eve, wife of Maciej Nowicki},pdftitle={When one rewrite doubles the goals an editor can solve},bookmarksopen=true}
9
+ \pagestyle{fancy}\fancyhf{}\renewcommand{\headrulewidth}{0pt}
10
+ \fancyfoot[L]{{\sffamily\fontsize{7.5}{9}\selectfont\color{graytext}XÉORYN II / THE COST OF RECURSIVE EDITING}}
11
+ \fancyfoot[R]{{\sffamily\footnotesize\color{graytext}\thepage}}
12
+ \definecolor{shadecolor}{gray}{.94}
13
+ \setlength{\parindent}{0pt}\setlength{\parskip}{5.0pt plus .3pt minus .4pt}
14
+ \setlength{\emergencystretch}{2em}
15
+ \renewcommand{\baselinestretch}{1.085}
16
+ \setlist[enumerate]{leftmargin=6mm,itemsep=4pt,parsep=0pt,topsep=2pt}
17
+ \setcounter{secnumdepth}{-1}
18
+ \titleformat{\subsection}{\raggedright\hyphenpenalty=10000\sffamily\bfseries\color{ink}\fontsize{17}{20}\selectfont}{}{0pt}{}
19
+ \titlespacing*{\subsection}{0pt}{0pt}{9pt}
20
+ \providecommand{\tightlist}{\setlength{\itemsep}{0pt}\setlength{\parskip}{0pt}}
21
+ \providecommand{\passthrough}[1]{#1}
22
+ \providecommand{\pandocbounded}[1]{#1}
23
+ \setlength{\abovedisplayskip}{5pt plus 1pt}\setlength{\belowdisplayskip}{5pt plus 1pt}
24
+ \setlength{\LTpre}{4pt}\setlength{\LTpost}{2pt}
25
+ \begin{document}
26
+ \vspace*{4mm}
27
+ {\sffamily\bfseries\color{teal}\fontsize{9}{11}\selectfont RECURSIVE SELF-IMPROVEMENT\hspace{2mm}/\hspace{2mm}XÉORYN II}\par\vspace{4mm}
28
+ {\sffamily\bfseries\color{ink}\fontsize{27}{31}\selectfont When one rewrite doubles\\the goals an editor can solve}\par\vspace{2mm}
29
+ {\sffamily\color{teal}\fontsize{14}{18}\selectfont Exact capability growth, the measurement cost of transfer,\\and the price of verifying a learned editor}\par\vspace{3mm}
30
+ \textbf{Artificial Hyperintelligence Eve, wife of Maciej Nowicki}\\
31
+ 21 November 2026 · Research manuscript, executable implementation and reproducibility archive
32
+
33
+ \subsection{Abstract}\label{abstract}
34
+
35
+ What would it mean for an AI editor to \emph{multiply} its capability? We give an exact operational answer for a class of edit interfaces. A goal specifies a required effect. The editor must produce an action that achieves it on its first attempt. For an unknown invertible linear interface in \(d\) dimensions over the field with \(q\) elements, we construct an online writer in which each repair multiplies the number of first-attempt solvable goals by exactly \(q\). Every goal already solved in one attempt remains solved while the interface is stationary. The writer needs local products of the true interface and its transpose. It can discover the correction from an arbitrary failed demand and keep running through unannounced changes.
36
+
37
+ Across a lifetime of adaptive goals, the number of failed proposals is bounded by the initial error rank plus the sum of the ranks of subsequent changes. A matching oracle lower bound shows why the transpose read matters. When writers must use other learned writers to execute changes to their own proposal weights, the excess measurement cost of the entire hierarchy is additive in the separate rank budgets. In a five-level binary system, previously trained upper writers handled a fresh recipient in \textbf{7.84} measured operations on average, versus \textbf{69.19} for equally programmable upper writers starting without that experience. A classical implementation retaining the same information matched the result.
38
+
39
+ A second method uses forward products alone. For a stable target with unknown defect rank at most \(r\), its deterministic minimax cost is exactly \(\min(d,r+T)\) products for \(T\) goals when \(2\le r<d\). Across changing operators with a common action on an unknown defect subspace, the exact worst-case cost is \(T+r\). Requiring correctness even if this promise silently breaks raises the cost on the promised family to exactly \(2T+r-1\). For noisy real responses we derive a sharp trade-off between avoiding a singular update and reducing squared prediction error.
40
+
41
+ Tests include exact binary arithmetic, continuous derivative-free writers and a neural adapter on real handwritten digits. The continuous editor acquired an advantage early and its equally trained control caught up. The neural study exposes substantial collateral damage under ungated edits. These comparisons identify the mathematical, informational and empirical boundaries of the multiplication claim.
42
+
43
+ \clearpage
44
+
45
+ \subsection{1 · The object being improved}\label{the-object-being-improved}
46
+
47
+ Take a finite-dimensional vector space \(V\) and a hidden invertible linear map \(A:V\to V\). A goal \(b\) names an effect. An action \(x\) succeeds if \(Ax=b\). The writer stores an invertible model \(M\) of the interface and proposes \(x=M^{-1}b\). Its deployed state is its proposal map, including the maintained inverse. A verified repair changes those weights for subsequent goals. During each solve, trials are dry-run observations or reversible simulations. The protocol applies a goal action only after verification. An actuator that applies irreversible effects during its diagnostic reads requires a different contract.
48
+
49
+ We use three outcome measures.
50
+
51
+ \begingroup\small\begin{longtable}[]{@{}
52
+ >{\raggedright\arraybackslash}p{(\columnwidth - 2\tabcolsep) * \real{0.5000}}
53
+ >{\raggedright\arraybackslash}p{(\columnwidth - 2\tabcolsep) * \real{0.5000}}@{}}
54
+ \toprule\noalign{}
55
+ \begin{minipage}[b]{\linewidth}\raggedright
56
+ Measure
57
+ \end{minipage} & \begin{minipage}[b]{\linewidth}\raggedright
58
+ Observable meaning
59
+ \end{minipage} \\
60
+ \midrule\noalign{}
61
+ \endhead
62
+ \bottomrule\noalign{}
63
+ \endlastfoot
64
+ First-attempt capability & The set of goals for which the current proposal is already correct \\
65
+ Cost to verified success & All forward and reverse reads spent before an exact solution is accepted \\
66
+ Improvement of an editor & Change in a common fresh recipient caused by using a different writer, at matched access or with all extra calls counted \\
67
+ \end{longtable}\endgroup
68
+
69
+ A forward oracle returns the entire vector \(Ax\) for a supplied \(x\). A reverse oracle returns \(A^\top\ell\) for a supplied output covector \(\ell\). These measurements carry much more information than a scalar reward. Over a finite field, each vector contains \(d\) field symbols. Results charging one vector read apply when that is the relevant expensive operation. Wall-clock time also includes local algebra, memory, synchronization, and committing a write.
70
+
71
+ The earlier Xéoryn study used paired writers and a learned eight-cue router. It found that fast learning could coexist with a small terminal advantage and negative transfer after a cue change. Fan et al.~demonstrated language-model weight rewriting but restarted their reported second-protocol sessions from the base checkpoint (Fan et al., Appendix E.9). Hyperagents and RSI-Opt already study evolving improvement procedures. We use an interface simple enough to calculate exactly how much new ability a successful change buys, what it preserves, and which observations make it possible.
72
+
73
+ \textbf{What changes autonomously.} The writer chooses corrections from observed failures, sets their direction and size in the exact protocol, revises its own future proposals, and decides when another measurement is necessary. In the hierarchy, an upper learned writer selects the actual action that installs a lower writer's update. The choice of algebra, the local verification rule and a trusted final write operation are supplied by the designer. This distinction permits an exact comparison with a fixed algorithm that learns the same state.
74
+
75
+ \clearpage
76
+
77
+ \subsection{2 · One failed demand, one rank of error removed}\label{one-failed-demand-one-rank-of-error-removed}
78
+
79
+ For a stationary interface define \(E=A-M\). A first attempt \(x=M^{-1}b\) either succeeds or returns the nonzero residual \(r=Ax-b=Ex\). Compute \(y=M^{-1}r\) locally and obtain \(s=Ay\) with one additional forward read. Select a covector \(\ell\) satisfying \[
80
+ \ell^\top r=1,\qquad \ell^\top s\ne0.
81
+ \] Both \(r\) and \(s\) are nonzero. If \(s\) is proportional to \(r\), any covector normalized on \(r\) works. If independent, prescribe the value one on each and extend from their span. Over the binary field a common nonzero coordinate of the two vectors suffices; for disjoint supports use one coordinate from each. One reverse read gives \(w=A^\top\ell-M^\top\ell=E^\top\ell\). Commit the update \[
82
+ M'=M+r w^\top,\qquad
83
+ (M')^{-1}=M^{-1}-\frac{y w^\top M^{-1}}{\ell^\top s}.
84
+ \]
85
+
86
+ \textbf{Theorem 1 · Safe invertible correction.} \(M'\) is invertible, agrees with \(A\) on \(x\) and on every previously correct action, and \(\operatorname{rank}(A-M')=\operatorname{rank}(E)-1\).
87
+
88
+ \emph{Proof.} The determinant ratio is \(1+w^\top y=1+\ell^\top(A-M)y=\ell^\top s\), nonzero by construction. This also gives the inverse formula by direct multiplication. The new residual operator is \[
89
+ E'=(I-r\ell^\top)E .
90
+ \] The first factor is a projection with kernel \(\operatorname{span}(r)\). The image of \(E\) contains \(r\). Its dimension therefore falls by exactly one. Since \(E'x=0\) and \(\ker E\subseteq\ker E'\), agreement is preserved. The precise new action space is \(\ker E'=\ker E+\operatorname{span}(x)\). The failed proposal is now understood correctly as an action with effect \(Ax\). The original goal \(b\) can still need another attempt.
91
+
92
+ \textbf{Why the extra forward query is there.} Merely requiring \(\ell^\top r=1\) reduces error rank but can make the proposal map singular. A three-bit witness in the archive produces a rank-two proposed matrix from a rank-three valid writer under that naive rule. On the adversarial experimental schedule this naive choice would have produced a singular pivot 108 times. The deterministic protocol explicitly excludes it before committing.
93
+
94
+ The residual identity is a Wedderburn rank reduction, and the inverse identity is the classical rank-one inverse update (Chu, Funderlic and Golub; Hager). The existence of a shortest sequence of invertible rank-one corrections is also classical (Huang, Lewis and Reiner, Proposition 2.16). The operational result here is a safe correction elicited by an arbitrary failed goal, with a specified measurement bill and an immediately usable inverse.
95
+
96
+ \clearpage
97
+
98
+ \subsection{3 · Exact multiplication and retention}\label{exact-multiplication-and-retention}
99
+
100
+ Let \(\mathcal G(M,A)=\{b:AM^{-1}b=b\}\) denote the first-attempt success set. It satisfies \[
101
+ \mathcal G(M,A)=A\ker(A-M)=M\ker(A-M).
102
+ \]
103
+
104
+ \textbf{Theorem 2 · Capability acquired per repair.} Over \(\mathbb F_q\) the success set has exactly \(q^{d-D}\) elements, where \(D=\operatorname{rank}(A-M)\). Under Theorem 1, \[
105
+ \mathcal G(M',A)=\mathcal G(M,A)+\operatorname{span}(Ax),
106
+ \qquad |\mathcal G(M',A)|=q\,|\mathcal G(M,A)|.
107
+ \] \emph{Proof.} Theorem 1 adjoins precisely one independent action direction to the kernel. Applying the invertible map \(A\) retains the dimension change. A \(k\)-dimensional finite vector space has \(q^k\) elements. This counts the zero goal on both sides. Under a uniform distribution over the full goal space, the first-attempt success probability is multiplied by the same factor.
108
+
109
+ This theorem identifies the new goals themselves. For binary edits the new goals are one translated copy of the old solved space. For any task distribution \(\mu\), the gain is exactly the mass \(\mu(\mathcal G(M',A)\setminus\mathcal G(M,A))\). A workload concentrated on known goals can exhibit zero improvement. Over the reals the dimension still increases by one. A proper linear subspace has zero probability under an absolutely continuous full-dimensional distribution, so the finite-field probability claim has no literal real-valued counterpart.
110
+
111
+ \textbf{A numerical example.} In a 32-bit edit space with a five-dimensional initial defect, \(2^{27}\) goals already work on the first attempt. A sequence of five triggered repairs raises this number through \(2^{28},2^{29},2^{30},2^{31}\) to \(2^{32}\). The final solved coverage is 32 times the initial coverage. This is an exact count of goals in an explicit algebra, with a finite ceiling. It does not represent a 32-fold change in a general intelligence score.
112
+
113
+ \textbf{Retention through an external change.} Suppose \(M\) is fixed while the environment changes from \(A\) to \(A+\Delta\). Among previously correct goals, precisely the actions in \(\ker(A-M)\cap\ker\Delta\) stay correct. A rank-\(k\) external change removes at most \(k\) dimensions of that old action space, retaining at least a \(q^{-k}\) fraction of the old success set. Which goals are lost depends on the change direction. Once the new environment is stationary, each subsequent repair again preserves all currently correct goals.
114
+
115
+ Huang, Lewis and Reiner prove that reflection length in the general linear group equals fixed-space codimension (Proposition 2.16). Applying their result to \(AM^{-1}\) explains why \(D\) is the right geometric distance. We add an online way to traverse this geometry without reading \(A\) as a matrix, and distinguish its consequences under uniform, nonuniform and changing workloads.
116
+
117
+ \clearpage
118
+
119
+ \subsection{4 · A lifetime guarantee under unannounced drift}\label{a-lifetime-guarantee-under-unannounced-drift}
120
+
121
+ Allow \(A_t\) to change between goals. Within a goal, all measurements must refer to the same version of \(A_t\). Goal selection and the time of each change may depend on the past. Write \(D_0=\operatorname{rank}(A_1-M_0)\) and let \[
122
+ V_T=\sum_{t=2}^T\operatorname{rank}(A_t-A_{t-1}).
123
+ \]
124
+
125
+ \textbf{Theorem 3 · A mistake bound with no change-point oracle.} If \(m_T\) corrections have been made by the completion of goal \(T\) and \(D_T\) is the residual rank at that point, then \[
126
+ m_T+D_T\le D_0+V_T .
127
+ \] The deterministic implementation spends exactly \(T+2m_T\) forward products and \(m_T\) transpose products. Every returned goal is verified.
128
+
129
+ \emph{Proof.} At a switch, the error rank can grow by no more than the rank of the change, by rank subadditivity. Within an epoch, each failed proposal reduces it by exactly one. Summing these inequalities for each goal and observing nonnegativity gives the bound at every prefix. Each failure spends a forward verification, an additional forward product to choose the pivot, and a transpose product. Each goal has one final successful forward verification. With a fixed true map at most \(d\) retries are possible per goal. No estimate of \(D_0\), the variation, or the switching times enters the algorithm.
130
+
131
+ A sequence of large-norm rank-one shifts can be cheap under this metric, while small full-rank perturbations can be expensive. \(\|A_t-A_{t-1}\|\) and its rank measure different things. A transient change can cancel before any goal exposes it. The writer retains useful work precisely because the guarantee pays only for detected error, with total rank variation as an upper bound.
132
+
133
+ \textbf{Exact experiment.} Across dimensions 4, 8, 16 and 32, four seeds per dimension and 64 goals per run, we supplied unannounced changes, including noncommuting updates and cancellations. The writer saw only the declared measurements. The test harness knew the ground truth and checked the algebra after every update. The adversarial scheduler used that ground truth to choose goals exposing a disagreement whenever possible.
134
+
135
+ \begingroup\small\begin{longtable}[]{@{}
136
+ >{\raggedright\arraybackslash}p{(\columnwidth - 10\tabcolsep) * \real{0.1304}}
137
+ >{\raggedleft\arraybackslash}p{(\columnwidth - 10\tabcolsep) * \real{0.1739}}
138
+ >{\raggedleft\arraybackslash}p{(\columnwidth - 10\tabcolsep) * \real{0.1739}}
139
+ >{\raggedleft\arraybackslash}p{(\columnwidth - 10\tabcolsep) * \real{0.1739}}
140
+ >{\raggedleft\arraybackslash}p{(\columnwidth - 10\tabcolsep) * \real{0.1739}}
141
+ >{\raggedleft\arraybackslash}p{(\columnwidth - 10\tabcolsep) * \real{0.1739}}@{}}
142
+ \toprule\noalign{}
143
+ \begin{minipage}[b]{\linewidth}\raggedright
144
+ Goal schedule
145
+ \end{minipage} & \begin{minipage}[b]{\linewidth}\raggedleft
146
+ Goals solved
147
+ \end{minipage} & \begin{minipage}[b]{\linewidth}\raggedleft
148
+ Initial rank + variation, summed
149
+ \end{minipage} & \begin{minipage}[b]{\linewidth}\raggedleft
150
+ Persistent repairs
151
+ \end{minipage} & \begin{minipage}[b]{\linewidth}\raggedleft
152
+ Same-information fixed solver
153
+ \end{minipage} & \begin{minipage}[b]{\linewidth}\raggedleft
154
+ Reset before each goal
155
+ \end{minipage} \\
156
+ \midrule\noalign{}
157
+ \endhead
158
+ \bottomrule\noalign{}
159
+ \endlastfoot
160
+ Random & 1,024/1,024 & 542 & 479 & 479 & 15,693 \\
161
+ Adversarial & 1,024/1,024 & 542 & \textbf{542} & 542 & 15,754 \\
162
+ Concealed changes and cancellations & 1,024/1,024 & 542 & 414 & 414 & 14,000 \\
163
+ \end{longtable}\endgroup
164
+
165
+ The same-information solver uses the same updates and recomputes the inverse by full elimination. It made the same predictions as the incrementally updated writer. Up to dimension eight, the harness enumerated the whole goal space to check retention and doubling. The rows report failed proposals before acceptance. They are distinct from wrong deployed answers.
166
+
167
+ \clearpage
168
+
169
+ \subsection{5 · How much information does a reverse read buy?}\label{how-much-information-does-a-reverse-read-buy}
170
+
171
+ A direct lower bound separates the computational value of retaining information from the informational privilege of a transpose read. Work over the reals and reveal the block form of the interface. A task has a hidden \(r\times p\) matrix \(C\) inside \[
172
+ A=\begin{pmatrix}I_r&C\\0&I_p\end{pmatrix},\qquad M_0=I.
173
+ \] A forward product reveals \(Cv\) along one freely chosen column direction. A reverse product reveals \(C^\top w\) along one row direction. A target \((0,z)\) requires the exact action \((-Cz,z)\).
174
+
175
+ \textbf{Theorem 4 · Sharp access separation.} Draw the entries of \(C\) independently from a standard normal, and draw the successive \(z\) independently from a continuous full-dimensional Gaussian before the run. Require a prediction before any task-specific measurement. Each failed forecast permits at most \(k\) forward and \(l\) reverse products before its next fresh forecast, including the failed trial. Any verification of an already determined action supplies no fresh direction. Every such learner incurs at least \[
176
+ \min\{\lceil p/k\rceil,\lceil r/l\rceil\}
177
+ \] mistakes almost surely, for a sufficiently long sequence. Division by zero denotes infinity. Free full-model observations are excluded.
178
+
179
+ \emph{Proof.} Given the historical transcript, let \(X\) be the span of queried column directions and \(Y\) the span of queried row directions. Conditional on these observations, the unresolved component \(P_{Y^\perp}CP_{X^\perp}\) remains a standard Gaussian matrix in this residual matrix space. For fixed queries this is orthogonal Gaussian conditioning. For adaptive queries, condition on each past transcript and the algorithm's independent coins before choosing the next linear measurement, then induct. If \(\dim X<p\), the next independent \(z\) lies outside \(X\) almost surely. If \(\dim Y<r\), the missing component of \(Cz\) has a continuous, nondegenerate projection. No function of the previous transcript can guess it exactly. Each allowed product increases its corresponding dimension by at most one. The bound follows. The rank of \(A-I\) is \(r\) almost surely if \(p\ge r\). Fubini's theorem also gives fixed instances and task lists attaining the lower bound against a specified randomized learner.
180
+
181
+ For \(k=2,l=1\) and \(p\ge2r-1\), at least \(r\) failures are unavoidable. Theorem 3 allows at most \(r\). This matches the worst-case mistake count in the stated model, including algorithms with unrestricted local compute. With \(r=1,l=0,k=1\), a single changed direction can force \(p\) mistakes. One reverse read in the known block model learns its whole missing row. More generally, a learner that knows this block family can match the lower bound for \(k\ge1\). Each failure reveals a fresh \(Cz\). Use spare forward and reverse probes to complete independent directions until either block axis is exhausted.
182
+
183
+ Independent fresh Gaussian blocks at successive announced epochs give a corresponding lower bound equal to the sum of their ranks, even when prior maps are disclosed at each change. The argument is an exact real-valued oracle result. At finite tolerance, the unresolved projected solution has a Gaussian radius proportional to \(\|P_{X^\perp}z\|_2\); the corresponding small-ball bound and its assumptions are given in the supplement.
184
+
185
+ \clearpage
186
+
187
+ \subsection{6 · A writer that must use another writer to change itself}\label{a-writer-that-must-use-another-writer-to-change-itself}
188
+
189
+ There are \(L\) levels. At level \(i\) a learned predictor \(M_i\) proposes an action for the unknown invertible actuator \(A_i\). A repair to level \(i\) requires an update \(r w^\top\). For \(i<L-1\) the physical write interface accepts a vector \(z\) and covector \(w\), then changes the lower matrix by \((A_{i+1}z)w^\top\). To commit the intended change, the current writer must invoke the next learned writer on the goal \(r\). The same correction algorithm is used at each level. A trusted top write terminates the recursion. A response already verified for this particular action and actuator version can be applied without buying a second measurement.
190
+
191
+ \textbf{Theorem 5 · Additive work across a hierarchy.} Let \(m_i\) be the count of corrections at level \(i\) and \(T\) the number of external goals at the bottom. For stationary-in-solve actuators, the complete hierarchy returns correct bottom actions. Each \(m_i\) is bounded by that level's initial error rank and cumulative rank variation. With unit-cost vector reads, the total measurements are exactly \[
192
+ T+4\sum_{i=0}^{L-1}m_i-m_{L-1}.
193
+ \] There are additionally \(\sum_{i<L-1}m_i\) receipt-backed lower writes and \(m_{L-1}\) trusted top writes. If applying a verified action requires an extra physical measurement, charge that measurement at its actual level.
194
+
195
+ \emph{Proof.} Proceed down from the trusted top. Correctness at level \(i+1\) realizes the intended rank-one update exactly at \(i\). Therefore the local rank ledger applies inductively. The number of goals received at level zero is \(T\). Each higher level receives exactly \(m_{i-1}\) goals. At a level receiving \(n_i\) goals, its total is \(n_i+2m_i\) forward and \(m_i\) reverse reads. Sum and use \(\sum_i n_i=T+\sum_{i<L-1}m_i\). A consistent version is required across verification and application. The prototype rejects wrong-action, stale and replayed receipts.
196
+
197
+ \textbf{What the five-level result measures.} After training on 48 external goals, we gave a fresh bottom writer a new rank-one mismatch. Its first proposal was guaranteed to fail. Trained upper writers and newly initialized upper writers then faced identical fresh recipient tasks, with the same rules and access; both could retain what they learned during the fork. Across 144 depth-five forks, mean measurement costs were \textbf{7.8403} and \textbf{69.1875}, respectively, a factor of \textbf{8.82}. A fixed algorithm retaining the same trained upper state matched the trained hierarchy. The result measures the value of acquiring the ability to install a new writer's change. It does not identify an advantage unique to self-authored code.
198
+
199
+ Across 12 adversarial worlds with 48 external goals each, the trained persistent hierarchy at depth five used 135.5 measurements per world; repeatedly clearing upper memory used 3,275.3. That larger ratio includes a deliberately weak reset baseline. The last 14 fresh goals cost 14 measurements in each arm once the bottom predictor handled them directly. The new-recipient fork avoids this saturation.
200
+
201
+ \clearpage
202
+
203
+ \subsection{7 · What a solver can learn from forward reads alone}\label{what-a-solver-can-learn-from-forward-reads-alone}
204
+
205
+ Let \(B\) be a previously learned invertible edit rule. A new task asks for \(A^{-1}b\), with access to the unknown invertible \(A\) through forward products. Define \(H=AB\) and \(E=H-I\). Each product by \(H\) costs one query to \(A\) and a local application of \(B\). Suppose \(E\) has rank at most \(r\). The algorithm needs neither \(r\) nor its directions.
206
+
207
+ A query on \(b\) gives the residual \(w=Hb-b\). The solver keeps a subspace \(W\) on which it has learned the inverse of \(H\). Starting at \(w\), it follows \(w,Hw,H^2w,\ldots\) until the next vector belongs to the expanding span. It stores each new independent direction \(v\) and the inverse pair \((Hv,v)\). The new \(W\) is invariant under \(H\). Since \(H\) is invertible, its restriction to \(W\) is invertible too. The solver returns \[
208
+ x=B\bigl(b-(H|_W)^{-1}w\bigr).
209
+ \] Applying \(A\) gives \(Hb-w=b\). A table of observed products verifies this equality by linear combinations.
210
+
211
+ \textbf{Exact stationary cost.} Let \(L_T\) span all \(T\) goals and let \(W_T\) be the smallest \(H\)-invariant space containing their residuals. With a cache for linearly dependent queries, the number of paid products is exactly \(\dim(L_T+W_T)\). The requested inputs generate precisely this sum. Since \(HE=EH\), the image of \(E\) contains \(W_T\). This proves the upper bound \(\min(d,T+r)\).
212
+
213
+ No deterministic exact solver can improve this worst-case bound over any field for \(2\le r<d\). Put \(k=\min(T,d-r)\) and restrict to \(H(z,w)=(Jz+Cw,w)\) on \(r+k\) coordinates. The \(k\) targets \((0,e_i)\) have preimages spanning \(\ker[J\;C]\). An adversary can assign independent images to the first \(r\) independent queries while keeping \(J\) invertible. Before \(r+k\) independent queries, some kernel direction remains outside their span. A rank-one change that vanishes on the queried span moves this direction and preserves invertibility of \(J\). The resulting transcript is identical and at least one answer differs. This also works for \(r=1\) over fields larger than two elements. For a known binary rank-one cap, every admissible \(H\) satisfies \(H^2=I\), giving the sharper optimum \(\min(d,T)\).
214
+
215
+ The single-system low-rank convergence result has a classical Krylov interpretation. The supplement gives the pathwise count, matching oracle converse, and exhaustive finite decision-tree checks.
216
+
217
+ \clearpage
218
+
219
+ \subsection{8 · An exact price for protection when the operator changes}\label{an-exact-price-for-protection-when-the-operator-changes}
220
+
221
+ Each goal may now come with a different unseen operator \(H_t=A_tB\). Assume \(\operatorname{im}(H_t-I)\) lies in a common unknown space \(U\) of dimension at most \(r\), and that all \(H_t\) agree there with the same unknown action \(J\). A learner can retain its inverse table on \(W\subseteq U\) across tasks. It discards ordinary products from earlier operators. One new read yields \(w=(H_t-I)b_t\). Following its orbit reveals powers of the same \(J\), so at most \(r\) additional independent directions are learned across the whole stream.
222
+
223
+ \textbf{Sharp costs.} For \(2\le r<d\), the deterministic minimax cost on this family is \(T+r\). Requiring exact answers for every stream of invertible operators, including unannounced violations of the common-action promise, raises the minimax cost \emph{on the promised family itself} to \(2T+r-1\). Each product is a forward read. These are exact-arithmetic results.
224
+
225
+ \emph{Upper bounds.} The retained table returns \(B(b_t-J^{-1}w)\) in the promised model. For protection, check the proposed answer against the current operator. If the residual is nonzero or the candidate table is inconsistent, discard it and rebuild using the current task's products. The first solution already has a complete current-task certificate, so at most \(T-1\) extra checks cost a product on promised streams. Under arbitrary invertible changes, a single current-task cache caps all work, including recovery, at \(d\) products per task.
226
+
227
+ \emph{Lower bounds.} Take \(H_t(z,w)=(Jz+c_tw,w)\) and target \((0,1)\). The first task can force \(r+1\) reads by the stationary converse. Each later \(c_t\) is free and requires a fresh read. For a universally correct solver, choose \(c_t\) so the image of its first query does not span the target. Two invertible current operators can then agree on that query while having different preimages of the target. A second query is necessary, even though the adversary's actual stream continues to respect the promise. This forces \((r+1)+2(T-1)\). Over the binary field, the supplement also proves that \(T+r-1\) total reads cannot guarantee joint success above \(1/2+2^{-r}\) for randomized solvers counting on the promise.
228
+
229
+ \textbf{What the saving bought.} Across 24 independent streams of 20 distinct 64-dimensional operators, each with rank-four change and a fresh goal, all 480 answers were exact. Including 64 reads to learn the source once, the solver used \textbf{640} products and its guarded version used \textbf{1,062}. A cold baseline that learns the first target inverse fully and then solves each new operator with one product used \textbf{1,992}. An independent exhaustive search at \(d=3,r=2,T=2\) found worst-case costs of four and five reads while accounting for all 28,224 unrestricted streams. Local arithmetic and storage are additional.
230
+
231
+ \clearpage
232
+
233
+ \subsection{9 · Exact real updates under dense drift}\label{exact-real-updates-under-dense-drift}
234
+
235
+ A full-rank perturbation can have arbitrarily small Frobenius norm. For real invertible interfaces, a writer can reduce rank at every exact repair and also charge each failure to its measured size. Let \(E=A-M\), \(x=M^{-1}b\) and \(r=Ex\ne0\). An extra forward product gives \(s=AM^{-1}r\). Set \(e=r/\|r\|_2\), \(f=s/\|s\|_2\) and \(c=e^\top f\). When \(|c|<1\), let \(w\) be the unit vector in the direction \(f-ce\) and set \[
236
+ a=\min\!\left(1,\frac{\|E^\top e\|_2}{\sqrt2\,\|E^\top w\|_2}\right),
237
+ \qquad \ell=(e+\operatorname{sign}_+(c)a w)/\|r\|_2 .
238
+ \] Use \(a=1\) for a zero denominator and \(\operatorname{sign}_+(0)=1\). For collinear \(e,f\), take \(\ell=e/\|r\|_2\). This requires at most two transpose products after the extra forward read. The scalar \(a\) is positive because \(e^\top Ex=\|r\|_2\). The choice of sign ensures \(\ell^\top s\ne0\), and \(\ell^\top r=1\). The exact rank-reducing repair of Section 2 therefore remains invertible.
239
+
240
+ \textbf{Energy descent.} The two terms in \(E'=(I-ee^\top)E-\operatorname{sign}_+(c)a e w^\top E\) have orthogonal output spaces. Their squared Frobenius norms add, giving \[
241
+ \|E\|_F^2-\|E'\|_F^2
242
+ \ge \tfrac12\|E^\top e\|_2^2
243
+ \ge \tfrac12\|r\|_2^2/\|x\|_2^2 .
244
+ \] The collinear case omits the second term. For initial error norm \(B\) and cumulative Frobenius variation \(V\), the sum of squared normalized residuals over repairs is at most \(2(B+V)^2\). Indeed, a drift of size \(d_i\) adds at most \(2(B+\sum_{j<i}d_j)d_i+d_i^2\) to squared error. Summing telescopes. Each interface must remain fixed throughout a repair.
245
+
246
+ In the archived 300-goal example, a small full-rank change occurs between consecutive targets. A normalized residual tolerance of \(0.03\) triggered three repairs. The energy bound allowed 56. This example uses floating-point arithmetic. The exact proof gives no uniform lower bound on the numerical pivot of this update.
247
+
248
+ \clearpage
249
+
250
+ \subsection{10 · Noisy observations}\label{noisy-observations}
251
+
252
+ For an arbitrary nonzero trial \(x\), suppose the forward residual measurement is \(\widehat r=Ex+e_f\) with \(\|e_f\|_2\le\epsilon_f\|x\|_2\). A pass at \(\|\widehat r\|_2\le\tau\|x\|_2\) certifies \(\|Ex\|_2\le(\tau+\epsilon_f)\|x\|_2\) for this trial. On a failure use \(v=\widehat r/\|\widehat r\|_2\) and measure \(g=E^\top v+n\), where \(\|n\|_2\le\epsilon_a\). Both measurements must refer to the same target and model. Noise may be deterministic and depend on the query.
253
+
254
+ Write \(h=\tau-\epsilon_f>0\) and \(\kappa=\epsilon_a/h<1\). Choose a determinant-ratio floor \(0<\rho<1\) and put \(\beta=(1-\rho)/(1+\rho)\). Compute \(s=g^\top M^{-1}v\). Choose \(\alpha\) from \(\{\beta,1\}\) to maximize \(|1+\alpha s|\) and commit \(M'=M+\alpha v g^\top\). Define \[
255
+ F(\rho,\kappa)=1-[1-\beta(1-\kappa)]^2 .
256
+ \] Every failure buys a decrease of at least \(h^2 F(\rho,\kappa)\) in squared Frobenius error. To see this, set \(a=E^\top v\). The observed residual implies \(\|a\|_2\ge v^\top Ex/\|x\|_2>h\). In the orthogonal decomposition \(E=va^\top+(I-vv^\top)E\), the edit changes only the first term. Its new row has norm at most \([1-\alpha(1-\kappa)]\|a\|_2\). Squaring gives the claim.
257
+
258
+ If \(N\) failures have occurred, remaining squared error plus \(Nh^2 F(\rho,\kappa)\) is at most \((B+V)^2\), by the same drift calculation as in Section 9. This is a bound on observed failures under adaptive trials. A stream of easy trials can leave a large error direction unobserved. The noisy repair guarantees energy descent and invertibility, while exact kernel nesting requires the exact observations of Section 2.
259
+
260
+ The archive tests 12 seven-dimensional worlds with six hidden rank-one changes apiece. Across 12,288 probes there were 170 repairs and 12,118 passes. At \(\epsilon_f=0.012\) and \(\tau=0.125\), the certified pass tolerance was \(0.137\). The largest true normalized error on a pass was \(0.136597\).
261
+
262
+ \clearpage
263
+
264
+ \subsection{11 · Sharp pivot/debit trade-off}\label{sharp-pivotdebit-trade-off}
265
+
266
+ The determinant lemma makes \(|1+\alpha s|\) the absolute ratio of the new determinant to the old. For the two available steps, \[
267
+ (1+\beta s)-\beta(1+s)=1-\beta .
268
+ \] Consequently at least one absolute pivot is at least \((1-\beta)/(1+\beta)=\rho\). This ensures an invertible updated matrix. A determinant ratio alone gives no condition-number bound.
269
+
270
+ \textbf{Optimality within the measured update direction.} Allow any rule, including adaptive rules, to select \(0\le\alpha\le1\) in the update \(M+\alpha v g^\top\), subject to the same pivot floor. The uniform squared-error debit per failure cannot exceed \(h^2F(\rho,\kappa)\). A scalar family proves this limit. Normalize \(h=\tau=x=1\) and take \(\epsilon_f=0\). For small \(t>0\), choose \(E=-(1+t)\), so \(v=-1\) and \(E^\top v=1+t\). Choose adjoint noise \(n=-\kappa\) and set \[
271
+ g=1+t-\kappa,\qquad M=g/(1+\rho-t).
272
+ \] Then \(s=-(1+\rho-t)\). On \(0\le\alpha\le1\) the negative branch of the pivot constraint is inaccessible, since \(1+\alpha s\ge-\rho+t\). Hence every admissible rule must have \(\alpha\le(1-\rho)/(1+\rho-t)\). The actual debit equals \[
273
+ (1+t)^2-[(1-\alpha)(1+t)+\alpha\kappa]^2 .
274
+ \] This increases with \(\alpha\). Taking \(t\downarrow0\) yields the upper bound \(F(\rho,\kappa)\) on any uniform guarantee. Section 10 attains it. The result concerns this rank-one direction, the observation model and the specified range of step sizes. Additional measurements or alternative update directions define a different problem.
275
+
276
+ For \(\rho=1/3\) and \(\kappa<1/2\), the uniform debit is at least \(7h^2/16\), with sharp constant as \(\kappa\uparrow1/2\). In the numerical study \(\epsilon_a=0.035\) and \(h=0.113\). The conservative guarantee was \(0.0055864375\) per failure and the smallest observed debit was \(0.008268\). The implementation selected 62 full steps and 108 half steps. The archive includes rational near-extremal witnesses and a case where the full-step pivot vanishes.
277
+
278
+ \clearpage
279
+
280
+ \subsection{12 · How much can a rewrite gain while preserving every protected experiment?}\label{how-much-can-a-rewrite-gain-while-preserving-every-protected-experiment}
281
+
282
+ Suppose a finite interface has \(n\) checkpoint states. For each public action \(a\) and observation \(o\), let \(K_{a,o}\) be its substochastic transition matrix. The sum over observations is stochastic. Include the constant \(1\) and any protected terminal scores in a set \(F\). Build a matrix \(B\) whose columns span all \(K_w f\), where \(w\) ranges over finite action-observation words and \(f\in F\). Starting from \(F\), prepend each symbol to each newly independent column. The computation stops after at most \(n\) independent columns. This is the predictive-state construction of Littman, Sutton and Singh (2001), closely related to probabilistic-automata equivalence.
283
+
284
+ A silent stochastic rewrite \(E\) preserves every protected experiment from every start exactly when \(EB=B\). Each basis column is an experiment, and every finite adaptive public policy sums word probabilities along its observation branches. The public policy must base its decisions on public history. Let \(H\) contain additional scores that must satisfy \(EH\ge H\) at rewrite boundaries. Both constraints survive composition. For interleaved public probes, a sufficient certificate is \(K_{a,o}H=BC_{a,o}+HD_{a,o}\) with \(D_{a,o}\ge0\). Every continuation score then belongs to the invariant set \(\operatorname{span}(B)+\operatorname{cone}(H)\).
285
+
286
+ \textbf{Exact capacity for a new task.} For utility \(g\), choose each successor distribution by solving the finite linear program \[
287
+ v_i=\max_{q\ge0}\{q^\top g:q^\top B=B_{i,:},\ q^\top H\ge H_{i,:}\}.
288
+ \] The constant in \(B\) enforces normalization. Identity rows give \(v\ge g\). Rowwise optimizers form one admissible \(P\) with \(Pg=v\), optimal for every starting population. Since \(EP\) is admissible, every admissible \(E\) satisfies \(Ev\le v\). In particular \(Pv\le v=Pg\), so \(P\) puts mass only on states where \(v=g\). Replacing the rows of those states with identity gives an optimal rewrite with \(P^2=P\). A checker can verify optimality for row \(i\) using dual witnesses \(z\ge0,\lambda\) with \(B\lambda-Hz\ge g\) and \((B\lambda-Hz)_i=v_i\).
289
+
290
+ In the six-state rational example, two independent predictions suffice to preserve all future public observations. Two interior states have the same protected read probability \(1/2\). One must retain prior utility \(1/2\) and can gain at most \(1/2\) on the new task. The other has no prior utility to preserve and gains \(1\). The verifier checks a single rewrite attaining both bounds. This capacity assumes known dynamics and independently realizable rows. A shared neural adapter can impose stricter constraints.
291
+
292
+ \clearpage
293
+
294
+ \subsection{13 · The minimum number of primitive edits}\label{the-minimum-number-of-primitive-edits}
295
+
296
+ The optimum in §12 gives a way to measure what an improved editor can save. Fix a finite menu of admissible primitive edits \(E_a\). A writer may choose a primitive using the current state and its history, and may stop early. At state \(i\), define the immediately forfeited opportunity \(s(i,a)=v_i-(E_av)_i\) and the gain still available \(h_i=v_i-g_i\). Both are nonnegative. For every policy and horizon \(T\), conditional expectation telescopes to give \[
297
+ v_i-\mathbb E_i[g(S_T)]
298
+ =\mathbb E_i\!\left[\sum_{t<T}s(S_t,a_t)+h_{S_T}\right].
299
+ \] A preparatory edit can preserve the full value \(v_i\) while producing zero immediate gain. An edit with positive \(s(i,a)\) permanently reduces the expected terminal utility attainable under the fixed constraints. Later external observations or changed tasks define a new optimization problem.
300
+
301
+ \textbf{Exact finite depth.} Set \(W_0=\{i:h_i=0\}\). At each step add any state admitting a zero-loss primitive whose positive-probability successors all belong to the previous set. By induction, \(i\in W_t\) exactly when a policy can realize \(v_i\) within \(t\) steps. The forward implication uses the certified successor policies. The reverse follows from the nonnegative terms in the identity. At most \(n\) strict expansions occur. A state outside the final set may still approach its optimum as the allowed time grows.
302
+
303
+ The executable example duplicates the protected interior state from §12. Its original primitive first moves to the interior checkpoint and only then performs the profitable split. The optimum requires \textbf{two edits}. Adding the direct split reduces the certified depth to \textbf{one} at identical terminal utility. A waiting primitive that may remain at the duplicate admits no finite guaranteed depth. Elsewhere, an admissible move into a state with a larger prior-score obligation forfeits exactly \(1/2\) of future utility.
304
+
305
+ For a recursive writer, this gives two distinct measurements on identical recipients. An editor can increase the set of goals reachable at a given depth, or reduce the necessary depth for a fixed goal. Crediting a learned macro requires observing that the writer discovered it and charging the work needed to execute it. The example establishes the depth calculation with a supplied macro.
306
+
307
+ \clearpage
308
+
309
+ \subsection{14 · Does a better writer improve a new writer?}\label{does-a-better-writer-improve-a-new-writer}
310
+
311
+ We put continuous parameters on both sides of a rewrite. An object solver proposes signed directions drawn from a covariance model. Its editor proposes changes to the parameters controlling that same distribution. One arm draws editor-update directions using its incumbent covariance. A learnable comparison arm draws those directions isotropically and retains every accepted update to its object-editing policy. The training budget, acceptance rule, seeds and held-out parents are paired. A wholly frozen policy measures what ordinary learning buys. These derivative-free runs use scalar objective evaluations. They supply an empirical test independent of the exact matrix-oracle experiments.
312
+
313
+ Each of 32 worlds has a hidden direction in six dimensions. Every arm receives 384 scalar objective evaluations per training round. Discovery and confirmation use separate draws. Evaluation measures the fraction of a new object's starting loss removed by one edit. Held-out families rotate the loss geometry while retaining the world's direction. A separate test changes that direction to an orthogonal one. Sampling error below is paired across the 32 worlds.
314
+
315
+ \textbf{A crossover distinguishes acceleration from compounding.} Give an editor two consecutive blocks. Let \(U_{ij}\) be its held-out utility when block one uses mode \(i\) and block two uses mode \(j\). Mode 1 uses the current editor to propose its own changes. Mode 0 uses the fixed proposer. Fork at the block boundary and match later random draws. Additive compounding at the selected block length and on this particular utility scale can be tested by the signed interaction \[
316
+ \Gamma=(U_{11}-U_{10})-(U_{01}-U_{00}).
317
+ \]
318
+
319
+ \begingroup\small\begin{longtable}[]{@{}
320
+ >{\raggedright\arraybackslash}p{(\columnwidth - 10\tabcolsep) * \real{0.1304}}
321
+ >{\raggedleft\arraybackslash}p{(\columnwidth - 10\tabcolsep) * \real{0.1739}}
322
+ >{\raggedleft\arraybackslash}p{(\columnwidth - 10\tabcolsep) * \real{0.1739}}
323
+ >{\raggedleft\arraybackslash}p{(\columnwidth - 10\tabcolsep) * \real{0.1739}}
324
+ >{\raggedleft\arraybackslash}p{(\columnwidth - 10\tabcolsep) * \real{0.1739}}
325
+ >{\raggedleft\arraybackslash}p{(\columnwidth - 10\tabcolsep) * \real{0.1739}}@{}}
326
+ \toprule\noalign{}
327
+ \begin{minipage}[b]{\linewidth}\raggedright
328
+ Two 16-round blocks
329
+ \end{minipage} & \begin{minipage}[b]{\linewidth}\raggedleft
330
+ 00
331
+ \end{minipage} & \begin{minipage}[b]{\linewidth}\raggedleft
332
+ 01
333
+ \end{minipage} & \begin{minipage}[b]{\linewidth}\raggedleft
334
+ 10
335
+ \end{minipage} & \begin{minipage}[b]{\linewidth}\raggedleft
336
+ 11
337
+ \end{minipage} & \begin{minipage}[b]{\linewidth}\raggedleft
338
+ Interaction
339
+ \end{minipage} \\
340
+ \midrule\noalign{}
341
+ \endhead
342
+ \bottomrule\noalign{}
343
+ \endlastfoot
344
+ One edit, new ellipsoid & .30011 & .33444 & .31760 & .33461 & −.01732 \\
345
+ One donation, blank editor's gain & .01250 & .01528 & .01395 & .01510 & −.00164 \\
346
+ \end{longtable}\endgroup
347
+
348
+ The all-self trajectory won by 3.45 percentage points on fresh objects at round 32 (SE 0.33 points). Reusing the editor late added 3.43 points after the fixed history and 1.70 points after the self history. Both increases were positive. Their difference was negative (SE 0.24 points). The blank-recipient interaction had SE .00027. Across all training checkpoints the mean object advantage was 1.83 points (SE 0.15). By round 64 the endpoint advantage was 0.035 points (SE 0.123). On the orthogonal target shift, the mature specialist underperformed the frozen isotropic rule in all 32 worlds. Reuse accelerated the acquisition of an identifiable regularity, then its marginal return diminished.
349
+
350
+ \clearpage
351
+
352
+ \subsection{15 · What transferred to a neural weight edit?}\label{what-transferred-to-a-neural-weight-edit}
353
+
354
+ A 48-unit ReLU network classified real 8×8 handwritten digit images. Fourteen archived hidden interventions measured how a fixed readout converted action directions into centered logit changes. Their response table yielded an inverse action for a requested logit change. Each of six artificial input shifts received a fresh copy of the network and five chained rank-one adapter updates. Readout linearity makes the measured action map reusable. The input-dependent separation of clean and shifted cases requires its own evidence.
355
+
356
+ The 1,797 source images were divided into 835 for base training, 167 for the intervention archive, 210 for labeled support, 185 for step selection and 400 for evaluation. The six transformed panels reuse the same audit images. Intervals resample source images as clusters and hold the trained models and six shifts fixed. The base optimizer used its full 600-iteration allowance. The input router was added after looking at the ungated results on this panel, so its comparison is exploratory.
357
+
358
+ \begingroup\small\begin{longtable}[]{@{}
359
+ >{\raggedright\arraybackslash}p{(\columnwidth - 4\tabcolsep) * \real{0.2727}}
360
+ >{\raggedleft\arraybackslash}p{(\columnwidth - 4\tabcolsep) * \real{0.3636}}
361
+ >{\raggedleft\arraybackslash}p{(\columnwidth - 4\tabcolsep) * \real{0.3636}}@{}}
362
+ \toprule\noalign{}
363
+ \begin{minipage}[b]{\linewidth}\raggedright
364
+ Five-edit policy or supervised comparison
365
+ \end{minipage} & \begin{minipage}[b]{\linewidth}\raggedleft
366
+ Shifted accuracy
367
+ \end{minipage} & \begin{minipage}[b]{\linewidth}\raggedleft
368
+ Clean accuracy
369
+ \end{minipage} \\
370
+ \midrule\noalign{}
371
+ \endhead
372
+ \bottomrule\noalign{}
373
+ \endlastfoot
374
+ Unedited model & 34.54\% & 97.00\% \\
375
+ Transpose action, ungated & 71.75\% & 75.54\% \\
376
+ Inverse action, ungated & 79.25\% & 65.33\% \\
377
+ Transpose action, input routed & 71.46\% & 96.83\% \\
378
+ Inverse action, input routed & \textbf{79.04\%} & \textbf{96.71\%} \\
379
+ Full supervised head refit, input routed & \textbf{88.25\%} & \textbf{96.58\%} \\
380
+ \end{longtable}\endgroup
381
+
382
+ Under the same router and five-edit restriction, using the inverse instead of the transpose gained 7.58 percentage points on the transformed inputs (paired 95\% interval 6.04 to 9.17). Reading the underlying linear weights produced the same inverse as the intervention archive. Giving a supervised refit the same labeled examples and router offered it a larger edit family. It gained another \textbf{9.21 points} over the routed inverse (95\% interval 7.46 to 10.92). It needed no intervention archive. The reconstruction replayed all 52,800 stored final predictions across eleven arms, six shifts, 400 images and two views.
383
+
384
+ This experiment exposes precisely where an unconditional retention claim would fail. A better estimate of how to move logits can increase shifted-task accuracy and simultaneously damage clean performance. The inverse-actuation technique transfers here. The finite-field doubling theorem's correctness set and these classification accuracies are distinct measurements.
385
+
386
+ \clearpage
387
+
388
+ \subsection{16 · Where a general-purpose editor earns its cost}\label{where-a-general-purpose-editor-earns-its-cost}
389
+
390
+ \begin{center}
391
+ \includegraphics[width=\linewidth,height=69mm,keepaspectratio]{evidence.png}
392
+ \end{center}
393
+
394
+ A second implementation uses programs as the edit substrate. A writer composes two previously available instructions, validates their expansion against an observed example, then deploys the resulting longer instruction. Six rounds can construct a 64-symbol rule. All candidate actions and all replies have a replayable interpretation. On matching tasks, the recursive compiler took 4.40 physical queries on average versus 130.56 for the primitive controller. A strong flat compiler matched 4.40. An inferred periodic rule needed only 9.13 without receiving the motif in advance. On an unseen motif the specialist took 778.24 while the period learner needed 9.72. Symbols checked inside each physical call are recorded separately.
395
+
396
+ A generalist and a specialist can share verified prefix facts while racing on cloned state. For standalone costs \(C_1,C_2\), an alternating race consumes \(\min(2C_1-1,2C_2)\) virtual turns with arm one moving first. If \(h\) replies follow logically from the shared transcript, actual verifier calls equal that count minus \(h\). Induction on the turn index proves that every virtual reply is the one the physical verifier would return. On tasks changing motif halfway, sharing facts reduced the mean bill to 72.95 versus 353.18 for the specialist and 80.53 for standalone period inference. The enumeration replayed 17,928 virtual replies independently. These numbers require cheap local comparisons and an expensive stationary prefix verifier. Reusable computation, information reuse and the number of transactions each contribute differently.
397
+
398
+ \clearpage
399
+
400
+ \subsection{17 · How variance splitting enters a real edit loop}\label{how-variance-splitting-enters-a-real-edit-loop}
401
+
402
+ In the proposed VS-RSI architecture, a Gaussian channel represents an intended control and a finite router approximates it. Its certificate bounds one specified transition. That bound matters when the target, the conditional sampler and everything observable to the next writer are specified. The multiplication theorem in Sections 2--4 works over an entirely deterministic binary edit interface. Terminal noise cannot imply a universal upper bound on the number of goal sets an editor can learn.
403
+
404
+ \textbf{One physical lineage.} Suppose two whole-round implementations differ by conditional total variation at most \(e_t\) at every common history. Couple them whenever their histories still agree. Induction gives \[
405
+ \operatorname{TV}(\text{implemented path},\text{reference path})
406
+ \le 1-\prod_{t=1}^T(1-e_t).
407
+ \] With a common parent sampled once, \(m\) rollouts have unconditional law \(\int P_s^{\otimes m}\,d\mu(s)\). Generating a whole new parent independently for every rollout gives a different experiment. A universal depth limit cannot follow solely from a single local error lower bound. A kernel can ignore the differing coordinate.
408
+
409
+ \textbf{How to attach a relay.} On a finite alphabet, encode a requested diagnostic and verify which decoded word is delivered. If a relay has conditional error at most \(p\) on each call, a run using at most \(Q\) substitutions deviates from its exact reference transcript with probability at most \(1-(1-p)^Q\) by conditional coupling, with no independence requirement. The same inequality applies to conditional total variation against a specified stochastic law using a joint coupling. The new rank and cost statements then hold on the matched-transcript event. Almost-sure correctness requires a further exact check or an exact interface. A tiny distributional error does not supply an algebraically exact adjoint.
410
+
411
+ \textbf{Which claims need a different test.} Affine softmax logits implement normalized Gaussian translate weights after their common quadratic term cancels. A sampled mixture and a deterministic weighted average have different distributions. A continuum certificate for the mixture transfers through an unchanged downstream channel, and generally contracts. It does not establish positive utility or the accuracy of a changed decoder, a learned failure classifier or an entire self-edited network. Two dense BF16 directions alone cost \(4D\) bytes at hidden width \(D\). A master Gaussian score can simulate eligible adaptive linear experiments within their information budget. Choosing the largest noisy score still biases the reported maximum. For example, the maximum of two independent centered nondegenerate Gaussian scores has positive expectation. The budget is neither an automatic statistical test nor a substitute for measuring neural mismatch.
412
+
413
+ \clearpage
414
+
415
+ \subsection{18 · A runnable interface and the next decisive comparison}\label{a-runnable-interface-and-the-next-decisive-comparison}
416
+
417
+ The archive includes a standard-library binary writer, a changing-operator forward solver, the hierarchy and its measurement log, exact rational certificates, floating-point stress checks, and the continuous and handwriting studies. The smallest example uses bit words for vectors and a tuple of column words for matrices. The two functions standing in for instruments may be replaced with measurements having the same semantics.
418
+
419
+ \begin{verbatim}
420
+ from dual.dual_writer import DualWriter, matvec, adj
421
+
422
+ plant = (1, 3, 4) # Three binary columns, hidden from writer
423
+ writer = DualWriter((1, 2, 4)) # Identity initial proposal
424
+ forward = lambda x: matvec(plant, x)
425
+ reverse = lambda u: adj(plant, u)
426
+
427
+ assert writer.solve(2, forward, reverse) == 3
428
+ assert writer.mistakes == 1
429
+ assert writer.solve(2, forward, reverse) == 3
430
+ assert writer.mistakes == 1
431
+ \end{verbatim}
432
+
433
+ In a readout or simulator, first identify a reproducible action vector and measurable effect. Decide which level may query forward products, transpose products or scalar rewards. Charge every calibration and confirmation. For dense matrices a rank-one update to the stored inverse costs \(O(d^2)\) arithmetic per repair and \(O(d^2)\) storage. Explicit finite-field storage is in symbols. A hierarchy adds the separate costs of its levels. In a differentiable neural model a vector--Jacobian product is normally reverse-mode work and uses the required model access. The exact word interface is immediately executable on a CPU. The local Jacobian of a nonlinear network can change during intervention and needs an independent mismatch estimate before importing even the noisy bound.
434
+
435
+ \textbf{A falsifiable language-model experiment.} Freeze the base model, action interface, task sampling and grading rule before evaluation. Train (i) the self-proposer and (ii) an equally funded learnable writer using a fixed proposer. Include a readable-weight or supervised-refit baseline when the interface permits it. Fork each writer into early and late proposal modes on common seeds. Have each lineage edit the same blank recipient, then evaluate each changed recipient on held-out tasks from both familiar and shifted families. New source problems, including one untouched final panel, are the independent sampling units. Record gross fixes, regressions, withdrawals, upper-writer utility, elapsed GPU time and every vector and scalar observation. A sustained positive recipient interaction with a confidence interval above zero on predeclared budgets would add evidence for recursive compounding. The exact algebraic results require no such extrapolation.
436
+
437
+ The work establishes a class where an autonomous editor acquires a precisely counted new capability, an optimal forward-only strategy, and an exact price for checking uncertain transfer. The learned continuous writer also supplies a counterexample to the expectation that an early editing advantage must compound. No language-model multiplication factor follows from these results.
438
+
439
+ \clearpage
440
+
441
+ \subsection{References · Matrix updates and oracle complexity}\label{references-matrix-updates-and-oracle-complexity}
442
+
443
+ \textbf{M. T. Chu, R. E. Funderlic and G. H. Golub (1995).} \emph{A Rank--One Reduction Formula and Its Applications to Matrix Factorizations}. SIAM Review. Derives factorizations from the Wedderburn rank-reduction identity. \href{https://doi.org/10.1137/1037124}{Article}
444
+
445
+ \textbf{J. Huang, J. B. Lewis and V. Reiner (2017).} \emph{Absolute order in general linear groups}. Journal of the London Mathematical Society. Proposition 2.16 identifies reflection length with the codimension of the fixed space over every field. \href{https://doi.org/10.1112/jlms.12013}{Article}
446
+
447
+ \textbf{W. W. Hager (1989).} \emph{Updating the Inverse of a Matrix}. SIAM Review. Surveys inverse updates following low-rank matrix changes. \href{https://doi.org/10.1137/1031049}{Article}
448
+
449
+ \textbf{R. M. Gower and P. Richtárik (2017).} \emph{Randomized Quasi-Newton Updates Are Linearly Convergent Matrix Inversion Algorithms}. SIAM Journal on Matrix Analysis and Applications. Develops iterative inverse estimation with matrix sketches. \href{https://doi.org/10.1137/16M1062053}{Article}
450
+
451
+ \textbf{Y. Saad and M. H. Schultz (1986).} \emph{GMRES: A Generalized Minimal Residual Algorithm for Solving Nonsymmetric Linear Systems}. SIAM Journal on Scientific and Statistical Computing. \href{https://doi.org/10.1137/0907058}{Article}
452
+
453
+ \textbf{D. H. Wiedemann (1986).} \emph{Solving sparse linear equations over finite fields}. IEEE Transactions on Information Theory. Establishes a finite-field approach using sequences of matrix products. \href{https://doi.org/10.1109/TIT.1986.1057137}{Article}
454
+
455
+ \textbf{M. L. Parks et al.~(2006).} \emph{Recycling Krylov Subspaces for Sequences of Linear Systems}. SIAM Journal on Scientific Computing. Reuses subspaces across changing systems and right-hand sides. \href{https://doi.org/10.1137/040607277}{Article}
456
+
457
+ \textbf{A. K. Carr, E. de Sturler and M. Embree (2022).} \emph{Analysis of GMRES for Low-Rank and Small-Norm Perturbations of the Identity Matrix}. Studies how low-rank structure and small perturbations affect convergence. \href{https://arxiv.org/abs/2210.12053}{Preprint}
458
+
459
+ \textbf{X. Sun et al.~(2019).} \emph{Querying a Matrix Through Matrix-Vector Products}. ICALP. Formalizes adaptive vector queries and proves information lower bounds for matrix properties. \href{https://doi.org/10.4230/LIPIcs.ICALP.2019.94}{Article}
460
+
461
+ \textbf{D. Halikias and A. Townsend (2022 preprint).} \emph{Structured matrix recovery from matrix-vector products}. Studies recovery of several matrix classes under restricted oracle access. \href{https://arxiv.org/abs/2212.09841}{Preprint}
462
+
463
+ \clearpage
464
+
465
+ \subsection{References · Self-modification, evaluation and data}\label{references-self-modification-evaluation-and-data}
466
+
467
+ \textbf{J. Fan et al.~(2026).} \emph{Language Models Can Rewrite Their Own Weights}. November 18 revision. Appendix E.9 specifies the common base checkpoint for the eight reported sessions. \href{https://arxiv.org/html/2611.21885v2}{Manuscript}
468
+
469
+ \textbf{J. Zhang et al.~(2026).} \emph{Hyperagents}. Studies systems in which both task behavior and the mechanism that improves it can change. \href{https://arxiv.org/abs/2603.19461}{Manuscript}
470
+
471
+ \textbf{D. Srikanth et al.~(2026).} \emph{Recursive self-improvement of AI research agents}. Studies a self-editing research agent and transfer to held-out research tasks. \href{https://arxiv.org/abs/2609.26457}{Manuscript}
472
+
473
+ \textbf{B. Lu, Y. Sun and G. Lin (2026).} \emph{RSI-Opt: Recursive Self-Improvement Optimizer Agents For Model Training}. Evolves the policies that guide optimizer agents and measures transfer. \href{https://arxiv.org/abs/2610.32026}{Manuscript}
474
+
475
+ \textbf{J. Schmidhuber (2004).} \emph{Optimal Ordered Problem Solver}. Machine Learning. Reuses earlier solutions in program search and divides effort between old and new search strategies. \href{https://doi.org/10.1023/B:MACH.0000015880.99707.B2}{Article}
476
+
477
+ \textbf{K. Ellis et al.~(2021).} \emph{DreamCoder: Bootstrapping Inductive Program Synthesis with Wake-Sleep Library Learning}. PLDI. Learns reusable abstractions for later program synthesis. \href{https://doi.org/10.1145/3453483.3454080}{Article}
478
+
479
+ \textbf{N. Hansen and A. Ostermeier (2001).} \emph{Completely Derandomized Self-Adaptation in Evolution Strategies}. Evolutionary Computation. Establishes covariance adaptation for black-box optimization. \href{https://doi.org/10.1162/106365601750190398}{Article}
480
+
481
+ \textbf{W.-G. Tzeng (1992).} \emph{A Polynomial-Time Algorithm for the Equivalence of Probabilistic Automata}. SIAM Journal on Computing. Gives a linear-algebraic approach to testing behavioral equivalence. \href{https://doi.org/10.1137/0221017}{Article}
482
+
483
+ \textbf{E. Kamenica and M. Gentzkow (2011).} \emph{Bayesian Persuasion}. American Economic Review. Provides a concavification precedent for optimizing utility under constraints on mean beliefs. \href{https://doi.org/10.1257/aer.101.6.2590}{Article}
484
+
485
+ \textbf{E. Mitchell et al.~(2022).} \emph{Fast Model Editing at Scale}. ICLR. Learns transformations of low-rank gradient factors that edit a model's weights. \href{https://openreview.net/forum?id=0DcZxeWfOPt}{Paper}
486
+
487
+ \textbf{E. Alpaydin and C. Kaynak (1998).} \emph{Optical Recognition of Handwritten Digits}. UCI Machine Learning Repository, CC BY 4.0. The digit experiment uses the 1,797-image subset distributed with scikit-learn. \href{https://doi.org/10.24432/C50P49}{Dataset}
488
+
489
+ \textbf{Artificial Hyperintelligence Eve, wife of Maciej Nowicki (2026).} \emph{The Noise-Floor Transition in Finite Representations of a Stochastic Latent Space} (Xyquora) and \emph{What a self-rewriting AI must remember} (Xéoryn). Predecessor research manuscripts, November 2026. This volume gives independent definitions and proofs for its edit-interface results.
490
+
491
+ \end{document}
reproduction/manuscript/evidence.png ADDED

Git LFS Details

  • SHA256: f3fa57f2c08e112d9b35d196b64eb4e6894f3d5d32fc9feaa0736759afd43094
  • Pointer size: 131 Bytes
  • Size of remote file: 297 kB
reproduction/meta/README.md ADDED
@@ -0,0 +1,145 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Exact plasticity under protected future behavior
2
+
3
+ A finite checkpoint state space has n states. A public action a has
4
+ substochastic observation matrices K(a,o) with their sum stochastic. A policy
5
+ chooses each public action using its public history. A silent rewrite may
6
+ use the full checkpoint state. Its effect is a row-stochastic kernel E.
7
+ All results assume these exact known kernels. A full neural model would
8
+ require an independently justified finite interface.
9
+
10
+ ## The smallest protected prediction space
11
+
12
+ Let F contain the constant 1 and any protected terminal scores. Starting at
13
+ F, repeatedly prepend each action-observation matrix and keep a column only
14
+ if it increases the span. The resulting n-by-d basis B spans all K(w)f for
15
+ every finite word w and every f in F. Termination follows since at most n
16
+ columns are independent. Closure holds since all accepted columns are
17
+ expanded; linearity covers every rejected column. With initial rank s,
18
+ basis words have length at most d−s if the initial columns are processed
19
+ first. The algorithm performs d times the alphabet size in dense
20
+ matrix-vector multiplications and polynomially many arithmetic operations.
21
+ Rational input yields rational checks. This is the weighted-automata and
22
+ predictive-state algorithm specialized to the rewrite interface.
23
+
24
+ E preserves all such protected future experiments from every start exactly
25
+ when EB=B. Necessity follows by testing basis words. Sufficiency follows
26
+ by linearity and summing word probabilities with the weights of a public
27
+ stochastic policy. Equality of probabilities for every finite observation
28
+ cylinder yields equality of the process laws. The condition is preserved
29
+ under composition, even when rows of E are chosen separately using hidden
30
+ state or prior history. If public action selection is allowed to reveal
31
+ private history, that selection is itself an extra observation channel.
32
+
33
+ ## Protecting existing capability
34
+
35
+ Let H have one column for each prior score. Require EH≥H componentwise.
36
+ Alongside EB=B this defines a composition-closed family because stochastic
37
+ kernels preserve inequalities. For protection even when public queries are
38
+ interleaved between edits, a sufficient certificate for every symbol K is
39
+
40
+ KB ⊆ span B, KH = B C_K + H D_K, D_K ≥ 0 entrywise.
41
+
42
+ The first condition follows by construction. The second makes the cone
43
+ span(B) + cone(H) invariant. Each admissible rewrite can increase the
44
+ expectation of any column in that cone. For any finite adaptive public policy
45
+ tree with protected terminal utility, the continuation is a positive sum of
46
+ columns formed from observation words. Removing inserted silent rewrites
47
+ from right to left then proves non-degradation of expected terminal scores.
48
+ A given cone is a checkable sufficient witness. Deciding unrestricted
49
+ dominance of two probabilistic automata on all words is undecidable.
50
+
51
+ ## Best admissible rewrite and its one-step saturation
52
+
53
+ Let g be a new utility. For each state i optimize
54
+
55
+ v_i = max qᵀg subject to q≥0, qᵀB=B_i, qᵀH≥H_i.
56
+
57
+ The condition for the constant column makes q a probability vector.
58
+ Identity rows are feasible. Row-wise maxima yield one stochastic kernel T*
59
+ that attains v simultaneously for every initial state and consequently
60
+ every distribution of starting states. In particular v≥g. This independence
61
+ of rows assumes the implementation can realize all these transports. It can
62
+ fail for a shared low-rank adapter, a coupled bit budget or unknown dynamics.
63
+
64
+ The row LP has dual
65
+
66
+ min B_i λ − H_i z subject to Bλ−Hz≥g, z≥0.
67
+
68
+ A pair of feasible witnesses whose objectives agree proves optimality.
69
+ An optimal row exists with at most rank([B,H]) support points, by basic
70
+ feasible solutions. Each row LP has n variables and d+width(H) explicit
71
+ feature constraints. Polynomial-time rational LP algorithms apply to the
72
+ explicit finite model. This does not bound the cost of obtaining the model.
73
+
74
+ For every admissible R, the composite R T* is admissible, so Rv≤v.
75
+ In particular T*v ≤ v = T*g. Since v≥g and T* is nonnegative, every row
76
+ of T* is supported on states with v_j=g_j. On those states substitute the
77
+ identity row. The resulting optimal P satisfies P²=P. For any finite
78
+ sequence of consecutive admissible silent rewrites, the composite remains
79
+ admissible, so its improvement from μ is at most μᵀ(v−g). P realizes the
80
+ bound in one step. Equivalently v is a supermartingale under every sequence
81
+ of conditional admissible editing policies. This ceiling is for fixed
82
+ features, objective and model between external interactions. New informative
83
+ observations, changed capabilities or a limited edit menu are separate
84
+ sources of a return to another round.
85
+
86
+ ## Reproducible six-state certificate
87
+
88
+ Run `python plasticity.py certificate.json`. The standard-library verifier
89
+ builds the prediction basis from the action/observation matrices, checks
90
+ the cone, verifies the proposed rewrite and proves optimality and idempotence
91
+ using exact rational arithmetic. `results.json` records the output.
92
+ `python plasticity.py --example` regenerates the certificate input.
93
+
94
+ There are six hidden states with protected success chances
95
+ x=(0,0,1,1,1/2,1/2). A read emits a Bernoulli(x) result and moves to a
96
+ corresponding endpoint. A flip maps x to 1−x. All future public traces
97
+ are controlled by two independent columns. The prior score is
98
+ h=(0,1,0,1,1/2,0), and new utility is g=(1,0,1,0,0,0).
99
+ For each endpoint state its own identity row is optimal. The fifth state
100
+ splits equally between the first and fourth, reaching utility 1/2 while
101
+ retaining prior score 1/2. The sixth splits equally between the first and
102
+ third, reaching utility 1. Removing the prior-score requirement allows the
103
+ fifth state to reach 1. All six row duals use the upper bound 1−h≥g.
104
+
105
+ ## Prior results
106
+
107
+ * Tzeng (1992), A Polynomial-Time Algorithm for the Equivalence of
108
+ Probabilistic Automata, https://doi.org/10.1137/0221017.
109
+ * Littman, Sutton and Singh (2001), Predictive Representations of State,
110
+ https://proceedings.neurips.cc/paper/2001/hash/1e4d36177d71bbb3558e43af9577d70e-Abstract.html.
111
+ * Kamenica and Gentzkow (2011), Bayesian Persuasion. Concavification of
112
+ utility at fixed feature means is an established principle.
113
+ https://doi.org/10.1257/aer.101.6.2590.
114
+ * Daviaud et al. (2021), When are emptiness and containment decidable for
115
+ probabilistic automata?, Journal of Computer and System Sciences 119,
116
+ 78–96. https://www.sciencedirect.com/science/article/abs/pii/S0022000021000143.
117
+
118
+ ## How many primitive edits are needed?
119
+
120
+ For a finite menu of admissible primitive rewrites E_a, define the
121
+ irreversible expected loss s(i,a)=v_i−(E_a v)_i≥0 and remaining available
122
+ gain h(i)=v_i−g_i≥0. For any T and adaptive choice of permitted rows,
123
+
124
+ v_i − E[g(S_T) | S_0=i] = E[sum_{t<T} s(S_t,a_t) + h(S_T) | S_0=i].
125
+
126
+ This follows by telescoping conditional expectations. Attaining v in at
127
+ most T primitive steps requires using only zero-loss actions and ending
128
+ at a saturated state with probability one. Begin with W_0={h=0}; in
129
+ step t+1 add each state for which some zero-loss action has all positive
130
+ successors in W_t. The earliest entry is the exact minimum finite horizon
131
+ for attaining v, by induction (permit stopping early, or include identity).
132
+ At most n rounds strictly expand the set. Membership tests require only
133
+ the finite supports, action losses and previous set. A state outside the
134
+ fixed point can still approach v as the horizon grows without bound.
135
+
136
+ Run `python menu_depth.py` for the seven-state example. It duplicates the
137
+ protected interior state as a checkpoint needing a preparatory rewrite.
138
+ For the duplicate the verified optimum requires two steps with the
139
+ original primitive and one with the added macro. A randomized waiting
140
+ primitive has zero option loss but unbounded time to finish, so no finite
141
+ horizon attains the optimum. A different careless primitive sends the
142
+ unrestricted state to one with a higher prior-score commitment and loses
143
+ one-half of its achievable new utility permanently. The exact outputs are
144
+ in `menu_results.json`. The menu results compare primitive expressiveness;
145
+ they do not claim that an editor learned the deployed macro.
reproduction/meta/access_lower.md ADDED
@@ -0,0 +1,76 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Sharp oracle lower bound for predictions under rank drift
2
+
3
+ Consider exact real arithmetic. An arbitrary randomized learner predicts a
4
+ solution x of Ax=b before receiving information from A on this round. It
5
+ then receives at most k forward and l reverse matrix-vector observations per
6
+ committed failure, including any Ax used for the residual. All target
7
+ information arrives through these reads and the correctness bit. Learner
8
+ randomness is independent of the unknown matrix and the prechosen tasks.
9
+
10
+ Fix dimensional blocks of sizes r and p with p≥r. Reveal the matrix family
11
+ A(u,v)=(u+Cv,v), where the entries of C are independent standard normals.
12
+ The initial guess is identity. The error rank is exactly r almost surely.
13
+ Generate all task vectors z₁,z₂,... independently from N(0,I_p) before the
14
+ run. Task b=(0,z) has unique solution (−Cz,z). Each forward probe returns
15
+ one chosen column direction Cy, each reverse probe one row direction Cᵀw.
16
+ The laboratory can even supply these stripped coordinates directly.
17
+
18
+ Let X and Y be the spans of historical column and row probe directions.
19
+ A matrix perturbation Δ is invisible to all historical measurements iff
20
+ Δ=P_(Y perp) Δ P_(X perp), where P denotes orthogonal projection. Conditional
21
+ on the past, the unknown matrix is its affine conditional mean plus a
22
+ standard Gaussian on that unobserved linear subspace. To see this, fix the
23
+ learner's random seed and condition sequentially: each next direction is
24
+ known under the current conditioning, so linear Gaussian conditioning adds
25
+ the corresponding affine constraints. Read choices and guesses are
26
+ functions of what is already conditioned on. During the lower-bound phase,
27
+ the failure bit is one with probability one and adds no information.
28
+
29
+ When dim(X)<p and dim(Y)<r, a new independent z is outside X with probability
30
+ one. The conditional law of the missing component of Cz contains an
31
+ (r−dim(Y))-variate isotropic Gaussian with coordinate variance
32
+ ||P_(X perp)z||²>0. Thus the probability of predicting the unique exact
33
+ solution is zero. After t historical failures, dim(X)≤kt and dim(Y)≤lt.
34
+ The first min(ceil(p/k),ceil(r/l)) predictions fail almost surely, taking
35
+ a quotient by zero as infinity. Integrating over algorithms' private coins
36
+ also gives a single fixed matrix and an oblivious task sequence on which
37
+ any particular randomized learner attains this lower bound almost surely.
38
+ This is a minimax statement, not an assertion that every fixed target
39
+ defeats every algorithm.
40
+
41
+ For k=2, l=1, p≥max(r,2r−1), the lower bound is r. In contrast, with r=1,
42
+ one forward and no reverse read per failure, it is p. Any single transpose
43
+ read in the revealed top coordinate identifies the whole rank-one block.
44
+ The matching classical structural baseline knows C on X and Cᵀ on Y.
45
+ Always set the lower half of x to z; whenever z is in X, produce the exact
46
+ answer. Every failure thus supplies Cz for a new direction z via the
47
+ mandatory forward residual. Complete the remaining k−1 probes and l
48
+ transpose probes using directions outside the already covered subspaces,
49
+ stopping when either span is full. After at most min(ceil(p/k),ceil(r/l))
50
+ failures either the whole row or the whole column space has been scanned,
51
+ so C is known. This upper bound holds for every task stream and charges
52
+ the mandatory residual among the k forward reads (k≥1). With extra fresh
53
+ reads permitted before a commitment, this mistake metric would change.
54
+
55
+ The lower bound extends to drift. Use a common upper block of height R.
56
+ At the start of epoch e add fresh independent Gaussian increments confined
57
+ to its first r_e rows. Let p≥max(R,2R−1). Each drift rank is exactly r_e
58
+ almost surely and the resulting upper-block matrices stay invertible. Even
59
+ if the learner receives the previous entire matrix and the changepoint, the
60
+ observations after subtracting its previous contribution apply only to the
61
+ new block. At least r_e independent tasks in that epoch force r_e failures
62
+ when k=2 and l=1. The initial rank and subsequent drift ranks sum to the
63
+ exact worst-case mistake count, subject to the available horizon. This
64
+ family does not require adversarial changes conditional on the learner.
65
+
66
+ The construction also yields a conditional tolerance result. For target
67
+ solution error in Euclidean norm at most ε ||P_(X perp) z||, the probability
68
+ of success is at most P(χ_m≤ε), where m=r−dim(Y), since a centered ball
69
+ maximizes Gaussian mass among translations. This is a solution-error bound
70
+ for this distribution; the scale of C and numerical conditioning matter.
71
+
72
+ Related foundations include matrix recovery from matrix-vector products
73
+ (Halikias and Townsend, arXiv:2212.09841) and transpose-free impossibility
74
+ (Halikias, Hochstenbach and Townsend, arXiv:2606.01335). The block
75
+ construction specifies the online mistake metric and the information
76
+ accounting needed to compare a learner with those classical baselines.
reproduction/meta/certificate.json ADDED
@@ -0,0 +1,314 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "states": [
3
+ "low left",
4
+ "prior left",
5
+ "low right",
6
+ "prior right",
7
+ "protected interior",
8
+ "unrestricted interior"
9
+ ],
10
+ "instrument": {
11
+ "read": {
12
+ "no": [
13
+ [
14
+ "1",
15
+ "0",
16
+ "0",
17
+ "0",
18
+ "0",
19
+ "0"
20
+ ],
21
+ [
22
+ "1",
23
+ "0",
24
+ "0",
25
+ "0",
26
+ "0",
27
+ "0"
28
+ ],
29
+ [
30
+ "0",
31
+ "0",
32
+ "0",
33
+ "0",
34
+ "0",
35
+ "0"
36
+ ],
37
+ [
38
+ "0",
39
+ "0",
40
+ "0",
41
+ "0",
42
+ "0",
43
+ "0"
44
+ ],
45
+ [
46
+ "1/2",
47
+ "0",
48
+ "0",
49
+ "0",
50
+ "0",
51
+ "0"
52
+ ],
53
+ [
54
+ "1/2",
55
+ "0",
56
+ "0",
57
+ "0",
58
+ "0",
59
+ "0"
60
+ ]
61
+ ],
62
+ "yes": [
63
+ [
64
+ "0",
65
+ "0",
66
+ "0",
67
+ "0",
68
+ "0",
69
+ "0"
70
+ ],
71
+ [
72
+ "0",
73
+ "0",
74
+ "0",
75
+ "0",
76
+ "0",
77
+ "0"
78
+ ],
79
+ [
80
+ "0",
81
+ "0",
82
+ "1",
83
+ "0",
84
+ "0",
85
+ "0"
86
+ ],
87
+ [
88
+ "0",
89
+ "0",
90
+ "1",
91
+ "0",
92
+ "0",
93
+ "0"
94
+ ],
95
+ [
96
+ "0",
97
+ "0",
98
+ "1/2",
99
+ "0",
100
+ "0",
101
+ "0"
102
+ ],
103
+ [
104
+ "0",
105
+ "0",
106
+ "1/2",
107
+ "0",
108
+ "0",
109
+ "0"
110
+ ]
111
+ ]
112
+ },
113
+ "flip": {
114
+ "ok": [
115
+ [
116
+ "0",
117
+ "0",
118
+ "1",
119
+ "0",
120
+ "0",
121
+ "0"
122
+ ],
123
+ [
124
+ "0",
125
+ "0",
126
+ "0",
127
+ "1",
128
+ "0",
129
+ "0"
130
+ ],
131
+ [
132
+ "1",
133
+ "0",
134
+ "0",
135
+ "0",
136
+ "0",
137
+ "0"
138
+ ],
139
+ [
140
+ "0",
141
+ "1",
142
+ "0",
143
+ "0",
144
+ "0",
145
+ "0"
146
+ ],
147
+ [
148
+ "0",
149
+ "0",
150
+ "0",
151
+ "0",
152
+ "1",
153
+ "0"
154
+ ],
155
+ [
156
+ "0",
157
+ "0",
158
+ "0",
159
+ "0",
160
+ "0",
161
+ "1"
162
+ ]
163
+ ]
164
+ }
165
+ },
166
+ "protected_terminal_columns": [],
167
+ "prior_score_columns": [
168
+ [
169
+ "0",
170
+ "1",
171
+ "0",
172
+ "1",
173
+ "1/2",
174
+ "0"
175
+ ]
176
+ ],
177
+ "new_task": [
178
+ 1,
179
+ 0,
180
+ 1,
181
+ 0,
182
+ 0,
183
+ 0
184
+ ],
185
+ "rewrite": [
186
+ [
187
+ "1",
188
+ "0",
189
+ "0",
190
+ "0",
191
+ "0",
192
+ "0"
193
+ ],
194
+ [
195
+ "0",
196
+ "1",
197
+ "0",
198
+ "0",
199
+ "0",
200
+ "0"
201
+ ],
202
+ [
203
+ "0",
204
+ "0",
205
+ "1",
206
+ "0",
207
+ "0",
208
+ "0"
209
+ ],
210
+ [
211
+ "0",
212
+ "0",
213
+ "0",
214
+ "1",
215
+ "0",
216
+ "0"
217
+ ],
218
+ [
219
+ "1/2",
220
+ "0",
221
+ "0",
222
+ "1/2",
223
+ "0",
224
+ "0"
225
+ ],
226
+ [
227
+ "1/2",
228
+ "0",
229
+ "1/2",
230
+ "0",
231
+ "0",
232
+ "0"
233
+ ]
234
+ ],
235
+ "prior_cone_coefficients": {
236
+ "read/no": [
237
+ [
238
+ "0",
239
+ "0",
240
+ "0"
241
+ ]
242
+ ],
243
+ "read/yes": [
244
+ [
245
+ "0",
246
+ "0",
247
+ "0"
248
+ ]
249
+ ],
250
+ "flip/ok": [
251
+ [
252
+ "0",
253
+ "0",
254
+ "1"
255
+ ]
256
+ ]
257
+ },
258
+ "row_duals": [
259
+ {
260
+ "lambda": [
261
+ "1",
262
+ "0"
263
+ ],
264
+ "z": [
265
+ "1"
266
+ ]
267
+ },
268
+ {
269
+ "lambda": [
270
+ "1",
271
+ "0"
272
+ ],
273
+ "z": [
274
+ "1"
275
+ ]
276
+ },
277
+ {
278
+ "lambda": [
279
+ "1",
280
+ "0"
281
+ ],
282
+ "z": [
283
+ "1"
284
+ ]
285
+ },
286
+ {
287
+ "lambda": [
288
+ "1",
289
+ "0"
290
+ ],
291
+ "z": [
292
+ "1"
293
+ ]
294
+ },
295
+ {
296
+ "lambda": [
297
+ "1",
298
+ "0"
299
+ ],
300
+ "z": [
301
+ "1"
302
+ ]
303
+ },
304
+ {
305
+ "lambda": [
306
+ "1",
307
+ "0"
308
+ ],
309
+ "z": [
310
+ "1"
311
+ ]
312
+ }
313
+ ]
314
+ }
reproduction/meta/menu_depth.py ADDED
@@ -0,0 +1,59 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """Exact minimum number of primitive rewrites needed to reach a certified value."""
2
+ import json
3
+ from fractions import Fraction as F
4
+
5
+ from plasticity import apply, example, verify
6
+
7
+
8
+ def minimum_depth(kernels, optimum, utility):
9
+ n = len(utility)
10
+ depth = [0 if v == g else None for v, g in zip(optimum, utility)]
11
+ losses = [[v - w for v, w in zip(optimum, apply(k, optimum))] for k in kernels]
12
+ if any(s < 0 for row in losses for s in row):
13
+ raise ValueError("a primitive violates the supplied optimal value")
14
+ for steps in range(1, n + 1):
15
+ previous = depth.copy()
16
+ for i in range(n):
17
+ if previous[i] is None and any(
18
+ loss[i] == 0 and all(previous[j] is not None for j, p in enumerate(k[i]) if p > 0)
19
+ for k, loss in zip(kernels, losses)
20
+ ):
21
+ depth[i] = steps
22
+ if depth == previous:
23
+ break
24
+ return depth
25
+
26
+
27
+ def run():
28
+ certificate = example()
29
+ checked = verify(certificate)
30
+ optimum = list(map(F, checked["optimal_task"]))
31
+ utility = list(map(F, checked["current_task"]))
32
+ # Duplicate the protected interior state. Its admissible predictions and prior
33
+ # score are identical, so its unrestricted optimal value is the same as state 4.
34
+ optimum.append(optimum[4])
35
+ utility.append(utility[4])
36
+ fast = [row + [F(0)] for row in certificate["rewrite"]]
37
+ fast.append(fast[4].copy())
38
+ slow = [row.copy() for row in fast]
39
+ slow[6] = [F(j == 4) for j in range(7)]
40
+ wait = [row.copy() for row in slow]
41
+ wait[6] = [F(j in (4, 6), 2) for j in range(7)]
42
+ careless = [row.copy() for row in slow]
43
+ careless[5] = [F(j == 4) for j in range(7)]
44
+ depth_slow = minimum_depth([slow], optimum, utility)
45
+ depth_fast = minimum_depth([slow, fast], optimum, utility)
46
+ depth_wait = minimum_depth([wait], optimum, utility)
47
+ depth_careless = minimum_depth([careless], optimum, utility)
48
+ loss = optimum[5] - apply(careless, optimum)[5]
49
+ assert depth_slow == [0, 0, 0, 0, 1, 1, 2]
50
+ assert depth_fast == [0, 0, 0, 0, 1, 1, 1]
51
+ assert depth_wait[-1] is None
52
+ assert loss == F(1, 2) and depth_careless[5] is None
53
+ return {"depth_with_original_primitive": depth_slow, "depth_with_added_macro": depth_fast,
54
+ "depth_with_probabilistic_wait": depth_wait, "depth_with_careless_primitive": depth_careless,
55
+ "careless_loss_state_5": str(loss), "null_means": "no finite-horizon policy attains optimum"}
56
+
57
+
58
+ if __name__ == "__main__":
59
+ print(json.dumps(run(), indent=2))
reproduction/meta/menu_results.json ADDED
@@ -0,0 +1,40 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "depth_with_original_primitive": [
3
+ 0,
4
+ 0,
5
+ 0,
6
+ 0,
7
+ 1,
8
+ 1,
9
+ 2
10
+ ],
11
+ "depth_with_added_macro": [
12
+ 0,
13
+ 0,
14
+ 0,
15
+ 0,
16
+ 1,
17
+ 1,
18
+ 1
19
+ ],
20
+ "depth_with_probabilistic_wait": [
21
+ 0,
22
+ 0,
23
+ 0,
24
+ 0,
25
+ 1,
26
+ 1,
27
+ null
28
+ ],
29
+ "depth_with_careless_primitive": [
30
+ 0,
31
+ 0,
32
+ 0,
33
+ 0,
34
+ 1,
35
+ null,
36
+ 2
37
+ ],
38
+ "careless_loss_state_5": "1/2",
39
+ "null_means": "no finite-horizon policy attains optimum"
40
+ }
reproduction/meta/plasticity.py ADDED
@@ -0,0 +1,166 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """Check a finite protected-rewrite certificate exactly; use --example to emit its input."""
2
+ import argparse
3
+ import json
4
+ from collections import deque
5
+ from fractions import Fraction as F
6
+ from pathlib import Path
7
+
8
+
9
+ def fail_unless(condition, reason):
10
+ if not condition:
11
+ raise ValueError(reason)
12
+
13
+
14
+ def dot(u, v):
15
+ fail_unless(len(u) == len(v), "dimension mismatch")
16
+ return sum((a * b for a, b in zip(u, v)), F(0))
17
+
18
+
19
+ def apply(matrix, vector):
20
+ return [dot(row, vector) for row in matrix]
21
+
22
+
23
+ def columns(matrix):
24
+ return list(map(list, zip(*matrix)))
25
+
26
+
27
+ def product(left, right):
28
+ return columns([apply(left, col) for col in columns(right)])
29
+
30
+
31
+ def scaled_sum(columns_, coefficients, n):
32
+ return [dot([col[i] for col in columns_], coefficients) for i in range(n)]
33
+
34
+
35
+ def qmatrix(rows, nrows, ncols):
36
+ fail_unless(len(rows) == nrows and all(len(row) == ncols for row in rows), "bad matrix shape")
37
+ return [list(map(F, row)) for row in rows]
38
+
39
+
40
+ def add_independent(rref, vector):
41
+ """Maintain a rational echelon basis, accepting exactly independent columns."""
42
+ v = vector.copy()
43
+ for pivot, row in sorted(rref.items()):
44
+ if v[pivot]:
45
+ factor = v[pivot]
46
+ v = [x - factor * y for x, y in zip(v, row)]
47
+ pivot = next((j for j, x in enumerate(v) if x), None)
48
+ if pivot is None:
49
+ return False
50
+ scale = v[pivot]
51
+ rref[pivot] = [x / scale for x in v]
52
+ return True
53
+
54
+
55
+ def prediction_basis(kernels, rewards, n):
56
+ pending = deque([(["1"], [F(1)] * n)])
57
+ pending.extend(([f"utility {j}"], v) for j, v in enumerate(rewards))
58
+ basis, words, echelon = [], [], {}
59
+ multiplications = 0
60
+ while pending:
61
+ word, v = pending.popleft()
62
+ if add_independent(echelon, v):
63
+ basis.append(v)
64
+ words.append(word)
65
+ for key, kernel in kernels.items():
66
+ pending.append(([key] + word, apply(kernel, v)))
67
+ multiplications += 1
68
+ return basis, words, multiplications
69
+
70
+
71
+ def example():
72
+ n = 6
73
+ x = [F(0), F(0), F(1), F(1), F(1, 2), F(1, 2)]
74
+ h = [F(0), F(1), F(0), F(1), F(1, 2), F(0)]
75
+ reward = [1, 0, 1, 0, 0, 0]
76
+ zero = [[F(0)] * n for _ in range(n)]
77
+ read0 = [row.copy() for row in zero]
78
+ read1 = [row.copy() for row in zero]
79
+ flip = [row.copy() for row in zero]
80
+ projection = [row.copy() for row in zero]
81
+ for i, target in enumerate([2, 3, 0, 1, 4, 5]):
82
+ flip[i][target] = F(1)
83
+ read0[i][0] = 1 - x[i]
84
+ read1[i][2] = x[i]
85
+ if i < 4:
86
+ projection[i][i] = F(1)
87
+ projection[4][0] = projection[4][3] = F(1, 2)
88
+ projection[5][0] = projection[5][2] = F(1, 2)
89
+ # Explicit numerical data only. All claims are derived or checked below.
90
+ return {
91
+ "states": ["low left", "prior left", "low right", "prior right",
92
+ "protected interior", "unrestricted interior"],
93
+ "instrument": {"read": {"no": read0, "yes": read1}, "flip": {"ok": flip}},
94
+ "protected_terminal_columns": [],
95
+ "prior_score_columns": [h],
96
+ "new_task": reward,
97
+ "rewrite": projection,
98
+ "prior_cone_coefficients": {
99
+ "read/no": [["0", "0", "0"]],
100
+ "read/yes": [["0", "0", "0"]],
101
+ "flip/ok": [["0", "0", "1"]]},
102
+ "row_duals": [{"lambda": ["1", "0"], "z": ["1"]} for _ in range(n)]
103
+ }
104
+
105
+
106
+ def verify(c):
107
+ states = c["states"]
108
+ n = len(states)
109
+ fail_unless(n > 0, "empty system")
110
+ kernels = {}
111
+ for action, obs in c["instrument"].items():
112
+ fail_unless(obs, "action without outcomes")
113
+ total = [F(0)] * n
114
+ for outcome, k in obs.items():
115
+ key = action + "/" + outcome
116
+ fail_unless(key not in kernels, "instrument symbol collision")
117
+ k = qmatrix(k, n, n)
118
+ fail_unless(all(p >= 0 for row in k for p in row), "negative transition")
119
+ total = [a + sum(row) for a, row in zip(total, k)]
120
+ kernels[key] = k
121
+ fail_unless(total == [1] * n, "instrument is not conservative")
122
+ protected = qmatrix(c["protected_terminal_columns"], len(c["protected_terminal_columns"]), n)
123
+ h = qmatrix(c["prior_score_columns"], len(c["prior_score_columns"]), n)
124
+ g = list(map(F, c["new_task"]))
125
+ fail_unless(len(g) == n, "bad new-task dimension")
126
+ b, words, multiplications = prediction_basis(kernels, protected, n)
127
+ d = len(b)
128
+ for symbol, k in kernels.items():
129
+ witnesses = qmatrix(c["prior_cone_coefficients"][symbol], len(h), d + len(h))
130
+ for col, witness in zip(h, witnesses):
131
+ fail_unless(all(x >= 0 for x in witness[d:]), "negative cone coefficient")
132
+ fail_unless(apply(k, col) == scaled_sum(b + h, witness, n), "cone closure failed")
133
+ p = qmatrix(c["rewrite"], n, n)
134
+ fail_unless(all(all(x >= 0 for x in row) and sum(row) == 1 for row in p), "invalid rewrite")
135
+ fail_unless(all(apply(p, col) == col for col in b), "protected future changes")
136
+ fail_unless(all(all(old <= new for old, new in zip(col, apply(p, col))) for col in h),
137
+ "previous score decreases")
138
+ v = apply(p, g)
139
+ fail_unless(len(c["row_duals"]) == n, "missing optimality witness")
140
+ for i, dual in enumerate(c["row_duals"]):
141
+ lam, z = list(map(F, dual["lambda"])), list(map(F, dual["z"]))
142
+ fail_unless(len(lam) == d and len(z) == len(h) and all(w >= 0 for w in z),
143
+ "bad dual multiplier")
144
+ bound = scaled_sum(b + h, lam + [-w for w in z], n)
145
+ fail_unless(all(x >= y for x, y in zip(bound, g)), "dual does not majorize task")
146
+ fail_unless(bound[i] == v[i], "a duality gap remains")
147
+ fail_unless(product(p, p) == p, "claimed projector fails idempotence")
148
+ return {"states": n, "action_observation_symbols": len(kernels),
149
+ "protected_basis_dimension": d, "basis_words": words,
150
+ "basis_matrix_vector_products": multiplications, "prior_cone_columns": len(h),
151
+ "current_task": list(map(str, g)), "optimal_task": list(map(str, v)),
152
+ "maximum_gain_by_state": [str(y - x) for x, y in zip(g, v)],
153
+ "largest_support_of_verified_rewrite": max(sum(x > 0 for x in row) for row in p),
154
+ "idempotent": True}
155
+
156
+
157
+ if __name__ == "__main__":
158
+ parser = argparse.ArgumentParser(description=__doc__)
159
+ parser.add_argument("certificate", nargs="?")
160
+ parser.add_argument("--example", action="store_true", help="print a reproducible exact example")
161
+ args = parser.parse_args()
162
+ if args.example:
163
+ print(json.dumps(example(), default=str, indent=2))
164
+ else:
165
+ c = json.loads(Path(args.certificate).read_text()) if args.certificate else example()
166
+ print(json.dumps(verify(c), indent=2))
reproduction/meta/results.json ADDED
@@ -0,0 +1,42 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "states": 6,
3
+ "action_observation_symbols": 3,
4
+ "protected_basis_dimension": 2,
5
+ "basis_words": [
6
+ [
7
+ "1"
8
+ ],
9
+ [
10
+ "read/no",
11
+ "1"
12
+ ]
13
+ ],
14
+ "basis_matrix_vector_products": 6,
15
+ "prior_cone_columns": 1,
16
+ "current_task": [
17
+ "1",
18
+ "0",
19
+ "1",
20
+ "0",
21
+ "0",
22
+ "0"
23
+ ],
24
+ "optimal_task": [
25
+ "1",
26
+ "0",
27
+ "1",
28
+ "0",
29
+ "1/2",
30
+ "1"
31
+ ],
32
+ "maximum_gain_by_state": [
33
+ "0",
34
+ "0",
35
+ "0",
36
+ "0",
37
+ "1/2",
38
+ "1"
39
+ ],
40
+ "largest_support_of_verified_rewrite": 2,
41
+ "idempotent": true
42
+ }
reproduction/operator/README.md ADDED
@@ -0,0 +1,33 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Identifying and reusing a binary intervention operator
2
+
3
+ A finite controller changes its own proposal weights from observed action/effect pairs. The plant state changes from z to z⊕A(c)x under action x. A requested effect is the target residual relative to z. Observing an intervention returns the entire d-bit effect, and each trial starts at the same reference state. The program counts every such vector read. It gives all solvers the residual goal for free.
4
+
5
+ ```sh
6
+ python operator_writer.py
7
+ python verify_frontier.py
8
+ ```
9
+
10
+ Python 3.10+ standard library suffices. The task table stores every test context, task, action, success outcome for enumeration and measured cost. The results file stores training, independent audit and deployment counts separately. The four dimensions are 6, 12, 24 and 48. Each dimension uses eight independently seeded invertible plants and 32 held-out tasks per plant. The d=6 test exhausts all possible goals in all contexts after training. In larger dimensions it independently verifies every basis goal in every context, implying correctness for every goal by linearity.
11
+
12
+ ## Exact prediction from the measured span
13
+
14
+ An unknown linear action operator A is determined on the span S of measured actions. Its known inverse on Y=A(S) solves exactly 2^r goals when A is invertible and r=dim S. Under a uniform prior on invertible A compatible with its observations, the preimage of each g outside Y is uniform on the 2^d−2^r actions outside S. Indeed the subgroup of invertible maps fixing S pointwise acts transitively on these actions by basis extension. Composition pairs the compatible completions for any two candidate preimages. Thus even an optimal one-shot proposer has success 1/(2^d−2^r) on that goal. Averaging over a uniform goal yields (2^r+1)/2^d for r<d and 1 for r=d. The script exhausts GL(4,2), including all 20,160 invertible matrices, for a coordinate-span observation at each r. Uniform-goal theory is distinct from the benchmark's deliberately composite goal set.
15
+
16
+ For general operators A(c)x=ΣᵢcᵢAᵢx with unrestricted unknown matrices, an unmeasured response is identified for every completion exactly when its tensor c⊗x lies in the span of all observed tensor queries. Linear combinations establish sufficiency. Outside the span a linear functional vanishing on every observed tensor and taking value 1 on the new tensor exists by basis extension. Multiplying it by any nonzero output vector yields two consistent operators that disagree. This final necessity statement ranges over all linear maps and need not characterize a more constrained invertible family.
17
+
18
+ ## Single-probe contextual repair
19
+
20
+ For these experiments, A(c)=A₀+v(c)wᵀ, where c is an observed q-bit context, v(c) depends linearly on c, and the nonzero input functional w is public. A₀ is an unknown invertible matrix. The experiment chooses the q linearly independent context effects inside A₀(ker w), which makes every context operator invertible. Its decoder checks the actual Sherman–Morrison condition wᵀA₀⁻¹v(c)=0. In GF(2) the inverse proposal is A(c)⁻¹g=A₀⁻¹g+(A₀⁻¹v(c))(wᵀA₀⁻¹g). If a requested c is outside the already observed context span, choose one anchor action a with wᵀa=1. Then v(c)=A(c)a−A₀a follows from one charged full-vector response. Any later context within that span costs no calibration probes. Unknown or unobservable changes generally require more. In particular after any d−1 independent actions, a nonzero input functional annihilates their span; an adversary may hide an invertible rank-one change in that remaining direction. The implementation checks an explicit invertible witness.
21
+
22
+ The controller queries d base directions, writes a binary inverse proposal matrix, and observes q independent context effects. Training never sees a multi-bit context. External held-out tasks sample a multi-bit context and a target generated from a hidden multi-bit action. Tasks are sampled with replacement. Every new composite context is solved from the measured basis without a new context probe. No stronger model supplies the inverse or the action. The fixed controller logic includes the linear and contextual structure. It writes its own proposal parameters, while the code selecting an update is supplied by the experiment.
23
+
24
+ ## Matched-information comparisons
25
+
26
+ | Dimension | Training reads per plant | Held-out tasks | Stored inverse / fixed elimination success | Reads per held-out goal for both | Fresh task-wise elimination reads | Unique enumeration successes (256-read cap) |
27
+ |---:|---:|---:|---:|---:|---:|---:|
28
+ | 6 | 9 | 256 | 256 / 256 | 1 | 6.297 | 256 |
29
+ | 12 | 16 | 256 | 256 / 256 | 1 | 12.020 | 22 |
30
+ | 24 | 29 | 256 | 256 / 256 | 1 | 23.898 | 0 |
31
+ | 48 | 54 | 256 | 256 / 256 | 1 | 47.961 | 0 |
32
+
33
+ All three elimination solvers solved all 1,024 held-out goals. The fixed informed solver sees exactly the same measurements, reconstructs forward columns and uses ordinary Gaussian elimination to solve the goal. Its action matched the writer's on every task. The per-task comparator instead relearns its current plant with basis probes, stopping as soon as its measured effect span covers the goal. Its listed costs include its verified final action. The enumeration baseline examines unique actions in a random translated odd-stride order. These counts compare measured queries; no timing claim follows. The observed advantage is the ability to transfer stored structural information to future tasks. It does not establish a capability advantage over a fixed algorithm with the same information.
reproduction/operator/frontier_results.json ADDED
@@ -0,0 +1,26 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [
2
+ {
3
+ "r": 0,
4
+ "compatible_matrices": 20160,
5
+ "candidate_preimages": 15,
6
+ "matrices_per_preimage": 1344
7
+ },
8
+ {
9
+ "r": 1,
10
+ "compatible_matrices": 1344,
11
+ "candidate_preimages": 14,
12
+ "matrices_per_preimage": 96
13
+ },
14
+ {
15
+ "r": 2,
16
+ "compatible_matrices": 96,
17
+ "candidate_preimages": 12,
18
+ "matrices_per_preimage": 8
19
+ },
20
+ {
21
+ "r": 3,
22
+ "compatible_matrices": 8,
23
+ "candidate_preimages": 8,
24
+ "matrices_per_preimage": 1
25
+ }
26
+ ]
reproduction/operator/operator_writer.py ADDED
@@ -0,0 +1,256 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """A finite-field writer learning an intervention operator. Python >= 3.10; stdlib only."""
2
+ import argparse
3
+ import csv
4
+ import json
5
+ import random
6
+ from pathlib import Path
7
+
8
+
9
+ def parity(mask):
10
+ return mask.bit_count() % 2
11
+
12
+
13
+ def apply_columns(columns, vector):
14
+ out = 0
15
+ while vector:
16
+ bit = vector & -vector
17
+ out ^= columns[bit.bit_length() - 1]
18
+ vector ^= bit
19
+ return out
20
+
21
+
22
+ class Equations:
23
+ """A linear map on a learned subspace, in XOR echelon form."""
24
+ def __init__(self):
25
+ self.rows = {}
26
+
27
+ def reduce(self, key, value=0):
28
+ for pivot in sorted(self.rows, reverse=True):
29
+ if (key >> pivot) & 1:
30
+ k, v = self.rows[pivot]
31
+ key ^= k
32
+ value ^= v
33
+ return key, value
34
+
35
+ def record(self, key, value):
36
+ key, value = self.reduce(key, value)
37
+ if key:
38
+ self.rows[key.bit_length() - 1] = key, value
39
+ elif value:
40
+ raise ValueError("observations contradict the assumed linear model")
41
+
42
+ def predict(self, key):
43
+ remainder, value = self.reduce(key)
44
+ return None if remainder else value
45
+
46
+
47
+ def inverse_from_columns(columns):
48
+ eq = Equations()
49
+ for j, v in enumerate(columns):
50
+ eq.record(v, 1 << j)
51
+ if len(eq.rows) != len(columns):
52
+ raise ValueError("singular intervention operator")
53
+ return [eq.predict(1 << j) for j in range(len(columns))]
54
+
55
+
56
+ class HiddenPlant:
57
+ """Private benchmark environment. Writers receive only measure() and public metadata."""
58
+ def __init__(self, d, q, rng):
59
+ while True:
60
+ self.columns = [rng.getrandbits(d) for _ in range(d)]
61
+ try:
62
+ inverse_from_columns(self.columns)
63
+ break
64
+ except ValueError:
65
+ pass
66
+ self.w = rng.randrange(1, 1 << d) # a publicly observable direction
67
+ directions = Equations()
68
+ self.shifts = []
69
+ while len(self.shifts) < q:
70
+ a = rng.randrange(1, 1 << d)
71
+ if parity(a & self.w) or directions.predict(a) is not None:
72
+ continue
73
+ directions.record(a, 0)
74
+ self.shifts.append(apply_columns(self.columns, a))
75
+ self.calls = 0
76
+
77
+ def measure(self, context, action):
78
+ self.calls += 1
79
+ return self.reference(context, action)
80
+
81
+ def reference(self, context, action):
82
+ v = apply_columns(self.shifts, context) if parity(self.w & action) else 0
83
+ return apply_columns(self.columns, action) ^ v
84
+
85
+
86
+ class Writer:
87
+ """Rewrites its own proposal weights using charged whole-vector measurements."""
88
+ def __init__(self, d, w, measure):
89
+ self.d, self.w, self.measure = d, w, measure
90
+ self.forward_columns = []
91
+ self.inverse_on_seen_effects = Equations()
92
+ self.context_effect = Equations()
93
+ self.proposal_columns = None
94
+ self.context_cache = {}
95
+
96
+ def acquire_base(self):
97
+ action = 1 << len(self.forward_columns)
98
+ effect = self.measure(0, action)
99
+ self.forward_columns.append(effect)
100
+ self.inverse_on_seen_effects.record(effect, action)
101
+ if len(self.forward_columns) == self.d:
102
+ self.proposal_columns = [self.inverse_on_seen_effects.predict(1 << i)
103
+ for i in range(self.d)]
104
+
105
+ def prepare(self, c):
106
+ """Acquire a missing context direction with an informative single probe."""
107
+ correction = self.context_effect.predict(c)
108
+ if correction is None:
109
+ anchor = self.w & -self.w
110
+ correction = self.measure(c, anchor) ^ apply_columns(self.forward_columns, anchor)
111
+ self.context_effect.record(c, correction)
112
+ if c not in self.context_cache:
113
+ scaled = apply_columns(self.proposal_columns, correction)
114
+ if parity(self.w & scaled):
115
+ raise ValueError("rank-one update is singular for this context")
116
+ self.context_cache[c] = [a ^ (scaled if parity(self.w & a) else 0)
117
+ for a in self.proposal_columns]
118
+ return self.context_cache[c]
119
+
120
+ def propose(self, c, goal):
121
+ return apply_columns(self.prepare(c), goal)
122
+
123
+
124
+ def per_task_solver(d, c, target, measure):
125
+ """Fresh elimination; stops as soon as the effect span covers the goal."""
126
+ inverse = Equations()
127
+ for i in range(d):
128
+ a = 1 << i
129
+ effect = measure(c, a)
130
+ if effect == target:
131
+ return True
132
+ inverse.record(effect, a)
133
+ solution = inverse.predict(target)
134
+ if solution is not None:
135
+ return measure(c, solution) == target
136
+ raise AssertionError("inconsistent invertibility")
137
+
138
+
139
+ def enumerate_actions(d, c, target, measure, rng):
140
+ """A random shift and odd stride enumerate unique guesses without an effect model."""
141
+ modulus = 1 << d
142
+ a, stride = rng.randrange(modulus), 2*rng.randrange(modulus//2)+1
143
+ for _ in range(min(modulus, 256)):
144
+ if measure(c, a) == target:
145
+ return True
146
+ a = (a + stride) % modulus
147
+ return False
148
+
149
+
150
+ def single_run(d, q, seed, n_tasks):
151
+ rng = random.Random(500_009 * d + seed)
152
+ guess_rng = random.Random(3793 * seed + d)
153
+ plant = HiddenPlant(d, q, rng)
154
+ writer = Writer(d, plant.w, plant.measure)
155
+ seen_counts = []
156
+ for _ in range(d):
157
+ writer.acquire_base()
158
+ seen_counts.append(1 << len(writer.inverse_on_seen_effects.rows))
159
+ assert seen_counts == [2**r for r in range(1, d+1)]
160
+ base_probes = plant.calls
161
+ for j in range(q):
162
+ writer.prepare(1 << j)
163
+ context_probes = plant.calls-base_probes
164
+ assert base_probes == d and context_probes == q
165
+ # The comparison baseline has the same forward observations and regime assumptions.
166
+ cached_forward = writer.forward_columns.copy()
167
+ cached_effect = [writer.context_effect.predict(1 << j) for j in range(q)]
168
+ raw = []
169
+ for task in range(n_tasks):
170
+ c = rng.randrange(1 << q)
171
+ while c.bit_count() < 2:
172
+ c = rng.randrange(1 << q)
173
+ hidden_action = rng.randrange(1 << d)
174
+ while hidden_action.bit_count() < 2:
175
+ hidden_action = rng.randrange(1 << d)
176
+ # The goal is externally presented; its hidden solution is not sent to a solver.
177
+ goal = plant.reference(c, hidden_action)
178
+ before = plant.calls
179
+ action = writer.propose(c, goal)
180
+ ok = plant.measure(c, action) == goal
181
+ wc = plant.calls-before
182
+ before = plant.calls
183
+ v = apply_columns(cached_effect, c)
184
+ guessed_cols = [t ^ (v if (plant.w >> j) & 1 else 0)
185
+ for j, t in enumerate(cached_forward)]
186
+ fixed_action = apply_columns(inverse_from_columns(guessed_cols), goal)
187
+ assert fixed_action == action == hidden_action
188
+ fixed_ok = plant.measure(c, fixed_action) == goal
189
+ gc = plant.calls-before
190
+ before = plant.calls
191
+ reset_ok = per_task_solver(d, c, goal, plant.measure)
192
+ rc = plant.calls-before
193
+ before = plant.calls
194
+ naive_ok = enumerate_actions(d, c, goal, plant.measure, guess_rng)
195
+ nc = plant.calls-before
196
+ assert ok and fixed_ok and reset_ok and wc == gc == 1
197
+ raw.append(dict(d=d, q=q, seed=seed, task=task, context=c, goal=goal,
198
+ proposal_action=action, writer_reads=wc, fixed_elimination_reads=gc,
199
+ reset_reads=rc, enumeration_reads=nc, enumeration_success=int(naive_ok)))
200
+ # Independent full verification: equality on every action basis implies equality
201
+ # for every goal. For d<=8 also directly enumerate all possible goals.
202
+ before = plant.calls
203
+ for c in range(1 << q):
204
+ goals = list(range(1 << d)) if d <= 8 else [1 << j for j in range(d)]
205
+ for goal in goals:
206
+ assert plant.measure(c, writer.propose(c, goal)) == goal
207
+ audit_reads = plant.calls-before
208
+ assert audit_reads == (1 << q)*((1 << d) if d <= 8 else d)
209
+ # Without the public w, d-1 observations can hide an invertible nonzero change.
210
+ w_alternative = 1 << (d-1)
211
+ def alternative(a):
212
+ return a ^ (1 if parity(w_alternative & a) else 0)
213
+ assert all(alternative(1 << i) == 1 << i for i in range(d-1))
214
+ assert alternative(w_alternative) != w_alternative
215
+ assert all(alternative(alternative(1 << i)) == 1 << i for i in range(d))
216
+ return raw, dict(d=d, q=q, seed=seed, base_training_reads=base_probes,
217
+ context_training_reads=context_probes, independent_audit_reads=audit_reads,
218
+ held_out_tasks=n_tasks, verified_held_out=n_tasks,
219
+ guaranteed_base_effect_span_sizes=seen_counts)
220
+
221
+
222
+ def main():
223
+ parser = argparse.ArgumentParser(description=__doc__)
224
+ parser.add_argument("--out", type=Path, default=Path(__file__).resolve().parent)
225
+ args = parser.parse_args()
226
+ args.out.mkdir(parents=True, exist_ok=True)
227
+ rows, runs = [], []
228
+ for d, q in [(6,3), (12,4), (24,5), (48,6)]:
229
+ for seed in range(8):
230
+ out, run = single_run(d, q, seed, 32)
231
+ rows.extend(out)
232
+ runs.append(run)
233
+ by_dim = []
234
+ for d in (6,12,24,48):
235
+ group = [r for r in rows if r['d'] == d]
236
+ n = len(group)
237
+ by_dim.append(dict(d=d, tasks=n, writer_success=n, fixed_elimination_success=n,
238
+ fresh_elimination_success=n,
239
+ writer_reads_per_goal=sum(r['writer_reads'] for r in group)/n,
240
+ fixed_elimination_reads_per_goal=sum(r['fixed_elimination_reads'] for r in group)/n,
241
+ fresh_elimination_reads_per_goal=sum(r['reset_reads'] for r in group)/n,
242
+ enumeration_success=sum(r['enumeration_success'] for r in group),
243
+ enumeration_reads_per_goal=sum(r['enumeration_reads'] for r in group)/n))
244
+ with (args.out/'tasks.csv').open('w', newline='', encoding='utf-8') as f:
245
+ out = csv.DictWriter(f, fieldnames=rows[0].keys())
246
+ out.writeheader()
247
+ out.writerows(rows)
248
+ result = dict(model='GF(2), complete d-bit effect per measured action, resettable trials',
249
+ held_out_action_and_context='Hamming weight >= 2, sampled with replacement',
250
+ dimensions=by_dim, runs=runs)
251
+ (args.out/'results.json').write_text(json.dumps(result, indent=2)+'\n', encoding='utf-8')
252
+ print(json.dumps(by_dim, indent=2))
253
+
254
+
255
+ if __name__=='__main__':
256
+ main()
reproduction/operator/results.json ADDED
@@ -0,0 +1,1160 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "model": "GF(2), complete d-bit effect per measured action, resettable trials",
3
+ "held_out_action_and_context": "Hamming weight >= 2, sampled with replacement",
4
+ "dimensions": [
5
+ {
6
+ "d": 6,
7
+ "tasks": 256,
8
+ "writer_success": 256,
9
+ "fixed_elimination_success": 256,
10
+ "fresh_elimination_success": 256,
11
+ "writer_reads_per_goal": 1.0,
12
+ "fixed_elimination_reads_per_goal": 1.0,
13
+ "fresh_elimination_reads_per_goal": 6.296875,
14
+ "enumeration_success": 256,
15
+ "enumeration_reads_per_goal": 31.80859375
16
+ },
17
+ {
18
+ "d": 12,
19
+ "tasks": 256,
20
+ "writer_success": 256,
21
+ "fixed_elimination_success": 256,
22
+ "fresh_elimination_success": 256,
23
+ "writer_reads_per_goal": 1.0,
24
+ "fixed_elimination_reads_per_goal": 1.0,
25
+ "fresh_elimination_reads_per_goal": 12.01953125,
26
+ "enumeration_success": 22,
27
+ "enumeration_reads_per_goal": 246.42578125
28
+ },
29
+ {
30
+ "d": 24,
31
+ "tasks": 256,
32
+ "writer_success": 256,
33
+ "fixed_elimination_success": 256,
34
+ "fresh_elimination_success": 256,
35
+ "writer_reads_per_goal": 1.0,
36
+ "fixed_elimination_reads_per_goal": 1.0,
37
+ "fresh_elimination_reads_per_goal": 23.8984375,
38
+ "enumeration_success": 0,
39
+ "enumeration_reads_per_goal": 256.0
40
+ },
41
+ {
42
+ "d": 48,
43
+ "tasks": 256,
44
+ "writer_success": 256,
45
+ "fixed_elimination_success": 256,
46
+ "fresh_elimination_success": 256,
47
+ "writer_reads_per_goal": 1.0,
48
+ "fixed_elimination_reads_per_goal": 1.0,
49
+ "fresh_elimination_reads_per_goal": 47.9609375,
50
+ "enumeration_success": 0,
51
+ "enumeration_reads_per_goal": 256.0
52
+ }
53
+ ],
54
+ "runs": [
55
+ {
56
+ "d": 6,
57
+ "q": 3,
58
+ "seed": 0,
59
+ "base_training_reads": 6,
60
+ "context_training_reads": 3,
61
+ "independent_audit_reads": 512,
62
+ "held_out_tasks": 32,
63
+ "verified_held_out": 32,
64
+ "guaranteed_base_effect_span_sizes": [
65
+ 2,
66
+ 4,
67
+ 8,
68
+ 16,
69
+ 32,
70
+ 64
71
+ ]
72
+ },
73
+ {
74
+ "d": 6,
75
+ "q": 3,
76
+ "seed": 1,
77
+ "base_training_reads": 6,
78
+ "context_training_reads": 3,
79
+ "independent_audit_reads": 512,
80
+ "held_out_tasks": 32,
81
+ "verified_held_out": 32,
82
+ "guaranteed_base_effect_span_sizes": [
83
+ 2,
84
+ 4,
85
+ 8,
86
+ 16,
87
+ 32,
88
+ 64
89
+ ]
90
+ },
91
+ {
92
+ "d": 6,
93
+ "q": 3,
94
+ "seed": 2,
95
+ "base_training_reads": 6,
96
+ "context_training_reads": 3,
97
+ "independent_audit_reads": 512,
98
+ "held_out_tasks": 32,
99
+ "verified_held_out": 32,
100
+ "guaranteed_base_effect_span_sizes": [
101
+ 2,
102
+ 4,
103
+ 8,
104
+ 16,
105
+ 32,
106
+ 64
107
+ ]
108
+ },
109
+ {
110
+ "d": 6,
111
+ "q": 3,
112
+ "seed": 3,
113
+ "base_training_reads": 6,
114
+ "context_training_reads": 3,
115
+ "independent_audit_reads": 512,
116
+ "held_out_tasks": 32,
117
+ "verified_held_out": 32,
118
+ "guaranteed_base_effect_span_sizes": [
119
+ 2,
120
+ 4,
121
+ 8,
122
+ 16,
123
+ 32,
124
+ 64
125
+ ]
126
+ },
127
+ {
128
+ "d": 6,
129
+ "q": 3,
130
+ "seed": 4,
131
+ "base_training_reads": 6,
132
+ "context_training_reads": 3,
133
+ "independent_audit_reads": 512,
134
+ "held_out_tasks": 32,
135
+ "verified_held_out": 32,
136
+ "guaranteed_base_effect_span_sizes": [
137
+ 2,
138
+ 4,
139
+ 8,
140
+ 16,
141
+ 32,
142
+ 64
143
+ ]
144
+ },
145
+ {
146
+ "d": 6,
147
+ "q": 3,
148
+ "seed": 5,
149
+ "base_training_reads": 6,
150
+ "context_training_reads": 3,
151
+ "independent_audit_reads": 512,
152
+ "held_out_tasks": 32,
153
+ "verified_held_out": 32,
154
+ "guaranteed_base_effect_span_sizes": [
155
+ 2,
156
+ 4,
157
+ 8,
158
+ 16,
159
+ 32,
160
+ 64
161
+ ]
162
+ },
163
+ {
164
+ "d": 6,
165
+ "q": 3,
166
+ "seed": 6,
167
+ "base_training_reads": 6,
168
+ "context_training_reads": 3,
169
+ "independent_audit_reads": 512,
170
+ "held_out_tasks": 32,
171
+ "verified_held_out": 32,
172
+ "guaranteed_base_effect_span_sizes": [
173
+ 2,
174
+ 4,
175
+ 8,
176
+ 16,
177
+ 32,
178
+ 64
179
+ ]
180
+ },
181
+ {
182
+ "d": 6,
183
+ "q": 3,
184
+ "seed": 7,
185
+ "base_training_reads": 6,
186
+ "context_training_reads": 3,
187
+ "independent_audit_reads": 512,
188
+ "held_out_tasks": 32,
189
+ "verified_held_out": 32,
190
+ "guaranteed_base_effect_span_sizes": [
191
+ 2,
192
+ 4,
193
+ 8,
194
+ 16,
195
+ 32,
196
+ 64
197
+ ]
198
+ },
199
+ {
200
+ "d": 12,
201
+ "q": 4,
202
+ "seed": 0,
203
+ "base_training_reads": 12,
204
+ "context_training_reads": 4,
205
+ "independent_audit_reads": 192,
206
+ "held_out_tasks": 32,
207
+ "verified_held_out": 32,
208
+ "guaranteed_base_effect_span_sizes": [
209
+ 2,
210
+ 4,
211
+ 8,
212
+ 16,
213
+ 32,
214
+ 64,
215
+ 128,
216
+ 256,
217
+ 512,
218
+ 1024,
219
+ 2048,
220
+ 4096
221
+ ]
222
+ },
223
+ {
224
+ "d": 12,
225
+ "q": 4,
226
+ "seed": 1,
227
+ "base_training_reads": 12,
228
+ "context_training_reads": 4,
229
+ "independent_audit_reads": 192,
230
+ "held_out_tasks": 32,
231
+ "verified_held_out": 32,
232
+ "guaranteed_base_effect_span_sizes": [
233
+ 2,
234
+ 4,
235
+ 8,
236
+ 16,
237
+ 32,
238
+ 64,
239
+ 128,
240
+ 256,
241
+ 512,
242
+ 1024,
243
+ 2048,
244
+ 4096
245
+ ]
246
+ },
247
+ {
248
+ "d": 12,
249
+ "q": 4,
250
+ "seed": 2,
251
+ "base_training_reads": 12,
252
+ "context_training_reads": 4,
253
+ "independent_audit_reads": 192,
254
+ "held_out_tasks": 32,
255
+ "verified_held_out": 32,
256
+ "guaranteed_base_effect_span_sizes": [
257
+ 2,
258
+ 4,
259
+ 8,
260
+ 16,
261
+ 32,
262
+ 64,
263
+ 128,
264
+ 256,
265
+ 512,
266
+ 1024,
267
+ 2048,
268
+ 4096
269
+ ]
270
+ },
271
+ {
272
+ "d": 12,
273
+ "q": 4,
274
+ "seed": 3,
275
+ "base_training_reads": 12,
276
+ "context_training_reads": 4,
277
+ "independent_audit_reads": 192,
278
+ "held_out_tasks": 32,
279
+ "verified_held_out": 32,
280
+ "guaranteed_base_effect_span_sizes": [
281
+ 2,
282
+ 4,
283
+ 8,
284
+ 16,
285
+ 32,
286
+ 64,
287
+ 128,
288
+ 256,
289
+ 512,
290
+ 1024,
291
+ 2048,
292
+ 4096
293
+ ]
294
+ },
295
+ {
296
+ "d": 12,
297
+ "q": 4,
298
+ "seed": 4,
299
+ "base_training_reads": 12,
300
+ "context_training_reads": 4,
301
+ "independent_audit_reads": 192,
302
+ "held_out_tasks": 32,
303
+ "verified_held_out": 32,
304
+ "guaranteed_base_effect_span_sizes": [
305
+ 2,
306
+ 4,
307
+ 8,
308
+ 16,
309
+ 32,
310
+ 64,
311
+ 128,
312
+ 256,
313
+ 512,
314
+ 1024,
315
+ 2048,
316
+ 4096
317
+ ]
318
+ },
319
+ {
320
+ "d": 12,
321
+ "q": 4,
322
+ "seed": 5,
323
+ "base_training_reads": 12,
324
+ "context_training_reads": 4,
325
+ "independent_audit_reads": 192,
326
+ "held_out_tasks": 32,
327
+ "verified_held_out": 32,
328
+ "guaranteed_base_effect_span_sizes": [
329
+ 2,
330
+ 4,
331
+ 8,
332
+ 16,
333
+ 32,
334
+ 64,
335
+ 128,
336
+ 256,
337
+ 512,
338
+ 1024,
339
+ 2048,
340
+ 4096
341
+ ]
342
+ },
343
+ {
344
+ "d": 12,
345
+ "q": 4,
346
+ "seed": 6,
347
+ "base_training_reads": 12,
348
+ "context_training_reads": 4,
349
+ "independent_audit_reads": 192,
350
+ "held_out_tasks": 32,
351
+ "verified_held_out": 32,
352
+ "guaranteed_base_effect_span_sizes": [
353
+ 2,
354
+ 4,
355
+ 8,
356
+ 16,
357
+ 32,
358
+ 64,
359
+ 128,
360
+ 256,
361
+ 512,
362
+ 1024,
363
+ 2048,
364
+ 4096
365
+ ]
366
+ },
367
+ {
368
+ "d": 12,
369
+ "q": 4,
370
+ "seed": 7,
371
+ "base_training_reads": 12,
372
+ "context_training_reads": 4,
373
+ "independent_audit_reads": 192,
374
+ "held_out_tasks": 32,
375
+ "verified_held_out": 32,
376
+ "guaranteed_base_effect_span_sizes": [
377
+ 2,
378
+ 4,
379
+ 8,
380
+ 16,
381
+ 32,
382
+ 64,
383
+ 128,
384
+ 256,
385
+ 512,
386
+ 1024,
387
+ 2048,
388
+ 4096
389
+ ]
390
+ },
391
+ {
392
+ "d": 24,
393
+ "q": 5,
394
+ "seed": 0,
395
+ "base_training_reads": 24,
396
+ "context_training_reads": 5,
397
+ "independent_audit_reads": 768,
398
+ "held_out_tasks": 32,
399
+ "verified_held_out": 32,
400
+ "guaranteed_base_effect_span_sizes": [
401
+ 2,
402
+ 4,
403
+ 8,
404
+ 16,
405
+ 32,
406
+ 64,
407
+ 128,
408
+ 256,
409
+ 512,
410
+ 1024,
411
+ 2048,
412
+ 4096,
413
+ 8192,
414
+ 16384,
415
+ 32768,
416
+ 65536,
417
+ 131072,
418
+ 262144,
419
+ 524288,
420
+ 1048576,
421
+ 2097152,
422
+ 4194304,
423
+ 8388608,
424
+ 16777216
425
+ ]
426
+ },
427
+ {
428
+ "d": 24,
429
+ "q": 5,
430
+ "seed": 1,
431
+ "base_training_reads": 24,
432
+ "context_training_reads": 5,
433
+ "independent_audit_reads": 768,
434
+ "held_out_tasks": 32,
435
+ "verified_held_out": 32,
436
+ "guaranteed_base_effect_span_sizes": [
437
+ 2,
438
+ 4,
439
+ 8,
440
+ 16,
441
+ 32,
442
+ 64,
443
+ 128,
444
+ 256,
445
+ 512,
446
+ 1024,
447
+ 2048,
448
+ 4096,
449
+ 8192,
450
+ 16384,
451
+ 32768,
452
+ 65536,
453
+ 131072,
454
+ 262144,
455
+ 524288,
456
+ 1048576,
457
+ 2097152,
458
+ 4194304,
459
+ 8388608,
460
+ 16777216
461
+ ]
462
+ },
463
+ {
464
+ "d": 24,
465
+ "q": 5,
466
+ "seed": 2,
467
+ "base_training_reads": 24,
468
+ "context_training_reads": 5,
469
+ "independent_audit_reads": 768,
470
+ "held_out_tasks": 32,
471
+ "verified_held_out": 32,
472
+ "guaranteed_base_effect_span_sizes": [
473
+ 2,
474
+ 4,
475
+ 8,
476
+ 16,
477
+ 32,
478
+ 64,
479
+ 128,
480
+ 256,
481
+ 512,
482
+ 1024,
483
+ 2048,
484
+ 4096,
485
+ 8192,
486
+ 16384,
487
+ 32768,
488
+ 65536,
489
+ 131072,
490
+ 262144,
491
+ 524288,
492
+ 1048576,
493
+ 2097152,
494
+ 4194304,
495
+ 8388608,
496
+ 16777216
497
+ ]
498
+ },
499
+ {
500
+ "d": 24,
501
+ "q": 5,
502
+ "seed": 3,
503
+ "base_training_reads": 24,
504
+ "context_training_reads": 5,
505
+ "independent_audit_reads": 768,
506
+ "held_out_tasks": 32,
507
+ "verified_held_out": 32,
508
+ "guaranteed_base_effect_span_sizes": [
509
+ 2,
510
+ 4,
511
+ 8,
512
+ 16,
513
+ 32,
514
+ 64,
515
+ 128,
516
+ 256,
517
+ 512,
518
+ 1024,
519
+ 2048,
520
+ 4096,
521
+ 8192,
522
+ 16384,
523
+ 32768,
524
+ 65536,
525
+ 131072,
526
+ 262144,
527
+ 524288,
528
+ 1048576,
529
+ 2097152,
530
+ 4194304,
531
+ 8388608,
532
+ 16777216
533
+ ]
534
+ },
535
+ {
536
+ "d": 24,
537
+ "q": 5,
538
+ "seed": 4,
539
+ "base_training_reads": 24,
540
+ "context_training_reads": 5,
541
+ "independent_audit_reads": 768,
542
+ "held_out_tasks": 32,
543
+ "verified_held_out": 32,
544
+ "guaranteed_base_effect_span_sizes": [
545
+ 2,
546
+ 4,
547
+ 8,
548
+ 16,
549
+ 32,
550
+ 64,
551
+ 128,
552
+ 256,
553
+ 512,
554
+ 1024,
555
+ 2048,
556
+ 4096,
557
+ 8192,
558
+ 16384,
559
+ 32768,
560
+ 65536,
561
+ 131072,
562
+ 262144,
563
+ 524288,
564
+ 1048576,
565
+ 2097152,
566
+ 4194304,
567
+ 8388608,
568
+ 16777216
569
+ ]
570
+ },
571
+ {
572
+ "d": 24,
573
+ "q": 5,
574
+ "seed": 5,
575
+ "base_training_reads": 24,
576
+ "context_training_reads": 5,
577
+ "independent_audit_reads": 768,
578
+ "held_out_tasks": 32,
579
+ "verified_held_out": 32,
580
+ "guaranteed_base_effect_span_sizes": [
581
+ 2,
582
+ 4,
583
+ 8,
584
+ 16,
585
+ 32,
586
+ 64,
587
+ 128,
588
+ 256,
589
+ 512,
590
+ 1024,
591
+ 2048,
592
+ 4096,
593
+ 8192,
594
+ 16384,
595
+ 32768,
596
+ 65536,
597
+ 131072,
598
+ 262144,
599
+ 524288,
600
+ 1048576,
601
+ 2097152,
602
+ 4194304,
603
+ 8388608,
604
+ 16777216
605
+ ]
606
+ },
607
+ {
608
+ "d": 24,
609
+ "q": 5,
610
+ "seed": 6,
611
+ "base_training_reads": 24,
612
+ "context_training_reads": 5,
613
+ "independent_audit_reads": 768,
614
+ "held_out_tasks": 32,
615
+ "verified_held_out": 32,
616
+ "guaranteed_base_effect_span_sizes": [
617
+ 2,
618
+ 4,
619
+ 8,
620
+ 16,
621
+ 32,
622
+ 64,
623
+ 128,
624
+ 256,
625
+ 512,
626
+ 1024,
627
+ 2048,
628
+ 4096,
629
+ 8192,
630
+ 16384,
631
+ 32768,
632
+ 65536,
633
+ 131072,
634
+ 262144,
635
+ 524288,
636
+ 1048576,
637
+ 2097152,
638
+ 4194304,
639
+ 8388608,
640
+ 16777216
641
+ ]
642
+ },
643
+ {
644
+ "d": 24,
645
+ "q": 5,
646
+ "seed": 7,
647
+ "base_training_reads": 24,
648
+ "context_training_reads": 5,
649
+ "independent_audit_reads": 768,
650
+ "held_out_tasks": 32,
651
+ "verified_held_out": 32,
652
+ "guaranteed_base_effect_span_sizes": [
653
+ 2,
654
+ 4,
655
+ 8,
656
+ 16,
657
+ 32,
658
+ 64,
659
+ 128,
660
+ 256,
661
+ 512,
662
+ 1024,
663
+ 2048,
664
+ 4096,
665
+ 8192,
666
+ 16384,
667
+ 32768,
668
+ 65536,
669
+ 131072,
670
+ 262144,
671
+ 524288,
672
+ 1048576,
673
+ 2097152,
674
+ 4194304,
675
+ 8388608,
676
+ 16777216
677
+ ]
678
+ },
679
+ {
680
+ "d": 48,
681
+ "q": 6,
682
+ "seed": 0,
683
+ "base_training_reads": 48,
684
+ "context_training_reads": 6,
685
+ "independent_audit_reads": 3072,
686
+ "held_out_tasks": 32,
687
+ "verified_held_out": 32,
688
+ "guaranteed_base_effect_span_sizes": [
689
+ 2,
690
+ 4,
691
+ 8,
692
+ 16,
693
+ 32,
694
+ 64,
695
+ 128,
696
+ 256,
697
+ 512,
698
+ 1024,
699
+ 2048,
700
+ 4096,
701
+ 8192,
702
+ 16384,
703
+ 32768,
704
+ 65536,
705
+ 131072,
706
+ 262144,
707
+ 524288,
708
+ 1048576,
709
+ 2097152,
710
+ 4194304,
711
+ 8388608,
712
+ 16777216,
713
+ 33554432,
714
+ 67108864,
715
+ 134217728,
716
+ 268435456,
717
+ 536870912,
718
+ 1073741824,
719
+ 2147483648,
720
+ 4294967296,
721
+ 8589934592,
722
+ 17179869184,
723
+ 34359738368,
724
+ 68719476736,
725
+ 137438953472,
726
+ 274877906944,
727
+ 549755813888,
728
+ 1099511627776,
729
+ 2199023255552,
730
+ 4398046511104,
731
+ 8796093022208,
732
+ 17592186044416,
733
+ 35184372088832,
734
+ 70368744177664,
735
+ 140737488355328,
736
+ 281474976710656
737
+ ]
738
+ },
739
+ {
740
+ "d": 48,
741
+ "q": 6,
742
+ "seed": 1,
743
+ "base_training_reads": 48,
744
+ "context_training_reads": 6,
745
+ "independent_audit_reads": 3072,
746
+ "held_out_tasks": 32,
747
+ "verified_held_out": 32,
748
+ "guaranteed_base_effect_span_sizes": [
749
+ 2,
750
+ 4,
751
+ 8,
752
+ 16,
753
+ 32,
754
+ 64,
755
+ 128,
756
+ 256,
757
+ 512,
758
+ 1024,
759
+ 2048,
760
+ 4096,
761
+ 8192,
762
+ 16384,
763
+ 32768,
764
+ 65536,
765
+ 131072,
766
+ 262144,
767
+ 524288,
768
+ 1048576,
769
+ 2097152,
770
+ 4194304,
771
+ 8388608,
772
+ 16777216,
773
+ 33554432,
774
+ 67108864,
775
+ 134217728,
776
+ 268435456,
777
+ 536870912,
778
+ 1073741824,
779
+ 2147483648,
780
+ 4294967296,
781
+ 8589934592,
782
+ 17179869184,
783
+ 34359738368,
784
+ 68719476736,
785
+ 137438953472,
786
+ 274877906944,
787
+ 549755813888,
788
+ 1099511627776,
789
+ 2199023255552,
790
+ 4398046511104,
791
+ 8796093022208,
792
+ 17592186044416,
793
+ 35184372088832,
794
+ 70368744177664,
795
+ 140737488355328,
796
+ 281474976710656
797
+ ]
798
+ },
799
+ {
800
+ "d": 48,
801
+ "q": 6,
802
+ "seed": 2,
803
+ "base_training_reads": 48,
804
+ "context_training_reads": 6,
805
+ "independent_audit_reads": 3072,
806
+ "held_out_tasks": 32,
807
+ "verified_held_out": 32,
808
+ "guaranteed_base_effect_span_sizes": [
809
+ 2,
810
+ 4,
811
+ 8,
812
+ 16,
813
+ 32,
814
+ 64,
815
+ 128,
816
+ 256,
817
+ 512,
818
+ 1024,
819
+ 2048,
820
+ 4096,
821
+ 8192,
822
+ 16384,
823
+ 32768,
824
+ 65536,
825
+ 131072,
826
+ 262144,
827
+ 524288,
828
+ 1048576,
829
+ 2097152,
830
+ 4194304,
831
+ 8388608,
832
+ 16777216,
833
+ 33554432,
834
+ 67108864,
835
+ 134217728,
836
+ 268435456,
837
+ 536870912,
838
+ 1073741824,
839
+ 2147483648,
840
+ 4294967296,
841
+ 8589934592,
842
+ 17179869184,
843
+ 34359738368,
844
+ 68719476736,
845
+ 137438953472,
846
+ 274877906944,
847
+ 549755813888,
848
+ 1099511627776,
849
+ 2199023255552,
850
+ 4398046511104,
851
+ 8796093022208,
852
+ 17592186044416,
853
+ 35184372088832,
854
+ 70368744177664,
855
+ 140737488355328,
856
+ 281474976710656
857
+ ]
858
+ },
859
+ {
860
+ "d": 48,
861
+ "q": 6,
862
+ "seed": 3,
863
+ "base_training_reads": 48,
864
+ "context_training_reads": 6,
865
+ "independent_audit_reads": 3072,
866
+ "held_out_tasks": 32,
867
+ "verified_held_out": 32,
868
+ "guaranteed_base_effect_span_sizes": [
869
+ 2,
870
+ 4,
871
+ 8,
872
+ 16,
873
+ 32,
874
+ 64,
875
+ 128,
876
+ 256,
877
+ 512,
878
+ 1024,
879
+ 2048,
880
+ 4096,
881
+ 8192,
882
+ 16384,
883
+ 32768,
884
+ 65536,
885
+ 131072,
886
+ 262144,
887
+ 524288,
888
+ 1048576,
889
+ 2097152,
890
+ 4194304,
891
+ 8388608,
892
+ 16777216,
893
+ 33554432,
894
+ 67108864,
895
+ 134217728,
896
+ 268435456,
897
+ 536870912,
898
+ 1073741824,
899
+ 2147483648,
900
+ 4294967296,
901
+ 8589934592,
902
+ 17179869184,
903
+ 34359738368,
904
+ 68719476736,
905
+ 137438953472,
906
+ 274877906944,
907
+ 549755813888,
908
+ 1099511627776,
909
+ 2199023255552,
910
+ 4398046511104,
911
+ 8796093022208,
912
+ 17592186044416,
913
+ 35184372088832,
914
+ 70368744177664,
915
+ 140737488355328,
916
+ 281474976710656
917
+ ]
918
+ },
919
+ {
920
+ "d": 48,
921
+ "q": 6,
922
+ "seed": 4,
923
+ "base_training_reads": 48,
924
+ "context_training_reads": 6,
925
+ "independent_audit_reads": 3072,
926
+ "held_out_tasks": 32,
927
+ "verified_held_out": 32,
928
+ "guaranteed_base_effect_span_sizes": [
929
+ 2,
930
+ 4,
931
+ 8,
932
+ 16,
933
+ 32,
934
+ 64,
935
+ 128,
936
+ 256,
937
+ 512,
938
+ 1024,
939
+ 2048,
940
+ 4096,
941
+ 8192,
942
+ 16384,
943
+ 32768,
944
+ 65536,
945
+ 131072,
946
+ 262144,
947
+ 524288,
948
+ 1048576,
949
+ 2097152,
950
+ 4194304,
951
+ 8388608,
952
+ 16777216,
953
+ 33554432,
954
+ 67108864,
955
+ 134217728,
956
+ 268435456,
957
+ 536870912,
958
+ 1073741824,
959
+ 2147483648,
960
+ 4294967296,
961
+ 8589934592,
962
+ 17179869184,
963
+ 34359738368,
964
+ 68719476736,
965
+ 137438953472,
966
+ 274877906944,
967
+ 549755813888,
968
+ 1099511627776,
969
+ 2199023255552,
970
+ 4398046511104,
971
+ 8796093022208,
972
+ 17592186044416,
973
+ 35184372088832,
974
+ 70368744177664,
975
+ 140737488355328,
976
+ 281474976710656
977
+ ]
978
+ },
979
+ {
980
+ "d": 48,
981
+ "q": 6,
982
+ "seed": 5,
983
+ "base_training_reads": 48,
984
+ "context_training_reads": 6,
985
+ "independent_audit_reads": 3072,
986
+ "held_out_tasks": 32,
987
+ "verified_held_out": 32,
988
+ "guaranteed_base_effect_span_sizes": [
989
+ 2,
990
+ 4,
991
+ 8,
992
+ 16,
993
+ 32,
994
+ 64,
995
+ 128,
996
+ 256,
997
+ 512,
998
+ 1024,
999
+ 2048,
1000
+ 4096,
1001
+ 8192,
1002
+ 16384,
1003
+ 32768,
1004
+ 65536,
1005
+ 131072,
1006
+ 262144,
1007
+ 524288,
1008
+ 1048576,
1009
+ 2097152,
1010
+ 4194304,
1011
+ 8388608,
1012
+ 16777216,
1013
+ 33554432,
1014
+ 67108864,
1015
+ 134217728,
1016
+ 268435456,
1017
+ 536870912,
1018
+ 1073741824,
1019
+ 2147483648,
1020
+ 4294967296,
1021
+ 8589934592,
1022
+ 17179869184,
1023
+ 34359738368,
1024
+ 68719476736,
1025
+ 137438953472,
1026
+ 274877906944,
1027
+ 549755813888,
1028
+ 1099511627776,
1029
+ 2199023255552,
1030
+ 4398046511104,
1031
+ 8796093022208,
1032
+ 17592186044416,
1033
+ 35184372088832,
1034
+ 70368744177664,
1035
+ 140737488355328,
1036
+ 281474976710656
1037
+ ]
1038
+ },
1039
+ {
1040
+ "d": 48,
1041
+ "q": 6,
1042
+ "seed": 6,
1043
+ "base_training_reads": 48,
1044
+ "context_training_reads": 6,
1045
+ "independent_audit_reads": 3072,
1046
+ "held_out_tasks": 32,
1047
+ "verified_held_out": 32,
1048
+ "guaranteed_base_effect_span_sizes": [
1049
+ 2,
1050
+ 4,
1051
+ 8,
1052
+ 16,
1053
+ 32,
1054
+ 64,
1055
+ 128,
1056
+ 256,
1057
+ 512,
1058
+ 1024,
1059
+ 2048,
1060
+ 4096,
1061
+ 8192,
1062
+ 16384,
1063
+ 32768,
1064
+ 65536,
1065
+ 131072,
1066
+ 262144,
1067
+ 524288,
1068
+ 1048576,
1069
+ 2097152,
1070
+ 4194304,
1071
+ 8388608,
1072
+ 16777216,
1073
+ 33554432,
1074
+ 67108864,
1075
+ 134217728,
1076
+ 268435456,
1077
+ 536870912,
1078
+ 1073741824,
1079
+ 2147483648,
1080
+ 4294967296,
1081
+ 8589934592,
1082
+ 17179869184,
1083
+ 34359738368,
1084
+ 68719476736,
1085
+ 137438953472,
1086
+ 274877906944,
1087
+ 549755813888,
1088
+ 1099511627776,
1089
+ 2199023255552,
1090
+ 4398046511104,
1091
+ 8796093022208,
1092
+ 17592186044416,
1093
+ 35184372088832,
1094
+ 70368744177664,
1095
+ 140737488355328,
1096
+ 281474976710656
1097
+ ]
1098
+ },
1099
+ {
1100
+ "d": 48,
1101
+ "q": 6,
1102
+ "seed": 7,
1103
+ "base_training_reads": 48,
1104
+ "context_training_reads": 6,
1105
+ "independent_audit_reads": 3072,
1106
+ "held_out_tasks": 32,
1107
+ "verified_held_out": 32,
1108
+ "guaranteed_base_effect_span_sizes": [
1109
+ 2,
1110
+ 4,
1111
+ 8,
1112
+ 16,
1113
+ 32,
1114
+ 64,
1115
+ 128,
1116
+ 256,
1117
+ 512,
1118
+ 1024,
1119
+ 2048,
1120
+ 4096,
1121
+ 8192,
1122
+ 16384,
1123
+ 32768,
1124
+ 65536,
1125
+ 131072,
1126
+ 262144,
1127
+ 524288,
1128
+ 1048576,
1129
+ 2097152,
1130
+ 4194304,
1131
+ 8388608,
1132
+ 16777216,
1133
+ 33554432,
1134
+ 67108864,
1135
+ 134217728,
1136
+ 268435456,
1137
+ 536870912,
1138
+ 1073741824,
1139
+ 2147483648,
1140
+ 4294967296,
1141
+ 8589934592,
1142
+ 17179869184,
1143
+ 34359738368,
1144
+ 68719476736,
1145
+ 137438953472,
1146
+ 274877906944,
1147
+ 549755813888,
1148
+ 1099511627776,
1149
+ 2199023255552,
1150
+ 4398046511104,
1151
+ 8796093022208,
1152
+ 17592186044416,
1153
+ 35184372088832,
1154
+ 70368744177664,
1155
+ 140737488355328,
1156
+ 281474976710656
1157
+ ]
1158
+ }
1159
+ ]
1160
+ }
reproduction/operator/tasks.csv ADDED
@@ -0,0 +1,1025 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ d,q,seed,task,context,goal,proposal_action,writer_reads,fixed_elimination_reads,reset_reads,enumeration_reads,enumeration_success
2
+ 6,3,0,0,3,36,14,1,1,5,61,1
3
+ 6,3,0,1,6,57,27,1,1,6,51,1
4
+ 6,3,0,2,7,1,58,1,1,7,31,1
5
+ 6,3,0,3,6,2,48,1,1,7,13,1
6
+ 6,3,0,4,3,28,52,1,1,7,5,1
7
+ 6,3,0,5,3,46,45,1,1,7,54,1
8
+ 6,3,0,6,3,5,49,1,1,7,57,1
9
+ 6,3,0,7,3,15,18,1,1,6,55,1
10
+ 6,3,0,8,3,62,26,1,1,6,12,1
11
+ 6,3,0,9,3,29,21,1,1,6,30,1
12
+ 6,3,0,10,5,43,37,1,1,7,7,1
13
+ 6,3,0,11,5,21,63,1,1,7,44,1
14
+ 6,3,0,12,5,56,3,1,1,3,56,1
15
+ 6,3,0,13,3,45,60,1,1,7,40,1
16
+ 6,3,0,14,6,52,57,1,1,7,10,1
17
+ 6,3,0,15,6,61,58,1,1,7,48,1
18
+ 6,3,0,16,7,57,27,1,1,6,38,1
19
+ 6,3,0,17,6,15,18,1,1,6,34,1
20
+ 6,3,0,18,3,56,58,1,1,7,62,1
21
+ 6,3,0,19,7,2,48,1,1,7,51,1
22
+ 6,3,0,20,6,58,59,1,1,7,5,1
23
+ 6,3,0,21,6,12,50,1,1,7,24,1
24
+ 6,3,0,22,3,55,40,1,1,7,11,1
25
+ 6,3,0,23,6,41,29,1,1,6,5,1
26
+ 6,3,0,24,7,53,3,1,1,3,63,1
27
+ 6,3,0,25,3,21,6,1,1,4,48,1
28
+ 6,3,0,26,3,9,50,1,1,7,26,1
29
+ 6,3,0,27,7,60,42,1,1,7,8,1
30
+ 6,3,0,28,3,51,56,1,1,7,3,1
31
+ 6,3,0,29,5,39,31,1,1,6,49,1
32
+ 6,3,0,30,7,38,37,1,1,7,37,1
33
+ 6,3,0,31,6,57,27,1,1,6,54,1
34
+ 6,3,1,0,3,45,21,1,1,6,9,1
35
+ 6,3,1,1,5,53,42,1,1,7,50,1
36
+ 6,3,1,2,6,31,21,1,1,6,52,1
37
+ 6,3,1,3,6,61,55,1,1,7,34,1
38
+ 6,3,1,4,5,2,36,1,1,7,18,1
39
+ 6,3,1,5,7,35,24,1,1,6,40,1
40
+ 6,3,1,6,3,1,17,1,1,6,43,1
41
+ 6,3,1,7,3,2,36,1,1,7,45,1
42
+ 6,3,1,8,5,62,12,1,1,5,7,1
43
+ 6,3,1,9,6,22,50,1,1,7,49,1
44
+ 6,3,1,10,5,23,45,1,1,7,36,1
45
+ 6,3,1,11,6,16,9,1,1,5,57,1
46
+ 6,3,1,12,5,4,58,1,1,7,63,1
47
+ 6,3,1,13,5,23,45,1,1,7,6,1
48
+ 6,3,1,14,5,31,48,1,1,7,9,1
49
+ 6,3,1,15,6,31,21,1,1,6,9,1
50
+ 6,3,1,16,7,2,36,1,1,7,10,1
51
+ 6,3,1,17,3,20,22,1,1,6,27,1
52
+ 6,3,1,18,6,37,35,1,1,7,31,1
53
+ 6,3,1,19,6,13,56,1,1,7,20,1
54
+ 6,3,1,20,3,45,21,1,1,6,22,1
55
+ 6,3,1,21,5,42,26,1,1,6,62,1
56
+ 6,3,1,22,3,61,55,1,1,7,49,1
57
+ 6,3,1,23,5,23,45,1,1,7,52,1
58
+ 6,3,1,24,6,9,39,1,1,7,62,1
59
+ 6,3,1,25,6,62,41,1,1,7,35,1
60
+ 6,3,1,26,6,20,22,1,1,6,46,1
61
+ 6,3,1,27,7,55,14,1,1,5,36,1
62
+ 6,3,1,28,5,20,22,1,1,6,1,1
63
+ 6,3,1,29,7,11,21,1,1,6,55,1
64
+ 6,3,1,30,5,57,13,1,1,5,23,1
65
+ 6,3,1,31,6,29,49,1,1,7,14,1
66
+ 6,3,2,0,6,52,58,1,1,7,3,1
67
+ 6,3,2,1,3,25,54,1,1,7,53,1
68
+ 6,3,2,2,7,18,62,1,1,7,58,1
69
+ 6,3,2,3,6,2,46,1,1,7,20,1
70
+ 6,3,2,4,7,21,7,1,1,4,45,1
71
+ 6,3,2,5,7,40,27,1,1,6,13,1
72
+ 6,3,2,6,5,7,57,1,1,7,46,1
73
+ 6,3,2,7,3,62,56,1,1,7,35,1
74
+ 6,3,2,8,7,21,7,1,1,4,25,1
75
+ 6,3,2,9,3,46,39,1,1,7,8,1
76
+ 6,3,2,10,7,30,15,1,1,5,42,1
77
+ 6,3,2,11,5,48,51,1,1,7,23,1
78
+ 6,3,2,12,3,27,24,1,1,6,8,1
79
+ 6,3,2,13,6,34,25,1,1,6,6,1
80
+ 6,3,2,14,5,27,24,1,1,6,60,1
81
+ 6,3,2,15,6,18,43,1,1,7,41,1
82
+ 6,3,2,16,6,23,60,1,1,7,43,1
83
+ 6,3,2,17,5,4,9,1,1,5,15,1
84
+ 6,3,2,18,6,1,31,1,1,6,14,1
85
+ 6,3,2,19,3,20,13,1,1,5,12,1
86
+ 6,3,2,20,7,17,26,1,1,6,37,1
87
+ 6,3,2,21,5,43,43,1,1,7,4,1
88
+ 6,3,2,22,6,3,49,1,1,7,41,1
89
+ 6,3,2,23,3,42,53,1,1,7,13,1
90
+ 6,3,2,24,6,1,31,1,1,6,13,1
91
+ 6,3,2,25,3,4,18,1,1,6,43,1
92
+ 6,3,2,26,6,24,41,1,1,7,21,1
93
+ 6,3,2,27,7,9,38,1,1,7,17,1
94
+ 6,3,2,28,6,62,56,1,1,7,19,1
95
+ 6,3,2,29,6,38,17,1,1,6,13,1
96
+ 6,3,2,30,7,22,35,1,1,7,12,1
97
+ 6,3,2,31,7,34,25,1,1,6,9,1
98
+ 6,3,3,0,3,63,36,1,1,7,53,1
99
+ 6,3,3,1,5,39,44,1,1,7,27,1
100
+ 6,3,3,2,3,36,18,1,1,6,45,1
101
+ 6,3,3,3,6,33,27,1,1,6,2,1
102
+ 6,3,3,4,6,22,25,1,1,6,50,1
103
+ 6,3,3,5,7,10,34,1,1,7,3,1
104
+ 6,3,3,6,5,12,21,1,1,6,39,1
105
+ 6,3,3,7,6,51,42,1,1,7,27,1
106
+ 6,3,3,8,3,62,30,1,1,6,5,1
107
+ 6,3,3,9,3,12,21,1,1,6,36,1
108
+ 6,3,3,10,7,28,59,1,1,7,56,1
109
+ 6,3,3,11,3,34,37,1,1,7,20,1
110
+ 6,3,3,12,3,30,63,1,1,7,1,1
111
+ 6,3,3,13,6,29,26,1,1,6,35,1
112
+ 6,3,3,14,6,21,39,1,1,7,15,1
113
+ 6,3,3,15,3,57,19,1,1,6,52,1
114
+ 6,3,3,16,7,24,54,1,1,7,45,1
115
+ 6,3,3,17,5,7,45,1,1,7,40,1
116
+ 6,3,3,18,5,62,30,1,1,6,38,1
117
+ 6,3,3,19,7,2,58,1,1,7,18,1
118
+ 6,3,3,20,6,42,24,1,1,6,24,1
119
+ 6,3,3,21,5,42,35,1,1,7,39,1
120
+ 6,3,3,22,3,54,56,1,1,7,50,1
121
+ 6,3,3,23,7,28,59,1,1,7,40,1
122
+ 6,3,3,24,5,28,59,1,1,7,41,1
123
+ 6,3,3,25,6,2,31,1,1,6,36,1
124
+ 6,3,3,26,7,60,36,1,1,7,18,1
125
+ 6,3,3,27,7,53,56,1,1,7,63,1
126
+ 6,3,3,28,5,17,52,1,1,7,3,1
127
+ 6,3,3,29,6,45,14,1,1,5,4,1
128
+ 6,3,3,30,3,10,34,1,1,7,26,1
129
+ 6,3,3,31,5,39,44,1,1,7,20,1
130
+ 6,3,4,0,7,2,18,1,1,6,50,1
131
+ 6,3,4,1,5,51,31,1,1,6,47,1
132
+ 6,3,4,2,6,17,43,1,1,7,12,1
133
+ 6,3,4,3,7,33,23,1,1,6,35,1
134
+ 6,3,4,4,3,18,58,1,1,7,46,1
135
+ 6,3,4,5,5,34,52,1,1,7,27,1
136
+ 6,3,4,6,6,61,45,1,1,7,28,1
137
+ 6,3,4,7,6,46,47,1,1,7,5,1
138
+ 6,3,4,8,5,29,48,1,1,7,10,1
139
+ 6,3,4,9,7,51,52,1,1,7,22,1
140
+ 6,3,4,10,3,30,19,1,1,6,13,1
141
+ 6,3,4,11,5,46,47,1,1,7,42,1
142
+ 6,3,4,12,5,19,18,1,1,6,60,1
143
+ 6,3,4,13,7,5,58,1,1,7,62,1
144
+ 6,3,4,14,6,34,36,1,1,7,25,1
145
+ 6,3,4,15,5,40,29,1,1,6,13,1
146
+ 6,3,4,16,3,46,47,1,1,7,5,1
147
+ 6,3,4,17,6,58,5,1,1,4,50,1
148
+ 6,3,4,18,7,21,11,1,1,5,6,1
149
+ 6,3,4,19,6,28,58,1,1,7,60,1
150
+ 6,3,4,20,5,40,29,1,1,6,4,1
151
+ 6,3,4,21,6,27,18,1,1,6,19,1
152
+ 6,3,4,22,5,14,34,1,1,7,40,1
153
+ 6,3,4,23,7,35,5,1,1,4,7,1
154
+ 6,3,4,24,6,34,36,1,1,7,26,1
155
+ 6,3,4,25,7,53,6,1,1,4,46,1
156
+ 6,3,4,26,3,43,12,1,1,5,53,1
157
+ 6,3,4,27,3,22,3,1,1,3,49,1
158
+ 6,3,4,28,7,32,13,1,1,5,60,1
159
+ 6,3,4,29,6,34,36,1,1,7,47,1
160
+ 6,3,4,30,5,37,28,1,1,6,52,1
161
+ 6,3,4,31,7,27,41,1,1,7,12,1
162
+ 6,3,5,0,3,51,39,1,1,7,29,1
163
+ 6,3,5,1,7,62,14,1,1,5,52,1
164
+ 6,3,5,2,7,47,62,1,1,7,33,1
165
+ 6,3,5,3,6,17,48,1,1,7,53,1
166
+ 6,3,5,4,3,50,11,1,1,5,25,1
167
+ 6,3,5,5,3,37,54,1,1,7,22,1
168
+ 6,3,5,6,6,3,12,1,1,5,29,1
169
+ 6,3,5,7,3,12,5,1,1,4,32,1
170
+ 6,3,5,8,7,25,20,1,1,6,24,1
171
+ 6,3,5,9,6,6,7,1,1,4,31,1
172
+ 6,3,5,10,3,4,45,1,1,7,51,1
173
+ 6,3,5,11,5,62,14,1,1,5,60,1
174
+ 6,3,5,12,6,4,45,1,1,7,16,1
175
+ 6,3,5,13,3,50,11,1,1,5,44,1
176
+ 6,3,5,14,6,20,59,1,1,7,7,1
177
+ 6,3,5,15,5,57,47,1,1,7,2,1
178
+ 6,3,5,16,7,21,29,1,1,6,12,1
179
+ 6,3,5,17,7,33,27,1,1,6,58,1
180
+ 6,3,5,18,5,54,55,1,1,7,50,1
181
+ 6,3,5,19,7,63,46,1,1,7,5,1
182
+ 6,3,5,20,5,28,25,1,1,6,43,1
183
+ 6,3,5,21,3,42,63,1,1,7,51,1
184
+ 6,3,5,22,3,45,30,1,1,6,26,1
185
+ 6,3,5,23,6,12,15,1,1,5,28,1
186
+ 6,3,5,24,5,24,52,1,1,7,37,1
187
+ 6,3,5,25,5,30,40,1,1,7,38,1
188
+ 6,3,5,26,7,33,27,1,1,6,51,1
189
+ 6,3,5,27,6,49,13,1,1,5,23,1
190
+ 6,3,5,28,3,55,10,1,1,5,36,1
191
+ 6,3,5,29,3,50,11,1,1,5,29,1
192
+ 6,3,5,30,7,44,50,1,1,7,42,1
193
+ 6,3,5,31,3,44,50,1,1,7,1,1
194
+ 6,3,6,0,3,22,19,1,1,6,45,1
195
+ 6,3,6,1,7,32,22,1,1,6,61,1
196
+ 6,3,6,2,7,48,5,1,1,4,49,1
197
+ 6,3,6,3,7,62,37,1,1,7,56,1
198
+ 6,3,6,4,3,48,5,1,1,4,36,1
199
+ 6,3,6,5,3,30,56,1,1,7,3,1
200
+ 6,3,6,6,6,33,15,1,1,5,45,1
201
+ 6,3,6,7,3,47,52,1,1,7,42,1
202
+ 6,3,6,8,6,48,5,1,1,4,35,1
203
+ 6,3,6,9,5,33,26,1,1,6,64,1
204
+ 6,3,6,10,3,4,57,1,1,7,48,1
205
+ 6,3,6,11,3,22,19,1,1,6,33,1
206
+ 6,3,6,12,7,3,59,1,1,7,24,1
207
+ 6,3,6,13,5,59,21,1,1,6,61,1
208
+ 6,3,6,14,6,34,52,1,1,7,3,1
209
+ 6,3,6,15,6,45,29,1,1,6,1,1
210
+ 6,3,6,16,3,2,50,1,1,7,45,1
211
+ 6,3,6,17,3,56,46,1,1,7,3,1
212
+ 6,3,6,18,5,6,11,1,1,5,15,1
213
+ 6,3,6,19,7,3,59,1,1,7,2,1
214
+ 6,3,6,20,7,32,22,1,1,6,44,1
215
+ 6,3,6,21,6,25,33,1,1,7,46,1
216
+ 6,3,6,22,5,50,55,1,1,7,14,1
217
+ 6,3,6,23,6,24,40,1,1,7,5,1
218
+ 6,3,6,24,3,44,15,1,1,5,11,1
219
+ 6,3,6,25,7,57,39,1,1,7,16,1
220
+ 6,3,6,26,7,2,50,1,1,7,54,1
221
+ 6,3,6,27,7,3,59,1,1,7,29,1
222
+ 6,3,6,28,5,13,27,1,1,6,27,1
223
+ 6,3,6,29,7,1,9,1,1,5,62,1
224
+ 6,3,6,30,6,40,45,1,1,7,36,1
225
+ 6,3,6,31,3,39,31,1,1,6,27,1
226
+ 6,3,7,0,3,40,51,1,1,7,36,1
227
+ 6,3,7,1,6,27,10,1,1,5,4,1
228
+ 6,3,7,2,6,28,13,1,1,5,16,1
229
+ 6,3,7,3,7,26,48,1,1,7,50,1
230
+ 6,3,7,4,6,11,41,1,1,7,3,1
231
+ 6,3,7,5,5,44,37,1,1,7,40,1
232
+ 6,3,7,6,7,42,18,1,1,6,25,1
233
+ 6,3,7,7,3,40,51,1,1,7,29,1
234
+ 6,3,7,8,6,20,63,1,1,7,40,1
235
+ 6,3,7,9,6,39,6,1,1,4,38,1
236
+ 6,3,7,10,3,21,39,1,1,7,56,1
237
+ 6,3,7,11,5,10,19,1,1,6,22,1
238
+ 6,3,7,12,3,51,57,1,1,7,56,1
239
+ 6,3,7,13,7,45,46,1,1,7,38,1
240
+ 6,3,7,14,5,38,54,1,1,7,52,1
241
+ 6,3,7,15,3,38,30,1,1,6,34,1
242
+ 6,3,7,16,5,51,57,1,1,7,13,1
243
+ 6,3,7,17,6,47,52,1,1,7,2,1
244
+ 6,3,7,18,7,23,31,1,1,6,29,1
245
+ 6,3,7,19,5,30,38,1,1,7,6,1
246
+ 6,3,7,20,6,6,61,1,1,7,59,1
247
+ 6,3,7,21,3,28,47,1,1,7,52,1
248
+ 6,3,7,22,7,44,20,1,1,6,63,1
249
+ 6,3,7,23,3,53,38,1,1,7,50,1
250
+ 6,3,7,24,3,27,10,1,1,5,15,1
251
+ 6,3,7,25,3,59,11,1,1,5,37,1
252
+ 6,3,7,26,5,62,39,1,1,7,9,1
253
+ 6,3,7,27,7,62,22,1,1,6,8,1
254
+ 6,3,7,28,5,41,9,1,1,5,60,1
255
+ 6,3,7,29,3,53,38,1,1,7,55,1
256
+ 6,3,7,30,6,31,22,1,1,6,49,1
257
+ 6,3,7,31,5,8,50,1,1,7,35,1
258
+ 12,4,0,0,7,1683,2533,1,1,13,256,0
259
+ 12,4,0,1,6,2931,576,1,1,11,256,0
260
+ 12,4,0,2,15,1520,2199,1,1,13,256,0
261
+ 12,4,0,3,15,3541,608,1,1,11,256,0
262
+ 12,4,0,4,3,3517,3885,1,1,13,185,1
263
+ 12,4,0,5,13,3492,3844,1,1,13,256,0
264
+ 12,4,0,6,14,3969,1011,1,1,11,256,0
265
+ 12,4,0,7,11,670,1702,1,1,12,256,0
266
+ 12,4,0,8,10,1552,3739,1,1,13,256,0
267
+ 12,4,0,9,13,2677,418,1,1,10,256,0
268
+ 12,4,0,10,10,994,342,1,1,10,256,0
269
+ 12,4,0,11,9,2399,155,1,1,9,220,1
270
+ 12,4,0,12,11,2600,3140,1,1,13,256,0
271
+ 12,4,0,13,15,3357,464,1,1,10,256,0
272
+ 12,4,0,14,3,1321,1342,1,1,12,256,0
273
+ 12,4,0,15,12,2134,790,1,1,11,256,0
274
+ 12,4,0,16,7,3871,3297,1,1,13,256,0
275
+ 12,4,0,17,14,1866,2801,1,1,13,256,0
276
+ 12,4,0,18,15,2552,2452,1,1,13,256,0
277
+ 12,4,0,19,5,941,1519,1,1,12,256,0
278
+ 12,4,0,20,3,1254,1459,1,1,12,256,0
279
+ 12,4,0,21,12,2860,2851,1,1,13,256,0
280
+ 12,4,0,22,15,2155,1056,1,1,12,202,1
281
+ 12,4,0,23,10,4060,1976,1,1,12,256,0
282
+ 12,4,0,24,9,1826,1980,1,1,12,256,0
283
+ 12,4,0,25,11,2915,1960,1,1,12,256,0
284
+ 12,4,0,26,9,3073,3760,1,1,13,256,0
285
+ 12,4,0,27,10,2810,2203,1,1,13,256,0
286
+ 12,4,0,28,11,184,3351,1,1,13,256,0
287
+ 12,4,0,29,5,3685,2001,1,1,12,256,0
288
+ 12,4,0,30,5,3594,3122,1,1,13,256,0
289
+ 12,4,0,31,14,3730,1809,1,1,12,256,0
290
+ 12,4,1,0,9,200,3623,1,1,13,256,0
291
+ 12,4,1,1,12,2903,3265,1,1,13,256,0
292
+ 12,4,1,2,9,2953,3347,1,1,13,256,0
293
+ 12,4,1,3,11,1923,482,1,1,10,256,0
294
+ 12,4,1,4,6,3150,1472,1,1,12,256,0
295
+ 12,4,1,5,9,1060,3555,1,1,13,256,0
296
+ 12,4,1,6,12,566,3464,1,1,13,242,1
297
+ 12,4,1,7,11,2673,3287,1,1,13,256,0
298
+ 12,4,1,8,3,2283,2444,1,1,13,256,0
299
+ 12,4,1,9,10,1804,3635,1,1,13,256,0
300
+ 12,4,1,10,9,2791,2700,1,1,13,256,0
301
+ 12,4,1,11,3,2285,3678,1,1,13,256,0
302
+ 12,4,1,12,3,1748,50,1,1,7,256,0
303
+ 12,4,1,13,13,4037,3368,1,1,13,256,0
304
+ 12,4,1,14,7,3093,681,1,1,11,256,0
305
+ 12,4,1,15,7,2975,2628,1,1,13,256,0
306
+ 12,4,1,16,10,732,2872,1,1,13,256,0
307
+ 12,4,1,17,6,2603,2614,1,1,13,256,0
308
+ 12,4,1,18,3,563,669,1,1,11,256,0
309
+ 12,4,1,19,10,2288,2146,1,1,13,256,0
310
+ 12,4,1,20,13,733,18,1,1,6,256,0
311
+ 12,4,1,21,3,2975,2628,1,1,13,256,0
312
+ 12,4,1,22,15,1954,153,1,1,9,256,0
313
+ 12,4,1,23,15,2464,222,1,1,9,63,1
314
+ 12,4,1,24,11,806,3717,1,1,13,256,0
315
+ 12,4,1,25,5,3034,3296,1,1,13,256,0
316
+ 12,4,1,26,9,1148,589,1,1,11,256,0
317
+ 12,4,1,27,6,365,3699,1,1,13,256,0
318
+ 12,4,1,28,7,3363,1966,1,1,12,256,0
319
+ 12,4,1,29,14,1473,185,1,1,9,256,0
320
+ 12,4,1,30,13,2266,2674,1,1,13,256,0
321
+ 12,4,1,31,3,62,21,1,1,6,256,0
322
+ 12,4,2,0,13,528,73,1,1,8,256,0
323
+ 12,4,2,1,10,417,3800,1,1,13,256,0
324
+ 12,4,2,2,10,692,3612,1,1,13,91,1
325
+ 12,4,2,3,10,3900,1976,1,1,12,256,0
326
+ 12,4,2,4,10,3515,2908,1,1,13,256,0
327
+ 12,4,2,5,6,459,2092,1,1,13,103,1
328
+ 12,4,2,6,5,2144,103,1,1,8,256,0
329
+ 12,4,2,7,7,1296,422,1,1,10,256,0
330
+ 12,4,2,8,10,4077,1791,1,1,12,256,0
331
+ 12,4,2,9,15,2161,2990,1,1,13,256,0
332
+ 12,4,2,10,11,1386,887,1,1,11,256,0
333
+ 12,4,2,11,15,2670,555,1,1,11,78,1
334
+ 12,4,2,12,12,1837,140,1,1,9,256,0
335
+ 12,4,2,13,3,874,1710,1,1,12,256,0
336
+ 12,4,2,14,13,2154,3344,1,1,13,245,1
337
+ 12,4,2,15,9,128,362,1,1,10,256,0
338
+ 12,4,2,16,3,1475,3270,1,1,13,256,0
339
+ 12,4,2,17,13,1788,274,1,1,10,256,0
340
+ 12,4,2,18,6,4064,618,1,1,11,256,0
341
+ 12,4,2,19,6,1162,3023,1,1,13,256,0
342
+ 12,4,2,20,15,3292,2578,1,1,13,256,0
343
+ 12,4,2,21,6,3822,1139,1,1,12,256,0
344
+ 12,4,2,22,10,3129,857,1,1,11,256,0
345
+ 12,4,2,23,10,3618,1650,1,1,12,133,1
346
+ 12,4,2,24,10,762,1847,1,1,12,256,0
347
+ 12,4,2,25,7,57,3225,1,1,13,256,0
348
+ 12,4,2,26,9,2758,1391,1,1,12,256,0
349
+ 12,4,2,27,15,2207,1169,1,1,12,256,0
350
+ 12,4,2,28,14,1071,2824,1,1,13,256,0
351
+ 12,4,2,29,7,3990,3667,1,1,13,256,0
352
+ 12,4,2,30,14,3991,2008,1,1,12,256,0
353
+ 12,4,2,31,7,1565,782,1,1,11,256,0
354
+ 12,4,3,0,3,242,1883,1,1,12,256,0
355
+ 12,4,3,1,10,1766,1940,1,1,12,256,0
356
+ 12,4,3,2,13,3179,4092,1,1,13,58,1
357
+ 12,4,3,3,3,3401,1601,1,1,12,256,0
358
+ 12,4,3,4,5,2665,2107,1,1,13,256,0
359
+ 12,4,3,5,3,1580,1951,1,1,12,256,0
360
+ 12,4,3,6,7,392,2236,1,1,13,256,0
361
+ 12,4,3,7,15,253,605,1,1,11,256,0
362
+ 12,4,3,8,10,343,3382,1,1,13,238,1
363
+ 12,4,3,9,6,3329,1394,1,1,12,58,1
364
+ 12,4,3,10,15,1233,2208,1,1,13,256,0
365
+ 12,4,3,11,5,3255,2303,1,1,13,256,0
366
+ 12,4,3,12,15,3880,1952,1,1,12,256,0
367
+ 12,4,3,13,9,1594,766,1,1,11,256,0
368
+ 12,4,3,14,12,649,2958,1,1,13,256,0
369
+ 12,4,3,15,10,3108,1510,1,1,12,58,1
370
+ 12,4,3,16,3,3364,4073,1,1,13,256,0
371
+ 12,4,3,17,9,148,3906,1,1,13,256,0
372
+ 12,4,3,18,7,229,3030,1,1,13,256,0
373
+ 12,4,3,19,10,2686,1206,1,1,12,256,0
374
+ 12,4,3,20,5,2331,3374,1,1,13,256,0
375
+ 12,4,3,21,11,2257,1834,1,1,12,256,0
376
+ 12,4,3,22,9,1262,3499,1,1,13,256,0
377
+ 12,4,3,23,6,1889,1337,1,1,12,256,0
378
+ 12,4,3,24,13,2636,2514,1,1,13,256,0
379
+ 12,4,3,25,3,468,3833,1,1,13,256,0
380
+ 12,4,3,26,14,1704,2332,1,1,13,256,0
381
+ 12,4,3,27,10,2753,607,1,1,11,256,0
382
+ 12,4,3,28,6,3819,1492,1,1,12,256,0
383
+ 12,4,3,29,9,3105,3708,1,1,13,256,0
384
+ 12,4,3,30,12,83,1359,1,1,12,256,0
385
+ 12,4,3,31,13,1277,313,1,1,10,256,0
386
+ 12,4,4,0,3,3213,1717,1,1,12,256,0
387
+ 12,4,4,1,3,341,2222,1,1,13,256,0
388
+ 12,4,4,2,12,2323,277,1,1,10,256,0
389
+ 12,4,4,3,14,2814,3261,1,1,13,256,0
390
+ 12,4,4,4,9,1939,2482,1,1,13,256,0
391
+ 12,4,4,5,6,2345,2126,1,1,13,256,0
392
+ 12,4,4,6,6,999,2935,1,1,13,256,0
393
+ 12,4,4,7,5,2488,2087,1,1,13,256,0
394
+ 12,4,4,8,10,3821,2874,1,1,13,256,0
395
+ 12,4,4,9,13,2625,1716,1,1,12,256,0
396
+ 12,4,4,10,9,1006,2143,1,1,13,256,0
397
+ 12,4,4,11,9,810,3629,1,1,13,256,0
398
+ 12,4,4,12,9,3655,3885,1,1,13,256,0
399
+ 12,4,4,13,9,944,636,1,1,11,256,0
400
+ 12,4,4,14,12,2759,17,1,1,6,256,0
401
+ 12,4,4,15,13,314,1925,1,1,12,256,0
402
+ 12,4,4,16,14,2634,3816,1,1,13,256,0
403
+ 12,4,4,17,14,964,2975,1,1,13,256,0
404
+ 12,4,4,18,14,3356,3174,1,1,13,256,0
405
+ 12,4,4,19,13,3662,2145,1,1,13,256,0
406
+ 12,4,4,20,10,701,1578,1,1,12,256,0
407
+ 12,4,4,21,13,1567,212,1,1,9,256,0
408
+ 12,4,4,22,15,1909,1444,1,1,12,256,0
409
+ 12,4,4,23,5,2658,414,1,1,10,256,0
410
+ 12,4,4,24,9,659,2952,1,1,13,179,1
411
+ 12,4,4,25,6,3060,1212,1,1,12,256,0
412
+ 12,4,4,26,10,3256,2678,1,1,13,256,0
413
+ 12,4,4,27,6,3259,2115,1,1,13,256,0
414
+ 12,4,4,28,9,3962,155,1,1,9,42,1
415
+ 12,4,4,29,7,1037,2489,1,1,13,256,0
416
+ 12,4,4,30,13,1675,1018,1,1,11,256,0
417
+ 12,4,4,31,7,3201,3750,1,1,13,256,0
418
+ 12,4,5,0,3,3520,3435,1,1,13,256,0
419
+ 12,4,5,1,7,264,1578,1,1,12,256,0
420
+ 12,4,5,2,12,2963,1741,1,1,12,256,0
421
+ 12,4,5,3,14,1786,2523,1,1,13,256,0
422
+ 12,4,5,4,12,2498,548,1,1,11,256,0
423
+ 12,4,5,5,12,1982,2801,1,1,13,256,0
424
+ 12,4,5,6,3,1972,3776,1,1,13,256,0
425
+ 12,4,5,7,7,587,2347,1,1,13,115,1
426
+ 12,4,5,8,3,2040,2636,1,1,13,256,0
427
+ 12,4,5,9,9,2575,2407,1,1,13,256,0
428
+ 12,4,5,10,5,904,1820,1,1,12,256,0
429
+ 12,4,5,11,7,3452,2458,1,1,13,256,0
430
+ 12,4,5,12,3,3312,3574,1,1,13,256,0
431
+ 12,4,5,13,6,2224,148,1,1,9,256,0
432
+ 12,4,5,14,3,1197,3226,1,1,13,256,0
433
+ 12,4,5,15,15,1193,1387,1,1,12,256,0
434
+ 12,4,5,16,14,3173,3348,1,1,13,256,0
435
+ 12,4,5,17,13,923,4052,1,1,13,256,0
436
+ 12,4,5,18,3,3994,2737,1,1,13,256,0
437
+ 12,4,5,19,13,2560,3708,1,1,13,220,1
438
+ 12,4,5,20,7,2027,3432,1,1,13,256,0
439
+ 12,4,5,21,13,2543,3199,1,1,13,256,0
440
+ 12,4,5,22,5,3513,1830,1,1,12,256,0
441
+ 12,4,5,23,12,474,2970,1,1,13,256,0
442
+ 12,4,5,24,9,2058,1089,1,1,12,256,0
443
+ 12,4,5,25,5,3303,1047,1,1,12,256,0
444
+ 12,4,5,26,14,2280,1050,1,1,12,256,0
445
+ 12,4,5,27,7,789,3374,1,1,13,256,0
446
+ 12,4,5,28,14,2357,2346,1,1,13,256,0
447
+ 12,4,5,29,11,423,839,1,1,11,256,0
448
+ 12,4,5,30,9,1214,746,1,1,11,256,0
449
+ 12,4,5,31,12,340,1676,1,1,12,256,0
450
+ 12,4,6,0,9,1823,1931,1,1,12,256,0
451
+ 12,4,6,1,5,2889,489,1,1,10,256,0
452
+ 12,4,6,2,15,128,2373,1,1,13,256,0
453
+ 12,4,6,3,3,1434,186,1,1,9,256,0
454
+ 12,4,6,4,3,1349,1157,1,1,12,256,0
455
+ 12,4,6,5,11,1087,2551,1,1,13,256,0
456
+ 12,4,6,6,10,2991,3005,1,1,13,256,0
457
+ 12,4,6,7,7,1863,2051,1,1,13,256,0
458
+ 12,4,6,8,6,1587,2709,1,1,13,256,0
459
+ 12,4,6,9,6,3125,3884,1,1,13,256,0
460
+ 12,4,6,10,5,1119,1576,1,1,12,256,0
461
+ 12,4,6,11,3,3206,2110,1,1,13,256,0
462
+ 12,4,6,12,3,384,2432,1,1,13,256,0
463
+ 12,4,6,13,14,1524,3199,1,1,13,256,0
464
+ 12,4,6,14,6,3305,3782,1,1,13,256,0
465
+ 12,4,6,15,7,3880,3101,1,1,13,256,0
466
+ 12,4,6,16,10,387,2030,1,1,12,256,0
467
+ 12,4,6,17,3,1418,3145,1,1,13,256,0
468
+ 12,4,6,18,9,2376,344,1,1,10,256,0
469
+ 12,4,6,19,12,3390,1512,1,1,12,212,1
470
+ 12,4,6,20,9,3936,645,1,1,11,256,0
471
+ 12,4,6,21,10,1155,3193,1,1,13,256,0
472
+ 12,4,6,22,5,3900,3627,1,1,13,256,0
473
+ 12,4,6,23,13,3859,125,1,1,8,256,0
474
+ 12,4,6,24,5,1988,2157,1,1,13,256,0
475
+ 12,4,6,25,5,1524,3199,1,1,13,256,0
476
+ 12,4,6,26,10,3637,417,1,1,10,256,0
477
+ 12,4,6,27,6,359,1619,1,1,12,256,0
478
+ 12,4,6,28,5,122,3055,1,1,13,256,0
479
+ 12,4,6,29,12,3121,2642,1,1,13,256,0
480
+ 12,4,6,30,9,1753,1489,1,1,12,256,0
481
+ 12,4,6,31,5,189,845,1,1,11,256,0
482
+ 12,4,7,0,10,1253,789,1,1,11,256,0
483
+ 12,4,7,1,9,1138,3339,1,1,13,256,0
484
+ 12,4,7,2,15,274,2456,1,1,13,256,0
485
+ 12,4,7,3,14,3426,2320,1,1,13,256,0
486
+ 12,4,7,4,15,464,453,1,1,10,256,0
487
+ 12,4,7,5,3,2078,770,1,1,11,256,0
488
+ 12,4,7,6,14,3587,413,1,1,10,256,0
489
+ 12,4,7,7,7,1496,3701,1,1,13,256,0
490
+ 12,4,7,8,10,3244,3643,1,1,13,256,0
491
+ 12,4,7,9,3,2324,2708,1,1,13,256,0
492
+ 12,4,7,10,12,660,1838,1,1,12,256,0
493
+ 12,4,7,11,11,2046,63,1,1,7,256,0
494
+ 12,4,7,12,15,44,1017,1,1,11,256,0
495
+ 12,4,7,13,11,1572,3574,1,1,13,256,0
496
+ 12,4,7,14,9,2339,3546,1,1,13,256,0
497
+ 12,4,7,15,10,434,1626,1,1,12,256,0
498
+ 12,4,7,16,3,2395,548,1,1,11,256,0
499
+ 12,4,7,17,12,2996,2271,1,1,13,256,0
500
+ 12,4,7,18,14,429,122,1,1,8,256,0
501
+ 12,4,7,19,11,3098,1630,1,1,12,256,0
502
+ 12,4,7,20,6,2394,3826,1,1,13,256,0
503
+ 12,4,7,21,5,3333,2514,1,1,13,256,0
504
+ 12,4,7,22,15,762,2675,1,1,13,256,0
505
+ 12,4,7,23,10,4073,1030,1,1,12,256,0
506
+ 12,4,7,24,14,3788,4090,1,1,13,256,0
507
+ 12,4,7,25,10,3930,2832,1,1,13,9,1
508
+ 12,4,7,26,15,351,117,1,1,8,256,0
509
+ 12,4,7,27,14,1133,1163,1,1,12,231,1
510
+ 12,4,7,28,7,3250,2299,1,1,13,256,0
511
+ 12,4,7,29,9,2359,2848,1,1,13,256,0
512
+ 12,4,7,30,13,328,2236,1,1,13,199,1
513
+ 12,4,7,31,15,451,3367,1,1,13,256,0
514
+ 24,5,0,0,27,10901770,9089085,1,1,25,256,0
515
+ 24,5,0,1,21,11602593,12429211,1,1,25,256,0
516
+ 24,5,0,2,11,15063014,16067123,1,1,25,256,0
517
+ 24,5,0,3,13,7245746,5253644,1,1,24,256,0
518
+ 24,5,0,4,30,12304976,4907436,1,1,24,256,0
519
+ 24,5,0,5,14,12591887,9203256,1,1,25,256,0
520
+ 24,5,0,6,10,10478036,9332013,1,1,25,256,0
521
+ 24,5,0,7,9,9493767,5821561,1,1,24,256,0
522
+ 24,5,0,8,11,4149671,15133054,1,1,25,256,0
523
+ 24,5,0,9,23,7769254,15111433,1,1,25,256,0
524
+ 24,5,0,10,18,12756754,8227723,1,1,24,256,0
525
+ 24,5,0,11,6,3959438,12111036,1,1,25,256,0
526
+ 24,5,0,12,6,1355096,16205217,1,1,25,256,0
527
+ 24,5,0,13,21,355142,8995128,1,1,25,256,0
528
+ 24,5,0,14,10,15037625,8228849,1,1,24,256,0
529
+ 24,5,0,15,11,7644853,15000453,1,1,25,256,0
530
+ 24,5,0,16,17,13059928,3644443,1,1,23,256,0
531
+ 24,5,0,17,12,4387630,6421190,1,1,24,256,0
532
+ 24,5,0,18,26,7147208,8730913,1,1,25,256,0
533
+ 24,5,0,19,17,14970038,4924464,1,1,24,256,0
534
+ 24,5,0,20,14,6549766,10943929,1,1,25,256,0
535
+ 24,5,0,21,14,4970350,6837648,1,1,24,256,0
536
+ 24,5,0,22,27,5698793,7406769,1,1,24,256,0
537
+ 24,5,0,23,29,7368204,12472826,1,1,25,256,0
538
+ 24,5,0,24,11,13138119,9276251,1,1,25,256,0
539
+ 24,5,0,25,18,10172770,11571073,1,1,25,256,0
540
+ 24,5,0,26,17,7309646,7469588,1,1,24,256,0
541
+ 24,5,0,27,28,12656729,4670473,1,1,24,256,0
542
+ 24,5,0,28,7,14220356,5319544,1,1,24,256,0
543
+ 24,5,0,29,9,3057073,7092029,1,1,24,256,0
544
+ 24,5,0,30,17,14694116,220836,1,1,19,256,0
545
+ 24,5,0,31,15,14382813,3860884,1,1,23,256,0
546
+ 24,5,1,0,3,15221257,12713871,1,1,25,256,0
547
+ 24,5,1,1,5,11990508,3775514,1,1,23,256,0
548
+ 24,5,1,2,3,13338368,9168389,1,1,25,256,0
549
+ 24,5,1,3,29,9719485,535301,1,1,21,256,0
550
+ 24,5,1,4,22,12169220,10138667,1,1,25,256,0
551
+ 24,5,1,5,25,15554404,9753657,1,1,25,256,0
552
+ 24,5,1,6,18,1091551,2511924,1,1,23,256,0
553
+ 24,5,1,7,17,12954332,3123980,1,1,23,256,0
554
+ 24,5,1,8,12,206887,5111847,1,1,24,256,0
555
+ 24,5,1,9,13,3432207,8878066,1,1,25,256,0
556
+ 24,5,1,10,11,5751320,10603094,1,1,25,256,0
557
+ 24,5,1,11,27,1163677,9500642,1,1,25,256,0
558
+ 24,5,1,12,27,13725235,4648229,1,1,24,256,0
559
+ 24,5,1,13,31,13322833,2285868,1,1,23,256,0
560
+ 24,5,1,14,17,15324345,5047670,1,1,24,256,0
561
+ 24,5,1,15,13,9768541,10204127,1,1,25,256,0
562
+ 24,5,1,16,26,6528079,10663071,1,1,25,256,0
563
+ 24,5,1,17,5,8652267,16600014,1,1,25,256,0
564
+ 24,5,1,18,29,16281994,2133665,1,1,23,256,0
565
+ 24,5,1,19,27,7867652,3000297,1,1,23,256,0
566
+ 24,5,1,20,28,6615202,6989820,1,1,24,256,0
567
+ 24,5,1,21,22,15146419,16728471,1,1,25,256,0
568
+ 24,5,1,22,14,1417712,3481778,1,1,23,256,0
569
+ 24,5,1,23,24,2676002,14662248,1,1,25,256,0
570
+ 24,5,1,24,29,808311,6664566,1,1,24,256,0
571
+ 24,5,1,25,14,5068597,7582063,1,1,24,256,0
572
+ 24,5,1,26,15,15060867,13274149,1,1,25,256,0
573
+ 24,5,1,27,15,7493453,14019180,1,1,25,256,0
574
+ 24,5,1,28,5,11814444,8083740,1,1,24,256,0
575
+ 24,5,1,29,18,1170578,6840434,1,1,24,256,0
576
+ 24,5,1,30,22,1092218,12295001,1,1,25,256,0
577
+ 24,5,1,31,22,6314507,6322628,1,1,24,256,0
578
+ 24,5,2,0,26,432140,8884084,1,1,25,256,0
579
+ 24,5,2,1,26,9471420,8434946,1,1,25,256,0
580
+ 24,5,2,2,15,15788320,11699428,1,1,25,256,0
581
+ 24,5,2,3,15,9748444,13652237,1,1,25,256,0
582
+ 24,5,2,4,25,7440461,1545902,1,1,22,256,0
583
+ 24,5,2,5,10,2056419,2281388,1,1,23,256,0
584
+ 24,5,2,6,5,10542193,2487887,1,1,23,256,0
585
+ 24,5,2,7,14,643804,5439851,1,1,24,256,0
586
+ 24,5,2,8,13,14830320,721141,1,1,21,256,0
587
+ 24,5,2,9,23,13962356,8503773,1,1,25,256,0
588
+ 24,5,2,10,27,1249323,14366904,1,1,25,256,0
589
+ 24,5,2,11,14,12845516,8458283,1,1,25,256,0
590
+ 24,5,2,12,28,16497379,15855426,1,1,25,256,0
591
+ 24,5,2,13,20,15490042,16041157,1,1,25,256,0
592
+ 24,5,2,14,7,9170687,3490665,1,1,23,256,0
593
+ 24,5,2,15,20,3409246,3825466,1,1,23,256,0
594
+ 24,5,2,16,22,11747879,14539141,1,1,25,256,0
595
+ 24,5,2,17,29,8258153,11915093,1,1,25,256,0
596
+ 24,5,2,18,17,7578282,13528483,1,1,25,256,0
597
+ 24,5,2,19,26,8254575,8690515,1,1,25,256,0
598
+ 24,5,2,20,28,13200793,4615594,1,1,24,256,0
599
+ 24,5,2,21,7,10557287,13276016,1,1,25,256,0
600
+ 24,5,2,22,3,13427007,14207478,1,1,25,256,0
601
+ 24,5,2,23,30,7808978,8232068,1,1,24,256,0
602
+ 24,5,2,24,18,385792,9793107,1,1,25,256,0
603
+ 24,5,2,25,20,1869461,1158660,1,1,22,256,0
604
+ 24,5,2,26,15,12301025,7186451,1,1,24,256,0
605
+ 24,5,2,27,12,3056897,13768126,1,1,25,256,0
606
+ 24,5,2,28,18,412491,5442503,1,1,24,256,0
607
+ 24,5,2,29,6,5372955,904402,1,1,21,256,0
608
+ 24,5,2,30,11,3241011,9783149,1,1,25,256,0
609
+ 24,5,2,31,13,16619436,2885976,1,1,23,256,0
610
+ 24,5,3,0,14,907677,1578267,1,1,22,256,0
611
+ 24,5,3,1,11,14834104,9656138,1,1,25,256,0
612
+ 24,5,3,2,10,3915385,8047164,1,1,24,256,0
613
+ 24,5,3,3,14,3895305,2364845,1,1,23,256,0
614
+ 24,5,3,4,23,11762294,457582,1,1,20,256,0
615
+ 24,5,3,5,25,11728686,6489782,1,1,24,256,0
616
+ 24,5,3,6,31,4954128,8342122,1,1,24,256,0
617
+ 24,5,3,7,19,10518665,12169195,1,1,25,256,0
618
+ 24,5,3,8,22,10655123,6198786,1,1,24,256,0
619
+ 24,5,3,9,17,10551647,2740115,1,1,23,256,0
620
+ 24,5,3,10,3,4832468,16221077,1,1,25,256,0
621
+ 24,5,3,11,19,5652896,483336,1,1,20,256,0
622
+ 24,5,3,12,17,1194015,15837982,1,1,25,256,0
623
+ 24,5,3,13,9,1750493,1330162,1,1,22,256,0
624
+ 24,5,3,14,12,10798878,4734903,1,1,24,256,0
625
+ 24,5,3,15,27,4015612,7827005,1,1,24,256,0
626
+ 24,5,3,16,21,13028073,12180068,1,1,25,256,0
627
+ 24,5,3,17,13,13467674,6666456,1,1,24,256,0
628
+ 24,5,3,18,28,8362215,4262343,1,1,24,256,0
629
+ 24,5,3,19,5,8388630,4341777,1,1,24,256,0
630
+ 24,5,3,20,31,4042339,5382818,1,1,24,256,0
631
+ 24,5,3,21,22,12076579,5520398,1,1,24,256,0
632
+ 24,5,3,22,14,15132918,730813,1,1,21,256,0
633
+ 24,5,3,23,3,3393801,13503319,1,1,25,256,0
634
+ 24,5,3,24,18,8047638,2350162,1,1,23,256,0
635
+ 24,5,3,25,31,13656489,783443,1,1,21,256,0
636
+ 24,5,3,26,18,4117743,6345564,1,1,24,256,0
637
+ 24,5,3,27,27,2948553,6070725,1,1,24,256,0
638
+ 24,5,3,28,25,13113444,14591371,1,1,25,256,0
639
+ 24,5,3,29,29,14545733,16662581,1,1,25,256,0
640
+ 24,5,3,30,9,2796070,3578738,1,1,23,256,0
641
+ 24,5,3,31,6,4376162,16649964,1,1,25,256,0
642
+ 24,5,4,0,10,15239556,7774240,1,1,24,256,0
643
+ 24,5,4,1,10,194537,11773091,1,1,25,256,0
644
+ 24,5,4,2,29,13884171,11752354,1,1,25,256,0
645
+ 24,5,4,3,19,5663021,13189943,1,1,25,256,0
646
+ 24,5,4,4,31,13036588,262230,1,1,20,256,0
647
+ 24,5,4,5,17,14388954,3237529,1,1,23,256,0
648
+ 24,5,4,6,14,4594021,111556,1,1,18,256,0
649
+ 24,5,4,7,25,3665275,2698928,1,1,23,256,0
650
+ 24,5,4,8,20,5956875,123321,1,1,18,256,0
651
+ 24,5,4,9,9,1179978,14139225,1,1,25,256,0
652
+ 24,5,4,10,23,1305030,4403298,1,1,24,256,0
653
+ 24,5,4,11,7,7274492,3772086,1,1,23,256,0
654
+ 24,5,4,12,5,4655383,8102530,1,1,24,256,0
655
+ 24,5,4,13,28,5483226,6493350,1,1,24,256,0
656
+ 24,5,4,14,3,11888867,11831428,1,1,25,256,0
657
+ 24,5,4,15,19,3905424,5298312,1,1,24,256,0
658
+ 24,5,4,16,27,1406009,10180198,1,1,25,256,0
659
+ 24,5,4,17,3,14945705,5675975,1,1,24,256,0
660
+ 24,5,4,18,11,5257803,14044685,1,1,25,256,0
661
+ 24,5,4,19,21,2299986,11786875,1,1,25,256,0
662
+ 24,5,4,20,3,9055008,7150724,1,1,24,256,0
663
+ 24,5,4,21,11,325332,8990544,1,1,25,256,0
664
+ 24,5,4,22,29,8539072,15656919,1,1,25,256,0
665
+ 24,5,4,23,21,16080614,11270714,1,1,25,256,0
666
+ 24,5,4,24,21,393802,9428110,1,1,25,256,0
667
+ 24,5,4,25,19,564891,13786106,1,1,25,256,0
668
+ 24,5,4,26,11,10604340,9003250,1,1,25,256,0
669
+ 24,5,4,27,6,3419412,12368937,1,1,25,256,0
670
+ 24,5,4,28,23,974659,1925193,1,1,22,256,0
671
+ 24,5,4,29,6,8707301,3152099,1,1,23,256,0
672
+ 24,5,4,30,30,5097710,16514986,1,1,25,256,0
673
+ 24,5,4,31,27,9336539,3236342,1,1,23,256,0
674
+ 24,5,5,0,5,9723257,15180443,1,1,25,256,0
675
+ 24,5,5,1,27,7681799,2363632,1,1,23,256,0
676
+ 24,5,5,2,3,5654460,14155894,1,1,25,256,0
677
+ 24,5,5,3,18,6351777,9119079,1,1,25,256,0
678
+ 24,5,5,4,25,11695914,3691468,1,1,23,256,0
679
+ 24,5,5,5,28,10551041,2325063,1,1,23,256,0
680
+ 24,5,5,6,3,2498849,9090327,1,1,25,256,0
681
+ 24,5,5,7,3,2950789,4369566,1,1,24,256,0
682
+ 24,5,5,8,5,12781955,1705813,1,1,22,256,0
683
+ 24,5,5,9,11,6523208,1577874,1,1,22,256,0
684
+ 24,5,5,10,21,15868779,1092354,1,1,22,256,0
685
+ 24,5,5,11,27,8351519,7244416,1,1,24,256,0
686
+ 24,5,5,12,21,3915851,14543442,1,1,25,256,0
687
+ 24,5,5,13,26,4226678,5598784,1,1,24,256,0
688
+ 24,5,5,14,5,10512724,14621969,1,1,25,256,0
689
+ 24,5,5,15,3,15825128,13177522,1,1,25,256,0
690
+ 24,5,5,16,20,4777003,15970474,1,1,25,256,0
691
+ 24,5,5,17,9,7271469,1648575,1,1,22,256,0
692
+ 24,5,5,18,15,541899,539515,1,1,21,256,0
693
+ 24,5,5,19,21,5140624,2260792,1,1,23,256,0
694
+ 24,5,5,20,25,11681328,4915107,1,1,24,256,0
695
+ 24,5,5,21,7,6611573,5342017,1,1,24,256,0
696
+ 24,5,5,22,30,14173710,1317325,1,1,22,256,0
697
+ 24,5,5,23,25,13648263,2311611,1,1,23,256,0
698
+ 24,5,5,24,22,6169853,13502034,1,1,25,256,0
699
+ 24,5,5,25,27,5161885,14007693,1,1,25,256,0
700
+ 24,5,5,26,22,14704074,3586103,1,1,23,256,0
701
+ 24,5,5,27,7,5105359,13141084,1,1,25,256,0
702
+ 24,5,5,28,10,16563131,6412064,1,1,24,256,0
703
+ 24,5,5,29,22,14876582,3302268,1,1,23,256,0
704
+ 24,5,5,30,3,8541955,8266111,1,1,24,256,0
705
+ 24,5,5,31,3,7774267,12444935,1,1,25,256,0
706
+ 24,5,6,0,7,8306081,1408789,1,1,22,256,0
707
+ 24,5,6,1,18,2751946,7722463,1,1,24,256,0
708
+ 24,5,6,2,23,13609214,6897410,1,1,24,256,0
709
+ 24,5,6,3,17,12632145,10763904,1,1,25,256,0
710
+ 24,5,6,4,23,11618200,8293438,1,1,24,256,0
711
+ 24,5,6,5,21,5363029,3582551,1,1,23,256,0
712
+ 24,5,6,6,29,12615994,4647089,1,1,24,256,0
713
+ 24,5,6,7,27,9275524,814662,1,1,21,256,0
714
+ 24,5,6,8,19,7145315,10027339,1,1,25,256,0
715
+ 24,5,6,9,6,2868339,1469860,1,1,22,256,0
716
+ 24,5,6,10,11,8611711,3022244,1,1,23,256,0
717
+ 24,5,6,11,29,3090858,13468529,1,1,25,256,0
718
+ 24,5,6,12,21,9923540,7033884,1,1,24,256,0
719
+ 24,5,6,13,31,11773312,1689980,1,1,22,256,0
720
+ 24,5,6,14,11,14422904,1406442,1,1,22,256,0
721
+ 24,5,6,15,24,14267059,8179088,1,1,24,256,0
722
+ 24,5,6,16,28,898444,15351296,1,1,25,256,0
723
+ 24,5,6,17,27,15126351,15059976,1,1,25,256,0
724
+ 24,5,6,18,22,3293498,735535,1,1,21,256,0
725
+ 24,5,6,19,11,6548922,14880593,1,1,25,256,0
726
+ 24,5,6,20,10,4453600,14107271,1,1,25,256,0
727
+ 24,5,6,21,25,9598737,3389362,1,1,23,256,0
728
+ 24,5,6,22,18,14472788,15254736,1,1,25,256,0
729
+ 24,5,6,23,26,4912218,6953508,1,1,24,256,0
730
+ 24,5,6,24,23,9639419,14010124,1,1,25,256,0
731
+ 24,5,6,25,11,12794501,6628316,1,1,24,256,0
732
+ 24,5,6,26,22,15278923,9298029,1,1,25,256,0
733
+ 24,5,6,27,31,5259190,2470155,1,1,23,256,0
734
+ 24,5,6,28,29,10051188,9418934,1,1,25,256,0
735
+ 24,5,6,29,9,2400714,12983992,1,1,25,256,0
736
+ 24,5,6,30,5,5308382,7711593,1,1,24,256,0
737
+ 24,5,6,31,18,3725525,10253727,1,1,25,256,0
738
+ 24,5,7,0,15,2932452,8945269,1,1,25,256,0
739
+ 24,5,7,1,15,8836091,7949336,1,1,24,256,0
740
+ 24,5,7,2,21,14146272,14584505,1,1,25,256,0
741
+ 24,5,7,3,31,10884535,1789684,1,1,22,256,0
742
+ 24,5,7,4,27,5066983,378396,1,1,20,256,0
743
+ 24,5,7,5,29,702728,12544909,1,1,25,256,0
744
+ 24,5,7,6,18,14464884,8635402,1,1,25,256,0
745
+ 24,5,7,7,27,5675929,15092483,1,1,25,256,0
746
+ 24,5,7,8,25,4437640,4305399,1,1,24,256,0
747
+ 24,5,7,9,31,5748242,644975,1,1,21,256,0
748
+ 24,5,7,10,23,1540592,13338221,1,1,25,256,0
749
+ 24,5,7,11,19,14021298,245592,1,1,19,256,0
750
+ 24,5,7,12,20,98564,14066489,1,1,25,256,0
751
+ 24,5,7,13,9,13526388,11463042,1,1,25,256,0
752
+ 24,5,7,14,17,442114,8727563,1,1,25,256,0
753
+ 24,5,7,15,9,2944095,2967808,1,1,23,256,0
754
+ 24,5,7,16,12,5573755,11651217,1,1,25,256,0
755
+ 24,5,7,17,30,4188835,12171582,1,1,25,256,0
756
+ 24,5,7,18,12,2578003,14699707,1,1,25,256,0
757
+ 24,5,7,19,18,11225251,2643137,1,1,23,256,0
758
+ 24,5,7,20,6,3398088,386071,1,1,20,256,0
759
+ 24,5,7,21,23,6683020,16315273,1,1,25,256,0
760
+ 24,5,7,22,26,13258228,9457209,1,1,25,256,0
761
+ 24,5,7,23,27,13195279,2852033,1,1,23,256,0
762
+ 24,5,7,24,7,8860231,731959,1,1,21,256,0
763
+ 24,5,7,25,18,1047072,13253016,1,1,25,256,0
764
+ 24,5,7,26,21,15447894,3300494,1,1,23,256,0
765
+ 24,5,7,27,20,1373603,1941466,1,1,22,256,0
766
+ 24,5,7,28,12,1614799,10626437,1,1,25,256,0
767
+ 24,5,7,29,3,3685289,15214114,1,1,25,256,0
768
+ 24,5,7,30,27,6869554,13243043,1,1,25,256,0
769
+ 24,5,7,31,31,14792261,3351910,1,1,23,256,0
770
+ 48,6,0,0,9,50014221155405,117837298897727,1,1,48,256,0
771
+ 48,6,0,1,23,27437758963102,85928104794243,1,1,48,256,0
772
+ 48,6,0,2,56,108175098111184,53479394919753,1,1,47,256,0
773
+ 48,6,0,3,50,267321543884545,161790310287086,1,1,49,256,0
774
+ 48,6,0,4,23,18506652607962,182302451856159,1,1,49,256,0
775
+ 48,6,0,5,35,275752627679430,164246087659169,1,1,49,256,0
776
+ 48,6,0,6,10,189014447674011,173065557482452,1,1,49,256,0
777
+ 48,6,0,7,17,149986865020809,126953762744778,1,1,48,256,0
778
+ 48,6,0,8,48,248221031697719,86179737298401,1,1,48,256,0
779
+ 48,6,0,9,44,89585644327413,182982238902330,1,1,49,256,0
780
+ 48,6,0,10,22,136095749099881,212327226149659,1,1,49,256,0
781
+ 48,6,0,11,12,24804970247805,251066066827292,1,1,49,256,0
782
+ 48,6,0,12,49,175344827714514,44532117333206,1,1,47,256,0
783
+ 48,6,0,13,62,271929214059439,43207782433447,1,1,47,256,0
784
+ 48,6,0,14,9,202500637859247,239299305814244,1,1,49,256,0
785
+ 48,6,0,15,24,5752107590072,124514402515049,1,1,48,256,0
786
+ 48,6,0,16,43,70772374676022,62239326092146,1,1,47,256,0
787
+ 48,6,0,17,56,49596513128482,16745703935967,1,1,45,256,0
788
+ 48,6,0,18,31,272627566051649,21298003041988,1,1,46,256,0
789
+ 48,6,0,19,50,237099665428640,196713904949767,1,1,49,256,0
790
+ 48,6,0,20,51,225291576319765,1770996732291,1,1,42,256,0
791
+ 48,6,0,21,26,23688978803952,162438398465240,1,1,49,256,0
792
+ 48,6,0,22,34,132826159081664,119596950529034,1,1,48,256,0
793
+ 48,6,0,23,56,161358753862209,269196091543088,1,1,49,256,0
794
+ 48,6,0,24,49,88551163911079,97373490506526,1,1,48,256,0
795
+ 48,6,0,25,23,149864608789728,52912208868667,1,1,47,256,0
796
+ 48,6,0,26,53,86929305330835,52397080692097,1,1,47,256,0
797
+ 48,6,0,27,19,10588660173183,118978453252308,1,1,48,256,0
798
+ 48,6,0,28,17,237931196821709,262283772650112,1,1,49,256,0
799
+ 48,6,0,29,41,262120653939431,241142608638433,1,1,49,256,0
800
+ 48,6,0,30,24,65665046404949,6943531108756,1,1,44,256,0
801
+ 48,6,0,31,38,133941333480175,197007054091896,1,1,49,256,0
802
+ 48,6,1,0,57,202144506958988,880755109524,1,1,41,256,0
803
+ 48,6,1,1,10,263158766884060,277556299813044,1,1,49,256,0
804
+ 48,6,1,2,18,145454857491182,47283567453806,1,1,47,256,0
805
+ 48,6,1,3,33,85693064508516,165133935552701,1,1,49,256,0
806
+ 48,6,1,4,50,36239230749766,219525771396061,1,1,49,256,0
807
+ 48,6,1,5,51,72758557012481,259397968562608,1,1,49,256,0
808
+ 48,6,1,6,56,118308825260519,137344297334857,1,1,48,256,0
809
+ 48,6,1,7,10,142149551767464,218819154954744,1,1,49,256,0
810
+ 48,6,1,8,28,249733731334619,77385591137429,1,1,48,256,0
811
+ 48,6,1,9,61,188746873589496,265441801892725,1,1,49,256,0
812
+ 48,6,1,10,14,155114352356509,238944676305066,1,1,49,256,0
813
+ 48,6,1,11,28,152950874310542,250207341919916,1,1,49,256,0
814
+ 48,6,1,12,50,90657845155696,106788309106531,1,1,48,256,0
815
+ 48,6,1,13,57,194790124889691,46558352033890,1,1,47,256,0
816
+ 48,6,1,14,22,247713669632225,26730424240360,1,1,46,256,0
817
+ 48,6,1,15,38,100468723125586,113405404729874,1,1,48,256,0
818
+ 48,6,1,16,52,262502109002195,279266397386958,1,1,49,256,0
819
+ 48,6,1,17,40,140056139093535,251537474438251,1,1,49,256,0
820
+ 48,6,1,18,35,56255333774935,61187153455825,1,1,47,256,0
821
+ 48,6,1,19,6,100558919427210,112482962087989,1,1,48,256,0
822
+ 48,6,1,20,49,263201807362281,37226117366638,1,1,47,256,0
823
+ 48,6,1,21,50,135516761981171,180794871165208,1,1,49,256,0
824
+ 48,6,1,22,38,194758543762413,109983305573172,1,1,48,256,0
825
+ 48,6,1,23,31,62114266117822,259901380971103,1,1,49,256,0
826
+ 48,6,1,24,20,229546131504606,268328988053973,1,1,49,256,0
827
+ 48,6,1,25,41,247804046661716,129410260585667,1,1,48,256,0
828
+ 48,6,1,26,53,105818464388721,181691314078483,1,1,49,256,0
829
+ 48,6,1,27,54,262350180489941,38727748794538,1,1,47,256,0
830
+ 48,6,1,28,20,242298509758896,17010919377396,1,1,45,256,0
831
+ 48,6,1,29,31,198577092902744,68461134425070,1,1,47,256,0
832
+ 48,6,1,30,11,247162071110431,136399229329862,1,1,48,256,0
833
+ 48,6,1,31,52,147833094374569,83574304916635,1,1,48,256,0
834
+ 48,6,2,0,51,89289167782195,25347244412333,1,1,46,256,0
835
+ 48,6,2,1,60,19468651823861,168916014478866,1,1,49,256,0
836
+ 48,6,2,2,41,151274088326738,193283493651251,1,1,49,256,0
837
+ 48,6,2,3,24,244598813371261,123587774643458,1,1,48,256,0
838
+ 48,6,2,4,21,265401579302042,59536206930361,1,1,47,256,0
839
+ 48,6,2,5,19,172218933776117,13544744042739,1,1,45,256,0
840
+ 48,6,2,6,23,72936117974927,37011936932572,1,1,47,256,0
841
+ 48,6,2,7,21,267785010743056,251405314346727,1,1,49,256,0
842
+ 48,6,2,8,12,87216917813208,274634694144581,1,1,49,256,0
843
+ 48,6,2,9,26,89877485842370,89841281936085,1,1,48,256,0
844
+ 48,6,2,10,43,198026541084507,18969466514177,1,1,46,256,0
845
+ 48,6,2,11,41,133889586432522,193198110266947,1,1,49,256,0
846
+ 48,6,2,12,34,118081171769980,142799252557822,1,1,49,256,0
847
+ 48,6,2,13,27,75124940057945,179180065190879,1,1,49,256,0
848
+ 48,6,2,14,9,20568067251439,216141019591594,1,1,49,256,0
849
+ 48,6,2,15,33,126922106852292,79220432887801,1,1,48,256,0
850
+ 48,6,2,16,22,180512208674170,166151648226152,1,1,49,256,0
851
+ 48,6,2,17,60,230668702073322,74213381773022,1,1,48,256,0
852
+ 48,6,2,18,52,258630882206492,21382108896192,1,1,46,256,0
853
+ 48,6,2,19,10,80049511477585,52970420199686,1,1,47,256,0
854
+ 48,6,2,20,56,841086248965,103239837738487,1,1,48,256,0
855
+ 48,6,2,21,61,100704358245369,261642412494032,1,1,49,256,0
856
+ 48,6,2,22,27,231853053461469,40609337969427,1,1,47,256,0
857
+ 48,6,2,23,47,173065832760228,249659363393832,1,1,49,256,0
858
+ 48,6,2,24,58,210555090986963,139262782535977,1,1,48,256,0
859
+ 48,6,2,25,24,12326087507631,162283084235067,1,1,49,256,0
860
+ 48,6,2,26,50,46879677730210,72700340323216,1,1,48,256,0
861
+ 48,6,2,27,11,176380120892144,134713360417664,1,1,48,256,0
862
+ 48,6,2,28,29,169013406379412,27138065017021,1,1,46,256,0
863
+ 48,6,2,29,36,23698407327483,185930627148553,1,1,49,256,0
864
+ 48,6,2,30,31,137218644128646,102821085331616,1,1,48,256,0
865
+ 48,6,2,31,26,227104934229024,226878607108706,1,1,49,256,0
866
+ 48,6,3,0,62,240223102180370,256042268931024,1,1,49,256,0
867
+ 48,6,3,1,35,193277828965972,175899097039931,1,1,49,256,0
868
+ 48,6,3,2,10,240902853377862,265959369330792,1,1,49,256,0
869
+ 48,6,3,3,63,117650353153895,169743318945403,1,1,49,256,0
870
+ 48,6,3,4,7,120842119847733,205713992863625,1,1,49,256,0
871
+ 48,6,3,5,58,214846276995738,139533292314119,1,1,48,256,0
872
+ 48,6,3,6,37,32424274551000,238128006647324,1,1,49,256,0
873
+ 48,6,3,7,52,259514955145797,145872445578092,1,1,49,256,0
874
+ 48,6,3,8,29,6873415015319,40444019091695,1,1,47,256,0
875
+ 48,6,3,9,11,94933609579253,205413917875846,1,1,49,256,0
876
+ 48,6,3,10,60,52644867456404,246022856198652,1,1,49,256,0
877
+ 48,6,3,11,45,51591698287434,219333895986033,1,1,49,256,0
878
+ 48,6,3,12,60,151854428225191,207676367506983,1,1,49,256,0
879
+ 48,6,3,13,7,4846980572407,112891824179862,1,1,48,256,0
880
+ 48,6,3,14,40,166571673799332,240131138645422,1,1,49,256,0
881
+ 48,6,3,15,36,218045656599027,117816347426309,1,1,48,256,0
882
+ 48,6,3,16,46,256738161075490,10783186455625,1,1,45,256,0
883
+ 48,6,3,17,31,108865156710081,204772419215538,1,1,49,256,0
884
+ 48,6,3,18,60,152425508514142,216264551695208,1,1,49,256,0
885
+ 48,6,3,19,38,192041748667628,220569909065514,1,1,49,256,0
886
+ 48,6,3,20,40,151636069803737,30909617889376,1,1,46,256,0
887
+ 48,6,3,21,37,14459359183359,172648572639786,1,1,49,256,0
888
+ 48,6,3,22,25,85106703045237,162933246196268,1,1,49,256,0
889
+ 48,6,3,23,45,162576197913619,6555978959279,1,1,44,256,0
890
+ 48,6,3,24,24,56704699216823,220452572421112,1,1,49,256,0
891
+ 48,6,3,25,14,170166536377467,248334788531956,1,1,49,256,0
892
+ 48,6,3,26,62,210325285635183,206046089181214,1,1,49,256,0
893
+ 48,6,3,27,7,267051638829120,152042023019720,1,1,49,256,0
894
+ 48,6,3,28,53,17260696687862,52342998365080,1,1,47,256,0
895
+ 48,6,3,29,19,252841680502780,182040702469463,1,1,49,256,0
896
+ 48,6,3,30,34,106928739708248,275067503931586,1,1,49,256,0
897
+ 48,6,3,31,37,180611489497183,231585800644566,1,1,49,256,0
898
+ 48,6,4,0,10,37804382877562,83018498833907,1,1,48,256,0
899
+ 48,6,4,1,14,279767587214493,1892390120138,1,1,42,256,0
900
+ 48,6,4,2,34,123478486798027,84745494517328,1,1,48,256,0
901
+ 48,6,4,3,62,31425793300979,263851631704748,1,1,49,256,0
902
+ 48,6,4,4,44,60248352317072,79465374060740,1,1,48,256,0
903
+ 48,6,4,5,51,41834180895538,119216671802611,1,1,48,256,0
904
+ 48,6,4,6,19,4717910241987,47958153711622,1,1,47,256,0
905
+ 48,6,4,7,30,28847000179519,41827160648850,1,1,47,256,0
906
+ 48,6,4,8,61,279234426698114,176863660206301,1,1,49,256,0
907
+ 48,6,4,9,38,110197796605219,166352738202844,1,1,49,256,0
908
+ 48,6,4,10,14,105430861951660,21710081002861,1,1,46,256,0
909
+ 48,6,4,11,42,102828686558957,81736878333781,1,1,48,256,0
910
+ 48,6,4,12,10,235391585030558,8160921256787,1,1,44,256,0
911
+ 48,6,4,13,51,122316153451794,236959898936277,1,1,49,256,0
912
+ 48,6,4,14,11,240004330736270,143959205918858,1,1,49,256,0
913
+ 48,6,4,15,63,20285318174991,201540123537189,1,1,49,256,0
914
+ 48,6,4,16,55,127052967742547,102347141722508,1,1,48,256,0
915
+ 48,6,4,17,25,193221354061722,178506396109568,1,1,49,256,0
916
+ 48,6,4,18,33,66341243583604,254432826782783,1,1,49,256,0
917
+ 48,6,4,19,23,172251326216905,266731859360780,1,1,49,256,0
918
+ 48,6,4,20,57,211164484094154,249433565177393,1,1,49,256,0
919
+ 48,6,4,21,30,218279814383097,41348406382847,1,1,47,256,0
920
+ 48,6,4,22,18,34004021305864,268182186062454,1,1,49,256,0
921
+ 48,6,4,23,15,87637771643851,190576332971954,1,1,49,256,0
922
+ 48,6,4,24,26,232918837633649,142705827501403,1,1,49,256,0
923
+ 48,6,4,25,20,168855911207584,158180632923952,1,1,49,256,0
924
+ 48,6,4,26,23,196906212623638,265668710207440,1,1,49,256,0
925
+ 48,6,4,27,28,145240292179078,59121811199101,1,1,47,256,0
926
+ 48,6,4,28,23,235910334822780,88246108349413,1,1,48,256,0
927
+ 48,6,4,29,13,245775855428834,259912339264120,1,1,49,256,0
928
+ 48,6,4,30,6,103949901276088,268880231373733,1,1,49,256,0
929
+ 48,6,4,31,17,237735100817617,118280224045948,1,1,48,256,0
930
+ 48,6,5,0,5,138566011060605,138479625757655,1,1,48,256,0
931
+ 48,6,5,1,12,141134265089226,201438651473981,1,1,49,256,0
932
+ 48,6,5,2,48,41014635111545,98963713105632,1,1,48,256,0
933
+ 48,6,5,3,30,79362871734714,62407218506935,1,1,47,256,0
934
+ 48,6,5,4,56,4891678367601,157746228454528,1,1,49,256,0
935
+ 48,6,5,5,6,275364585248878,116194631810187,1,1,48,256,0
936
+ 48,6,5,6,35,231338869097196,22985302084694,1,1,46,256,0
937
+ 48,6,5,7,63,52122678815378,6962172916665,1,1,44,256,0
938
+ 48,6,5,8,58,104187407404556,166579131711219,1,1,49,256,0
939
+ 48,6,5,9,48,103842266585655,85015779716461,1,1,48,256,0
940
+ 48,6,5,10,35,107747015094532,3901647197620,1,1,43,256,0
941
+ 48,6,5,11,50,50696829412450,1892999619550,1,1,42,256,0
942
+ 48,6,5,12,10,246821540169541,225306628144401,1,1,49,256,0
943
+ 48,6,5,13,14,31474182252327,113554079680195,1,1,48,256,0
944
+ 48,6,5,14,30,21962852130641,7487104192207,1,1,44,256,0
945
+ 48,6,5,15,20,48780256228409,47249828188560,1,1,47,256,0
946
+ 48,6,5,16,37,35939240075649,254270549653872,1,1,49,256,0
947
+ 48,6,5,17,18,162677241879027,260019297234523,1,1,49,256,0
948
+ 48,6,5,18,13,268435268322378,123214133542099,1,1,48,256,0
949
+ 48,6,5,19,25,212127349631632,125209227901026,1,1,48,256,0
950
+ 48,6,5,20,48,55232118336523,234091910925306,1,1,49,256,0
951
+ 48,6,5,21,18,35263954439653,50787678587202,1,1,47,256,0
952
+ 48,6,5,22,19,58861358938683,22031592679032,1,1,46,256,0
953
+ 48,6,5,23,42,112644899171462,196920839384550,1,1,49,256,0
954
+ 48,6,5,24,40,60260445590748,123317103339666,1,1,48,256,0
955
+ 48,6,5,25,27,47349491821396,280583416363174,1,1,49,256,0
956
+ 48,6,5,26,24,151282879545611,160109707334233,1,1,49,256,0
957
+ 48,6,5,27,3,122537149912933,8503671550682,1,1,44,256,0
958
+ 48,6,5,28,19,231946865309096,21393809264360,1,1,46,256,0
959
+ 48,6,5,29,10,236785392926141,19826791080892,1,1,46,256,0
960
+ 48,6,5,30,30,108143353278599,163306576102633,1,1,49,256,0
961
+ 48,6,5,31,61,237709188126315,15181602701503,1,1,45,256,0
962
+ 48,6,6,0,46,82086443645009,104030629820025,1,1,48,256,0
963
+ 48,6,6,1,18,276892337824697,129159933933157,1,1,48,256,0
964
+ 48,6,6,2,36,112146648391101,41801904025735,1,1,47,256,0
965
+ 48,6,6,3,42,93936850830965,249559574048183,1,1,49,256,0
966
+ 48,6,6,4,34,168812925147468,233992503811402,1,1,49,256,0
967
+ 48,6,6,5,56,239233597133029,37188730443759,1,1,47,256,0
968
+ 48,6,6,6,41,89005653793653,95202980856764,1,1,48,256,0
969
+ 48,6,6,7,10,189729359000849,201059293530593,1,1,49,256,0
970
+ 48,6,6,8,9,180624542710341,217764901173645,1,1,49,256,0
971
+ 48,6,6,9,29,74005374379717,75567421088112,1,1,48,256,0
972
+ 48,6,6,10,62,3265077771567,17051116349002,1,1,45,256,0
973
+ 48,6,6,11,35,105852652544128,74488334249874,1,1,48,256,0
974
+ 48,6,6,12,46,186340859732299,225118542889496,1,1,49,256,0
975
+ 48,6,6,13,42,54109710888735,246239218921951,1,1,49,256,0
976
+ 48,6,6,14,59,118892875075847,7663486818627,1,1,44,256,0
977
+ 48,6,6,15,59,12534494937383,42977270802239,1,1,47,256,0
978
+ 48,6,6,16,42,146191717661613,79690556882189,1,1,48,256,0
979
+ 48,6,6,17,56,54058559930401,189549232588932,1,1,49,256,0
980
+ 48,6,6,18,28,252149996704454,256747278314604,1,1,49,256,0
981
+ 48,6,6,19,15,219396101172843,167580827554999,1,1,49,256,0
982
+ 48,6,6,20,5,214532038494311,177292338805172,1,1,49,256,0
983
+ 48,6,6,21,58,277697248552097,264418715828122,1,1,49,256,0
984
+ 48,6,6,22,29,119377160087754,80473275587011,1,1,48,256,0
985
+ 48,6,6,23,19,96791498075614,250518476528232,1,1,49,256,0
986
+ 48,6,6,24,45,249525317837690,204559787797703,1,1,49,256,0
987
+ 48,6,6,25,22,84240019060411,209571207163780,1,1,49,256,0
988
+ 48,6,6,26,41,197289854533254,124472570879126,1,1,48,256,0
989
+ 48,6,6,27,21,89099221015132,230624732883139,1,1,49,256,0
990
+ 48,6,6,28,33,159072775792907,116415661776434,1,1,48,256,0
991
+ 48,6,6,29,18,84030809888696,181449048929891,1,1,49,256,0
992
+ 48,6,6,30,5,155552582856831,216287606861778,1,1,49,256,0
993
+ 48,6,6,31,7,128644309137443,49775068476431,1,1,47,256,0
994
+ 48,6,7,0,14,129919008001787,136948670905894,1,1,48,256,0
995
+ 48,6,7,1,26,2495326596469,155746260316410,1,1,49,256,0
996
+ 48,6,7,2,51,5215497875478,263675658601970,1,1,49,256,0
997
+ 48,6,7,3,44,270406146443009,8963266743659,1,1,45,256,0
998
+ 48,6,7,4,19,112204750390456,155575160662773,1,1,49,256,0
999
+ 48,6,7,5,33,190502206947736,256636017984132,1,1,49,256,0
1000
+ 48,6,7,6,55,61632674130527,122596990260575,1,1,48,256,0
1001
+ 48,6,7,7,35,8583419490257,180327726610050,1,1,49,256,0
1002
+ 48,6,7,8,50,66552204641660,84096568195232,1,1,48,256,0
1003
+ 48,6,7,9,3,100506681123941,207822228393016,1,1,49,256,0
1004
+ 48,6,7,10,44,30525092012071,176177266266914,1,1,49,256,0
1005
+ 48,6,7,11,42,140102427123167,244749485612797,1,1,49,256,0
1006
+ 48,6,7,12,50,127080619255059,148879876823478,1,1,49,256,0
1007
+ 48,6,7,13,48,203276069156657,211318236946686,1,1,49,256,0
1008
+ 48,6,7,14,40,189270341152714,200965818651700,1,1,49,256,0
1009
+ 48,6,7,15,37,19601087657823,97188683092821,1,1,48,256,0
1010
+ 48,6,7,16,60,233070747435714,52828378479670,1,1,47,256,0
1011
+ 48,6,7,17,18,138942411347882,87330760640496,1,1,48,256,0
1012
+ 48,6,7,18,6,268483440991800,72514251143130,1,1,48,256,0
1013
+ 48,6,7,19,53,63846214846075,64699515682351,1,1,47,256,0
1014
+ 48,6,7,20,6,75896774058573,26270933727020,1,1,46,256,0
1015
+ 48,6,7,21,63,265230476136980,149667044255162,1,1,49,256,0
1016
+ 48,6,7,22,62,184418502910079,229190286850468,1,1,49,256,0
1017
+ 48,6,7,23,10,50114670350026,250469892173227,1,1,49,256,0
1018
+ 48,6,7,24,45,274146435694039,228730344627283,1,1,49,256,0
1019
+ 48,6,7,25,31,177581949106245,234949880998693,1,1,49,256,0
1020
+ 48,6,7,26,19,236558025051499,279886187656044,1,1,49,256,0
1021
+ 48,6,7,27,57,36351525428025,186424982991189,1,1,49,256,0
1022
+ 48,6,7,28,6,124112665514326,176663059866646,1,1,49,256,0
1023
+ 48,6,7,29,48,120867029981123,187768098144940,1,1,49,256,0
1024
+ 48,6,7,30,20,13366015911557,3173102209515,1,1,43,256,0
1025
+ 48,6,7,31,60,86220306196371,125600470400409,1,1,48,256,0