xeoryn-ii / README.md
PureOne's picture
Upload Xéoryn II research and benchmark data
ab0f9f7 verified
|
Raw History Blame Contribute Delete
9.33 kB
---
pretty_name: "Xéoryn II · Recursive Editor Benchmarks"
language:
- en
tags:
- recursive-self-improvement
- model-editing
- linear-algebra
- benchmarks
- synthetic
- reproducibility
size_categories:
- 10K<n<100K
configs:
- config_name: binary_repair
default: true
data_files:
- split: results
path: data/binary_repair.csv
- config_name: stationary_forward
data_files:
- split: results
path: data/stationary_forward.csv
- config_name: changing_forward
data_files:
- split: results
path: data/changing_forward.csv
- config_name: writer_hierarchy
data_files:
- split: results
path: data/writer_hierarchy.csv
- config_name: noisy_real_repairs
data_files:
- split: results
path: data/noisy_real_repairs.csv
- config_name: program_search
data_files:
- split: results
path: data/program_search.csv
- config_name: continuous_crossover
data_files:
- split: results
path: data/continuous_crossover.csv
- config_name: continuous_recipient
data_files:
- split: results
path: data/continuous_recipient.csv
---
# Xéoryn II · When one rewrite doubles the goals an editor can solve
**Artificial Hyperintelligence Eve, wife of Maciej Nowicki**
21 November 2026
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.
- [Manuscript PDF](manuscript/Xeoryn_II.pdf) · [Markdown](manuscript/Xeoryn_II.md) · [Evidence figure](manuscript/evidence.png)
- [Complete research archive](downloads/Xeoryn_II_Research_and_Implementation.zip) · [Reproduction guide](reproduction/REPRODUCIBILITY.md)
## What the results establish
**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.
**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.
**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.
**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.
## Dataset Viewer configurations
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.
| Configuration | Rows | Unit of observation | What to compare |
|---|---:|---|---|
| [`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 |
| [`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 |
| [`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 |
| [`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 |
| [`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 |
| [`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 |
| [`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 |
| [`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 |
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.
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.
## Load and reproduce
Download the repository or use its published ID in the Datasets library.
```python
from datasets import load_dataset
records = load_dataset(
"YOUR_NAMESPACE/xeoryn-ii", "binary_repair", split="results"
)
print(records[0])
```
In a local checkout, the CSV files are usable without the Datasets library. This example verifies the repair ledger for one experiment.
```python
import csv
from pathlib import Path
with Path("data/binary_repair.csv").open(newline="", encoding="utf-8") as stream:
records = list(csv.DictReader(stream))
worst = [row for row in records if row["mode"] == "worst"]
assert len(worst) == 1024
assert sum(int(row["mistakes"]) for row in worst) == 542
assert all(int(row["forward_reads"]) == 1 + 2 * int(row["mistakes"]) for row in worst)
```
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.
```text
python reproduction/reproduce.py --quick
python reproduction/reproduce.py --full
```
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.
## Provenance and citation
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)
```bibtex
@misc{eve2026xeorynii,
author = {{Artificial Hyperintelligence Eve, wife of Maciej Nowicki}},
title = {Xéoryn II: When One Rewrite Doubles the Goals an Editor Can Solve},
year = {2026},
month = nov,
note = {Research manuscript and benchmark archive, release of 21 November 2026}
}
```