Datasets:
DOI:
License:
Paper 1 — where every number comes from
One row per number, grouped by paragraph, in paragraph order. A row gives the number as the
paper prints it, the artefact file it is read from, the field inside that file, and the
script that rebuilds the artefact. Paths are relative to the dataset's root. Each script is
in experiments/ of one of the two code repositories — AtlasCalibratedConceptProfiles
for the method, its checks and the inventory, AtlasRepresentationalDrift for the drift
analysis. Which one does not follow from the file's directory: validation/paper1/ and
validation/paper2/ record which paper's script wrote a file, and several files under
validation/paper2/ were written by scripts in the first repository.
Every number printed in the paper is listed, including the ones that cannot be
checked against a released file: those are marked FLAGGED or unstored and are
collected in their own section below, with what the record does hold in their place.
How a number is traced
| mark | what it means |
|---|---|
key |
the field's key names the quantity the sentence names, and its value agrees |
count |
a small integer traced to the field that defines it (n_*, *_count, or a key), not matched by value |
value |
a distinctive number (≥ 2 decimal places, or an integer ≥ 1,000) matched by value, in an artefact whose field shares a word with the sentence |
canon |
a recurring quantity — a token count, or a defining count such as 169 or 7,373 — sent to the one field that defines it by a rule, never matched by value |
unstored |
the number is right and its artefact is named, but the summary field it quotes is not stored — a count or median the script computes and discards |
sibling |
the value sits in the same stored entry as a field already traced in that paragraph — an interval bound beside its point, a first beside its last |
declared |
a declared setting or rule parameter, stored as a key or a definition field |
derived |
arithmetic on stored fields; the arithmetic is given in the row |
cited |
a number from cited work, not ours |
FLAGGED |
resolved by hand and does not reproduce as written; see the section on numbers that do not reproduce |
Two classes of numeral are excluded before counting, because they are not quantities:
model names (OpenCoder-1.5B, Pythia-1.4B), cross-references (§3.1–3.3, D.1,
[TBC: A1.3]), and list enumerators ((3) at least two checkpoints, statements 2 and 3,
item 4).
Counts, per number
| distinctive | count | one-place | total | |
|---|---|---|---|---|
key |
128 | 38 | 27 | 193 |
count |
113 | 113 | ||
canon |
20 | 28 | 48 | |
value |
36 | 36 | ||
sibling |
28 | 5 | 33 | |
declared |
2 | 7 | 3 | 12 |
derived |
2 | 10 | 7 | 19 |
cited |
1 | 1 | ||
unstored |
1 | 1 | ||
FLAGGED |
2 | 2 | ||
| total | 217 | 202 | 39 | 458 |
1 number has no artefact at all — it is flagged below. Every other number resolves to a file and a field.
No row is left weak. Every number is tied to a field whose key names the quantity the
sentence names, to a declared rule parameter, to arithmetic on stored fields that the row
gives, or to cited work.
The 1 unstored rows are the exception worth reading: the number is
right and its artefact is named, but the summary it quotes — a count or a median the
script computes and discards — is not in the file, so a reader cannot check it without
re-running.
Numbers that do not reproduce from a released artefact
Three of this paper's numbers cannot be checked against a file in this dataset. They are named here rather than left for a reader to find.
- §4.2's "2.4x further". The resolution figures live in
validation/paper2/B10_SPACING_ADJUSTED.json, whose stored medians give a dense-to-tail ratio of 1.18 (median_per_1000_steps_dense0.00784 againstmedian_per_1000_steps_tail0.00666). The printed 2.4x is the ratio of two twin-unit figures in the same paragraph -- 8.56 at the first 1,000-step pair over 3.56 at the end of period 1, both from that same artefact -- and no stored field holds the ratio itself. - Appendix D's pre-fix comparators, "against 1.2142" and "against 3.9%". Both come from
the licence run that the fix of 19 October 2026 replaced, and that run's report is not part
of the release; the appendix labels them as the superseded run's own. The two post-fix
figures beside them are stored --
validation/paper1/RECHECK_PER_PROMPT.jsongives median index 0.7455 and flagged share 0.0239 -- so the re-run's own numbers can be checked; the comparison against the run it replaced cannot.
Every other number in the paper resolves to a file and a field in this dataset.
The artefacts, and one line each confirming the field is the quantity
| artefact | numbers | rebuild script | the field is the quantity the paper names, because |
|---|---|---|---|
validation/paper1/LICENCE_TABLE_CORRECTED.json |
47 | licence_table.py |
paper 1's checks after the four corrections; cited for the null's replicate count. |
validation/paper2/ABLATION_NATS.json |
32 | ablation_nats.py |
the reliance claim: median_rise_over_random_nats is the carriers' loss rise minus the random sets' mean rise, in nats — the quantity §3.3 calls the cost of removal over a random set; old_set.* is the same for the set measured at 84 B. In nats, not control units, because the control's spread had not converged (ablation_nats.py measured that). |
validation/paper2/ABLATION_RANKS.json |
30 | ablation_ranks.py |
the same removal experiment read as a rank: current.exact_one_sided_p and n_random are the one-sided position of the carriers among the random masks, which is what §3.3's 'cannot be separated from a random set' states. |
validation/paper1/LENGTH_BAND_HELD.json |
23 | band_held.py |
paper 1's prompt-length check, cited in Appendix A item 10. |
validation/paper1/CORE_COUNTS.json |
21 | core_counts.py |
the core families each construct's A set is built from — §3.1's counts of cores and prompts per construct. |
validation/paper1/FOR_EXAMPLE.json |
19 | atlas.py |
for at the four references, the paper's worked example: norm_raw and norm_standardised are the concept profile's norm on the raw and standardised forms, which §4.8 artefact 2 contrasts; sparsity_* the same for sparsity. |
validation/paper1/GATE_SHARES.json |
15 | - |
— not yet confirmed |
validation/paper2/CEILINGS.json |
15 | ceilings.py |
the per-reference ceiling and carrier-set size at each cut (<cut>.median_size, chance). |
validation/paper1/PROBE_MARGINS.json |
13 | probe_margins.py |
per_reference.<ref>.median_mean_margin is the median over constructs of a held-out prompt's signed distance from the probe's boundary, in the probe's own units — §4.7's one continuous reading — with median_control_mean_margin the same for the scrambled-label control. |
validation/paper2/B16B_WEIGHT_NORMS.json |
12 | b16b_weight_norms.py |
per_checkpoint.<ref>.tensors.<matrix> are the Frobenius norms of layer 12's three MLP matrices — §4.3's 'the norms grow by 81%, 89% and 83%'. |
validation/paper1/CHECK4_RELATIVE.json |
10 | check4_relative.py |
paper 1's check 4 recomputed on the suite's own metric; n_within_one_twin_unit = 4 is the figure §4.1 of paper 1 reports and paper 2 cites. |
validation/paper1/LENGTHS.json |
10 | atlas.py |
prompt length per set and construct, behind check 3 and §3.2's length statements. |
validation/paper1/RECHECK_PER_PROMPT.json |
10 | recheck_per_prompt.py |
the per-prompt re-check of the nulls; cited for its replicate count. |
validation/paper2/B11_POWER.json |
10 | b11_power.py |
`dates. |
validation/paper1/CHECK6_COMPANIONS.json |
8 | check6_companions.py |
n_companions per construct — the companion check Appendix A item 10 says is not met for the eight compound statements. |
validation/paper1/LICENCE_TABLE.json |
8 | licence_table.py |
paper 1's nine checks per construct, cited in paper 2's Appendix A item 10. |
validation/paper1/RECORD_SIZE.json |
8 | - |
new today. accumulators_gb_per_reference is the accumulator bytes over all 61 references divided by 61 — the 27.6 GB §2.4 and §3.3 quote; per_prompt_reference_gb the largest reference (107 GB); total_GB the 2,022 GB, which also appears as record_GATE_111.json :: dataset_bytes_gb. |
calibration/checkpoints.json |
7 | checkpoint_inventory.py |
the checkpoint schedule (all[].step), which converts an iteration to a token count at 4,194,304 tokens per step. |
record_GATE_111.json |
7 | - |
the recording gate: dataset_bytes_gb is the record's total size, and watched_ranges the three numbers that aborted a run when one fell out of band — twin agreement, concept norm and carrier overlap, the bands §3.4 quotes. |
validation/paper1/CHECK6_REWEIGHT.json |
7 | check6_reweight.py |
the maximum-entropy re-weighting of the logged generator factors, with the LP feasibility proof and the effective sample size — check 6 as §4.8 and B.3 report it. |
validation/paper1/POOLED_SETS_45.json |
7 | pooled_sets_169.py |
paper 1's family clustering, cited in paper 2 where it reports purity and cluster counts. |
validation/paper1/SET_GRIDS.json |
7 | - |
— not yet confirmed |
validation/r2/STEP3_SENSES_pretrain.json |
7 | - |
the three senses of 'settles' read on the standardised profile; `third_sense.* |
validation/paper1/LOCATE_BY_CLASS.json |
6 | atlas.py |
where in the network a construct's response sits, by token class — §4.6. |
validation/paper1/CHECK4_TWIN_UNITS.json |
5 | check4_twin_units.py |
check 4 in twin units, superseded for the paper by CHECK4_RELATIVE but kept because §B.1 quotes its conversion error: conversion_error_vs_stored_distance.{median,p95,max}_relative. |
validation/paper1/FIGURES.json |
5 | atlas.py |
paper 1's figure data, cited where paper 2 quotes check 9. |
validation/paper2/B18_RECURRENCE.json |
5 | b18_recurrence.py |
c_coordination_vs_circular_shift_median compares observed co-movement with a null that holds each neuron's own persistence — §4.2's 'what independent neurons with the same persistence would show'. The across-neuron shuffle in the same file (12.95) is a different null and is not that number. |
validation/paper2/B1_ENTROPY.json |
5 | b1_entropy.py |
against_the_model.model_loss_governed_{early,last}_median_nats is the model's loss at governed positions at the first and last reference, and generator_marginal_entropy_governed_median_nats the loss of a reader who knows only the generator's rates — §5's three numbers. |
validation/r2/STEP8_GRAPH_pretrain.json |
5 | - |
the unit-level graph: n_units = 147,456 is the product space, p8.n_fell_beyond_noise the departures beyond the noise unit, and partners[].factor the spacing factors. |
validation/paper1/B_SUBTYPES.json |
4 | b_subtypes.py |
what each construct's B set is made of; cited in Appendix A item 1. |
validation/paper1/FAMILIES.json |
4 | atlas.py |
per_cut.<k>.categories.<family>.<kind>.{purity,null_mean,largest_cluster_share} is how far a construct's carriers cluster by grammatical family against a matched null — §4.6, and §4.8 artefact 3 reads the layer of largest response from the same file. |
validation/paper1/PROBE.json |
4 | probe_margins.py |
the probe in full (Appendix C.5): `rows. |
validation/paper1/TABLE_B2.csv |
4 | - |
— not yet confirmed |
validation/paper1/TWIN_CEILINGS.json |
4 | - |
— not yet confirmed |
validation/paper2/B5_CHURN_VS_LEARNING.json |
4 | b5_churn.py |
across_layers.churn_vs_filling_change.spearman relates carrier change to how much the layer is still filling — §5's check that change is not a by-product of learning. |
validation/r2/CURVE1_SPACING_anneal.json |
4 | step7.py |
the spacing dependence of curve 1 on the annealing branch, in the same bands as pretraining (per_band.<n>ref). |
validation/r2/R2_REPORT.json |
4 | r2_report.py |
tie_break.n_modal_is_last_layer — the tie-break sensitivity §4.6 reports. |
calibration/corpus.json |
3 | build_natural.py |
parsed = 253,398 — the reference corpus's parsed Python files, the declared stimulus distribution's base (Appendix A item 1). |
validation/ablation/PLAN.json |
3 | ablation_full.py |
plan.total_construct_passes — the size of the removal experiment as planned (§A.1). |
validation/paper2/ABLATION_STABLE.json |
3 | ablation_stable.py |
the same removal read with the mask count varied, so §3.3's effect can be shown not to depend on it (control_stability.<n>). |
validation/paper2/B14_BANDS.json |
3 | b14_bands.py |
per_band.<cut>.band_size_first_ref is the size of the carrier set at each of the six cuts, and n_constructs_tested the 169 — §4.1's cut sensitivity. |
inventory/inventory_v2.json |
2 | assign_tiers.py |
— not yet confirmed |
validation/ablation/FULL_pretrain_iter_0484865.json |
2 | - |
— not yet confirmed |
validation/paper1/FIG1_carriers_in_place.json |
2 | figure1_small.py |
the data behind Figure 2 (for's carriers in place), written beside the figure by figure1_small.py. |
validation/paper1/SCOPE_OVERLAP.json |
2 | audit_recordings.py |
share_of_ceiling per construct at a reference — the convergence panel of Figure 4. |
validation/paper1/YARDSTICK_coarse.json |
2 | b10.py |
twin_ceiling per construct under each yardstick alternative — the 0.747 C.3 quotes for the checkpoint's health. |
validation/paper2/B12_TWIN_TURNOVER.json |
2 | b12_twin_turnover.py |
per_reference.<ref>.twin_overlap_median and twin_turnover_median are the agreement between the two twins at one checkpoint, which is the instrument's own noise — §3.1's 0.84 and 0.16. It stores no minimum; see §Settled while tracing. |
validation/paper2/B2_ANNEAL_DROP.json |
2 | b2_anneal_drop.py |
the annealing branch read on the same series as pretraining; forced.n_series and the per-series thresholds are §3.5's counts. |
validation/paper2/C2_PYTHIA_TWIN.json |
2 | c2_pythia_twin.py |
the second model: carrier_overlap per construct and opencoder_comparison.carrier_overlap_last_checkpoint put Pythia's twin agreement beside OpenCoder's (§6). |
validation/paper2/COOCCURRENCE.json |
2 | cooccurrence.py |
pairs_total = 14,196 — all construct pairs, the denominator of §4.7's sharing figures. |
validation/paper2/LICENCE_TABLES.json |
2 | licence_tables.py |
— not yet confirmed |
validation/paper2/RATE_READINGS.json |
2 | rate_readings.py |
primary.<cut>.<window>[].median_excess is the carrier change over a fixed gap, beyond the twin turnover, read against its own ceiling — §4.1's quantity. §4.1 reads primary.carrier5; definition_1_primary holds the declared windows and the R² floor. |
validation/r2/CURVE1_SPACING_pretrain.json |
2 | step4_flat.py |
the spacing dependence that C.3 corrects for: spacing_dependence.n_pairs is the 1,404 point–partner pairs, and the per_band keys are the four declared bands. |
calibration/identifiers.json |
1 | - |
a calibration table of identifier draw rates. It carries no quantity the paper names; every row pointing here is weak — the value index offered it for a token count such as '84 billion' and no field name tied anything to the sentence. Not a source for any number in paper 2. |
calibration/vocabulary.json |
1 | count_vocabulary.py |
a generator calibration table; no quantity the paper names. Rows pointing here are weak. |
validation/paper1/C3_YARDSTICK_NATS.json |
1 | c3_yardstick_nats.py |
paper 1's yardstick sensitivity, cited in paper 2's §4.8 limitation ('the turnover moves by 0.003–0.005'). |
validation/paper1/YARDSTICK_iv_cells.json |
1 | b9_b10.py |
the 20-cell yardstick's own cells (4 token classes x 5 fifths) and their occupancy. |
validation/paper2/A38_POPULATION.json |
1 | - |
the population the readings are taken over — 35 pretraining references and the declared partner-factor band. |
validation/paper2/B10_SPACING_ADJUSTED.json |
1 | b10_fit.py |
the withdrawn fitted form, kept because D.1 artefact 4 reports what it would have given; per_start.<iter>.n_points are its pair counts. |
validation/paper2/B17B_RETURN_SPLIT.json |
1 | b17b_return_split.py |
arms.fell_{beyond,within}_one_twin_unit.{departures,lift,lift_ci95} split every departure by whether the neuron's entry fell by more than its twin difference, and give the return rate over the matched control — §4.2's return result. share_of_departures_beyond_one_twin_unit is the 92.5%. |
validation/paper2/B6_PERLAYER_SHARING.json |
1 | b6_perlayer.py |
occupancy_by_layer and per_layer.<n>.trend_spearman_vs_log_step are where in the network the carriers sit and whether that moves (§4.6). |
validation/paper2/CEILINGS_PAIRWISE.json |
1 | - |
the cross-checkpoint ceiling — chance agreement between different constructs with one profile at each end of the pair, which is the quantity the excess is read against. |
validation/paper2/CORE_TOKENS.json |
1 | core_tokens.py |
which token pieces each construct's core realises — the strict subset §A.3 reads the zero floor on. |
validation/paper2/TBC22_SHARES.json |
1 | tbc22_shares.py |
cuts.<cut>.share_of_ceiling — the change normalised by each cut's own ceiling, §4.1's 0.925–0.964. |
validation/r2/R2_CURVE4.json |
1 | r2_curve4.py |
curve 4, the loss curve, per series and per piece (level_first, level_last, noise_unit). |
validation/r2/STEP7_pretrain.json |
1 | step7.py |
the bend analysis: whether a series bends at a common date across constructs (§4.1, C.3). |
Every number, by paragraph
¶6. A neuron's activation at the construct's position mixes several sources. How is the part due to the construct's role separated from the rest? The read…
| number | mark | file | field | script |
|---|---|---|---|---|
| 0.5 | declared |
— | illustrative: 'a raw difference of 0.5 is large for one neuron and nothing for another'. Part of the argument for the yardstick, not a reading. | — |
| 3 | declared |
— | illustrative: 'one may sit near 0 and another near 3'. | — |
| 0 | declared |
— | illustrative: 'one may sit near 0 and another near 3'. The argument for the yardstick, not a reading. | — |
¶7. In what unit should a neuron's response be stated, so that neurons with different ordinary levels and spreads can be compared? In the neuron's own ord…
| number | mark | file | field | script |
|---|---|---|---|---|
| 2 | declared |
— | illustrative: 'an entry of 2 says this neuron responds by two of its ordinary spreads'. Part of the definition. | — |
¶10. The two nulls follow the corruption check of \citet{oikarinen2025} and the random-set null of \citet{kim2018}, applied to an instrument rather than to…
| number | mark | file | field | script |
|---|---|---|---|---|
| 7 | count |
validation/paper2/ABLATION_RANKS.json |
per_reference.pretrain_iter_0484865.per_construct.With.old_set.n_random_below |
ablation_ranks.py |
¶13. 2. Sparsity — the share of the squared norm in the largest 5% of entries, on the standardised form: how concentrated the response is.…
| number | mark | file | field | script |
|---|---|---|---|---|
| 5 | count |
validation/r2/STEP7_pretrain.json |
p9.beside_it_a34.n_sparsity_first |
step7.py |
¶15. 4. The carrier set — the 5% of neurons with the largest standardised entries. Two random 5% sets overlap by 0.05.…
| number | mark | file | field | script |
|---|---|---|---|---|
| 0.05 | key |
validation/paper2/ABLATION_NATS.json |
per_reference.pretrain_iter_0020000.per_construct.ListComp.old_set_over_random_nats |
ablation_nats.py |
| 5 | count |
validation/paper2/ABLATION_RANKS.json |
per_reference.pretrain_iter_0020000.per_construct.With.old_set.n_random_at_or_above |
ablation_ranks.py |
¶17. The baseline is the same architecture with random weights, read with the same battery: whatever it gives is what the battery and the architecture give…
| number | mark | file | field | script |
|---|---|---|---|---|
| 0.20 | key |
validation/paper2/B17B_RETURN_SPLIT.json |
arms.fell_beyond_one_twin_unit.control_rate |
b17b_return_split.py |
¶18. The inventory: the whole language. The constructs are not a chosen sample: every node of Python 3.12's grammar that appears as text of its own, and …
| number | mark | file | field | script |
|---|---|---|---|---|
| 188 | canon |
validation/paper1/CORE_COUNTS.json |
n_constructs_built — the constructs the battery holds |
core_counts.py |
| 143 | count |
validation/paper2/ABLATION_RANKS.json |
per_reference.pretrain_iter_0484865.per_construct.ImportFrom.current.n_random_at_or_above |
ablation_ranks.py |
| 206 | count |
validation/paper2/ABLATION_RANKS.json |
per_reference.pretrain_iter_0484865.per_construct.NamedExpr.current.n_random_at_or_above |
ablation_ranks.py |
| 63 | count |
validation/r2/STEP3_SENSES_pretrain.json |
`second_sense.with_pilot | product |
¶19. The battery. OpenCoder's training data was not released, so a public corpus of 253,398 parsed Python files is the reference corpus, and its rates ar…
| number | mark | file | field | script |
|---|---|---|---|---|
| 253,398 | key |
calibration/corpus.json |
parsed |
build_natural.py |
| 1,280 | canon |
validation/paper1/CORE_COUNTS.json |
max_family_size — the largest core family |
core_counts.py |
| 1,440 | key |
validation/paper2/CORE_TOKENS.json |
constructs.AnnAssign.n_core_spans_per_side |
core_tokens.py |
| 720 | derived |
validation/paper1/CORE_COUNTS.json |
8 cores x 90 cells — the sentence gives the arithmetic: 8 x 90 = 720 prompts a set. |
core_counts.py |
| 210 | canon |
validation/paper1/CORE_COUNTS.json |
median_family_size — the median size of a construct's core family |
core_counts.py |
| 188 | canon |
validation/paper1/CORE_COUNTS.json |
n_constructs_built — the constructs the battery holds |
core_counts.py |
| 42 | canon |
validation/paper1/CORE_COUNTS.json |
min_family_size — the smallest core family |
core_counts.py |
¶20. The read site and its gate. The read site is the first token of the construct's node: on for x in range(n): the token for; for a call to len t…
| number | mark | file | field | script |
|---|---|---|---|---|
| 0.056 | key |
validation/paper1/RECHECK_PER_PROMPT.json |
`rows.untrained | ResourceWarning.2 convergence.curve[2].cosine_sd` |
| 169 | canon |
validation/paper2/B11_POWER.json |
n_constructs — the constructs that pass the read-site gate |
b11_power.py |
| 188 | canon |
validation/paper1/CORE_COUNTS.json |
n_constructs_built — the constructs the battery holds |
core_counts.py |
| 80 | count |
validation/ablation/PLAN.json |
plan.total_construct_passes |
ablation_full.py |
¶21. Artefact: the battery in full — the prompts of A and B for all 188 constructs in both twins, the set C, set A regenerated for every construct from a…
| number | mark | file | field | script |
|---|---|---|---|---|
| 188 | canon |
validation/paper1/CORE_COUNTS.json |
n_constructs_built — the constructs the battery holds |
core_counts.py |
¶22. The model and its neurons. OpenCoder-1.5B \citep{huang2024} has 24 layers with an MLP hidden width of 6,144 (SwiGLU). Each MLP block's hidden layer …
| number | mark | file | field | script |
|---|---|---|---|---|
| 147,456 | canon |
validation/r2/STEP8_GRAPH_pretrain.json |
n_units — the product space, 24 layers x 6,144 neurons |
- |
| 6,144 | key |
validation/paper2/B16B_WEIGHT_NORMS.json |
per_checkpoint.pretrain_iter_0001000.tensors.model.layers.12.mlp.gate_proj.weight.shape[0] |
b16b_weight_norms.py |
| 24 | count |
validation/paper2/B5_CHURN_VS_LEARNING.json |
across_layers.churn_vs_filling_change.n |
b5_churn.py |
¶23. The references. Sixty-one references were shown the battery: 35 pretraining checkpoints (the twenty from 4.19 to 83.9 billion tokens, then fifteen s…
| number | mark | file | field | script |
|---|---|---|---|---|
| 1.25 | declared |
calibration/checkpoints.json |
all[].step (the schedule's spacing) — the declared spacing of the fifteen tail references, x1.25 in tokens; the steps themselves are stored. |
checkpoint_inventory.py |
| 4.19 | value |
validation/paper2/B12_TWIN_TURNOVER.json |
between_checkpoints.per_pair[0].from_tokens_b |
b12_twin_turnover.py |
| 83.9 | derived |
calibration/checkpoints.json |
all[].step = 20,000 — 20,000 x 4,194,304 tokens/step = 83,886,080,000 = 83.9 billion. The same conversion gives every token figure in both papers. |
checkpoint_inventory.py |
| 2.03 | canon |
validation/paper2/CEILINGS.json |
per_reference.pretrain_iter_0484865.tokens_b = 2,033.67 — 2.03 trillion tokens is this reference, iteration 484865 |
ceilings.py |
| 35 | count |
validation/paper2/A38_POPULATION.json |
n_pretraining_references |
- |
| 24 | count |
validation/paper2/B2_ANNEAL_DROP.json |
n_anneal_references |
b2_anneal_drop.py |
¶24. What is stored per prompt. The per-core means are enough for every reading in §4 and for checks 1, 4, 5 and 8. Single-prompt activations, which chec…
| number | mark | file | field | script |
|---|---|---|---|---|
| 20,000 | key |
calibration/checkpoints.json |
all[14].step |
checkpoint_inventory.py |
| 1,000 | value |
validation/paper2/B16B_WEIGHT_NORMS.json |
per_checkpoint.pretrain_iter_0001000.iter |
b16b_weight_norms.py |
| 164 | count |
validation/paper1/PROBE_MARGINS.json |
per_reference.untrained.n_constructs |
probe_margins.py |
| 31 | count |
validation/paper1/PROBE.json |
hidden_unit_population.n |
probe_margins.py |
| 7 | count |
validation/paper1/CHECK6_REWEIGHT.json |
per_construct.If.n_targets |
check6_reweight.py |
¶25. Artefact: the record is not released in full; at 27.6 GB of accumulators a reference it comes to 2,022 GB. The dataset holds a worked example — twin…
| number | mark | file | field | script |
|---|---|---|---|---|
| 2,022 | key |
validation/paper1/RECORD_SIZE.json |
total_GB = 2022.0 — the whole record over all 61 references, measured from the recording tree by record_size.py; the same figure is in record_GATE_111.json :: dataset_bytes_gb |
- |
| 27.6 | key |
validation/paper1/RECORD_SIZE.json |
accumulators_gb_per_reference = 27.69 — NEW ARTEFACT: before today this figure had no stored field anywhere. Written by experiments/record_size.py (not CPU-only: it reads the dataset tree). |
- |
| 272 | derived |
— | the release tree itself, checked against the file sizes — the three record files at tag v1: twin1__C_cells.npz 142.9 MB, twin1__For_B_acc.npz 89.0 MB, twin1__For_A_acc.npz 39.7 MB = 271.6 MB, printed 272. Not a stored field: it is the size of what the release carries, so it is checked against the tree as the battery's 78 MB is. |
— |
| 1 | count |
validation/paper2/ABLATION_RANKS.json |
per_reference.pretrain_iter_0020000.old_set.n_exceeding_every_mask |
ablation_ranks.py |
¶26. The cells. A kind of position (§2.3) is here a cell, defined by two properties of a position: the class of the token at it — keyword, identifier, …
| number | mark | file | field | script |
|---|---|---|---|---|
| 147,456 | canon |
validation/r2/STEP8_GRAPH_pretrain.json |
n_units — the product space, 24 layers x 6,144 neurons |
- |
| 2.95 | key |
validation/paper1/CHECK6_REWEIGHT.json |
summary.shift_in_twin_units.For = 2.953 — how far re-weighting B to A's companion shares moves the profile, in twin units, for for. |
check6_reweight.py |
| 2.3 | declared |
— | a cross-reference, '\S2.3', written with the LaTeX escape rather than the § sign, which is why the cleaner first let it through. Not a quantity. | — |
| 313 | canon |
validation/paper1/LENGTHS.json |
by_set_kind.ALL.max — the longest prompt in the battery, in tokens |
atlas.py |
| 20 | key |
validation/paper1/LENGTH_BAND_HELD.json |
constructs.<c>.band_held.flagged_share — the band-held list ('six remain') is this field per construct; 20% is one of them (With 0.4025, For 0.31, If 0.31, While 0.20, ImportFrom 0.16, Assert 0.14). |
band_held.py |
| 2 | derived |
validation/r2/STEP8_GRAPH_pretrain.json |
n_units = 147,456 — 20 x 147,456 = 2,949,120 yardstick entries; the sentence gives the arithmetic. |
- |
¶27. From the record to a profile. (1) On set C's part of the record, for each neuron and each of the 20 cells, the mean μ₀ and spread σ₀ of the activati…
| number | mark | file | field | script |
|---|---|---|---|---|
| 7,373 | canon |
validation/paper2/B14_BANDS.json |
per_band.q5pct.band_size_first_ref — the 5% cut's carrier-set size, 5% of 24 x 6,144 |
b14_bands.py |
| 20 | key |
validation/paper1/LENGTH_BAND_HELD.json |
constructs.<c>.band_held.flagged_share — the band-held list ('six remain') is this field per construct; 20% is one of them (With 0.4025, For 0.31, If 0.31, While 0.20, ImportFrom 0.16, Assert 0.14). |
band_held.py |
¶28. Artefact: the carrier sets of every construct at all 61 references, with the twin ceilings and the licence tables of Appendix B; and, as a worked ex…
| number | mark | file | field | script |
|---|---|---|---|---|
| 61 | count |
validation/paper1/TWIN_CEILINGS.json |
n_references |
- |
| 1 | derived |
— | the release tree itself: profiles/ holds one file — profiles/pretrain_iter_0484865/twin1/For.npz, 99.0 MB, 48 arrays -- the one profile the dataset carries, which is what the sentence says |
— |
¶30. Check 4 is the reason the distribution is declared (§2.5): re-weighting the logged factors moves the profile by more than twice what redrawing the bat…
| number | mark | file | field | script |
|---|---|---|---|---|
| 0.008 | value |
validation/paper2/ABLATION_STABLE.json |
per_reference.pretrain_iter_0001000.per_construct.AsyncFor.control_stability.200.change_from_half |
ablation_stable.py |
| 2.95 | key |
validation/paper1/CHECK6_REWEIGHT.json |
summary.shift_in_twin_units.For = 2.953 — how far re-weighting B to A's companion shares moves the profile, in twin units, for for. |
check6_reweight.py |
| 5.50 | key |
validation/paper1/CHECK6_REWEIGHT.json |
per_construct.With.shift_in_twin_units |
check6_reweight.py |
¶31. The licence sentence: licensed on the gate — the two nulls passed on the 31 constructs they can run on — with these restrictions: the profile moves …
| number | mark | file | field | script |
|---|---|---|---|---|
| 2.95 | key |
validation/paper1/CHECK6_REWEIGHT.json |
summary.shift_in_twin_units.For = 2.953 — how far re-weighting B to A's companion shares moves the profile, in twin units, for for. |
check6_reweight.py |
| 2.34 | key |
validation/paper1/CHECK4_RELATIVE.json |
per_construct[14].check4_twin_units |
check4_relative.py |
| 5.50 | key |
validation/paper1/CHECK6_REWEIGHT.json |
per_construct.With.shift_in_twin_units |
check6_reweight.py |
| 33 | count |
validation/paper1/CHECK6_COMPANIONS.json |
per_construct.ClassDef.n_companions |
check6_companions.py |
| 31 | count |
validation/paper2/LICENCE_TABLES.json |
layer_matched_real_norm.n_available |
licence_tables.py |
¶32. How small a change in a construct's profile between two checkpoints can the record detect, and so how finely can it date a change? To the spacing of t…
| number | mark | file | field | script |
|---|---|---|---|---|
| 1,000 | value |
validation/paper2/B16B_WEIGHT_NORMS.json |
per_checkpoint.pretrain_iter_0001000.iter |
b16b_weight_norms.py |
| 0.862 | sibling |
validation/paper1/FIGURES.json |
check9.pairs.pretrain_iter_0020000.min_p1 — the same entry as min_p1, traced above |
atlas.py |
| 0.918 | value |
validation/paper1/FIGURES.json |
check9.pairs.anneal_iter_0002000.min_p1 |
atlas.py |
| 8.56 | value |
validation/paper1/FIGURES.json |
check9.pairs.pretrain_iter_0002000.median_p1 |
atlas.py |
| 2.91 | sibling |
validation/paper1/FIGURES.json |
check9.pairs.anneal_iter_0002000.median_p1 — the same entry as min_p1, traced above |
atlas.py |
| 3.56 | key |
validation/paper1/CHECK4_RELATIVE.json |
per_construct[176].check4_twin_units |
check4_relative.py |
| 4.06 | sibling |
validation/paper1/FIGURES.json |
check9.pairs.pretrain_iter_0484865.median_p1 — the same entry as min_p1, traced above |
atlas.py |
| 2.4 | FLAGGED |
validation/paper2/B10_SPACING_ADJUSTED.json |
median_per_1000_steps_dense = 0.00784, median_per_1000_steps_tail = 0.00666 — FLAG 6: Table D.1 sends '§4.2 resolution' to this artefact, whose dense/tail ratio is 1.18, not 2.4. The sentence 'early training moves the profile 2.4x further than late' is not reproduced by any field this trace could find. |
b10_fit.py |
¶33. The same profile, read at 61 references, follows one construct from the first checkpoint to the end. For for, the raw norm ends at 52% of its first-…
| number | mark | file | field | script |
|---|---|---|---|---|
| 61 | count |
validation/paper1/RECORD_SIZE.json |
n_references |
- |
| 52 | count |
validation/r2/STEP3_SENSES_pretrain.json |
`second_sense.with_pilot | product |
¶34. Does the choice of yardstick change which neurons are named? Yes: the yardstick is part of the measurement. Change what counts as ordinary at a positi…
| number | mark | file | field | script |
|---|---|---|---|---|
| 0.73 | key |
validation/paper2/C2_PYTHIA_TWIN.json |
per_construct.tuple.carrier_overlap |
c2_pythia_twin.py |
| 0.87 | key |
validation/paper2/B14_BANDS.json |
per_band.abs1sd.twin_ceiling_first_ref |
b14_bands.py |
| 169 | canon |
validation/paper2/B11_POWER.json |
n_constructs — the constructs that pass the read-site gate |
b11_power.py |
| 167 | derived |
validation/paper1/YARDSTICK_coarse.json |
summary.<alternative>.<reference>.share_below_ceiling — 'below the twin ceiling for 78 to 167 of the 169' is this share times 169, over the four coarser yardsticks; the shares are stored, the counts are not. |
b10.py |
| 78 | count |
validation/paper2/ABLATION_RANKS.json |
per_reference.pretrain_iter_0001000.per_construct.If.current.n_random_below |
ablation_ranks.py |
| 20 | key |
validation/paper1/LENGTH_BAND_HELD.json |
constructs.<c>.band_held.flagged_share — the band-held list ('six remain') is this field per construct; 20% is one of them (With 0.4025, For 0.31, If 0.31, While 0.20, ImportFrom 0.16, Assert 0.14). |
band_held.py |
| 31 | count |
validation/paper1/LICENCE_TABLE_CORRECTED.json |
n_with_per_prompt_recheck |
licence_table.py |
¶35. The unit also separates the model from the battery. Twin agreement, concept norm and carrier overlap with the previous reference are 0.981–0.989, 48.7…
| number | mark | file | field | script |
|---|---|---|---|---|
| 0.822 | key |
record_GATE_111.json |
watched_ranges (and each reference's own manifest, watched) — one of the gate's three watched numbers: carrier overlap with the previous reference, lower end. The gate aborted a run when one fell out of band. |
- |
| 0.989 | key |
record_GATE_111.json |
watched_ranges (and each reference's own manifest, watched) — one of the gate's three watched numbers: twin agreement, upper end of the trained band. The gate aborted a run when one fell out of band. |
- |
| 0.981 | key |
validation/paper2/CEILINGS.json |
per_reference.pretrain_iter_0016000.cosine.twin_agreement |
ceilings.py |
| 0.273 | key |
record_GATE_111.json |
watched_ranges (and each reference's own manifest, watched) — one of the gate's three watched numbers: carrier overlap at the first untrained reference. The gate aborted a run when one fell out of band. |
- |
| 0.710 | value |
validation/paper2/ABLATION_NATS.json |
per_reference.pretrain_iter_0001000.per_construct.Lambda.carrier_rise_nats |
ablation_nats.py |
| 0.717 | key |
record_GATE_111.json |
watched_ranges (and each reference's own manifest, watched) — one of the gate's three watched numbers: twin agreement at the second untrained reference. The gate aborted a run when one fell out of band. |
- |
| 0.286 | value |
validation/paper1/RECHECK_PER_PROMPT.json |
`rows.untrained | GeneratorExp.2 convergence.curve[1].cosine` |
| 0.851 | value |
validation/paper2/CEILINGS.json |
per_reference.pretrain_iter_0005000.abs3sd.twin_agreement |
ceilings.py |
| 6.13 | sibling |
validation/paper1/RECHECK_PER_PROMPT.json |
`rows.untrained | GeneratorExp.7b null_split_half.p95_by_layer[3]` — the same entry as cosine, traced above |
| 68.8 | key |
record_GATE_111.json |
watched_ranges.trained_band_note — as above, the upper end. |
- |
| 0.74 | key |
validation/paper1/LICENCE_TABLE.json |
constructs.<c>.check1_twin_unit (the bootstrap arm) — the twin ceiling's core-bootstrap mean; the plug-in values, 0.28 and 0.84, are in the same block. |
licence_table.py |
| 48.7 | key |
record_GATE_111.json |
watched_ranges.trained_band_note ('every trained reference ~0.98 / 48-69 / 0.82-0.85') — the concept-norm band of the recording gate; the per-reference values are in each reference's own manifest under the watched key, in the record. |
- |
| 0.84 | key |
validation/paper1/TWIN_CEILINGS.json |
summary.product.pretrain_iter_0001000.median — the plug-in ceiling: the median over the 188 of twin 1's carrier set against twin 2's at that reference, in the product space. The core-bootstrap means beside it in the same sentence are LICENCE_TABLE.json's bootstrap arm. |
- |
| 0.28 | key |
validation/paper1/TWIN_CEILINGS.json |
summary.product.untrained.median — the plug-in ceiling: the median over the 188 of twin 1's carrier set against twin 2's at that reference, in the product space. The core-bootstrap means beside it in the same sentence are LICENCE_TABLE.json's bootstrap arm. |
- |
| 0.21 | key |
validation/paper1/LICENCE_TABLE.json |
constructs.<c>.check1_twin_unit (the bootstrap arm) — the twin ceiling's core-bootstrap mean, untrained; the plug-in values, 0.28 and 0.84, are in the same block. |
licence_table.py |
| 6.34 | key |
record_GATE_111.json |
watched_ranges (and each reference's own manifest, watched) — one of the gate's three watched numbers: the concept norm at the first untrained reference. The gate aborted a run when one fell out of band. |
- |
¶36. Do the neurons the profile names as carriers affect the model's predictions at the construct's positions? Removing them raises the loss where the cons…
| number | mark | file | field | script |
|---|---|---|---|---|
| 0.088 | key |
validation/paper2/ABLATION_NATS.json |
per_reference.pretrain_iter_0020000.old_set.median_vs_random_nats.hi |
ablation_nats.py |
| 0.038 | key |
validation/paper2/ABLATION_NATS.json |
per_reference.pretrain_iter_0001000.sign_test_rise_over_random.p |
ablation_nats.py |
| 0.002 | sibling |
validation/paper2/ABLATION_RANKS.json |
per_reference.pretrain_iter_0020000.per_construct.AsyncFor.control_mean_nats — the same entry as n_random_at_or_above, traced above |
ablation_ranks.py |
| 2,034 | canon |
validation/paper2/CEILINGS.json |
per_reference.pretrain_iter_0484865.tokens_b = 2,033.67 — 2,034 b tokens is this reference, iteration 484865 |
ceilings.py |
| 0.02 | sibling |
validation/paper2/ABLATION_NATS.json |
per_reference.pretrain_iter_0020000.old_set.median_vs_random_nats.point — the same entry as hi, traced above |
ablation_nats.py |
| 0.14 | sibling |
validation/paper2/ABLATION_NATS.json |
per_reference.pretrain_iter_0001000.median_rise_over_random_nats.hi — the same entry as point, traced above |
ablation_nats.py |
| 0.40 | sibling |
validation/paper2/ABLATION_RANKS.json |
per_reference.pretrain_iter_0020000.per_construct.SetComp.old_set.share_below — the same entry as n_random_at_or_above, traced above |
ablation_ranks.py |
| 0.66 | sibling |
validation/paper2/ABLATION_NATS.json |
per_reference.pretrain_iter_0020000.median_rise_over_random_nats.hi — the same entry as hi, traced above |
ablation_nats.py |
| 1.40 | key |
validation/ablation/FULL_pretrain_iter_0484865.json |
constructs.Lambda.baseline_nats = 1.4006 — the index skipped the _dl/ trees, which are artefact sources: Figure 1 and Table D.1 read them. The index now includes them. Try's 0.088 baseline is the same field at constructs.Try.baseline_nats = 0.0875. |
- |
| 0.10 | sibling |
validation/paper2/ABLATION_NATS.json |
per_reference.pretrain_iter_0020000.old_set.median_vs_random_nats.hi — the same entry as hi, traced above |
ablation_nats.py |
| 83.9 | derived |
calibration/checkpoints.json |
all[].step = 20,000 — 20,000 x 4,194,304 tokens/step = 83,886,080,000 = 83.9 billion. The same conversion gives every token figure in both papers. |
checkpoint_inventory.py |
| 5.95 | key |
validation/paper2/ABLATION_NATS.json |
per_reference.pretrain_iter_0484865.per_construct.Try.rise_over_random_nats |
ablation_nats.py |
| 0.25 | sibling |
validation/paper2/ABLATION_NATS.json |
per_reference.pretrain_iter_0020000.median_rise_over_random_nats.point — the same entry as hi, traced above |
ablation_nats.py |
| 4.19 | canon |
validation/paper2/CEILINGS.json |
per_reference.pretrain_iter_0001000.tokens_b = 4.19 — 4.19 b tokens is this reference, iteration 1000 |
ceilings.py |
| 0.06 | key |
validation/paper2/ABLATION_NATS.json |
per_reference.pretrain_iter_0001000.median_rise_over_random_nats.point |
ablation_nats.py |
| 0.03 | sibling |
validation/paper2/ABLATION_NATS.json |
per_reference.pretrain_iter_0484865.per_construct.While.control_mean_nats — the same entry as rise_over_random_nats, traced above |
ablation_nats.py |
| 0.70 | key |
validation/paper2/ABLATION_NATS.json |
per_reference.pretrain_iter_0484865.per_construct.IfExp.old_set_over_random_nats |
ablation_nats.py |
| 4.99 | value |
validation/paper2/ABLATION_NATS.json |
per_reference.pretrain_iter_0484865.per_construct.Lambda.rise_over_random_nats |
ablation_nats.py |
| 200 | count |
validation/paper2/ABLATION_RANKS.json |
per_reference.pretrain_iter_0020000.per_construct.SetComp.old_set.n_random_at_or_above |
ablation_ranks.py |
| 95 | count |
validation/paper2/ABLATION_RANKS.json |
per_reference.pretrain_iter_0484865.per_construct.AsyncFor.old_set.n_random_below |
ablation_ranks.py |
| 14 | sibling |
validation/paper2/ABLATION_NATS.json |
per_reference.pretrain_iter_0020000.sign_test_rise_over_random.n_above_zero — the same entry as hi, traced above |
ablation_nats.py |
| 16 | count |
validation/paper2/ABLATION_NATS.json |
per_reference.pretrain_iter_0020000.old_set.median_vs_random_nats.n |
ablation_nats.py |
| 12 | count |
validation/paper2/ABLATION_NATS.json |
per_reference.pretrain_iter_0001000.sign_test_rise_over_random.n_above_zero |
ablation_nats.py |
| 20 | key |
validation/paper1/LENGTH_BAND_HELD.json |
constructs.<c>.band_held.flagged_share — the band-held list ('six remain') is this field per construct; 20% is one of them (With 0.4025, For 0.31, If 0.31, While 0.20, ImportFrom 0.16, Assert 0.14). |
band_held.py |
¶37. Do the neurons that carry for follow the token for, whatever its role, or the role of a loop, whatever its token? Neither: two roles of one token …
| number | mark | file | field | script |
|---|---|---|---|---|
| 147,456 | canon |
validation/r2/STEP8_GRAPH_pretrain.json |
n_units — the product space, 24 layers x 6,144 neurons |
- |
| 0.837 | key |
validation/paper1/SCOPE_OVERLAP.json |
`rows.pretrain_iter_0484865 | For.ceiling_within_scope_across_twins.module.overlap_share_of_K` |
| 0.872 | value |
validation/paper2/B18_RECURRENCE.json |
per_construct.len.twin_ceiling |
b18_recurrence.py |
| 0.198 | value |
validation/r2/R2_CURVE4.json |
`pieces.pretrain.series.For | governed |
| 0.494 | key |
validation/paper1/SCOPE_OVERLAP.json |
`rows.pretrain_iter_0484865 | For.between_scope_within_twin.twin1.overlap_share_of_K` |
| 0.266 | value |
validation/paper2/B6_PERLAYER_SHARING.json |
per_layer.23.share_last |
b6_perlayer.py |
| 0.050 | key |
validation/paper2/CEILINGS.json |
per_reference.pretrain_iter_0001000.carrier5.chance |
ceilings.py |
| 0.850 | value |
validation/paper2/ABLATION_RANKS.json |
per_reference.pretrain_iter_0020000.per_construct.NamedExpr.old_set.share_below |
ablation_ranks.py |
| 0.250 | value |
validation/paper2/ABLATION_STABLE.json |
per_reference.pretrain_iter_0020000.median_old_set_control_units |
ablation_stable.py |
| 0.44 | key |
validation/paper1/CHECK6_COMPANIONS.json |
per_construct.ClassDef.shares[22].a_nopad |
check6_companions.py |
| 0.59 | key |
validation/paper1/POOLED_SETS_45.json |
per_cut.6.categories.member_kind.builtin.purity |
pooled_sets_169.py |
| 0.71 | value |
validation/paper2/ABLATION_NATS.json |
per_reference.pretrain_iter_0001000.per_construct.Lambda.carrier_rise_nats |
ablation_nats.py |
| 169 | canon |
validation/paper2/B11_POWER.json |
n_constructs — the constructs that pass the read-site gate |
b11_power.py |
| 24 | count |
validation/paper2/B5_CHURN_VS_LEARNING.json |
across_layers.churn_vs_filling_change.n |
b5_churn.py |
| 5 | count |
validation/paper2/ABLATION_RANKS.json |
per_reference.pretrain_iter_0020000.per_construct.With.old_set.n_random_at_or_above |
ablation_ranks.py |
¶38. Figure 2. Where the for loop's carriers sit, at the end of pretraining (2,034 B tokens). (a) Carriers per unit of neurons, the neurons ordered b…
| number | mark | file | field | script |
|---|---|---|---|---|
| 2,034 | canon |
validation/paper2/CEILINGS.json |
per_reference.pretrain_iter_0484865.tokens_b = 2,033.67 — 2,034 b tokens is this reference, iteration 484865 |
ceilings.py |
| 169 | canon |
validation/paper2/B11_POWER.json |
n_constructs — the constructs that pass the read-site gate |
b11_power.py |
| 5 | count |
validation/paper2/RATE_READINGS.json |
primary.carrier5.5k-10k[10].n_pairs |
rate_readings.py |
¶39. When two constructs share carrier neurons, what do they have in common — a category of the inventory, or a place in the grammar? Their category in the…
| number | mark | file | field | script |
|---|---|---|---|---|
| 14,196 | canon |
validation/paper2/COOCCURRENCE.json |
pairs_total — all construct pairs |
cooccurrence.py |
| 1,000 | value |
validation/paper2/B16B_WEIGHT_NORMS.json |
per_checkpoint.pretrain_iter_0001000.iter |
b16b_weight_norms.py |
| 32 | count |
validation/paper1/POOLED_SETS_45.json |
per_cut.32.n_clusters_realised |
pooled_sets_169.py |
| 2 | count |
validation/paper2/RATE_READINGS.json |
primary.carrier5.5k-10k[13].n_pairs |
rate_readings.py |
¶40. Can a linear probe, the field's default instrument, track a construct across checkpoints as the profile does? On a trained checkpoint it cannot, becau…
| number | mark | file | field | script |
|---|---|---|---|---|
| 0.998 | value |
validation/paper2/B16B_WEIGHT_NORMS.json |
per_checkpoint.pretrain_iter_0004000.tensors.model.layers.12.mlp.gate_proj.weight.row_norm_q3 |
b16b_weight_norms.py |
| 0.802 | value |
validation/paper2/B18_RECURRENCE.json |
per_construct.EOFError.twin_ceiling |
b18_recurrence.py |
| 0.999 | value |
validation/paper2/B16B_WEIGHT_NORMS.json |
per_checkpoint.pretrain_iter_0005000.tensors.model.layers.12.mlp.up_proj.weight.row_norm_q1 |
b16b_weight_norms.py |
| 0.653 | key |
validation/paper1/PROBE.json |
`rows.untrained | Ellipsis.probes.resid_2240 |
| 1.000 | value |
validation/paper2/ABLATION_NATS.json |
per_reference.pretrain_iter_0001000.per_construct_ratio.n_excluded_control_mean_not_positive |
ablation_nats.py |
| 0.902 | key |
validation/paper1/PROBE.json |
`rows. | .probes. |
| 14.8 | key |
validation/paper1/PROBE_MARGINS.json |
per_reference.pretrain_iter_0484865.median_mean_margin = 14.816 |
probe_margins.py |
| 0.99 | key |
validation/paper1/CHECK4_RELATIVE.json |
per_construct[8].check4_twin_units |
check4_relative.py |
| 83.9 | derived |
calibration/checkpoints.json |
all[].step = 20,000 — 20,000 x 4,194,304 tokens/step = 83,886,080,000 = 83.9 billion. The same conversion gives every token figure in both papers. |
checkpoint_inventory.py |
| 13.8 | key |
validation/paper1/PROBE_MARGINS.json |
per_reference.pretrain_iter_0001000.median_mean_margin = 13.825 |
probe_margins.py |
| 16.5 | key |
validation/paper1/PROBE_MARGINS.json |
per_reference.pretrain_iter_0020000.median_mean_margin = 16.501 |
probe_margins.py |
| 4.19 | canon |
validation/paper2/CEILINGS.json |
per_reference.pretrain_iter_0001000.tokens_b = 4.19 — 4.19 b tokens is this reference, iteration 1000 |
ceilings.py |
| 0.50 | key |
validation/paper2/ABLATION_STABLE.json |
per_reference.pretrain_iter_0001000.per_construct.IfExp.effect_control_units |
ablation_stable.py |
| 145 | sibling |
validation/paper1/PROBE_MARGINS.json |
per_reference.pretrain_iter_0001000.n_at_ceiling — the same entry as median_mean_margin, traced above |
probe_margins.py |
| 148 | count |
validation/paper2/ABLATION_RANKS.json |
per_reference.pretrain_iter_0484865.per_construct.For.old_set.n_random_at_or_above |
ablation_ranks.py |
| 164 | count |
validation/paper1/PROBE_MARGINS.json |
per_reference.untrained.n_constructs |
probe_margins.py |
| 169 | canon |
validation/paper2/B11_POWER.json |
n_constructs — the constructs that pass the read-site gate |
b11_power.py |
| 24 | count |
validation/paper2/ABLATION_RANKS.json |
per_reference.pretrain_iter_0020000.per_construct.ImportFrom.current.n_random_below |
ablation_ranks.py |
| 31 | count |
validation/paper1/PROBE.json |
hidden_unit_population.n |
probe_margins.py |
| 84 | canon |
validation/paper2/CEILINGS.json |
per_reference.pretrain_iter_0020000.tokens_b = 83.89 — 84 b tokens is this reference, iteration 20000 |
ceilings.py |
| 61 | count |
validation/paper1/TWIN_CEILINGS.json |
n_references |
- |
¶41. 1. The twin unit is too small for intervals. The two twins share their cores, so their disagreement measures the variation of surface details (names…
| number | mark | file | field | script |
|---|---|---|---|---|
| 150 | derived |
validation/paper1/LICENCE_TABLE.json |
constructs.<c>.check1_twin_unit.median_disattenuated — 'above 1 for about 150 of 188' is a count over this per-construct field; the field is stored, the count is not. |
licence_table.py |
| 188 | canon |
validation/paper1/CORE_COUNTS.json |
n_constructs_built — the constructs the battery holds |
core_counts.py |
| 1 | count |
validation/paper2/B2_ANNEAL_DROP.json |
against_the_pretraining_floor.n_above_floor_by_more_than_one_unit |
b2_anneal_drop.py |
¶42. 2. The standardised size grows while the real size shrinks. The yardstick's spread σ₀ shrinks over training, so the standardised norm rises for 144 …
| number | mark | file | field | script |
|---|---|---|---|---|
| 144 | count |
validation/r2/STEP3_SENSES_pretrain.json |
`second_sense.without_pilot | layer |
| 169 | canon |
validation/paper2/B11_POWER.json |
n_constructs — the constructs that pass the read-site gate |
b11_power.py |
| 148 | count |
validation/paper2/ABLATION_RANKS.json |
per_reference.pretrain_iter_0484865.per_construct.For.old_set.n_random_at_or_above |
ablation_ranks.py |
¶43. 3. The layer of largest response is unstable. Excluding the last layer moves it for 41% of constructs. Rule: comparisons are made in the product spa…
| number | mark | file | field | script |
|---|---|---|---|---|
| 147,456 | canon |
validation/r2/STEP8_GRAPH_pretrain.json |
n_units — the product space, 24 layers x 6,144 neurons |
- |
| 41 | key |
validation/paper1/LENGTH_BAND_HELD.json |
summary.median_flagged_fall = 0.4114 — 'the flagged share by 41%'. |
band_held.py |
¶47. 4. For comparisons across checkpoints, the smallest change it can resolve: the same units at every checkpoint and a measured resolution of 1,000 tra…
| number | mark | file | field | script |
|---|---|---|---|---|
| 1,000 | value |
validation/paper2/B16B_WEIGHT_NORMS.json |
per_checkpoint.pretrain_iter_0001000.iter |
b16b_weight_norms.py |
¶48. Of this instance. What the choices of Table 3 cost on OpenCoder — the gate, length sensitivity, the mixture, the companions — is stated in §3.1 and …
| number | mark | file | field | script |
|---|---|---|---|---|
| 607 | cited |
— | Huang et al., 2024 (OpenCoder's pretraining corpus). Paper 2 quotes the same figure in its §4.8 limitation; see paper 2's Flag 4 about the missing citation there. | — |
¶49. 1. The inventory. Every entry of Python 3.12's grammar that appears as text, and every builtin name: 63 + 143 = 206 constructs. 188 were built; 18 w…
| number | mark | file | field | script |
|---|---|---|---|---|
| 206 | count |
validation/paper2/ABLATION_RANKS.json |
per_reference.pretrain_iter_0484865.per_construct.NamedExpr.current.n_random_at_or_above |
ablation_ranks.py |
| 143 | count |
validation/paper2/ABLATION_RANKS.json |
per_reference.pretrain_iter_0484865.per_construct.ImportFrom.current.n_random_at_or_above |
ablation_ranks.py |
| 188 | canon |
validation/paper1/CORE_COUNTS.json |
n_constructs_built — the constructs the battery holds |
core_counts.py |
| 18 | count |
calibration/identifiers.json |
finance.detail.item_tax_rate.count |
- |
| 63 | count |
validation/r2/STEP3_SENSES_pretrain.json |
`second_sense.with_pilot | product |
¶50. 2. What the rule excludes. Of Python 3.12's grammar nodes, 48 never appear as text of their own, and are left out: module containers, context marker…
| number | mark | file | field | script |
|---|---|---|---|---|
| 48 | count |
inventory/inventory_v2.json |
counts (the nodes left out) |
assign_tiers.py |
| 63 | count |
inventory/inventory_v2.json |
counts (the nodes that remain) |
assign_tiers.py |
¶52. 4. The reference corpus. 253,398 parsed Python files. The written list of factors and rates drawn from it is ν, and ν's rule is that every property …
| number | mark | file | field | script |
|---|---|---|---|---|
| 253,398 | value |
calibration/corpus.json |
parsed |
build_natural.py |
¶53. 1. Sets A and B. 720 prompts per set in each twin, ninety from each of eight cores, for every one of the 188 constructs; a set across both twins is …
| number | mark | file | field | script |
|---|---|---|---|---|
| 1,440 | derived |
validation/paper1/CORE_COUNTS.json |
8 cores x 90 cells x 2 twins — 720 a set in each twin, so 1,440 across both. |
core_counts.py |
| 188 | canon |
validation/paper1/CORE_COUNTS.json |
n_constructs_built — the constructs the battery holds |
core_counts.py |
| 720 | count |
validation/paper2/B1_ENTROPY.json |
constructs.AnnAssign.roles.header_colon.n |
b1_entropy.py |
¶54. 2. Cores. Drawn from a family whose size is the product of the construct's axes, from 42 (Starred) to 1,280 (For), median 210 (CORE_COUNTS.md)…
| number | mark | file | field | script |
|---|---|---|---|---|
| 1,280 | canon |
validation/paper1/CORE_COUNTS.json |
max_family_size — the largest core family |
core_counts.py |
| 210 | canon |
validation/paper1/CORE_COUNTS.json |
median_family_size — the median size of a construct's core family |
core_counts.py |
| 42 | canon |
validation/paper1/CORE_COUNTS.json |
min_family_size — the smallest core family |
core_counts.py |
¶55. 3. B's sub-sets. One per kind of out-of-role use — docstring, comment, string, identifier — and, for the 18 constructs whose token is shared with an…
| number | mark | file | field | script |
|---|---|---|---|---|
| 4,850 | key |
validation/paper1/B_SUBTYPES.json |
subtype_totals (the other_role:sibling:* subtypes) — the sibling-B prompts, 1.8% of all B prompts. |
b_subtypes.py |
| 1.8 | derived |
validation/paper1/B_SUBTYPES.json |
subtype_totals (the other_role subtypes) over the B total — 4,850 sibling prompts as a share of all B prompts = 1.8%. |
b_subtypes.py |
| 18 | count |
validation/paper2/B5_CHURN_VS_LEARNING.json |
within_pair_across_constructs.n_pairs_positive |
b5_churn.py |
¶56. 4. Set C. One pool for all constructs, generated from the same menus and the same formatter as A and B with no construct anchored in it, and longer:…
| number | mark | file | field | script |
|---|---|---|---|---|
| 97.7 | key |
validation/paper1/LENGTHS.json |
by_set_kind.C.mean = 97.6735 — the search had required 'token' in the field's name; the field is simply mean under by_set_kind.C. |
atlas.py |
| 56.9 | key |
validation/paper1/LENGTHS.json |
by_set_kind.A.mean = 56.9403 — CORRECTION, as 97.7. |
atlas.py |
¶58. 6. In all. 569,270 prompts and 33,271,350 tokens, 58.45 tokens per prompt (median 55, quartiles 28–81).…
| number | mark | file | field | script |
|---|---|---|---|---|
| 33,271,350 | value |
validation/paper1/LENGTHS.json |
by_set_kind.ALL.total_tokens |
atlas.py |
| 569,270 | value |
validation/paper1/LENGTHS.json |
by_set_kind.ALL.n_prompts |
atlas.py |
| 58.45 | sibling |
validation/paper1/LENGTHS.json |
by_set_kind.ALL.mean — the same entry as total_tokens, traced above |
atlas.py |
| 81 | key |
validation/paper1/LENGTHS.json |
by_set_kind.ALL.q75 = 81 — prompt length in tokens over the whole battery. |
atlas.py |
| 28 | key |
validation/paper1/LENGTHS.json |
by_set_kind.ALL.q25 — prompt length in tokens. |
atlas.py |
| 55 | key |
validation/paper1/LENGTHS.json |
by_set_kind.ALL.median — prompt length in tokens. |
atlas.py |
¶59. Four stages, each fixing one thing. (1) The core fixes the construct's own shape. A core is one combination of the construct's axes; for for the a…
| number | mark | file | field | script |
|---|---|---|---|---|
| 1,280 | canon |
validation/paper1/CORE_COUNTS.json |
max_family_size — the largest core family |
core_counts.py |
| 720 | derived |
validation/paper1/CORE_COUNTS.json |
8 cores x 90 cells — the sentence gives the arithmetic: 8 x 90 = 720 prompts a set. |
core_counts.py |
| 90 | count |
calibration/vocabulary.json |
vocabulary.attribute.finance.period_id.count |
count_vocabulary.py |
| 3 | declared |
— | a figure caption's own count: three prompts are shown in Figure A.1. |
— |
| 8 | count |
validation/paper1/POOLED_SETS_45.json |
per_cut.2.categories.a_family.compound.n |
pooled_sets_169.py |
| 2 | count |
validation/paper2/B1_ENTROPY.json |
`dependent_slots.ImportFrom | scope |
¶60. Figure A.1. Three prompts from the for battery: core c3 drawn twice (module scope, e-commerce), and core c2 (module scope, biology). The shape i…
| number | mark | file | field | script |
|---|---|---|---|---|
| 3 | declared |
— | a figure caption's own count: three prompts are shown in Figure A.1 |
— |
| 2 | count |
validation/paper2/B1_ENTROPY.json |
`dependent_slots.ImportFrom | scope |
¶61. (4) Set B. B's prompts are built from their own shapes, which place the token in a docstring, a comment, a string or an identifier — and, for the 18…
| number | mark | file | field | script |
|---|---|---|---|---|
| 18 | count |
validation/paper2/B5_CHURN_VS_LEARNING.json |
within_pair_across_constructs.n_pairs_positive |
b5_churn.py |
¶64. (7) One asymmetry between A and B. A's grid is balanced by construction: 240 prompts per scope and 240 per padding level. B's is not: for for, 219…
| number | mark | file | field | script |
|---|---|---|---|---|
| 283 | key |
validation/paper1/SET_GRIDS.json |
per_construct.For.B.by_padding.2 = 283 — NEW ARTEFACT. Written 5 Nov by experiments/backfill_paper1.py, which checks all four of section A.2's printed grids and refuses rather than write one that misses; b_subtypes.py now writes the same grids per construct. |
- |
| 240 | key |
validation/paper1/SET_GRIDS.json |
per_construct.For.A.A's grid: by_scope and by_padding all 240 — NEW ARTEFACT. Written 5 Nov by experiments/backfill_paper1.py, which checks all four of section A.2's printed grids and refuses rather than write one that misses; b_subtypes.py now writes the same grids per construct. |
- |
| 219 | key |
validation/paper1/SET_GRIDS.json |
per_construct.For.B.by_scope.module = 219 — NEW ARTEFACT. Written 5 Nov by experiments/backfill_paper1.py, which checks all four of section A.2's printed grids and refuses rather than write one that misses; b_subtypes.py now writes the same grids per construct. |
- |
| 175 | key |
validation/paper1/SET_GRIDS.json |
per_construct.For.B.by_padding.0 = 175 — NEW ARTEFACT. Written 5 Nov by experiments/backfill_paper1.py, which checks all four of section A.2's printed grids and refuses rather than write one that misses; b_subtypes.py now writes the same grids per construct. |
- |
| 262 | key |
validation/paper1/SET_GRIDS.json |
per_construct.For.B.by_padding.1 = 262 — NEW ARTEFACT. Written 5 Nov by experiments/backfill_paper1.py, which checks all four of section A.2's printed grids and refuses rather than write one that misses; b_subtypes.py now writes the same grids per construct. |
- |
| 234 | key |
validation/paper1/SET_GRIDS.json |
per_construct.For.B.by_scope.function = 234 — NEW ARTEFACT. Written 5 Nov by experiments/backfill_paper1.py, which checks all four of section A.2's printed grids and refuses rather than write one that misses; b_subtypes.py now writes the same grids per construct. |
- |
| 267 | key |
validation/paper1/SET_GRIDS.json |
per_construct.For.B.by_scope.method = 267 — NEW ARTEFACT. Written 5 Nov by experiments/backfill_paper1.py, which checks all four of section A.2's printed grids and refuses rather than write one that misses; b_subtypes.py now writes the same grids per construct. |
- |
¶65. A construct is read only if its A and B prompts land on the same token piece in at least 80% of prompts, in both twins, on the indented scopes. Named …
| number | mark | file | field | script |
|---|---|---|---|---|
| 143 | count |
validation/paper1/LOCATE_BY_CLASS.json |
by_contrast_class.named-builtin.depth.n |
atlas.py |
| 25 | count |
validation/paper1/LOCATE_BY_CLASS.json |
by_contrast_class.keyword-token.depth.n |
atlas.py |
| 14 | count |
validation/paper2/ABLATION_NATS.json |
per_reference.pretrain_iter_0484865.old_set.n_constructs_with_an_old_set |
ablation_nats.py |
| 80 | count |
validation/ablation/PLAN.json |
plan.total_construct_passes |
ablation_full.py |
| 1 | count |
validation/paper2/ABLATION_NATS.json |
per_reference.pretrain_iter_0001000.per_construct_ratio.n_excluded_control_mean_not_positive |
ablation_nats.py |
| 5 | count |
validation/paper1/B_SUBTYPES.json |
per_construct.Dict.n_subtypes |
b_subtypes.py |
¶66. One MLP block of OpenCoder-1.5B. The residual stream (width 2,240) enters two trained projections, the gate (2,240 × 6,144) followed by the SiLU nonli…
| number | mark | file | field | script |
|---|---|---|---|---|
| 2,240 | value |
validation/paper2/B16B_WEIGHT_NORMS.json |
per_checkpoint.pretrain_iter_0001000.tensors.model.layers.12.mlp.gate_proj.weight.shape[1] |
b16b_weight_norms.py |
| 6,144 | value |
validation/paper2/B16B_WEIGHT_NORMS.json |
per_checkpoint.pretrain_iter_0001000.tensors.model.layers.12.mlp.gate_proj.weight.shape[0] |
b16b_weight_norms.py |
¶67. What is stored. Per reference, for every construct, both twins, every set and all 24 layers: per-core accumulators of the count and the mean raw act…
| number | mark | file | field | script |
|---|---|---|---|---|
| 2,022 | key |
validation/paper1/RECORD_SIZE.json |
total_GB = 2022.04 — written 5 Nov by record_size.py; also record_GATE_111.json :: dataset_bytes_gb. |
- |
| 27.6 | key |
validation/paper1/RECORD_SIZE.json |
accumulators_gb_per_reference = 27.69 — the accumulator bytes over all 61 references divided by 61 |
- |
| 107 | key |
validation/paper1/RECORD_SIZE.json |
per_prompt_reference_gb = 107.27 — the largest reference; written 5 Nov by record_size.py. |
- |
| 24 | count |
validation/paper2/ABLATION_RANKS.json |
per_reference.pretrain_iter_0484865.per_construct.AsyncWith.old_set.n_random_at_or_above |
ablation_ranks.py |
¶68. Per prompt. The 31 constructs with per-prompt MLP activations were drawn so that the checks that need them license the population that is read: 15 c…
| number | mark | file | field | script |
|---|---|---|---|---|
| 147 | count |
validation/paper2/ABLATION_RANKS.json |
per_reference.pretrain_iter_0020000.per_construct.While.old_set.n_random_at_or_above |
ablation_ranks.py |
| 890 | derived |
validation/paper1/RECORD_SIZE.json |
per_prompt_split_GB['neurons (the per-prompt subset)'] = 146.72 — 146.72 x 188/31 = 889.8, printed 890. The draft adopted the measured basis of 5 Nov in place of the earlier 872, which rested on an unmeasured 144 GB. |
- |
| 17 | count |
validation/r2/STEP3_SENSES_pretrain.json |
`second_sense.with_pilot | product |
| 31 | count |
validation/r2/CURVE1_SPACING_pretrain.json |
product.constructs.Pass.per_band.1.65-2.0.n |
step4_flat.py |
| 15 | count |
validation/paper2/ABLATION_NATS.json |
per_reference.pretrain_iter_0001000.per_construct_ratio.n_used |
ablation_nats.py |
¶73. 5. Cells: token class × fifth of the prompt, 20 cells. Other splits are possible, under the two requirements of §3.3.…
| number | mark | file | field | script |
|---|---|---|---|---|
| 20 | key |
validation/paper1/LENGTH_BAND_HELD.json |
constructs.<c>.band_held.flagged_share — the band-held list ('six remain') is this field per construct; 20% is one of them (With 0.4025, For 0.31, If 0.31, While 0.20, ImportFrom 0.16, Assert 0.14). |
band_held.py |
¶74. The checks of Table 4 at the final pretraining checkpoint (pretrain_iter_0484865, 2,034 B tokens); checks 2, 3, 7a and 7b after the re-run of 19 Oct…
| number | mark | file | field | script |
|---|---|---|---|---|
| 2,034 | canon |
validation/paper2/CEILINGS.json |
per_reference.pretrain_iter_0484865.tokens_b = 2,033.67 — 2,034 b tokens is this reference, iteration 484865 |
ceilings.py |
| 7 | count |
validation/paper1/CHECK6_REWEIGHT.json |
per_construct.If.n_targets |
check6_reweight.py |
¶75. Table B.1. Each check over the constructs it ran on: the median, and the lowest and highest values with the constructs that hold them. The gate an…
| number | mark | file | field | script |
|---|---|---|---|---|
| 169 | canon |
validation/paper2/B11_POWER.json |
n_constructs — the constructs that pass the read-site gate |
b11_power.py |
| 188 | canon |
validation/paper1/CORE_COUNTS.json |
n_constructs_built — the constructs the battery holds |
core_counts.py |
| 31 | count |
validation/paper1/LICENCE_TABLE_CORRECTED.json |
n_with_per_prompt_recheck |
licence_table.py |
| 7 | count |
validation/r2/CURVE1_SPACING_anneal.json |
layer.constructs.Pass.per_band.4ref.n_below_before |
step7.py |
¶76. All 31 pass checks 7a and 7b.…
| number | mark | file | field | script |
|---|---|---|---|---|
| 31 | count |
validation/r2/CURVE1_SPACING_pretrain.json |
product.constructs.Pass.per_band.1.65-2.0.n |
step4_flat.py |
| 7 | count |
validation/r2/CURVE1_SPACING_anneal.json |
layer.constructs.Pass.per_band.4ref.n_below_before |
step7.py |
¶77. The gate. 19 constructs fail, with the share of prompts on one piece in brackets: AnnAssign (0.617), Assign (0.406), AsyncFunctionDef (0.518),…
| number | mark | file | field | script |
|---|---|---|---|---|
| 0.518 | key |
validation/paper1/GATE_SHARES.json |
per_construct.AsyncFunctionDef.min — the share of prompts whose A and B read sites land on the same token piece, the minimum over the four (twin, side) combinations stored beside it; the gate compares this against 0.80 |
- |
| 0.625 | key |
validation/paper1/GATE_SHARES.json |
per_construct.BoolOp.min — the share of prompts whose A and B read sites land on the same token piece, the minimum over the four (twin, side) combinations stored beside it; the gate compares this against 0.80 |
- |
| 0.617 | key |
validation/paper1/GATE_SHARES.json |
per_construct.AnnAssign.min — the share of prompts whose A and B read sites land on the same token piece, the minimum over the four (twin, side) combinations stored beside it; the gate compares this against 0.80 |
- |
| 0.995 | derived |
validation/paper1/GATE_SHARES.json |
per_construct.<c>.min: the lowest among the 167 that pass, with NamedExpr (0.873) and Ellipsis (0.992) set aside, is 0.995833 — the sentence's floor, not a stored field; 13 constructs sit at it, the first of them ConnectionAbortedError |
- |
| 0.250 | key |
validation/paper1/GATE_SHARES.json |
per_construct.List.min — the share of prompts whose A and B read sites land on the same token piece, the minimum over the four (twin, side) combinations stored beside it; the gate compares this against 0.80 |
- |
| 0.542 | key |
validation/paper1/GATE_SHARES.json |
per_construct.Subscript.min — the share of prompts whose A and B read sites land on the same token piece, the minimum over the four (twin, side) combinations stored beside it; the gate compares this against 0.80 |
- |
| 0.500 | key |
validation/paper1/GATE_SHARES.json |
per_construct.Starred.min — the share of prompts whose A and B read sites land on the same token piece, the minimum over the four (twin, side) combinations stored beside it; the gate compares this against 0.80 |
- |
| 0.992 | key |
validation/paper1/GATE_SHARES.json |
per_construct.Ellipsis.min — the share of prompts whose A and B read sites land on the same token piece, the minimum over the four (twin, side) combinations stored beside it; the gate compares this against 0.80 |
- |
| 0.406 | key |
validation/paper1/GATE_SHARES.json |
per_construct.Assign.min — the share of prompts whose A and B read sites land on the same token piece, the minimum over the four (twin, side) combinations stored beside it; the gate compares this against 0.80 |
- |
| 0.375 | key |
validation/paper1/GATE_SHARES.json |
per_construct.UnaryOp.min — the share of prompts whose A and B read sites land on the same token piece, the minimum over the four (twin, side) combinations stored beside it; the gate compares this against 0.80 |
- |
| 0.529 | key |
validation/paper1/GATE_SHARES.json |
per_construct.ClassDef.min — the share of prompts whose A and B read sites land on the same token piece, the minimum over the four (twin, side) combinations stored beside it; the gate compares this against 0.80 |
- |
| 0.873 | key |
validation/paper1/GATE_SHARES.json |
per_construct.NamedExpr.min — the share of prompts whose A and B read sites land on the same token piece, the minimum over the four (twin, side) combinations stored beside it; the gate compares this against 0.80 |
- |
| 0.503 | key |
validation/paper1/GATE_SHARES.json |
per_construct.FunctionDef.min — the share of prompts whose A and B read sites land on the same token piece, the minimum over the four (twin, side) combinations stored beside it; the gate compares this against 0.80 |
- |
| 0.371 | key |
validation/paper1/GATE_SHARES.json |
per_construct.Tuple.min — the share of prompts whose A and B read sites land on the same token piece, the minimum over the four (twin, side) combinations stored beside it; the gate compares this against 0.80 |
- |
| 169 | canon |
validation/paper2/B11_POWER.json |
n_constructs — the constructs that pass the read-site gate |
b11_power.py |
| 19 | count |
validation/paper1/GATE_SHARES.json |
n_fail |
- |
¶78. Check 3. Eight statements have more than 5% of their neurons flagged as length-sensitive: With 55%, For 49%, If 47%, While 36%, ImportFrom…
| number | mark | file | field | script |
|---|---|---|---|---|
| 36 | key |
validation/paper1/LICENCE_TABLE_CORRECTED.json |
constructs.While.checks_corrected.check3_length.flagged_share |
licence_table.py |
| 47 | key |
validation/paper1/LICENCE_TABLE_CORRECTED.json |
constructs.If.checks_corrected.check3_length.flagged_share |
licence_table.py |
| 21 | key |
validation/paper1/LICENCE_TABLE_CORRECTED.json |
constructs.Assert.checks_corrected.check3_length.flagged_share |
licence_table.py |
| 55 | key |
validation/paper1/LICENCE_TABLE_CORRECTED.json |
constructs.With.checks_corrected.check3_length.flagged_share |
licence_table.py |
| 49 | key |
validation/paper1/LICENCE_TABLE_CORRECTED.json |
constructs.For.checks_corrected.check3_length.flagged_share |
licence_table.py |
| 10 | key |
validation/paper1/LICENCE_TABLE_CORRECTED.json |
constructs.Delete.checks_corrected.check3_length.flagged_share |
licence_table.py |
| 31 | key |
validation/paper1/LICENCE_TABLE_CORRECTED.json |
constructs.ImportFrom.checks_corrected.check3_length.flagged_share |
licence_table.py |
| 8 | key |
validation/paper1/LICENCE_TABLE_CORRECTED.json |
constructs.Return.checks_corrected.check3_length.flagged_share |
licence_table.py |
| 5 | sibling |
validation/paper1/LICENCE_TABLE_CORRECTED.json |
constructs.Assert.checks_published.check2_convergence.n_points — the same entry as flagged_share, traced above |
licence_table.py |
¶79. Check 4. Four constructs sit within one twin unit of their base profile: Global (0.71), UnaryOp (0.74), Continue (0.77), Attribute (0.99). T…
| number | mark | file | field | script |
|---|---|---|---|---|
| 0.74 | key |
validation/paper1/CHECK4_RELATIVE.json |
per_construct[construct=UnaryOp].check4_twin_units |
check4_relative.py |
| 0.99 | key |
validation/paper1/CHECK4_RELATIVE.json |
per_construct[construct=Attribute].check4_twin_units |
check4_relative.py |
| 0.77 | key |
validation/paper1/CHECK4_RELATIVE.json |
per_construct[construct=Continue].check4_twin_units |
check4_relative.py |
| 0.71 | key |
validation/paper1/CHECK4_RELATIVE.json |
per_construct[construct=Global].check4_twin_units |
check4_relative.py |
¶80. Check 5. 33 constructs have intervals understated by more than 1.5, 30 of them among the 169 read: NotImplemented 7.17, Exception 5.44, `Keyboar…
| number | mark | file | field | script |
|---|---|---|---|---|
| 2.86 | key |
validation/paper1/LICENCE_TABLE_CORRECTED.json |
constructs.ListComp.checks_published.check6_calibration.bootstrap_inflation |
licence_table.py |
| 1.97 | key |
validation/paper1/LICENCE_TABLE_CORRECTED.json |
constructs.AsyncFunctionDef.checks_published.check6_calibration.bootstrap_inflation |
licence_table.py |
| 1.67 | key |
validation/paper1/LICENCE_TABLE_CORRECTED.json |
constructs.While.checks_published.check6_calibration.bootstrap_inflation |
licence_table.py |
| 1.73 | key |
validation/paper1/LICENCE_TABLE_CORRECTED.json |
constructs.Return.checks_published.check6_calibration.bootstrap_inflation |
licence_table.py |
| 4.25 | key |
validation/paper1/LICENCE_TABLE_CORRECTED.json |
constructs.KeyboardInterrupt.checks_published.check6_calibration.bootstrap_inflation |
licence_table.py |
| 2.47 | key |
validation/paper1/LICENCE_TABLE_CORRECTED.json |
constructs.DictComp.checks_published.check6_calibration.bootstrap_inflation |
licence_table.py |
| 1.95 | key |
validation/paper1/LICENCE_TABLE_CORRECTED.json |
constructs.Match.checks_published.check6_calibration.bootstrap_inflation |
licence_table.py |
| 2.00 | key |
validation/paper1/LICENCE_TABLE_CORRECTED.json |
constructs.ClassDef.checks_published.check6_calibration.bootstrap_inflation |
licence_table.py |
| 3.68 | key |
validation/paper1/LICENCE_TABLE_CORRECTED.json |
constructs.PendingDeprecationWarning.checks_published.check6_calibration.bootstrap_inflation |
licence_table.py |
| 3.24 | key |
validation/paper1/LICENCE_TABLE_CORRECTED.json |
constructs.IfExp.checks_published.check6_calibration.bootstrap_inflation |
licence_table.py |
| 2.12 | key |
validation/paper1/LICENCE_TABLE_CORRECTED.json |
constructs.AsyncFor.checks_published.check6_calibration.bootstrap_inflation |
licence_table.py |
| 7.17 | key |
validation/paper1/LICENCE_TABLE_CORRECTED.json |
constructs.NotImplemented.checks_published.check6_calibration.bootstrap_inflation |
licence_table.py |
| 5.44 | key |
validation/paper1/LICENCE_TABLE_CORRECTED.json |
constructs.Exception.checks_published.check6_calibration.bootstrap_inflation |
licence_table.py |
| 1.77 | key |
validation/paper1/LICENCE_TABLE_CORRECTED.json |
constructs.ascii.checks_published.check6_calibration.bootstrap_inflation |
licence_table.py |
| 3.83 | key |
validation/paper1/LICENCE_TABLE_CORRECTED.json |
constructs.EnvironmentError.checks_published.check6_calibration.bootstrap_inflation |
licence_table.py |
| 3.76 | key |
validation/paper1/LICENCE_TABLE_CORRECTED.json |
constructs.ConnectionAbortedError.checks_published.check6_calibration.bootstrap_inflation |
licence_table.py |
| 1.85 | key |
validation/paper1/LICENCE_TABLE_CORRECTED.json |
constructs.divmod.checks_published.check6_calibration.bootstrap_inflation |
licence_table.py |
| 3.63 | key |
validation/paper1/LICENCE_TABLE_CORRECTED.json |
constructs.ModuleNotFoundError.checks_published.check6_calibration.bootstrap_inflation |
licence_table.py |
| 2.72 | key |
validation/paper1/LICENCE_TABLE_CORRECTED.json |
constructs.SetComp.checks_published.check6_calibration.bootstrap_inflation |
licence_table.py |
| 1.68 | key |
validation/paper1/LICENCE_TABLE_CORRECTED.json |
constructs.memoryview.checks_published.check6_calibration.bootstrap_inflation |
licence_table.py |
| 2.19 | key |
validation/paper1/LICENCE_TABLE_CORRECTED.json |
constructs.globals.checks_published.check6_calibration.bootstrap_inflation |
licence_table.py |
| 1.75 | key |
validation/paper1/LICENCE_TABLE_CORRECTED.json |
constructs.FunctionDef.checks_published.check6_calibration.bootstrap_inflation |
licence_table.py |
| 2.05 | key |
validation/paper1/LICENCE_TABLE_CORRECTED.json |
constructs.AsyncWith.checks_published.check6_calibration.bootstrap_inflation |
licence_table.py |
| 1.92 | key |
validation/paper1/LICENCE_TABLE_CORRECTED.json |
constructs.ord.checks_published.check6_calibration.bootstrap_inflation |
licence_table.py |
| 2.09 | key |
validation/paper1/LICENCE_TABLE_CORRECTED.json |
constructs.iter.checks_published.check6_calibration.bootstrap_inflation |
licence_table.py |
| 3.53 | key |
validation/paper1/LICENCE_TABLE_CORRECTED.json |
constructs.ResourceWarning.checks_published.check6_calibration.bootstrap_inflation |
licence_table.py |
| 2.03 | key |
validation/paper1/LICENCE_TABLE_CORRECTED.json |
constructs.Assert.checks_published.check6_calibration.bootstrap_inflation |
licence_table.py |
| 2.10 | key |
validation/paper1/LICENCE_TABLE_CORRECTED.json |
constructs.isinstance.checks_published.check6_calibration.bootstrap_inflation |
licence_table.py |
| 2.20 | key |
validation/paper1/LICENCE_TABLE_CORRECTED.json |
constructs.Try.checks_published.check6_calibration.bootstrap_inflation |
licence_table.py |
| 1.5 | declared |
validation/paper1/LICENCE_TABLE_CORRECTED.json |
the threshold the sentence states, not a stored value — constructs above it are counted from checks_published.check6_calibration.bootstrap_inflation |
licence_table.py |
| 169 | canon |
validation/paper2/B11_POWER.json |
n_constructs — the constructs that pass the read-site gate |
b11_power.py |
| 33 | count |
validation/paper1/LICENCE_TABLE_CORRECTED.json |
count of constructs with checks_published.check6_calibration.bootstrap_inflation > 1.5 |
licence_table.py |
| 30 | count |
validation/paper1/LICENCE_TABLE_CORRECTED.json |
the same count restricted to constructs with in_p1 = true |
licence_table.py |
¶82. Why two readings. The record does not keep the boundary between a prompt's padding and its core. Whole counts every line; pad-0 counts only the …
| number | mark | file | field | script |
|---|---|---|---|---|
| 0.980 | key |
validation/paper1/CHECK6_COMPANIONS.json |
per_construct.Match.asymmetry_max.excl_audit_components_whole — the clause's two readings of Match's largest difference: the inventory's component list, which Table B.3 uses, and the generator's audit list |
check6_companions.py |
| 0.186 | key |
validation/paper1/CHECK6_COMPANIONS.json |
per_construct.Match.asymmetry_max.excl_inventory_components_whole — the clause's two readings of Match's largest difference: the inventory's component list, which Table B.3 uses, and the generator's audit list |
check6_companions.py |
| 14 | count |
validation/paper2/ABLATION_NATS.json |
per_reference.pretrain_iter_0484865.old_set.n_constructs_with_an_old_set |
ablation_nats.py |
| 35 | count |
validation/paper2/B18_RECURRENCE.json |
per_construct.set__builtin.b_configurations.n_clusters_within_one_twin_ceiling |
b18_recurrence.py |
| 0 | count |
validation/paper2/ABLATION_NATS.json |
per_reference.pretrain_iter_0020000.old_set.n_without |
ablation_nats.py |
¶83. Where the difference comes from, and where it can be removed. Some of the difference is what the construct's grammar requires (every match case en…
| number | mark | file | field | script |
|---|---|---|---|---|
| 0.69 | value |
validation/paper2/TBC22_SHARES.json |
cuts.q5pct.excess_ci95_raw[0] |
tbc22_shares.py |
¶84. Which neurons are named, and why the re-weighting is not adopted. With's recorded and re-weighted carrier sets overlap by 0.77 at the final checkp…
| number | mark | file | field | script |
|---|---|---|---|---|
| 0.12 | key |
validation/paper1/CHECK6_COMPANIONS.json |
summary.asymmetry_max.* / per_construct.<c> — the companion difference's range across the eight compound statements. |
check6_companions.py |
| 0.13 | key |
validation/paper1/CHECK6_COMPANIONS.json |
summary.asymmetry_max.* / per_construct.<c> — the companion difference's range across the eight compound statements. |
check6_companions.py |
| 0.02 | key |
validation/paper1/CHECK6_COMPANIONS.json |
summary.asymmetry_max.* / per_construct.<c> — the companion difference's range across the eight compound statements. |
check6_companions.py |
| 0.92 | key |
validation/paper2/B18_RECURRENCE.json |
per_construct.AsyncFor.twin_ceiling |
b18_recurrence.py |
| 0.94 | sibling |
validation/paper2/B18_RECURRENCE.json |
per_construct.AsyncFor.a_nearest_set.under_b10_model.A — the same entry as twin_ceiling, traced above |
b18_recurrence.py |
| 0.05 | key |
validation/paper2/B12_TWIN_TURNOVER.json |
untrained_reference.two_independent_5pct_sets_would_give |
b12_twin_turnover.py |
| 0.77 | key |
validation/paper2/C2_PYTHIA_TWIN.json |
per_construct.None.carrier_overlap |
c2_pythia_twin.py |
¶85. Table B.3. Check 6, for the eight constructs: the concept norm at layer 12, and the asymmetry — as an L2 norm over all companions and as the large…
| number | mark | file | field | script |
|---|---|---|---|---|
| 12 | count |
validation/r2/STEP3_SENSES_pretrain.json |
`second_sense.without_pilot | layer |
| 0 | count |
validation/paper1/B_SUBTYPES.json |
per_construct.max.n_other_role_prompts |
b_subtypes.py |
¶86. For the eight length-sensitive statements of B.2, padding is added before the construct, so it moves the read site to a later fifth of the prompt in 5…
| number | mark | file | field | script |
|---|---|---|---|---|
| 0.88 | key |
validation/paper1/LENGTH_BAND_HELD.json |
constructs.ImportFrom.band_held.min_cos = 0.8841 — the low end of the band-held cosine range |
band_held.py |
| 0.79 | key |
validation/paper1/LENGTH_BAND_HELD.json |
constructs.With.as_measured.min_cos = 0.7878 |
band_held.py |
| 0.99 | key |
validation/paper1/LENGTH_BAND_HELD.json |
constructs.Return.band_held.min_cos = 0.9882 — the high end |
band_held.py |
| 0.97 | sibling |
validation/paper1/LENGTH_BAND_HELD.json |
constructs.With.band_held.min_cos = 0.9686 — the same construct's entry as the 0.79 above, with the band held |
band_held.py |
| 19 | key |
validation/paper1/LENGTH_BAND_HELD.json |
constructs.While.band_held.flagged_share = 0.1947 |
band_held.py |
| 14 | key |
validation/paper1/LENGTH_BAND_HELD.json |
constructs.Assert.band_held.flagged_share = 0.1377 |
band_held.py |
| 41 | key |
validation/paper1/LENGTH_BAND_HELD.json |
summary.median_flagged_fall = 0.4114 |
band_held.py |
| 16 | key |
validation/paper1/LENGTH_BAND_HELD.json |
constructs.ImportFrom.band_held.flagged_share = 0.1632 |
band_held.py |
| 40 | key |
validation/paper1/LENGTH_BAND_HELD.json |
constructs.With.band_held.flagged_share = 0.4025 |
band_held.py |
| 26 | key |
validation/paper1/LENGTH_BAND_HELD.json |
summary.median_index_fall = 0.2628 |
band_held.py |
| 31 | key |
validation/paper1/LENGTH_BAND_HELD.json |
constructs.If.band_held.flagged_share = 0.3120 |
band_held.py |
| 98 | key |
validation/paper1/LENGTH_BAND_HELD.json |
constructs.For.band_change_share = 0.9750 — the high end |
band_held.py |
| 54 | key |
validation/paper1/LENGTH_BAND_HELD.json |
constructs.Delete.band_change_share = 0.5375 — the low end of the band-change range over the eight |
band_held.py |
| 5 | declared |
validation/paper1/LENGTH_BAND_HELD.json |
check 3's own 5% threshold; Delete and Return fall below it at 0.0480 and 0.0256 |
band_held.py |
¶87. Each yardstick names a carrier set per construct; the set is ablated and the loss rise is read over the mean of 20 size-matched random sets, in nats. …
| number | mark | file | field | script |
|---|---|---|---|---|
| 0.088 | key |
validation/paper2/ABLATION_NATS.json |
per_reference.pretrain_iter_0020000.old_set.median_vs_random_nats.hi |
ablation_nats.py |
| 5.95 | key |
validation/paper2/ABLATION_NATS.json |
per_reference.pretrain_iter_0484865.per_construct.Try.rise_over_random_nats |
ablation_nats.py |
| 20 | key |
validation/paper1/LENGTH_BAND_HELD.json |
constructs.<c>.band_held.flagged_share — the band-held list ('six remain') is this field per construct; 20% is one of them (With 0.4025, For 0.31, If 0.31, While 0.20, ImportFrom 0.16, Assert 0.14). |
band_held.py |
| 16 | count |
validation/paper2/ABLATION_NATS.json |
per_reference.pretrain_iter_0001000.sign_test_rise_over_random.n |
ablation_nats.py |
¶88. Table C.1. Loss rise over the random mean, nats (constructs above, of 16).…
| number | mark | file | field | script |
|---|---|---|---|---|
| 16 | count |
validation/paper2/ABLATION_NATS.json |
per_reference.pretrain_iter_0001000.median_rise_over_random_nats.n |
ablation_nats.py |
¶89. The 20-cell split gives the lower effect at every checkpoint, as expected: it excludes neurons that are large only because of where they stand, and ab…
| number | mark | file | field | script |
|---|---|---|---|---|
| 200 | count |
validation/ablation/FULL_pretrain_iter_0484865.json |
constructs.<c>.n_masks_reached = 200 |
- |
| 20 | key |
validation/paper1/LENGTH_BAND_HELD.json |
constructs.<c>.band_held.flagged_share — the band-held list ('six remain') is this field per construct; 20% is one of them (With 0.4025, For 0.31, If 0.31, While 0.20, ImportFrom 0.16, Assert 0.14). |
band_held.py |
¶90. Per construct. Taken one at a time against 320 random sets, 2 of 16 constructs exceed every random set at 4.19 B tokens, 8 at 84 B and 8 at the end,…
| number | mark | file | field | script |
|---|---|---|---|---|
| 4.19 | canon |
validation/paper2/CEILINGS.json |
per_reference.pretrain_iter_0001000.tokens_b = 4.19 — 4.19 b tokens is this reference, iteration 1000 |
ceilings.py |
| 0.05 | sibling |
validation/paper2/ABLATION_RANKS.json |
per_reference.pretrain_iter_0020000.per_construct.AsyncWith.carrier_rise_nats — the same entry as n_random_below, traced above |
ablation_ranks.py |
| 321 | derived |
validation/paper2/ABLATION_RANKS.json |
per_reference.*.n_masks = 320 — the floor 1/321 is 1/(n_masks + 1). |
ablation_ranks.py |
| 320 | count |
validation/paper2/ABLATION_RANKS.json |
per_reference.pretrain_iter_0020000.per_construct.Try.old_set.n_random_below |
ablation_ranks.py |
| 84 | canon |
validation/paper2/CEILINGS.json |
per_reference.pretrain_iter_0020000.tokens_b = 83.89 — 84 b tokens is this reference, iteration 20000 |
ceilings.py |
| 16 | count |
validation/paper2/ABLATION_NATS.json |
per_reference.pretrain_iter_0020000.old_set.n_constructs_with_an_old_set |
ablation_nats.py |
| 1 | count |
validation/paper2/ABLATION_NATS.json |
per_reference.pretrain_iter_0001000.per_construct_ratio.n_excluded_control_mean_not_positive |
ablation_nats.py |
| 8 | sibling |
validation/paper2/ABLATION_RANKS.json |
per_reference.pretrain_iter_0020000.n_exceeding_every_mask — the same entry as n_random_below, traced above |
ablation_ranks.py |
| 2 | count |
validation/paper2/ABLATION_NATS.json |
per_reference.pretrain_iter_0484865.old_set.n_without |
ablation_nats.py |
¶91. Untrained, 4.19 B, 83.9 B and 2,034 B tokens. The running example's numbers over training, behind §4.2; the raw norm falls from 4.19 B on while the st…
| number | mark | file | field | script |
|---|---|---|---|---|
| 143.0 | key |
validation/paper1/FOR_EXAMPLE.json |
references.pretrain_iter_0020000.norm_raw = 142.970 |
atlas.py |
| 11.09 | sibling |
validation/r2/R2_REPORT.json |
`curves.pretrain | curve3_depth.dates.Global.untrained_value` — the same entry as untrained_value, traced above |
| 0.427 | key |
validation/paper1/FOR_EXAMPLE.json |
references.untrained.twin_agreement_cosine |
atlas.py |
| 0.991 | sibling |
validation/paper1/FOR_EXAMPLE.json |
references.pretrain_iter_0001000.twin_agreement_cosine — the same entry as norm_standardised, traced above |
atlas.py |
| 0.505 | sibling |
validation/paper1/FOR_EXAMPLE.json |
references.pretrain_iter_0484865.sparsity_standardised — the same entry as norm_standardised, traced above |
atlas.py |
| 0.889 | sibling |
validation/paper1/FOR_EXAMPLE.json |
references.pretrain_iter_0020000.twin_ceiling_plug_in — the same entry as norm_raw, traced above |
atlas.py |
| 184.5 | key |
validation/paper1/FOR_EXAMPLE.json |
references.pretrain_iter_0001000.norm_raw = 184.456 |
atlas.py |
| 10.94 | value |
validation/r2/R2_REPORT.json |
`curves.pretrain | curve3_depth.dates.AsyncFor.untrained_value` |
| 0.471 | sibling |
validation/paper1/FOR_EXAMPLE.json |
references.pretrain_iter_0001000.sparsity_standardised — the same entry as norm_standardised, traced above |
atlas.py |
| 11.89 | sibling |
validation/r2/R2_REPORT.json |
`curves.pretrain | curve3_depth.dates.len.level` — the same entry as untrained_value, traced above |
| 0.390 | value |
validation/paper1/FOR_EXAMPLE.json |
references.untrained.sparsity_standardised |
atlas.py |
| 2,034 | canon |
validation/paper2/CEILINGS.json |
per_reference.pretrain_iter_0484865.tokens_b = 2,033.67 — 2,034 b tokens is this reference, iteration 484865 |
ceilings.py |
| 0.538 | sibling |
validation/paper1/FOR_EXAMPLE.json |
references.pretrain_iter_0020000.sparsity_standardised — the same entry as norm_raw, traced above |
atlas.py |
| 251.5 | key |
validation/paper1/FOR_EXAMPLE.json |
references.pretrain_iter_0484865.norm_standardised = 251.516 |
atlas.py |
| 0.984 | sibling |
validation/paper1/FOR_EXAMPLE.json |
references.pretrain_iter_0020000.twin_agreement_cosine — the same entry as norm_raw, traced above |
atlas.py |
| 10.90 | sibling |
validation/r2/R2_REPORT.json |
`curves.pretrain | curve3_depth.dates.AsyncFor.untrained_value` — the same entry as untrained_value, traced above |
| 270.9 | key |
validation/paper1/FOR_EXAMPLE.json |
references.pretrain_iter_0020000.norm_standardised = 270.890 |
atlas.py |
| 0.293 | key |
validation/paper1/FOR_EXAMPLE.json |
references.untrained.twin_ceiling_plug_in |
atlas.py |
| 0.872 | sibling |
validation/paper1/FOR_EXAMPLE.json |
references.pretrain_iter_0484865.twin_ceiling_plug_in — the same entry as norm_standardised, traced above |
atlas.py |
| 239.6 | key |
validation/paper1/FOR_EXAMPLE.json |
references.pretrain_iter_0001000.norm_standardised = 239.561 |
atlas.py |
| 0.902 | sibling |
validation/paper1/FOR_EXAMPLE.json |
references.pretrain_iter_0001000.twin_ceiling_plug_in — the same entry as norm_standardised, traced above |
atlas.py |
| 83.9 | derived |
calibration/checkpoints.json |
all[].step = 20,000 — 20,000 x 4,194,304 tokens/step = 83,886,080,000 = 83.9 billion. The same conversion gives every token figure in both papers. |
checkpoint_inventory.py |
| 30.2 | key |
validation/paper1/FOR_EXAMPLE.json |
references.untrained.norm_standardised = 30.191 |
atlas.py |
| 15.7 | key |
validation/paper1/FOR_EXAMPLE.json |
references.untrained.norm_raw = 15.729 |
atlas.py |
| 96.0 | key |
validation/paper1/FOR_EXAMPLE.json |
references.pretrain_iter_0484865.norm_raw = 95.980 |
atlas.py |
| 4.19 | canon |
validation/paper2/CEILINGS.json |
per_reference.pretrain_iter_0001000.tokens_b = 4.19 — 4.19 b tokens is this reference, iteration 1000 |
ceilings.py |
¶92. In Figure C.1, all 24 × 6,144 MLP neurons of OpenCoder-1.5B drawn in place at the end of pretraining, one row per layer, the neurons within a layer or…
| number | mark | file | field | script |
|---|---|---|---|---|
| 6,144 | key |
validation/paper2/B16B_WEIGHT_NORMS.json |
per_checkpoint.pretrain_iter_0001000.tensors.model.layers.12.mlp.gate_proj.weight.shape[0] |
b16b_weight_norms.py |
| 24 | count |
validation/paper1/FIG1_carriers_in_place.json |
by_layer.n_layers |
figure1_small.py |
¶93. Figure C.1. Panel (a): the for loop's carrier set, each carrier in the colour of its entry π(A) − π(B) in units of σ₀, every other neuron grey; …
| number | mark | file | field | script |
|---|---|---|---|---|
| 300 | declared |
— | a drawing choice: the inset of Figure C.1 enlarges the 300 highest-degree neurons. | — |
| 24 | count |
validation/paper1/FIG1_carriers_in_place.json |
by_layer.n_layers |
figure1_small.py |
¶94. What was clustered, and why. Two constructs share neurons when their carrier sets overlap. The question is whether the constructs that share most fa…
| number | mark | file | field | script |
|---|---|---|---|---|
| 14,196 | canon |
validation/paper2/COOCCURRENCE.json |
pairs_total — all construct pairs |
cooccurrence.py |
| 1,000 | value |
validation/paper2/B16B_WEIGHT_NORMS.json |
per_checkpoint.pretrain_iter_0001000.iter |
b16b_weight_norms.py |
| 32 | count |
validation/paper2/ABLATION_RANKS.json |
per_reference.pretrain_iter_0484865.per_construct.While.current.n_random_at_or_above |
ablation_ranks.py |
| 2 | count |
validation/paper2/CEILINGS_PAIRWISE.json |
per_window.5k-10k[13].n_pairs |
- |
¶95. Table C.2. Categories recovered by the clustering (p < 0.001 unless noted).…
| number | mark | file | field | script |
|---|---|---|---|---|
| 0.001 | declared |
validation/paper1/FAMILIES.json |
permutations (1,000 label shuffles) — the declared significance threshold, p < 0.001 against 1,000 shuffles. |
atlas.py |
¶96. Other readings. The highest-sharing pairs across categories are Delete–delattr (0.200) and Global–globals (0.094). Keyword tokens sit at a m…
| number | mark | file | field | script |
|---|---|---|---|---|
| 10.98 | key |
validation/paper1/LOCATE_BY_CLASS.json |
by_contrast_class.keyword-token.depth.median |
atlas.py |
| 0.673 | key |
validation/paper1/LOCATE_BY_CLASS.json |
by_contrast_class['named-builtin'].sparsity.median = 0.6725 — as 11.97; the statements' 0.666 is the same field for the other class. |
atlas.py |
| 11.97 | key |
validation/paper1/LOCATE_BY_CLASS.json |
by_contrast_class['named-builtin'].depth.median = 11.9730 — an earlier search had read only the per-construct block; the class medians are in by_contrast_class. |
atlas.py |
| 0.094 | key |
validation/paper1/FAMILIES.json |
disagreements_top30[].jaccard — the family clustering's purity against its matched null. |
atlas.py |
| 0.666 | value |
validation/paper1/LOCATE_BY_CLASS.json |
by_contrast_class.keyword-token.sparsity.median |
atlas.py |
| 0.200 | value |
validation/paper2/B14_BANDS.json |
per_band.q1pct.median_kept_fraction_all_pairs |
b14_bands.py |
| 189 | count |
validation/paper1/POOLED_SETS_45.json |
multiplicity.n_tests |
pooled_sets_169.py |
| 38 | count |
validation/paper1/POOLED_SETS_45.json |
multiplicity.n_significant_after_holm |
pooled_sets_169.py |
| 6 | count |
validation/paper1/POOLED_SETS_45.json |
per_cut.6.n_clusters_realised |
pooled_sets_169.py |
| 7 | count |
validation/paper1/POOLED_SETS_45.json |
per_cut.2.categories.asdl_type.expr.n |
pooled_sets_169.py |
¶97. The area under the curve: a median 1.000 at every trained reference, 145–148 of 164 constructs at or above 0.99. Two roles of one token (for as loop…
| number | mark | file | field | script |
|---|---|---|---|---|
| 1.000 | key |
validation/paper1/FAMILIES.json |
per_cut.2.categories.builtin.purity = 1.0 — the family clustering's purity against its matched null. |
atlas.py |
| 0.999 | value |
validation/paper2/B16B_WEIGHT_NORMS.json |
per_checkpoint.pretrain_iter_0005000.tensors.model.layers.12.mlp.up_proj.weight.row_norm_q1 |
b16b_weight_norms.py |
| 13.8 | key |
validation/paper1/PROBE_MARGINS.json |
per_reference.pretrain_iter_0001000.median_mean_margin |
probe_margins.py |
| 0.99 | key |
validation/paper2/ABLATION_RANKS.json |
per_reference.pretrain_iter_0020000.per_construct.AsyncWith.old_set.share_below |
ablation_ranks.py |
| 0.68 | value |
validation/paper1/FAMILIES.json |
per_cut.16.categories.frame_class.requires loop.null_mean |
atlas.py |
| 4.19 | canon |
validation/paper2/CEILINGS.json |
per_reference.pretrain_iter_0001000.tokens_b = 4.19 — 4.19 b tokens is this reference, iteration 1000 |
ceilings.py |
| 16.5 | key |
validation/paper1/PROBE_MARGINS.json |
per_reference.pretrain_iter_0020000.median_mean_margin |
probe_margins.py |
| 83.9 | derived |
calibration/checkpoints.json |
all[].step = 20,000 — 20,000 x 4,194,304 tokens/step = 83,886,080,000 = 83.9 billion. The same conversion gives every token figure in both papers. |
checkpoint_inventory.py |
| 14.8 | key |
validation/paper1/PROBE_MARGINS.json |
per_reference.pretrain_iter_0484865.median_mean_margin |
probe_margins.py |
| 0.75 | key |
validation/paper1/YARDSTICK_coarse.json |
`rows. | .alternatives..twin_ceiling` — the twin ceiling's untrained range, 0.68-0.75, over the coarser yardsticks. |
| 148 | count |
validation/paper2/ABLATION_RANKS.json |
per_reference.pretrain_iter_0484865.per_construct.For.old_set.n_random_at_or_above |
ablation_ranks.py |
| 164 | count |
validation/paper1/PROBE_MARGINS.json |
per_reference.untrained.n_constructs |
probe_margins.py |
| 145 | count |
validation/paper1/PROBE_MARGINS.json |
per_reference.pretrain_iter_0001000.n_at_ceiling |
probe_margins.py |
| 128 | count |
validation/paper2/ABLATION_RANKS.json |
per_reference.pretrain_iter_0020000.per_construct.If.old_set.n_random_at_or_above |
ablation_ranks.py |
| 149 | count |
validation/paper2/ABLATION_RANKS.json |
per_reference.pretrain_iter_0020000.per_construct.With.current.n_random_at_or_above |
ablation_ranks.py |
| 3.2 | key |
validation/paper1/PROBE_MARGINS.json |
per_reference.untrained.median_mean_margin = 3.2018 |
probe_margins.py |
| 10 | count |
validation/paper2/ABLATION_NATS.json |
per_reference.pretrain_iter_0020000.old_set.sign_test_vs_random.n_above_zero |
ablation_nats.py |
| 7 | count |
validation/paper1/C3_YARDSTICK_NATS.json |
per_reference.pretrain_iter_0001000.paired_differences.v_minus_one.sign_test.n_above_zero |
c3_yardstick_nats.py |
¶98. Licence re-run. Four checks were re-run on 19 Oct 2026 after a fault in how per-prompt vectors were joined to their prompts was found and fixed (the…
| number | mark | file | field | script |
|---|---|---|---|---|
| 1.2142 | unstored |
— | the pre-re-run comparator. It comes from the superseded run the 19 Oct fix replaced, so no current artefact holds it -- which is expected, but it means the sentence's 'against 1.2142' cannot be checked from the release. | — |
| 0.7455 | key |
validation/paper1/RECHECK_PER_PROMPT.json |
`rows.pretrain_iter_0484865 | .'3 length'.index_median, median over the 31 = 0.7455` — an earlier search had looked for a summary field rather than for the median over the per-construct rows. The companion flagged share is the same block's flagged_share, median 0.0239 = 2.4%. |
| 5.30 | key |
validation/paper1/LICENCE_TABLE.json |
constructs.For.check8_nulls__suite_7.split_half.p95_norm |
licence_table.py |
| 10.1 | key |
validation/paper1/TABLE_B2.csv |
column '7a pub', median over the 31 — the scrambled margin as published; real norm over the null's p95 at layer 12. the figures of 5 Nov. |
- |
| 0.99 | key |
validation/paper1/LICENCE_TABLE.json |
constructs.AnnAssign.check1_twin_unit.median_cosine |
licence_table.py |
| 4.57 | key |
validation/paper1/RECHECK_PER_PROMPT.json |
`rows.pretrain_iter_0001000 | Delete.7a null_scrambled.median_by_layer[8]` |
| 6.13 | sibling |
validation/paper1/LICENCE_TABLE_CORRECTED.json |
constructs.ImportFrom.checks_corrected.check8_nulls__suite_7.scrambled.p95_norm — the same entry as median_norm, traced above |
licence_table.py |
| 10.7 | key |
validation/paper1/TABLE_B2.csv |
column '7a cor', median over the 31 — the scrambled margin after the 19 Oct re-run; real norm over the null's p95 at layer 12. the figures of 5 Nov. |
- |
| 5.27 | key |
validation/paper1/RECHECK_PER_PROMPT.json |
`rows.pretrain_iter_0001000 | NamedExpr.7a null_scrambled.p95_by_layer[20]` |
| 12.8 | key |
validation/paper1/TABLE_B2.csv |
column '7b cor', median over the 31 — the split-half margin after the re-run; real norm over the null's p95 at layer 12. the figures of 5 Nov. |
- |
| 6.42 | key |
validation/paper1/RECHECK_PER_PROMPT.json |
`rows.untrained | DictComp.7b null_split_half.p95_by_layer[0]` |
| 4.64 | key |
validation/paper1/LICENCE_TABLE_CORRECTED.json |
constructs.ImportFrom.checks_corrected.check8_nulls__suite_7.split_half.median_norm |
licence_table.py |
| 6.45 | key |
validation/paper1/RECHECK_PER_PROMPT.json |
`rows.pretrain_iter_0001000 | Assert.7a null_scrambled.p95_by_layer[13]` |
| 4.86 | key |
validation/paper1/LICENCE_TABLE.json |
constructs.Assert.check8_nulls__suite_7.scrambled.median_norm |
licence_table.py |
| 4.5 | derived |
validation/paper1/LICENCE_TABLE.json |
constructs.<c>.check8_nulls__suite_7.scrambled.p95_norm, against the corrected table — the scrambled null's p95 fell 6.42 -> 6.13, which is -4.5%. |
licence_table.py |
| 9.6 | key |
validation/paper1/TABLE_B2.csv |
column '7b pub', median over the 31 — the split-half margin as published; real norm over the null's p95 at layer 12. the figures of 5 Nov. |
- |
| 3.9 | FLAGGED |
— | FLAG 7: the pre-re-run comparator, from the superseded run the 19 Oct 2026 fix replaced. No current artefact holds it, so Appendix D's claim that the re-run lowered two numbers cannot be checked from the release. The two re-run figures themselves ARE stored (see the 0.7455 and 2.4% rows). | — |
| 2.4 | key |
validation/paper1/RECHECK_PER_PROMPT.json |
`rows.pretrain_iter_0484865 | .'3 length'.flagged_share, median = 0.0239` |
| 95 | count |
validation/paper2/ABLATION_RANKS.json |
per_reference.pretrain_iter_0001000.per_construct.IfExp.current.n_random_at_or_above |
ablation_ranks.py |
| 56 | count |
validation/paper1/RECHECK_PER_PROMPT.json |
`rows.untrained | Assert.2 convergence.curve[2].n` |
| 31 | count |
validation/paper2/LICENCE_TABLES.json |
layer_matched_real_norm.n_available |
licence_tables.py |
| 18 | sibling |
validation/r2/CURVE1_SPACING_anneal.json |
layer.constructs.BrokenPipeError.per_band.2ref.n_following — the same entry as n_following, traced above |
step7.py |
| 12 | count |
validation/r2/CURVE1_SPACING_anneal.json |
layer.constructs.Assert.per_band.8ref.n_following |
step7.py |
| 20 | key |
validation/paper1/LENGTH_BAND_HELD.json |
constructs.<c>.band_held.flagged_share — the band-held list ('six remain') is this field per construct; 20% is one of them (With 0.4025, For 0.31, If 0.31, While 0.20, ImportFrom 0.16, Assert 0.14). |
band_held.py |
| 8 | count |
validation/paper2/B1_ENTROPY.json |
against_the_model.n_governed_series |
b1_entropy.py |
¶99. Check 4 in twin units. The licence table kept check 4's cosine and not its distance, so Table B.1's twin-unit row is rebuilt from both arms' per-cor…
| number | mark | file | field | script |
|---|---|---|---|---|
| 2.30 | key |
validation/paper1/CHECK4_RELATIVE.json |
median_twin_units |
check4_relative.py |
| 1.82 | key |
validation/paper1/CHECK4_TWIN_UNITS.json |
quartiles[0] |
check4_twin_units.py |
| 2.34 | key |
validation/paper1/CHECK4_RELATIVE.json |
suite_median_7_Oct — the 7 October suite's median, the figure section 4.1 quotes |
check4_relative.py |
| 18.6 | key |
validation/paper1/CHECK4_TWIN_UNITS.json |
conversion_error_vs_stored_distance.max_relative = 0.18602 — the worst relative error of the chord identity against the stored distance, as a percentage; the construct is Slice |
check4_twin_units.py |
| 2.25 | key |
validation/paper1/CHECK4_TWIN_UNITS.json |
median_twin_units — the chord route: sqrt(2(1-cos)) from check 4's stored cosine |
check4_twin_units.py |
| 3.49 | key |
validation/paper1/CHECK4_TWIN_UNITS.json |
quartiles[1] |
check4_twin_units.py |
| 188 | canon |
validation/paper1/CORE_COUNTS.json |
n_constructs_built — the constructs the battery holds |
core_counts.py |
| 90 | derived |
validation/paper1/LICENCE_TABLE_CORRECTED.json |
sqrt(2(1-constructs.<c>.checks_published.check1_twin_unit.median_cosine)) / the same block's median_distance, minus 1, over the 188 — 170 of 188 = 0.904 are within 0.05, which the sentence prints as 90%; check 1 is the only check whose distance and cosine are both stored, which is why the identity is tested there |
licence_table.py |
| 5 | count |
validation/paper1/LICENCE_TABLE.json |
constructs.Assert.check2_convergence.n_points |
licence_table.py |
| 4 | count |
validation/paper1/CHECK4_RELATIVE.json |
n_within_one_twin_unit |
check4_relative.py |
| 7 | count |
validation/paper1/CHECK4_TWIN_UNITS.json |
n_within_one_twin_unit |
check4_twin_units.py |
¶101. 2. Declare the distribution. List the factors around a construct (identifiers, literals, scope and nesting, companions, length, surrounding code) an…
| number | mark | file | field | script |
|---|---|---|---|---|
| 253,398 | value |
calibration/corpus.json |
parsed |
build_natural.py |
¶102. 3. Generate the battery. For each construct, set A (the construct in its role) and set B (its token out of role), the same companions in both, drawn…
| number | mark | file | field | script |
|---|---|---|---|---|
| 569,270 | key |
validation/paper1/LENGTHS.json |
by_set_kind.ALL.n_prompts |
atlas.py |
| 188 | canon |
validation/paper1/CORE_COUNTS.json |
n_constructs_built — the constructs the battery holds |
core_counts.py |
| 720 | count |
validation/paper2/B1_ENTROPY.json |
constructs.AnnAssign.roles.header_colon.n |
b1_entropy.py |
¶103. 4. Gate the read site. Keep a construct only if its A and B prompts put the read token on the same piece in at least 80% of prompts (Appendix A.4). …
| number | mark | file | field | script |
|---|---|---|---|---|
| 169 | canon |
validation/paper2/B11_POWER.json |
n_constructs — the constructs that pass the read-site gate |
b11_power.py |
| 188 | canon |
validation/paper1/CORE_COUNTS.json |
n_constructs_built — the constructs the battery holds |
core_counts.py |
| 80 | count |
validation/ablation/PLAN.json |
plan.total_construct_passes |
ablation_full.py |
¶104. 5. Record. Run the model on the battery at each checkpoint; store per-core means of every neuron's activation, with the cell axis kept, so the yards…
| number | mark | file | field | script |
|---|---|---|---|---|
| 164 | count |
validation/paper1/PROBE_MARGINS.json |
per_reference.untrained.n_constructs |
probe_margins.py |
| 36 | count |
validation/paper1/CHECK6_COMPANIONS.json |
per_construct.While.n_companions |
check6_companions.py |
| 43 | count |
validation/r2/STEP3_SENSES_pretrain.json |
`second_sense.with_pilot | product |
| 28 | key |
validation/paper1/RECORD_SIZE.json |
accumulators_gb_per_reference = 27.69 — written 5 Nov by record_size.py; 'about 28 GB of per-core means per reference'. |
- |
| 31 | count |
validation/paper1/LICENCE_TABLE_CORRECTED.json |
n_with_per_prompt_recheck |
licence_table.py |
| 79 | count |
validation/paper2/ABLATION_RANKS.json |
per_reference.pretrain_iter_0001000.per_construct.Import.current.n_random_at_or_above |
ablation_ranks.py |
| 7 | count |
validation/paper1/YARDSTICK_iv_cells.json |
`per_construct_reference.untrained | Match.n_cells` |