Datasets:

---

piotrwilam

/

AtlasOpenCoder

Like 0

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.

| 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`).

| | 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.

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_dense` 0.00784 against

`median_per_1000_steps_tail` 0.00666). 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.json` gives 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.

| 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.<shape>|<date>.detection_rate` is the share of 338 injected draws in which the corrected rule found the settling it was given — §3.4's power. The injection dates are named in the keys, so '84 billion to 1.5 trillion' is provenanced by the key, not by a stored value. |