REBUILD.md

What rebuilds from this release, and what does not

Checked by running each script in a clean directory containing only this dataset and the two code repositories, with ATLAS_DATA pointing here and no other access.

The last run: 6 November 2026, at tag v1. Two fresh clones, this dataset downloaded at the tag into an empty ATLAS_DATA, a new virtual environment built from the tool repository's own pyproject.toml, and an empty Hugging Face cache — so nothing was served from a machine that had seen the data before. The run of 5 November was repeated on 6 November after the repository a script fetches from became an environment variable (ATLAS_DATASET), which touched eighteen released scripts: all 106 still import, the seven that run from the release still run, and the nine figures are still byte-identical.

The versions it ran under, because byte-identical output is version-sensitive and matplotlib changes its rendering between minor releases: Python 3.12.13, numpy 2.5.4, scipy 1.18.1, matplotlib 3.11.2, huggingface_hub 2.4.0. The figures below are byte-identical under those; under others, compare the figures' JSON sidecars, which carry the numbers.

How to put the two repositories on the path

Both repositories' experiments/ directories are plain script folders, not installed packages: pip install -e installs profilebattery, not experiments. Paper 2's analysis scripts import five helpers that live in paper 1's repository (atlas.py, construct_token_loss.py, b10_fit.py, b10v.py, concept_store.py), so both repository roots go on PYTHONPATH:

export ATLAS_DATA=/path/to/AtlasOpenCoder
export PYTHONPATH=/path/to/AtlasRepresentationalDrift:/path/to/AtlasCalibratedConceptProfiles

With that, all 106 released scripts import. Without the second entry, 45 of the drift repository's 50 scripts fail on a missing module.

Rebuilds from the release alone

Each of these was run in the clean tree and wrote its artefact.

script	repository	what it makes	time
figure_method.py	tool	Figure 1 (paper 1)	1 s
figure1_small.py	tool	Figure 2 (paper 1)	1 s
figure1.py	tool	Figure 1's neuron map	2 s
figure_c1.py	tool	Figure C.1 (paper 1)	1 s
paper2_figures.py	drift	Figures 1–4 (paper 2)	1 s
twin_ceilings.py	tool	TWIN_CEILINGS.json: every construct's twin ceiling at all 61 references, from validation/carriers/	11 s
check6_companions.py	tool	the companion check and Table B.3's asymmetry columns	1 s
core_counts.py, b_subtypes.py	tool	§3.1's core families, B's sub-sets, each set's grid by scope and padding
backfill_paper1.py	tool	SET_GRIDS.json, checked against all four printed grids
probe_margins.py	tool	§4.7's probe margins
c3_yardstick_nats.py	tool	Appendix C.1
b10_fit.py	tool	the spacing model
backfill_summaries.py	drift	the four summaries paper 2 quotes, each checked against the printed figure
b12_twin_turnover.py, b18_recurrence.py	drift	re-read the carrier sets at every reference, which are released for all 61

The nine figures of both papers come out byte-identical to the ones in the PDFs — the same SHA-256 for each, from the five figure scripts above.

Needs the record, which is not released

The record is 2,022 GB over 61 references and this dataset carries one worked example, so these scripts cannot run here. Each says so and writes nothing rather than producing a file that looks complete:

script	what it needs	what it says
band_held.py	record/<ref>/twin1__<c>_A.npz for the eight length-sensitive statements	"read 0 of 8 constructs … the public dataset does not carry [the per-prompt record]"
check4_relative.py	record/<ref>/{twin1,alt}__<c>_A_acc.npz for all 188	"read 0 of 188 constructs … Nothing written"
licence_tables.py	profiles/<ref>/twin1/<c>.npz for the 31 per-prompt constructs, for Table B.1's two null ratios	names the first missing profile and points at NULL_RATIOS.json, where the ratios themselves are released
rate_readings.py	the carrier index across references in the recorder's own layout	its output RATE_READINGS.json is released, so the numbers are checkable
record_size.py	the dataset repository's whole file tree, because that tree is the record	its output RECORD_SIZE.json is released
assign_tiers.py	the prior study's compositional_level.csv, named on the command line	argparse says the argument is required
build_battery.py	the generator — and could not reproduce the battery anyway: the per-prompt seeds are salted
run_s15.py, run_s16.py	OpenCoder's checkpoints and a GPU

So every number those scripts stand behind is checkable in the released artefacts even where the script that wrote them cannot be re-run here. The one exception is named in each paper's manifest.

One script disagrees with its own released input

retention_table.py (drift) reads validation/ablation/FULL_pretrain_iter_0484865.json, which is released, but raises KeyError: 'drifted' on For and If: in that file those two carry drift_unavailable ("this is the first reference of the run"), while the published TABLE_D1_retention has all 16 rows. The table and its input disagree, so the script cannot reproduce the released table.