docs/phase1_status.md

# Phase 1 status: data before models

Status date: 2026-08-12 America/Los_Angeles.

## Current decision

Use natural repository code as the primary loss-scaling corpus and keep task data
in separately reported roles:

| Role | Sources | Decision |
|---|---|---|
| Base pretraining and repository loss | The Stack v3 TRAIN | Primary source; admit only resolved, allowlisted permissive files in the main condition. |
| Same-problem auxiliary ablation | Project CodeNet | Eligible only after whole-problem-family splitting and global deduplication. |
| Task/SFT data | McEval-Instruct; conditional Nebius task/trajectory subsets | Never merge silently into the base-pretraining condition. |
| Functional evaluation | McEval plus a later untouched temporal/private set | Quarantine complete task families from every training source. |
| Provenance/contamination firewall | SWE-rebench V2, PRs, leaderboard, and repository pointers | Use repo, commit, PR, patch, tests, and task IDs to exclude overlap. |
| Metadata-only/legal quarantine | MultiPL-E; CuTe DSL until permission | Publish pointers and exclusion decisions, not training payload. |

Nebius does not distribute one reusable full-repository archive per task. Its
records contain repository and base-commit pointers, patches, tests, issue text,
and execution metadata. Reconstructed repositories are useful for execution and
contamination checks; adding them again to Stack v3 would usually duplicate core
code.

## Provisional matrix panel

The audit panel contains 24 mutually exclusive language buckets: the exact 20
SWE-rebench V2 languages plus Haskell, Fortran, COBOL, and Zig. This creates 576
directed cells (including diagonals) and 276 unordered bilingual pairs. CuTe DSL is
a separate Python-embedded adaptation target, not a 25th mutually exclusive file
language.

The panel is not final until permissive post-filter tokens, unique repository
families, clone concentration, parser/compiler success, and evaluation coverage are
measured. See `registry/audit_panel.json`.

## First reproducible measurements

The Stack v3 source is pinned to revision
`a34523fd426a429b7686a5e5786496bf2e1b1172`. The official statistics show why the
project must distinguish raw availability from training-eligible availability:

| Language | Raw estimated tokens | Permissive-labeled estimated tokens | Permissive share |
|---|---:|---:|---:|
| Python | 200.865B | 12.586B | 6.27% |
| C++ | 801.913B | 32.642B | 4.07% |
| Rust | 19.135B | 4.363B | 22.80% |
| Haskell | 2.063B | 107.9M | 5.23% |
| OCaml | 1.343B | 57.5M | 4.28% |
| Julia | 1.851B | 43.7M | 2.36% |
| Elixir | 1.148B | 159.9M | 13.93% |
| Zig | 909M | 33.3M | 3.67% |
| Fortran | 7.055B | 135.9M | 1.93% |
| COBOL | 530M | 7.1M | 1.34% |

These are still upper bounds before the project SPDX allowlist, vendor/generated
filtering, parser audit, benchmark firewall, repository-family grouping, and global
deduplication.

A record-level pilot has now been run on pinned Stack v3 shard `part-00000`. The
416,084,426-byte Parquet file contains 21,137 repository rows and 241,665 files;
every declared file count matched the nested payload. Under the deliberately
incomplete pilot gate (one of the 24 panel languages, Stack's `permissive` label,
and non-vendored), 5,054 content IDs survive and none repeats in the shard. Some
low-resource languages have no surviving file in this particular shard, so it is
strictly a pipeline validation—not a basis for extrapolating final totals. The
full 8,192-shard scan and explicit detected-license allowlist remain pending.

Pinned SWE-rebench V2 contains 32,079 tasks across 3,617 repositories and 20
language labels. Its distribution is highly uneven: Python has 7,243 tasks over
691 repos, Go 6,144 over 967, C++ 182 over 17, OCaml 58 over 18, and Lua 39 over
seven. It has no Fortran, COBOL, Haskell, or Zig task slice.

Pinned McEval contains exactly 16,031 rows over 40 language labels and six task
surfaces. Several completion surfaces are variants of the same canonical problem;
they must be grouped before any split. Pinned McEval-Instruct contains 35,943 rows
over 36 case-normalized labels, not the roughly 110K examples reported for the
original construction. Artifact versions therefore matter.

The pinned SWE-rebench leaderboard payload contains 860 unique instances. Its 15
monthly partitions contain the same 860 instance IDs as the aggregate `test`
partition, so counting every advertised split produces 1,720 rows but no extra
tasks. The pinned Parquet schema also lacks the `license_name` field described by
the dataset card. It remains a frozen evaluation/firewall source, never a training
source.

The remaining pinned Nebius task artifacts are also locally complete and
checksummed. SWE-rebench V2 PRs contains 126,300 rows but only 122,910 unique
repo/pull IDs; 1,920 IDs repeat, producing 3,390 surplus rows, and the payload has
no language column. The OpenHands release contains 67,074 unique trajectories over
6,306 issues and 1,823 repositories; 3,792 issues have at least one successful
trajectory. These are post-training/task artifacts, not additional full-repository
pretraining corpora.

The complete Project CodeNet 1.0.0 archive is now pinned by SHA-256. A census of
all 4,053 per-problem metadata files reproduces exactly 13,916,868 submissions,
7,460,588 accepted submissions, and 55 mapped language/category labels. It also
materializes 55,230 problem-language cells for leakage-safe family selection.
Notably, Fortran has 15,059 submissions across 1,088 problems, whereas COBOL has
1,277 across 340; these are useful controlled-transfer examples but tiny compared
with C++'s 8,008,527 submissions. The paper confirms CDLA-Permissive-2.0, although
the downloaded archive omits a packaged license notice; any derivative release
must add the agreement text.

The pinned mHumanEval commit does not contain the advertised public artifact. A
complete streaming audit found only 17 partially populated programming languages
and 544,911 logical rows, versus the claimed 25-language, 836,400-row grid. Eight
claimed languages are wholly absent: Bash, D, Julia, Kotlin, MATLAB, Visual Basic,
Fortran, and COBOL. Consequently mHumanEval cannot supply the remembered public
COBOL/Fortran evaluation data at this revision. The complete archive hash,
per-language census, and every discrepancy are published in the provenance
repository.

## Exact ATLAS scale constraint

The ATLAS full grid spans 20 reported sizes from 9,044,352 to 8,452,190,208
parameters. Its transfer matrix uses the 2,034,422,912-parameter scale-34 model,
50/50 bilingual sampling, and a 42B-token reference horizon. See
`docs/atlas_replication.md` for every architecture dimension.

For 24 languages there are 276 unordered bilingual runs. Under ATLAS's BTS
definition, the zero-transfer point is `2 * 42B = 84B` total bilingual tokens; a
full direct grid reaching that point would therefore be 23.184T tokens, before
monolingual baselines and scale experiments, and negative-interference curves can
require more. ATLAS itself measured a subset and estimated the larger matrix using
continual-pretraining adaptation signals; Code-ATLAS will label every cell as
measured or estimated.

## Split lock

Do not create final train/validation/test IDs until all admitted sources have been
globally deduplicated. The loss-test target follows ATLAS exactly per language:
`min(20,000,000 reference tokens, 20% of unique eligible tokens)`. A new validation
target is `min(2,000,000, 10%)`. Repository/clone/problem connected components are
indivisible. See `docs/split_protocol.md`.

## Publication

The public release is divided into provenance, permissive payload, parallel
auxiliary, and evaluation-manifest repositories. Restricted sources are visible as
metadata-only exclusions. The provenance repository is public at
https://huggingface.co/datasets/arpandeepk/code-atlas-provenance. The permissive
payload repository will not be created until record-level license and contamination
gates pass.