Historical Roots 30 EN

Historical root access for all2,400 selected tasks — experiment30

All2,400 frozen pre-fix snapshots have accessible root-tree metadata. This extends the previous53-repository sample to every selected snapshot.2,230 contain root license-name candidates;170 require nested-path inspection. This is not completed license-text review or model evaluation.

Scope and results

Following the precommitted plan, recompute experiment29's selection from pinned projections. Verify2,400 entries and membership SHA-256 fa58c4e51ae307963eb587813e49adf70d987dcb4893f6ee1aa6050bc6b80a5a, then query each exact base_commit root tree. Retain every selected request, including the53 long inputs.

Item	Result
Target / accessible snapshots	2,400 / 2,400
Unavailable snapshots	0
Snapshots with root license candidates	2,230
Snapshots needing nested-path inspection	170
Metadata response bytes	14,752,218
Deduplicated candidate objects	143

The170 are163 matplotlib/matplotlib snapshots and7 tokio-rs/axum snapshots. Further inspection of previously sampled roots shows a LICENSE directory in Matplotlib and a multi-package layout in Axum. Absence of a matching root file does not mean absence of a license. Nested lookup is a subsequent step.

The LICENSE/COPYING/NOTICE filename rule is unchanged from experiment29.143 counts distinct (repository, path, blob ID) tuples, not143 license types. Object references and occurrence counts stay in a private local manifest. The public repository aggregates contain no individual task or commit IDs.

Execution and caching

Two workers own separate input/result slots. Bound requests to30 seconds, responses to2MiB and error capture to4KiB; space each worker's request starts by at least250ms. Existing GitHub CLI authentication handles credentials. Credentials and raw errors never enter public results.

Only validated successful responses enter the cache through a temporary file and atomic rename. Reject incomplete trees, invalid object types/modes/IDs, duplicate paths and oversized responses. Do not replace failures with easier tasks. Reuse the original53 sample responses. This cache is source metadata, not license clearance.

An --offline replay makes no GitHub requests. Missing cache entries remain unavailable observations. All2,400 entries were present here; the aggregate and private candidate manifest are byte-identical to the collection run.

go run ./cmd/riido-rootcoverage --out .cache/historical-roots-new
# After collection, reproduce into a new directory without network requests.
go run ./cmd/riido-rootcoverage --offline --out .cache/historical-roots-offline-new

Pinned projections are required. Output directories must be new. Keep candidates.json and the source cache local; do not upload them to Git/HF.

Verification, resources and interpretation

Injected-response tests cover cache reuse, failure non-persistence, malformed identities and bounded output. Test candidate deduplication and per-repository success/failure/missing-candidate accounting. Full Go race tests, vet and formatting passed; the original53-sample report still reproduces. An actual2,400-entry offline run under the race detector also matches the original aggregate.

One cache-only M4 Pro CPU replay took0.93s wall,0.44s user,0.10s sys and32,440,320 bytes peak RSS (about30.9MiB). This does not measure the entire initial network collection, model inference or GPU costs.14.8MB of metadata does not estimate full source or container storage.

The result establishes root-list access at every selected commit. No code or license contents were downloaded and recursive catalogs were not checked. Candidate files still require text and third-party-condition review. LicensesReviewed and ProductionReady remainfalse.

Next inspect nested license paths for170 snapshots, verify deduplicated candidate contents and document source conditions. Long-input handling and actual2,400-task evaluation remain pending. No training, inference, new HF release, default activation or use of existing final reserves occurred.

PR #53 · CI

Follow-up: experiment 31 resolves the 170 nested layouts and verifies candidate text for all 2,400 snapshots. Complete legal review and model evaluation remain separate.