# Pacific Compute public evidence codebook — V2

Snapshot: 2026-09-01, source commit 17df23f. V2 editorial review: 2026-09-06 UTC.
Data licence: CC BY 4.0. Credit Pacific Compute and retain the snapshot and methods.

## Units and denominator

19 coded records: 18 release-specific records and one Qwen3.6 family record. 21 additional families are queued, producing 40 rows. These are not a representative sample of all open models.
The 42 high-popularity families split into 17 represented, 4 below the frontier-scope rule, and 21 queued.

## Evidence tiers

T3: third-party dangerous-capability evaluation, with a public method and relevant results. Independence is from the model developer, not from every funder or affiliation. Not comprehensive coverage, replication, peer review, certification, or a safe-model verdict.
T2: developer-run dangerous-capability evaluation.
T1: developer documentation, including model cards, safety claims and partial ordinary safety tests, without a qualifying dangerous-capability report found in the bounded review. Internal source tiers 1–3 roll up here.
T0: completed source search found no qualifying safety/evaluation disclosure. Not proof that none exists.
unreviewed: evaluation search incomplete; never counted as T0.

Domain values: yes = source supports the domain; partial = narrower, ambiguous or not checkpoint-specific; no = none found in the dated search; unknown = not reviewed. A yes in one domain does not cover the others.

## Record-review status

verified: legacy source label for an internal AI-assisted second pass on 2026-09-01, displayed as “Internal second pass”. NOT a third-party audit.
coded: initial AI-assisted source coding.
needs second reviewer: a substantive source change still needs a second pass.
unreviewed: no completed evaluation coding.

No external auditor has signed off on this dataset. The source-link check is downloadable separately; HTTP success does not establish factual correctness.

## Fields

id: stable page anchor. source_row_id: original CSV ID, null for derived queue.
model / developer: labels in the source snapshot.
release_date: displayed record date; corrected Kimi K3 weight-release date is 2026-07-27.
original_snapshot_date / date_basis: preserve the source date and explain changes. Queued rows use repository-creation dates, not verified release dates.
record_scope: release or family.
evidence_tier / review_status: distinct coding dimensions above.
sources: original documents, with the redirected Qwen3.6 GitHub link excluded.
evaluation_reports: named evaluator, title, tested checkpoint/scope, relevant domains and report URL. Reports are selected support, not an exhaustive bibliography.
review: internal process, external_audit=false and a public caveat.

## Launch series

192 families, API pull 2026-09-01, 42 organisation IDs / 41 labs, 6,977 repositories. 250/500/1,000 likes are nested popularity thresholds, not capability classes. Counts in the last 365 days: 113/76/42.
Family dates use earliest repository creation; peak likes at the pull set the band. Exclude quantisations and specified derivatives; preserve product tiers such as Flash and Pro. The chart includes 24 complete months through August 2026, excluding September's partial month.
Historical raw API responses are not archived here. Re-running today's API is a new observation, not reproduction of the old like counts.

## Known review work

Split Qwen3.6 into checkpoint records; recode external benchmark-only evidence consistently; finish second review for MiniMax-M3; validate release dates and licence distinctions; commission an independent human audit with documented conflicts and a signed disposition for each row. Do not change tier counts merely because a URL fails to load.

Corrections: info@pacificcompute.org, with model, field and primary source.