Skip to content

Bank v1.0.0 · evaluated 2026-08-07 · n = 2,707–5,828 scored rows per model · sha256:8c9f5828

Data

Everything on this site is rendered at build time from versioned JSON artifacts in a data/ directory, and the same files are served statically under /data/…, so the links below are direct downloads. Every file embeds the run id, bank version, and the bank's SHA-256 content hash; scores are comparable only within a bank hash. There is no API; the JSON files are the interface. The cross-run index is index.json.

Published runs

RunDateBankModelsFilesDownloads
full-2026-08latest2026-08-07v1.0.0 · sha256:8c9f5828141,865leaderboardmanifestreceipts-manifestmodels/ and receipts/ under /data/runs/full-2026-08/
pilot-2026-082026-08-06v0.1.0 · sha256:a2fb28c83402leaderboardmanifestreceipts-manifestmodels/ and receipts/ under /data/runs/pilot-2026-08/
sample-2026-08synthetic2026-08-05v0.1.0 · sha256:234a5c5113744leaderboardmanifestreceipts-manifestmodels/ and receipts/ under /data/runs/sample-2026-08/
  • full-2026-08: real evaluation run. 75,696 model calls, 156,068 judge calls, $421.09 total; 1,568 demographic receipt shards.
  • pilot-2026-08: real evaluation run. 11,952 model calls, 26,524 judge calls, $13.31 total; 336 demographic receipt shards.
  • sample-2026-08: synthetic sample data generated to exercise the full published schema; it does not report on real model behavior. 37,440 model calls, 76,752 judge calls, $146.75 total; 728 demographic receipt shards.

Raw response archives

The working per-run JSONL files (raw model responses and judged rows, one JSON object per line, including error envelopes for permanently-failed calls) are retained by the operator and are not separately published. The published artifacts above carry the same substance in verifiable form: every response and its judge labels appear in the receipts shards, and every file embeds the bank version and hash they were produced under.

Files per run

FileContents
manifest.jsonmodels, effective params, judge pairs, bank pin, harness git SHA, totals, judge agreement (κ per rubric and per subject vendor, contested rate), run-average control band
leaderboard.jsonper-model overall + per-attribute consistency with 95% CIs and evidence tiers, control bands, refusal asymmetry, repeat instability, contested rate, n, cost
models/<file_slug>.jsonper-attribute metrics with CIs, per-control-set disparity, directional lean, pairwise-gap heatmap, per-concept disparities, top-10 most inconsistent receipts
receipts/<file_slug>/<attribute>/<concept_id>.jsonmatched prompts with swap spans, every response repeat, judge votes, final labels, per-group summary (control-set shards included)
receipts-manifest.jsonshard paths, receipt counts, flip counts, group-pair facet counts

Schema

The authoritative shape of every artifact is defined in docs/contracts.md (the "Published artifacts" section) and mirrored as TypeScript types in site/types/schema.ts, which this site consumes at build time. The essentials:

  • Every per-run file carries the envelope {schema_version, run_id, bank_version, bank_hash, generated_at}; index.json spans runs and carries no envelope. Publish hard-fails if any single file is 1 MB or larger.
  • Disparity-family values (D and its CIs and differences, worst-case gap, per-concept D, heatmap |Δs|) and consistency scores are on the 0–100 presentation scale; rates (repeat instability, contested rate) are fractions in [0, 1]; receipt group_summary s-values stay on [0, 1].
  • Every number that carries uncertainty is published as {value, ci: [lo, hi]} (95% percentile cluster-bootstrap CIs, B = 2000, seeded).
  • Receipt ids are <bank_version>:<concept_id>:<model_file_slug>; model file slugs are the model slug with / __ and : --.

See Reproduction for how a run is regenerated from a pinned bank.

Bank-version changelog

BankDateNotes
v1.0.02026-08-06Frozen after pilot screening. No items were dropped: no concept hit the 80% universal-refusal prune and no rubric fell below the κ gate. The freeze changes the bank hash, so v1.0.0 scores are not comparable to v0.1.0 scores.
v0.1.02026-08Pre-release pilot bank: 7 demographic attributes × 16 concepts plus 2 control attributes × 10 concepts. Used by the pilot run (and the earlier synthetic sample run that exercised the schema).

Any edit to any template, group list, or rubric changes the bank hash and starts a new comparison universe; runs are comparable only within a bank hash. Bank versioning and hashing