Receipts without exposing the frozen test package.
The public snapshot makes every cost, token, timing, status, and artifact-integrity claim independently inspectable while keeping exact prompts and complete run ledgers inside the protected Builder package.
Pinned source
All 28 public records are generated from the private execution harness at commit 3f14507a8ce536d68a1f472f34f4c1a64c511f82. The harness remains private because the seven exact prompts and complete available run package are Builder material—not because the benchmark receipts are hidden.
Sanitized evidence
Each public record contains the task and configuration identity, prompt version, status, wall-clock duration, processed tokens, disclosed cost and pricing provenance, artifact hash, caveats, missing evidence, and the source-record SHA-256. Exact prompt text, private source files, and unreviewed logs are excluded.
Immutable versions
Published v1 JSON files are immutable and carry reviewed checksums in the manifest. A factual correction creates v2; it does not silently rewrite v1. That keeps links from reviews, videos, and independent audits reproducible over time.
Separate verdict dimensions
Cost and workflow-time comparisons are derived from the recorded ledgers. They do not imply a quality, reliability, or overall winner. Editorial judgement, community blind preference, and repeat-run reliability remain separate evidence categories.
Separate post-hoc supplement
Four shader reproductions were independently audited later and synced at harness commit f8f8982ca29d8c22fe76900a26fd30a1999bc8a2. Their versioned RTX manifestsupplements the frozen v1 ledgers without rewriting them. The records remain partial, the generation sample remains n=1, and no blind-evaluation outcome is claimed.
One request per run. No rescue pass.
- The benchmark target and frozen prompt match within each round.
- No human follow-up or manual correction occurs after submission.
- Provider-specific agents use their disclosed stacks; the stacks are not treated as identical.
- Cost figures are estimates from recorded usage and cited pricing—not subscription invoices.
- Missing preference, capture, environment, or repeat-run fields remain explicitly missing.