The claim registry
No headline number on this site is typed by hand. Each one is generated from a registry entry that references the run IDs it was derived from, the aggregation used, the exact scope (system, runtime, model artifact, workload revision) the number is true for, and the limitations that bound it. A retracted claim cannot render at all — the build fails.
Superseded and retracted claims stay in the registry with their status; corrections are logged on the errata page.
Evidence bundles
A published qualification result ships with a bundle containing:
- environment manifests — hardware fingerprint, firmware, kernel, driver, image digests;
- workload identity — revision, item hashes, load model, cache policy, quality gates;
- metrics — derived measurements with the queries that produced them;
- request-level data — per-request timings, including failed requests;
- telemetry — thermal, memory and utilization time series where captured;
- logs and checksums for every file in the bundle.
Invalid and failed runs are preserved and marked, not deleted — a rerun gets a new run ID and the original stays in the record.
Redaction for private engagements
Private bundles go through a deterministic redaction pipeline before any public derivative exists: sensitive values are removed, a secret scan runs, a human review follows, and the public derivative gets its own checksum. What was redacted is declared; the redaction itself is part of the record.
Evidence levels
| Level | Meaning |
|---|---|
| legacy | Collected before methodology v1; kept for history, never used for current claims. |
| lab_single_run | One valid run in the lab. The weakest current level; labeled as such. |
| lab_repeated | Repeated valid runs on one unit; headline numbers require at least this. |
| lab_unit_replicated | Reproduced on a second, commercially identical unit. |
| external_reproduced | Reproduced by an operator outside the lab from the published recipe. |
| vendor_disclosed | A vendor-supplied figure, reported as such — never merged with lab measurements. |
| estimated | A derived or modeled figure, labeled as an estimate; never a headline. |
What checksums do and do not prove
Checksums prove a bundle has not changed since publication. They do not prove the operator ran the test honestly — no hash can. That is exactly why evidence levels and external reproduction exist: the strongest claim this lab can make is one someone else has reproduced from the recipe.
Availability
Public bundles ship together with the first published tested configurations. Until then this page documents the format, not an archive.