One system

Six tools that only mean something together

Each answers a different question about whether an evaluation can be trusted, and each one's number below is recomputed from bytes pinned by sha256 to the issuer commit named on that panel. Nothing here is typed by hand, because a page of copied numbers is true on the day it is written and quietly wrong afterwards.

If a source stops matching its pin it renders STALE and loses its number; if it cannot be obtained at all it renders ABSENT. Neither one keeps the value it printed yesterday, or the word verified means nothing.

generated by suite/build.py (build 99673458, manifest d77f54e8) · sources pinned as of 2026-10-13, each to the issuer commit on its panel and not to current main · fetch any link and hash it: it must match the sha256 beside it

evalmut

Does your eval suite actually check anything?

91.3%

mutation score: 91.3%
caught / applied: 42 / 46
holes found: 4
declined, not guessed: 223

evalmut dogfood mutation score against its own corpus
pinned to docs/dogfood_gradecore.json
sha256 4cccd5e0 · evalmut@d92f24f · evalmut_dogfood@1

reference-fleet

Which defect classes does a suite catch, against a known answer key?

13/18

suite x member results: 18
suite archetypes: 3
pairs with any detection: 13 of 18
fleet commit: 07c6697

Suite-by-member detection results against a seeded answer key
pinned to board/results.json
sha256 c00213d1 · reference-fleet@53f931f · fleet_board@1

model-drift

Did a model move, or is that inside the noise floor?

LLMs in the last run: 17
LLM series ever charted: 19 (2 retired)
suite size: 35
last run at the pinned commit: 2026-10-13
control series: 1

Models tracked by the frozen daily drift suite
pinned to dashboard/drift_board.json
sha256 a0879c39 · model-drift@d4e7115 · drift_metrics@1

agent-certlab

Can this agent repair seeded defects under a stated policy?

6/6

seeded defects fixed: 6 of 6
capability: repair-seeded-defects:intervals — claude-code-headless fixed
under policy: 6 of 6
tests green: 6 of 6

Seeded defects repaired under a stated certification policy
pinned to certifications/claude-code-cloud-2026-08-14/vac.json
sha256 ab9994fe · agent-certlab@9bc0019 · certlab_bundle@2

crashkit

Do the graders bite, with no LLM judge anywhere?

declared checks in bundle: 8
capability: crashkit's no-LLM-judge battery grading is deterministic end-to-end an

Deterministic checks declared in the crashkit evidence bundle
pinned to vac/vac.json
sha256 95998310 · crashkit@878a0ef · crashkit_bundle@1

vac-protocol

Can a stranger verify these bundles offline?

bundles in registry: 12
pending: 0
registry version: 0.1

VAC registry acceptance state, including bundles the verifier refused
pinned to registry.json
sha256 443e126d · vac-protocol@69c9510 · vac_registry@1