Drill into one run: its committed ledger line, joined with the local findings tier and the committed sets where they exist on this machine — per-item detail, provenance, and an honest fallback wherever detail is absent.
Measurements only (attestation lines join as the attested pill). Detail depth per run depends on which findings files exist on this machine — the column on the right says which.
Runs whose local detail files exist but which never wrote (or pre-date) a committed ledger line — shown rather than hidden, so the local tier and the ledger can be reconciled.
The committed corpus this gate grades against — every item with its expected capability, arguments and rubric. Expand a row for the checks and notes.
The tool schemas exactly as production sends them to the model —
dumped from capabilities.prompt_surface.tool_schemas(), the same surface
prompt_sha hashes (spec 14 §5). Per-set usage counts show which eval items target each.
Sibling pages: the Eval Ledger walks one model over time ·
the Footprint Ladder walks models within one prompt era ·
Anatomy of a Gate Run explains the mechanism behind all three ·
the Run Composer composes the command for the next run, and never runs it.
This page needs python3 evals/dashboard/serve.py --local-details (loopback only — the
findings tier never mirrors and never serves off-machine). Narration hand-review and attestation stay
with evals/harness/review_narration.py; this page renders, it never writes.