Oracle
History, comparison, and verification for recorded verdicts. Every number on this page is either computed live in this browser from fetched or uploaded data, or is one of the recorded trials below. No model grades itself.
Evaluation
Computed in your browser from trial-pack.json / bench data — no model grades itself.
Three code-scored checks. Nothing here is asserted by prose: each metric is recomputed by this page's code from the fetched data.
Checks have not run yet — press Run evaluation checks.
Trial-pack verification
Each trial hash is recomputed over the canonical core, and the pack hash is recomputed over the joined trial hashes.
Tally recomputation
Recomputed from the miss data, then compared with the tally recomputed from the bench data.
Coverage
For each recorded trial, which evidence fields are present and which are absent, as counts only.
Bench panel
Computed in your browser from trial-pack.json — the frozen copy of bench/results.json. No model grades itself.
Wall time per recorded trial, with the median and the range across trials. With so few trials, no percentile is shown: a percentile on a tiny n would imply precision the data does not carry.
Loading bench panel…
Verdict history
Computed in your browser from trial-pack.json / bench data — no model grades itself.
Verdicts run in this browser are recorded here, on this device only. Nothing is prefilled: the only pre-existing rows are the recorded trials below.
No verdicts recorded in this browser yet — run a verdict in the app, or view the recorded trials below.
Recorded trials, 29 Oct 2026
Recorded, not live. These rows are the trial-pack trials, shown separately from browser history.
Before and after comparison
Computed in your browser from trial-pack.json / bench data — no model grades itself.
Pick any two entries — recorded trials, history entries, or one of each — and see them side by side. Differences are arithmetic display only.
Comparison has not run yet — press Compare.
Upload and verify
Computed in your browser from trial-pack.json / bench data — no model grades itself.
Upload a JSON file holding a verdict object, a workspace object, or a single receipt. Uploaded data never leaves this browser — no network call is made for uploads.
No file uploaded yet.