Local record inspection · No live inference
Bring your own decision records.
Compare conditions for the same model and item. Align candidate probabilities by semantic key and trace each value to the physical record you supplied.
Your JSONL stays in this browser tab. No upload, network request with your data, analytics, or browser storage. Refresh or “Clear data” removes the app’s in-memory copy.
Choose a JSONL file
One physical record per line, up to 25 MiB or 20,000 records. Supply logits plus temperature, or explicitly declared full candidate probabilities. Read the v1 schema.
Or paste JSONL
Inspect a recorded engineering case: download 48 retained diagnostic records, then choose that file above. These are model outputs on synthetic inputs, with two numerical-tolerance failures and unchanged labels. Read the case and its limits.
Choose a file or try the clearly synthetic example.
Input checks
Choose a comparison
Candidate comparison
Probabilities are conditional on the declared candidate set. Δ is B − A in percentage points. They do not measure the probability of a correct answer.
| Semantic key | A | B | Δ (pp) |
|---|
Original records and provenance
Source identities and file hashes are supplied by the author. The importer checks consistency, not source authenticity. Derived/shared-member recipes are unsupported in v1.
This utility checks the format and compares supplied distributions. It does not assess model quality without a reference, verify semantic mappings, run a model, or establish human usability. “Full” means the author’s declared candidate set; an undisclosed missing candidate cannot be detected.