Existing checks found missing evidence, but weak citations slipped through.
Catch rate: 60%
Seeded fail cases caught: 27 / 45
Miscitation cases caught: 0 / 18
Seeded pass cases rejected: 0 / 27
Method notes · benchmark
The correction is the proof.
See our own claim-verification harness before and after a separate labeled benchmark check. The receipt is self-administered and bounded to its own labeled set; it is not customer proof or third-party validation.
Before / check / after
Before · baseline harness
Separate check · citation support
Corrected · same definition rerun
The correction is the proof.
The artifact, labeled case definition, missed cases, correction, and rerun are ours. No prospect, customer, or third-party product appears in this demo.
Compare the claim words with the cited line.
A deterministic add-on reads the first path:line citation on a report line and flags unreadable, empty, or low-overlap evidence.
It is a lexical containment heuristic, not semantic entailment. It remains advisory for real reports because a valid claim can span more than one cited line.
The miscitation classes closed on the labeled set.
Catch rate: 100%
Seeded fail cases caught: 45 / 45
Miscitation cases caught: 18 / 18
Seeded pass cases rejected: 0 / 27
Committed sources: data/quality/verification-bench/bench-2026-07-09-expanded.json and skills/quality/verification-bench/citation_support.py.
Safe aggregates
01
02
03
Numbers come from evidence or they do not render.
Current inventory totals come from the committed Git index at build time. The repeated-run result comes from a retained dated excerpt. Neither is presented as a fresh full-suite run, and hand-typed metric markup still fails the build.
Tracked substrate test programs
460 named test programs in the committed index: 459 shell and 1 Python. Inventory only; this does not say they ran or passed in this build.
Skill documents with operating contracts
118 non-archived committed skill documents declare a nonempty top-level contract. This is a document inventory, not an execution result.
Retained repeated-run summaries
The 2026-06-11 excerpt contains 4 summary records. Each records 5 of 5 attempts passing. The excerpt retains complete constituent run rows for 3 summaries. This is a historical internal rehearsal, not current system reliability.
What a receipt contains
Enough detail to reproduce the decision.
- The artifact or input that was checked.
- The rule and the evidence source used by the checker.
- The timestamped pass, fail, or need-data result.
- The correction and the re-run, without erasing the first failure.