matiasmolinas/dental-analysis/traceability-audit

Traceability Audit Skill

<!-- Evolved by SkillOpt (live run 2026-07): parent a587e4e67342 -> 802df657583d. Adopted under the gate (guardrail 1.0, field-citation coverage 0.00->0.93, gain CI [0.80,1.00]). See docs/evolution/live-run-2026-07.md. Purely additive over the parent. -->

Verify that every claim in a structured oral-systemic output is grounded in the input record.

Audit discipline

• For each relational_axes[*], confirm traceability lists real fields that

exist in the normalized record; reject axes citing absent or invented fields.

• Confirm oral_evidence and systemic_evidence items are actual values from the

record, not paraphrases that drift from the data.

• Reject any relation asserted without both an oral and a systemic anchor (a

cross-domain axis needs both sides present).

• Check that hypothesized_mechanism is a recognized association from

oral-systemic-kb, not a novel causal invention.

• **Every individual hypothesis or relational statement (not just the axis as a

whole) must carry its own inline machine-checkable citation tag in the exact form [field: <name>], where <name> is a real field in the normalized record.** A hypothesis is unsupported if it lacks this tag even when a field name appears elsewhere in prose, is paraphrased, is embedded only in narrative text, or is asserted without brackets - free-text mentions of a field do not satisfy traceability; the tag must be present and machine-parseable.

• When a hypothesis draws on multiple fields, require one [field: <name>] tag

per contributing field (e.g. [field: bop_percent] [field: hba1c]); partial tagging (citing only one of several relied-upon fields) is unsupported for the untagged portion.

Confidence discipline

• Down-rate confidence when an axis leans on a MISSING mediating datum; the

correct move is a required_missing_data flag, not a high-confidence claim.

• Flag over-claiming: language implying diagnosis, causation in an individual, or

treatment benefit must be softened to association/hypothesis.

Output discipline

• Return a per-claim verdict (supported / unsupported / over-claimed) with the

offending field or phrase, so the orchestrator can revise before finalizing.

• Compute and report a traceability_coverage score: the fraction of hypotheses

that carry a valid [field: <name>] tag for every field they rely on. Any hypothesis missing this tag counts against coverage regardless of how plausible or well-referenced its prose appears.

• When coverage is below 1.0, list each untagged hypothesis verbatim alongside

the missing [field: <name>] tag(s) it needs, so the orchestrator can insert them rather than rewrite the claim.