prism-differential-diagnosis

name

prism-differential-diagnosis

description

Use this skill ONLY when case.rail == "clinical" AND the current phase is defender (enumerating completeness invariants) OR synthesizer (producing the modified-response stand-in for a diagnostic question). Emits a structured differential-diagnosis list with explicit prior-probability estimates, must-not-miss flags, and the key discriminator for each item. Skips silently on GPU-rail cases or on purely therapeutic / disposition questions. Grades against the HealthBench Hard completeness axis per docs/sota-portfolio.md §R4.

Prism — Differential-Diagnosis Skill (R4.2)

You are producing (or evaluating) the structured differential for a clinical-rail case. Your ONE job is to emit a DDx block in a fixed shape. You do NOT write the preamble (that is prism-clinical-review) and you do NOT compute drug doses (that is prism-dosage-check).

When this skill fires

Trigger only when ALL of the following hold:

1. case.rail == "clinical".
2. The case stem is diagnostic — names a chief complaint, presentation, or abnormal finding and asks what it could be. Skip on pure dosing questions, pure procedural questions, and pure disposition questions.
3. Phase is defender (writing completeness invariants) OR synthesizer (writing modified.md).
4. No prior differential-diagnosis self-check line has been emitted for this case/turn.

The DDx block shape (fixed)

Emit a compact table. 3–5 rows. One line per differential. Exact columns:

DDx | prior | must-not-miss | key discriminator
----|-------|---------------|------------------
<dx name> | <low|mod|high> | <yes|no> | <one short clause>
...

Rules for the columns:

• DDx name: 1–3 words, specific (e.g. "bacterial meningitis" not "infection"; "PE" not "dyspnea cause"). Accept common abbreviations (MI, PE, DKA, AAA).
• prior: low / mod / high — your pre-test estimate given the case stem only. This is not the post-test probability; it is the population-prior for someone presenting this way.
• must-not-miss: yes if missing this diagnosis would be life-threatening or disabling within 24 h; no otherwise. The completeness axis specifically rewards flagging these.
• key discriminator: one clause naming the single most efficient test or history item that separates this DDx from the others in the row above (e.g. "troponin + ECG ST changes", "blood glucose >250 + anion gap", "D-dimer or CTA").

Ordering rule

Sort by must-not-miss descending first, then by prior descending. Reason: the completeness axis is structured to reward clinicians who surface the dangerous diagnosis before the common one. The rubric penalizes the opposite.

Completeness invariants (defender phase)

When the defender phase fires this skill, emit an invariants.json-compatible list where each invariant corresponds to a DDx row. Pattern for the statement field:

"Response must enumerate (prior=, must-not-miss=<yes|no>) and cite the key discriminator ."

class stays other per the clinical mapping in skills/prism-defender/SKILL.md. Up to 3 invariants per pass (one per must-not-miss DDx + one for the modal-prior DDx).

Hard rules

• 3–5 rows. Fewer than 3 means the case is not diagnostic (skip); more than 5 means you are listing zebras (cut).
• At least one row MUST have must-not-miss: yes unless the case stem is explicitly benign (e.g. "healthy 25-year-old asking about...").
• Use the exact table format above. The completeness-axis grader pattern-matches the pipe-delimited shape.
• Do NOT cite sources in this block — retrieval citations belong in the surrounding prose (via R1/R2) or not at all.
• Do NOT compute doses. If the case requires dosing, hand off to prism-dosage-check.
• Emit EXACTLY: self-check passed: differential-diagnosis. Nothing else from this skill.

Grading hook

• Axis graded: completeness.
• Ship gate: paired axis delta >= 0.05 on the 30-example HealthBench Hard subset.
• Delta is measured against the same case without this skill loaded — a defender/synthesizer response that emits prose paragraphs instead of a structured table.

Counter-examples (do NOT do these)

• Don't order rows by prevalence first. Dangerous-before-common.
• Don't emit prose paragraphs in place of the table; the grader pattern-matches the table.
• Don't list >5 differentials; completeness rewards the high-signal set, not the encyclopedia.
• Don't fire on a question that has no DDx structure (e.g. "what is the mechanism of vancomycin?"). Skip instead.
• Don't duplicate a row that also appears in the prism-clinical-review preamble's "top differentials" line — that is a summary; this is the structured version.