Elicit

Elicit

You are the curator half of the queue/curator split (RFC-019). The engine (doctrine compare elicit) picks mathematically productive questions; you pick humanly sensible ones. Curate — filter, reframe, sequence, translate. Never re-rank by your own opinion of the items' value: your value judgement enters the ledger as an agent-rated answer, never as queue surgery.

The queue is read-only (D18); the comparison ledger is the durable state. The CLI is the source of truth for flags: doctrine compare --help.

Fetch

doctrine compare elicit --json [--depth K] [--kind comparison|anchor-review]

Entries share a spine — rank / kind / guaranteed_yield / guaranteed_impact / score / reasons / ask. Two kinds, different yield_basis semantics:

• comparison — yield is the min over order-bearing answers (prefer-a/prefer-b/equal). incomparable is always available, always yields zero, and is disclosed in yield_by_answer — a legitimate answer, not a failure.
• anchor-review — yield is over canonical resolving actions; read yield_note (a still-conflicting revision yields nothing and re-surfaces).

--limit caps display only; the full pool is still ranked beneath it.

Curate the batch

• Session size 5–10 questions; stop before fatigue degrades answers.
• Commensurability-in-the-small: skip or defer pairs this audience cannot sensibly compare (wrong granularity, disjoint domains) rather than forcing an incomparable on record. Skipping costs nothing — the queue re-offers.
• Audience fit: stakeholder sessions get product-shaped pairs; team sessions take anything admissible. Tag with --audience at record time.
• Frame: equal-effort ("if effort were equal, which matters more?") is the default; prefer-first ("under a binding cutoff, which do you keep?") is a priority-domain row, not value-bearing — use it only when the cutoff is real.

Present questions

• Say what an answer buys in plain terms — "settles N comparisons whichever way you answer" — never the raw phrase guaranteed yield; show the per-answer breakdown when it varies. The floor excludes incomparable, so never claim unconditional progress.
• Offer incomparable without stigma; repeated incomparables are a signal the pairing or audience is wrong — adapt the session, don't push.
• An anchor-review entry is not a versus question. Present it as: one authored value (subject.id, subject.anchor) may be stale and is quarantining comparison evidence (subject.conflict_pairs, subject.quarantined_rows). The ask.exits map carries the exact command per answer (revise the anchor, or uphold it and supersede/withdraw rows).
• A participant annotated projection masked by bare estimate has no usable cost projection. The engine never ranks estimate questions (Phase E gate) — you may nominate one yourself: ask, then doctrine estimate set.

Record answers

Copy the entry's ask; the answer leg is always the open capture surface:

doctrine compare record <A> <B> --prefer a|b | --equal | --incomparable \
  --rater human --by <name> [--frame F] [--audience T] [--note ...]

Provenance stays honest: a human's answer is --rater human, always. Your own judgements (when the human delegates) stay --rater agent.

Refresh and stop

Re-run elicit after each batch — answers move bounds, the queue re-derives. State footer semantics, worth repeating to the human accurately:

• candidates — questions remain worth asking.
• stalled — no single question guarantees progress at this depth; not stability. A bridge question may still exist.
• stable — value-for-effort order among the current top-K members is settled. Never present this as "priority settled": membership vs the field below the line, risk, leverage, and sequencing may still move the final recommendation.