challenge-review

Challenge Review

The adversarial refinement loop for analysis spaces: who challenges, with what context, how findings return, and how they enter the space.

When to Use

• Loaded by the business-analysis entry before a domain gate, and by the analysis-challenger and domain-expert roles it spawns.

The Loop

1. Cast the panel: pick lenses from the lens bank sized to the domain, and cast topic-specific expert profiles; record both in the round record with a one-line why each.
2. Fan out: one fresh-context, read-only spawn per lens or profile (analysis-challenger role; domain-expert for named questions). Every spawn prompt carries the constitution body, read from the file printed by "$RUN" path "$TEAM" constitution.md (dispatcher per the develop flow's state contract) and pasted verbatim: {{constitution}}
3. Triage: the analyst, the space's single writer, disposes every finding: covered / fix / assumption / question / rejected, each with its resolving target ids; the round record holds the full table.
4. Audit the burial paths: one fresh-context spawn re-judges only the covered and rejected dispositions against their cited evidence; disagreements are flagged in the record and surface at the gate.
5. Apply fixes, run the compiler's check and render, close the record locked. Verdict converged means zero blocking findings this round.
6. Rounds: 1 is mandatory per domain; run another only while the last round produced blocking findings; hard cap 3; residue becomes open questions at the gate. Cross-domain lenses run once at space level before the space closes.

Core Rules

• Context isolation is the point: challengers get FILES ONLY (the target subtree read fully; the space overview, glossary and generated registry summary-only), never the authoring conversation. DON'T brief a challenger on what the analysis "means"; the documents must carry it.
• Spawned members are read-only by constitution and by capability; every space write in the loop is the analyst's.
• Severity travels untouched: the challenger's blocking/minor verdict is recorded as returned. A blocking finding forces the next round even when fixed this round; fresh eyes verify fixes.
• Expert output enters the space only as assumption rows (source labeled), open questions, or draft edits awaiting the gate; a proposal becomes a fact only through the owner's ruling.
• DON'T pad rounds: a small domain takes a small panel (the lens bank names the floor); zero findings is a legal, welcome round.

References

• lens-bank: the standard lenses with their question stems and panel sizing. Read when casting a round's panel.
• expert-casting: how to write a grounded expert profile for a topic and what to ask it. Read when a domain needs practitioner knowledge the owner cannot supply.
• triage: disposition semantics, the triage audit, and the round record's shape. Read when findings return from a round or when closing a record.