| name | reviewing-actuarial-ux-syntheses |
| description | Use when comparing, reviewing, reconciling, adjudicating, or finalizing multiple completed actuarial UX synthesis passes, especially Prediction Lab research syntheses, before using findings for product or design decisions. |
Reviewing actuarial UX syntheses
Core principle
This is a terminal evidence audit, not a vote among summaries. Agreement across synthesis passes measures interpretive stability; evidence strength still comes from unique raw reports, source fingerprints, source classes, directness, relevance, recency, and credible challenge.
Input boundary
- Read
references/meta-synthesis-protocol.md and references/final-output-template.md completely.
- Use a caller-supplied campaign directory when present. Otherwise resolve the user's Desktop and use
RESEARCH-LAB/prediction-lab-actuarial-ux/. Request narrowly scoped filesystem approval if that existing folder is sandbox-protected; use repository fallback research/actuarial-ux/prediction-lab-actuarial-ux/ only if approval is denied or Desktop remains unavailable.
- Discover completed layer-two pass directories under
synthesis/. Recognize artifact roles from content and structure, not exact filenames alone. Freeze and hash the input manifest before analysis. Never add a pass that appears after the freeze.
- Exclude
final/, previous layer-three outputs, partial output directories, and raw runs/ reports from the primary input set. Do not browse the web or introduce new external evidence. If only one valid synthesis exists, perform a pass-quality audit but make no cross-pass convergence claim.
- Do not read a previous final before the current review is frozen. Read one only when the user explicitly requests a legacy delta audit; it contributes zero evidence and imports nothing without raw lineage.
Review workflow
- Inventory every pass and artifact. Assign chronological aliases
P01, P02, and so on; record path, hashes, model, date, input-report hashes, blindness, contamination, and missing artifacts.
- Assess each pass using the protocol's separate quality dimensions. Retain incomplete and contaminated passes, but qualify their use. A pass that read another synthesis contributes zero independent interpretive convergence. An untraceable claim cannot affect final confidence until lineage is repaired.
- Namespace every local identifier:
P01/CAN-006, P02/MS-014, P01/OUT-001. Local IDs and matching prose are not global identity.
- Normalize atomic propositions and underlying sources. Merge only equivalent claims. Preserve
supports, extends, narrows, qualifies, contradicts, alternative cause, and distinct segment relationships.
- Build the reversible chain: final claim → namespaced synthesis claims → raw finding IDs → run-local source IDs → global source fingerprints. Inspect raw reports only to repair or verify missing lineage, material contradictions, or high-impact claims; do not redo layer one or two.
- Recompute evidence confidence independently of pass agreement. Report interpretive convergence, raw-report independence, source independence, source-class diversity, segment coverage, and challenges as separate dimensions. Never average local confidence labels or use majority rule.
- Produce every artifact in the final-output template. Every accepted, rejected, split, merged, contradicted, or quarantined synthesis claim must be accounted for.
- Capture completion time in UTC and save under
final/YYYY-MM-DDTHH-mm-ssZ--<review-id>--<model>/. Normalize identifiers to lowercase kebab-case; append --02, then --03 on collision. Never overwrite or leave the review only in chat. Return the exact absolute output paths and any persistence limitation.
Completion gate
- Every frozen synthesis pass and recognized claim is inventoried.
- Every final factual claim traces to raw evidence or is explicitly quarantined.
- Pass convergence is never presented as source corroboration.
- Contamination, missing lineage, contradictions, minority claims, and negative evidence remain visible.
- Counts reconcile and every dedupe decision is reversible.
- Product/design implications distinguish established requirements, contested interpretations, and hypotheses.
- No web evidence, previous final, or post-freeze synthesis silently entered the corpus.
Anti-patterns
- “Five of six syntheses agree, therefore confidence is strong”
- Treating repeated raw reports or source families as new evidence
- Matching local
CAN-* or MS-* IDs across passes by number
- Silently discarding incomplete, contaminated, minority, or untraceable material
- Reading previous finals before freezing an independent review
- Re-running the original research instead of auditing synthesis behavior
- Producing one polished summary without lineage, accounting, and disagreement artifacts