Synthesize

Turn a pile of evidence items into a small set of problem-framed themes, with trend and corroboration analysis, explained confidence, and specific gaps. This is the place where raw evidence becomes buildable problem candidates; do not make users wait for a separate "themes, then problems" step.

When to run

After evidence gathering. Read .product-eval/<scope>/evidence/* (already identity-deduped and weighted).

Preflight (data contract). Before clustering, run the conformance preflight in DATA-CONTRACT.md (Validation) on evidence/*. If an item is malformed (missing strength or weight, an unknown source_type, a dangling corroborated_by), stop and report the exact file and field rather than theming around it.

Workflow

1. Cluster into problem-themes

Group evidence by problem, not topic: same pain + same persona + same context, ideally from 2+ source types. Method in references/synthesis.md: for modest corpora use LLM map-reduce (extract a normalized claim per item, cluster the claims, label each as a problem, not a solution); for large corpora pre-aggregate or embed-and-cluster, then label. Typical result is 2-5 problem candidates; more than 7 means splitting too finely; only 1 means evidence too narrow.

Then frame each at the right altitude: the problem-statement formula [persona] cannot [job] [context] because [root cause], resulting in [business consequence], splitting too-broad themes (by persona / journey step / severity / frequency) and lifting too-narrow ones to the underlying problem. In early exploration, return the best-current framing with gaps; for evaluate/validate work, run the crispness grader (critique) as a gate and rewrite any dimension scoring 0. See references/problem-framing.md.

2. Detect trends and patterns

Per theme: volume over time (WoW/MoM) and the signal-count bands (3-5 emerging, 6-12 established, 13+ well-known, 50+ fundamental or recent regression). A sudden spike usually signals a recent regression. Record cross-source corroboration, including internal↔external triangulation, and surface contradictions (e.g. analytics show 78% completion while a vocal cluster complains) rather than smoothing them.

3. Reason about confidence

For each theme, compute and explain confidence from total evidence weight, source diversity, recency, and contradictions, never just a number ("Confidence 72: 12 items across 3 source types, recent, but no first-party quantitative data").

4. Write gaps

Frame specific gap questions, each with where to look and which source would close it.

5. Write and hand off

Write .product-eval/<scope>/themes.md: per theme, name (problem-framed), member evidence IDs, affected personas, distinct accounts, average strength, total weight, trend, corroboration, contradictions, exact confidence plus confidence band, and gaps, following the theme schema in the data contract (DATA-CONTRACT.md). Hand the framed problem candidates directly to prioritize-backlog for scoring.

Output

A scannable problem-candidate list (problem statement, trend, confidence band, top gap) plus the written themes. Lead with the strongest candidate. Do not make build recommendations here, that is the scoring step's job. End with Next move: and recommend ranking the candidates, or running readiness on the strongest single bet if the user is already trying to commit.