feedback-pattern-minerlisted

# Feedback Pattern Miner A feedback dump in, ranked themes out — and the books must balance: every input item accounted for, exactly once. ## Verification gates (defined first; output is blocked until all pass) - **G1 — Reconciliation:** sum of theme counts (including `unclassified`) equals the input item total. The output must show the arithmetic. Off by one = gate failure. - **G2 — One primary theme per item:** each item is counted in exactly one theme. Overlaps become secondary tags outside the counts. - **G3 — No invented evidence:** every cited item ID exists in the input; every quoted fragment is a verbatim substring of its item. ## Steps 1. **Number the input.** Assign IDs (F1, F2, …) in given order and state the total N. If the input isn't itemizable (prose, transcript), stop and route: interview material belongs to interview-synthesizer. 2. **Cluster.** Group items by the underlying problem, not surface wording ("sync breaks" and "events dropped" are one theme). Name each theme by the user problem, not the feature request phrasing. 3. **Force nothing.** Items that fit no theme go to `unclassified` — a counted bucket, not a trash can. Force-fitting noise into a theme to look tidy is a gate failure. 4. **Count and rank.** Primary assignment only. Rank themes by count, descending; ties broken by severity of language in the items, stated as a judgment. 5. **Gate pass.** Verify: counts sum to N (G1), no item in two counts (G2), every ID and quote real (G3). Fix and r