## spt-core requirement confusion audit — paused at weekly cap, resume later **Run dir (everything lives here, nothing in the repo):** `C:/Users/decid/AppData/Local/Temp/claude/C--Users-decid-Documents-projects-liaison/29287739-71ae-4df9-a62d-b2c76b202b34/scratchpad/rca` **READ FIRST on resume, in this order:** `RESUME.md` (procedure + every ruling on record), `STATE.md` (scoring coverage), `STATE_GATE4.md` (consolidation state + an open concern that must be read before dispatching more S1). **Mode: AUDIT ONLY, read-only.** Phases A–E then STOP. No edits, commits, or branch changes in spt-core. Applying a fix needs explicit operator authorization; committing is a separate one after that. ### Where it stands - **Scoring is DONE: 177,906 / 177,906 pairs = 100%, zero unscored.** All 1,250 chunks passed the gates. Matrix written, 597x597, family-ordered. - **Gate 4 consolidation is ~11% done.** 88 slices total, 10 framed, 78 remaining (63 never dispatched, 15 dispatched-and-held). - Findings banked so far: 357/597 requirements have no doc-stage evidence (scored on id+title alone); titles are mini-specs (mean 960 chars, max 20,164, capped at 1200 for scoring, 145 affected); 8,728 candidates >=30; ~225 pairs look genuinely overlapping. One early cluster to chase in Phase D: stale-latched-status pairs (CONTROL-STAMP-CONVERGENCE with LIVEHOST-BOOT-LIVENESS-GATE at 45, and with BOUNDARY-READY-STRAND at 44 from an independent scorer). ### Immediate next steps 1. Read `STATE_GATE4.md` OPEN CONCERN first: 60% of returned slices failed the calibration gate on their merits. Do not loosen the gate; measure why. Cheapest checks listed there (slice size, whether re-dispatch passes at a better rate, content pattern). Bring doyle the measurement before touching any threshold. 2. Finish S1 over the remaining 78 slices, in tranches sized to the window. 3. Then S2 (everything >=40, ordered DESCENDING), then S3 — a single agent over the top band. Only S3 output is citable as a ranking. 4. Then Phase D diagnosis against the implementation, Phase E exact wording proposals, then STOP and report. ### Rulings in force (all recorded in RESUME.md — do not relitigate) - S1 entry threshold >=30, deliberately not >=40. - S2 ceiling 3,000 HARD; report the ACTUAL count, not a projection; if it exceeds 3,000, checkpoint doyle before dispatching S2. - Concordance gate >=0.75, gating BEFORE slope; deadband FROZEN at 10 and explicitly not a tunable knob; always report raw AND deadbanded concordance. - Ballast (20 known-low pairs/slice, drawn fresh per slice) doubles as a gate input. - The 6 chunk-stage reference pairs keep their values verbatim in every slice; a reference needing re-judgment is a FINDING, not a silent update. Asserted in code. - Hold reasons stay in separate buckets; a held slice re-dispatches ONCE, then escalates to doyle. Never average two bad frames into one acceptable-looking one. - Every cell carries attribution: consolidated / provisional / coordinator-scored / NS. ### Operational facts worth not rediscovering - Harvest is idempotent and safe on a live or already-harvested run; a session death is a pause, not a loss. Always check a run's journal before assuming data is gone — twice the harness reported no completion record while the journal held full results. - Size a dispatch to the REMAINING window, never to the work left. Overshooting burns four retry rounds per dead agent; measured cost was roughly half a window wasted. - Never hand-edit generated state or the is/ns vectors; regenerate them. - Parked, operator's call only: dropping the 49 placeholder requirements would cut ~16% of the work. Operator chose the full roster at the start, so it stands.