Total-Recall — Program Status & CTO Deliverables

Live snapshot 2026-07-24 · MAIN coordinator · every number below is cross-model-verified or marked. Companion: verified-measurement dossier at system-overview-34y.pages.dev.

1 · The one thing to understand: schema vs reality

The vision = RESPONSE = SELECT-not-FUSE( Gemma endpoint partial-recall ~60–80% ∪ our layers ) → total recall. Correction (operator was right): the composition is BUILT-BUT-UNWIRED, not unbuilt — fetchBase, runRecallPipeline, ContentRouterSelect, unionRecall all exist in code, but router.search never calls them. Live today ≈ ⅓ of the schema: LIVEcapture + standalone local recall · NOT WIREDendpoint compose · L3 perception · snowflake spine. Next build = wire + turn on (claude-revive's keystone lane), gated on rotating the endpoint token (was cleartext in git).

2 · What's LIVE & LANDED

LIVEStore-fill capture
2,653 records = 200 seed + 2,390 real captures (self-dogfooded from this session). 47+ secret-refusals, 0 stored. Service ok.
LANDED2 engineering fixes
origin/main 816b113fef, full FSM 0-bypass: O(N²) ingest → linear 156×@80k; event-loop HOL 107ms→1.5ms. Not deployed yet.

3 · CTO deliverable — "how do we do recall scoring?"

Perspective (answered)
Today = SELECT-not-FUSE: per candidate take the best plane signal (max), not an average — a strong specialist isn't drowned out; then a PPR-field re-rank. The ceiling: scoring is static & query-blind — same planes every query, fixed weights. The 0→1: (a) query-conditioned SELECT (a router that trusts the right planes per question-type), (b) a credit loop that learns those weights from real outcomes, (c) plane calibration. Can't learn scoring without ground-truth outcomes → needs the filled store (now ✅) + outcome signal. That's why it's staged.

4 · CTO deliverable — extremely challenging self-recall benchmark

BUILT + AUDITED28 cases (v1.1)
Each grounded in our real work (20/20 zero fabrication, adversarially audited + Grok-confirmed). Engineered to attack the axes today's scorer is blind to: exclusion ("what we rejected"), role ("which model caught the bug"), temporal-supersession, honest-status decoy (strongest lexical match is the wrong/synthetic claim), lexical-gap, specificity-under-a-hot-term — +8 harder cases each with a decoy that lexically outranks the gold. Every case has a null-control (shuffle → must collapse to chance). Score against the isolated store, not the 232k unified endpoint (else content-absence masquerades as difficulty).

5 · Benchmarks confirmed (6/6 cross-model 2-approve) + per-layer plan

Machinery: latency bimodal — point-lookup sub-ms & flat to 1M, broad scan O(N)~0.7–1.3s@1M; concurrency knee K≈8, event-loop HOL proven; ingest O(N²) root-caused → 140× lever. Real-data efficacy: perception reproduced (0.594≈curated, NC-clean); PPR-field 0.916 recall@10, NC=0.013 (committed corpus; version caveat). Coverage-Law honest DID-NOT-RUN (corpus missing).

6 · Next & pending