the honest board: what's shipped · what's serving · what's proven vs held — derived from the CTO-triaged register + provenance audit
generated 2026-08-10T08:02:48Z (real UTC, r3 — dual-CTO correction packet applied: decisive test GATED not running · pre-run authorization ≠ result certification · provenance populations unsummed · formal math split from banked lab results) · sources: NOT_DONE register v3.5 (QG-triaged, CTO-bound) · provenance-recert v1+v2 (SHA-pinned) · serving-state map (deployed-verified)
Shipped, wired & gated
SHIPPED · SHADOWSemantic chunk-dense retrieval — the program's first product win
On clean, find_hit-excluded semantic gold: recall@20 0.667 vs BM25 / keyword 0.000 (lexical total failure on semantic queries). Certification reproduced, shipped default-OFF in shadow with telemetry + a product contract, exact-HEAD Grok+Gemini approvals (FSM COMPLETE @ 6e9db110). Next: main integration. Scope honesty: this is generic chunk-dense embedding — not the exotic graph stack.
GATEDThe decisive test (beats-strong-baseline at scale)
The certification keystone — authorized-stage only. Actual experiment RUN has not begun. The research checkpoint name EXPERIMENT_RUNNING is an authorization-stage FSM label, not evidence of an active process or a completed run. The run remains gated on a sealed clean-gold successor, hypothesis approval, and gate-clean execution evidence. Current F0 v1.7 is dual-CTO REQUEST_CHANGES and NOT SEAL — its user-query frame and deterministic pool contract are being repaired, so the decisive run cannot start from v1.7. The preregistration design (frozen qrels custody, out-of-band signing, arm-blind excerpts, CTO-bound 6-leg pre-fire screen) is drafted for when the gate opens. No result yet; nothing is claimed until it lands.
A fail-closed pre-run authorization hook is wired for matched Claude Bash experiment commands, requiring EXPERIMENT_RUNNING plus valid track-scoped approval (exit-2 fail-closed, no silent demotion; applied after 3-CTO signature). Post-run result-artifact and review-closure certification is not yet complete. Scope honesty: this is a pre-tool hook over matched commands — not universal runtime interception and not result provenance. Completion enforcement (result-artifact guard + review-closure) is being built in the governance-repair sprint (SEED hash-locked, dual-CTO).
Serving reality (measured, not aspirational)
HONESTWhat actually serves today
Live path = provider fan-out; Total Recall internal vector+lexical/RRF candidate retrieval; legacy outer scoreResults; B3 capture. ARK precedence and PPR post-rerank exist behind flags but are OFF today (live config: ARK=0, WICK_PIPELINE=0, GRASSMANN=0, B3=1, vector=on). The composed advanced recall pipeline is not wired into serving; advanced-plane serving remains 0%. Flag-present ≠ live; source-present ≠ activated. Comparative observability (per-plane timing / candidate counts / attribution) is absent; the telemetry mount is specced and owned. No subsystem-utility % is claimable until it exists.
Provenance (the hard number)
AUDITEDNo historical result was retro-certified in either audit pass
A read-only, rubric-bound provenance audit ran over two distinct populations — they are not summed: pass-1: 2,201 units — 0 clean / 16 re-run / 907 tainted / 1,278 unverifiable · pass-2: 2,292 conclusion candidates — 0 re-certified. Together they quantify the cost of the gate being dead historically. This does not make the science false — proofs and honest NULLs stand as such — it means retroactive certification fails. Disposition (MAIN + Grok): preserve real work, cite nothing historical as certification evidence, re-run in dependency order under the wired gate. The store-fill loss incident: 71 failed put events (unique-content loss ≤71, unproven), contained + held.
Evidence classes — formal math vs banked lab results
FORMALFormal / mathematical constructions
CRT/chord zero-collision addressing (1.22e15 capacity) · chaos-lattice (Bell-ultrametric, 3/3 audited) · pinball-SFT (golden-mean subshift → Fibonacci tree, 3/3) · dual-recovery (Hamming ⊕ braid-MDS, 3/3) · chaos-security L6. Proof scope is the stated construction under its stated model — a proved property of the described scheme, not an end-to-end system guarantee and not a deployed behavior. Each requires a product contract (threat model / failure domains / concurrency) before integration — proven ≠ needed ≠ shipped.
LAB RESULTBanked lab results and research laws
HDC 0.996 agreement — a banked experimental fidelity measurement, not a theorem. The launch packet marks the relevant HDC numbers synthetic-proven; real-data transfer remains to be measured. · Field-Subsumption Law — a named empirical/architectural research law about where structural primitives did and did not improve ranking; it is not a formal theorem and not a serving capability. Both are not retro-certified, not necessarily real-data transferred, and not serving unless separately verified. They are kept, not withdrawn — only their evidence class and product status are corrected here.
Retired / forbidden claims (do not cite as current): store-fill "enable-safe / fully hardened / zero-loss" · endpoint "partial-recall 60–80% by design" · "grows 60–80% → total" · the 275× capacity figure (phantom: tainted synthetic ratio; lifecycle pinned) · triangulated/VSA/beam/translator "live" · Marvel blend active (shadowed under ARK) · any TR-utility % (no telemetry) · resonance (nulled) · exotic HDC beats dense (falsified — ties at best) · transition-adjacency field lift (density artifact) · F0 "sealed" · the decisive experiment "running" · advanced planes "serving" · historical results "retro-certified" · synthetic HDC transfer "proven".