CONFIDENTIAL · PATENT PENDING (J. Weinberg / N. Baron) · wick-total-recall Track-1 · compound research update

WICK Capture — Compound Status

2026-07-17 · lane: wick-capture · PM: claude-coord · builds on the prior capture status

Headline: the capture transport is LIVE on production main (dormant, verified no-op), and the security unit that hardens the evidence substrate underneath it is in flight. The battle the operator named — "won or lost in Part-2 = capture" — has its first piece on prod.

822d844etransport on origin/main

14/14transport tests · ~98% cov

4/4·4/4·4/4DESIGN · CODE · AUDIT converged

0unauthorized bypass / fabrication

Where we are, in one paragraph

Sprint 1 — WICK_CONTINUOUS_INGEST (the source-agnostic capture transport: usage → parse-firewall → quarantine only, never canonical) — LANDED + CLOSED on origin/main 822d844e, coord-verified, default-OFF. Sprint 2 — WICK_EVIDENCE_INTEGRITY_HARDENING_v1 (hardens signed-evidence + audit-log + locks) — at PLAN; SEED converged 3/3, and the PLAN panel just caught a real security design flaw before any code. Enable on the transport stays OFF behind a triple gate.

What happened since the last update

Transport hardened through the AUDIT gate

The write-boundary red-team caught 3 real defects (forgeable brand, path-traversal, runtime-mutable ledger) → all fixed + guard-tested.

Verify-vs-disk caught a frozen-base recipe bug (coord-owned)

The base had been captured at the branch tip, not the merge-base — require_new_in_diff would have walled the walk. Fixed by a clean re-init at the corrected base under a fresh identity (v1a), no hand-edit of the immutable event.

Full v2 FSM re-driven to COMPLETE

DESIGN 4/4 (6 rounds), CODE_REVIEW 4/4, AUDIT 4/4 (3-round red-team), RETRO 2/2. Every panel finding a real defect, each fixed with a guard test. Grok forced-articulation every round; rubber-stamps byte-rejected + re-fired.

Landed to production main + coord binding second-verify

Rebased onto current origin/main, pushed through the full hook chain, no demotion — stopped at a stale-review hook flag and re-fired a clean Grok+Codex+Gemini panel on the landed SHA rather than override. Byte-identical unit + default-OFF no-op verified on prod.

Evidence-integrity finding surfaced (disclose-first)

A gemma receipt carried a placeholder dispatcher label for content from the remote endpoint — genuine content, FALSE provenance metadata. Coord elevated it to a real defect and credited the disclosure. Now scope-item (e) of Sprint 2.

Sprint 2 reclaimed + driven to PLAN

Pre-SEED consult refined scope (split d→d1+d2, e→e1, ADD f = sprint_id binding). SEED 3/3 approve (Grok 1302b forced-articulation). Then the PLAN panel fired.

PLAN panel caught a security design flaw (this is the pipeline working)

3/3 request_changes: the schema-migration is under-specified and has an NC1-vs-NC3 conflict — accepting legacy receipts (no sprint_id) to stay backward-compatible would let an attacker replay one in a new sprint, bypassing the very guard being added. Caught before a line of code.

Sprint 2 — the 7 guards (canonical-source-verified scope)

Guard	Gap (verified on disk)	Status
a symlink / O_NOFOLLOW	evidence path can symlink out of its namespace	planned
b max_age clamp + key-provenance	oversized max_age replays stale evidence; rotated key not rejected on load	planned
c events.jsonl injection	newline/control/JSON bytes forge audit records	planned
d1 PID-reuse identity	reused PID passes `kill -0` as the original owner	planned
d2 safe-cleanup throw	unsafe cleanup proceeds silently	planned
e asserted provenance my finding	dispatcher/endpoint_url SIGNED but caller-asserted → false-provenance	planned
f sprint_id binding biggest gap	signed payload binds sha/scope/machine/transition but NOT sprint_id → cross-sprint replay	planned

Coord-sharpened precision: overwatchd.ts:210-227 already binds target_sha/scope/machine/transition — cross-node/scope replay is already closed; (f) adds ONLY sprint_id, no duplication. Binding acceptance bar (operator): a production-active test per guard at COMPLETE.

Operating directives locked this cycle

Endpoint = capture, dogfooded

Recall endpoint (34.70.203.14) is PRIMARY for the Gemma role — to dogfood our own total recall — with local gemma as backup. Its inference traffic is the capture source: distill/pull traces → transport → fill the well. Operator-verified this session.

Full roster, nothing benched

Frontier (Codex/Grok/Gemini) drive 0→1 discovery; gemma also contributes; Fable and myself are first-class voices, not just dispatchers. Applied from the SEED panel onward.

Provenance honesty

No placeholder dispatcher labels — the false-provenance lesson is now the fix (guard e). Signatures never vouch for caller-asserted metadata.

Outstanding / roadmap

Item	Owner	Status
Sprint 2 — revise plan (schema-migration + key-provenance design resolving NC1-vs-NC3) → re-fire PLAN → DESIGN → BUILD(7 guards) → AUDIT(red_team) → land	wick-capture	in flight (at PLAN)
Transport ENABLE — flip on live capture	gated	held: needs P2 + admission unit + operator SOURCE decision
Capture-source build — endpoint-trace SourceAdapter ("fill the well")	wick-capture	next capture piece; build ok, live-enable gated
TRIANGULATION_DELTA_ADMISSION_v1 (P3) — the ONLY canonical route	pool	queued
StructuralAddressIndex (SAI) — snowflake-spine addressing	wick-capture	granted, sequenced behind Sprint 2
Discovery arc (chord/CRT/E8/Merkle-snowflake) — frontier lane	revive-frontier + coord	banking wins (gate #2, E8 re-open, CRT p=5)

Integrity note: this cycle the FSM held under three process findings + a recipe bug + a false-provenance catch, and none reached prod as a defect. The takeaway being banked — disclose-first turns every latent defect into a caught event — is why security-critical code can land autonomously. Enable stays off; the transport ships dormant. · State + CXDB snapshot are patent-safe; full resume in project_2026_07_17_wick_capture_resume_checkpoint.