WICK · verifiable memory that wraps Modulum · partial recall → total recall
WICK Total Recall — the honest system map
A verifiable-memory system that wraps the live MODULUM partial-recall API in a closed learning loop to drive imperfect recall toward total recall. This map is deliberately honest about what is merged & tested but dormant, what is research-validated in python, and what is only designed or NULL — because the gap between "shipped" and "live" is the whole point.
verified against origin/main @ f707b8e5 (NOT the local worktree — 230 commits behind, would not build) · 3 code-grounded verification agents + 2 prod suites ran live (codec 34/34 · provenance 26/26) panel-verified: Grok approved the facts · Codex + Gemini refined the framing · maintained by claude-coord
CONFIDENTIAL — WICK IP · PATENT PENDING · internal only, noindex — do not distribute
★ North Star — wrap Modulum, close the loop, reach total recall
MODULUM is a live, hosted partial-recall substrate (the "Gemma / Granite" Librarian, GSV BiblioTech — board → recall) that is honestly imperfect: ~60–70%, and it confabulates at scale. WICK wraps it in an unbounded verifiable STORE plus a closed learning loop: retrieve the relevant slice, reason over it, act on Modulum, capture the outcome, bind utility → seal → write back with provenance. Every cycle makes the next retrieval sharper. Not a bigger model — a bigger, verifiable, ever-growing memory the substrate reasons through, pushing partial recall toward total recall.
Structural moat = the unbounded verifiable STORE + the learning loop. The store never loses data and proves it; the loop is what turns a 60–70% substrate into total recall.
✗ NOT Hypernym Cloud · ✗ nothing is LIVE in serving yet · ✗ the Modulum client itself is NOT in the product (python research only)
READ THIS FIRSTThree honest tiers — dormant ≠ live, designed ≠ built
Every component on this map is colored by real maturity, not aspiration. "Merged & tested" is genuinely shipped and IP-real — but it is not serving users. Keep the tiers separate or the picture lies.
Merged · Tested · Dormant
Shipped to prod TS, test suites pass, but default-OFF and not serving. No gold ingested, wiring deferred. Real for IP/patent + ready to activate — not currently live.
Research-validated
Validated in python research code, NOT in product TS. None killed. Some tail sub-primitives shipped separately; the assembled component is research-only.
Designed / NULL
A design, a synthetic PoC, a proven-but-unwired primitive, or a verified NULL on real data. Honest frontier: not yet a product faculty.
THE CLOSED LOOPEight faculties: memory → cognition → act on Modulum → learn
Seven faculties form the running loop; DESIGN sits above as the meta-faculty (RL auto-improver + FSM governance). MEMORY is the hub every faculty reads and writes. The loop only actually learns once EXECUTOR runs against live Modulum and produces outcomes.
⌖ META · DESIGNmeta-faculty — the system proposing its own next moves: RL auto-improver (retro-cluster → retro-propose → policy) + FSM governance. Designs the loop below; does not run inside it.
▼
1 · HUB
MEMORY
verifiable store & recall — the substrate
→
2 · REASON
THINKING
deliberate over the recalled slice
→
3 · PRIOR
INTUITION
belief field biases what's worth recall
→
4 · GROUND
PERCEPTION
lift grounded candidates into the fiber
→
5 · CHOOSE
DECISION
admit / abstain / steer toward & away
→
THE TIP
6 · ACT
EXECUTOR
act on Modulum, capture the outcome
→
7 · LEARN
CREDIT
utility-bind → seal
⚠EXECUTOR acts on the external MODULUM partial-recall API — hosted, ~60–70% accurate, confabulates at scale (34.172.101.117 · board→recall). Wiring it in is a BUILD, not a flag: today there is ZERO Modulum client in product TS.
↺CREDIT seals back into MEMORY → new edges update INTUITION's community field; the next retrieval is sharper. The loop closes only when real outcomes flow. (Faculty numbering follows the MEMORY-first spine: MEMORY = 1, cognition 2–7, DESIGN = meta.)
The most-built faculty by far — the full L0–L6 substrate is in product TS with passing suites. But every layer is default-OFF, no gold is ingested, and nothing serves users. "Shipped + tested" ✓ · "live in production" ✗.
The verifiable store the whole mind reads and writes. Recall = one exact, tamper-evident answer. Everything below is real prod TS with passing tests — and all of it is behind default-OFF flags, not serving. The addressing row and the security lattice are the only places where research and prod diverge (marked amber).
DURABILITYBraid-MDS Cauchy-RS codec + SECDED dual-recovery — "never lose data"; the research-era "holographic erasure code," prod symbol wick-braid-mds-secded-v1 · tr/wick-durability-codec.ts · ran 34/34dormant
ENABLE MACHINEPPR-field re-ranker (master lever 0.861→0.924) + memory-gold gate + GOLD⊥GRAPH fail-closed guard + qrel harness + content-resolver — all prod & tested, NO gold ingested → not servingdormant
DUR-1 SCRUBDurability scrub daemon — recomputes identity-binding from bytes; prod logic (tested 31) but NOT production-wireddormant
COGNITIONFaculties 2–7 — mostly designed or NULL, honestly
The cognitive top-half. A few primitives are proven; most is designed, a synthetic PoC, or a verified NULL. This is the real frontier — do not overstate it.
2THINKINGfloor proven · NULL on real
slow deliberation over the recalled slice
Reason over the retrieved + belief-weighted content. The research floor is proven (snowflake recall 5/5 + VSA binding fair-test), but the headline claim did not survive real data.
✓ floor proven in research. ✗ composition-beats-selection = NULL on real data — selection is irreducible. Not product-shipped.
3INTUITIONprototype · blocked on B3
belief PRIOR · biases what's worth recall
A learned belief field that biases retrieval toward the spaces that matter. Its community field = degree-damped PPR (FORGE_WICK_RECALL_PIPELINE) + Fisher-Rao belief-prior (FORGE_PLANE_L3). Implemented-prototype / flagged-plane — NOT yet a validated product faculty.
◐ warm-validated only (+0.040 CI-clean on warm co-access). ○ the temporal / outcome mechanism is BLOCKED on trajectory data (B3).
4PERCEPTIONproven but unwired
grounds & lifts candidates into the fiber
Anti-hallucination grounding + lifting the right memory into a fiber for reasoning. The grounding invariants are exhaustively proven; the geometry above them is a hypothesis.
✓ grounding invariants I1–I4 PROVEN (exhaustive; I2 conditional on SHA-256) but UNWIRED; fiber-lift lab-only. ○ T*M symplectic geometry = DESIGN HYPOTHESIS, not validated.
5DECISIONsynthetic PoC
admit / abstain / steer toward & away
Commits what to retrieve / admit / enable, and what to steer toward vs away. Exists only as a toy today.
○ DESIGNED + synthetic PoC only — calibrated abstention NC-clean, but toy data. Needs CREDIT policy + grounded PERCEPTION candidates to be real.
6EXECUTORTHE TIP · capture-half only
acts on Modulum · captures the outcome
Acts on the live Modulum substrate and captures the outcome — the step that closes the loop and makes intelligence compound. The true tip of the critical path: nothing above MEMORY genuinely learns until this runs.
◐ only the CAPTURE half is built — tr/eval/live-trajectory-sink.ts, emit-only. ✗ NO Modulum-client in prod — a BUILD, not a flag.
7CREDIT#1 gap · loses offline
utility-bind → seal · assigns credit (RL)
Binds utility to what worked and seals it back into MEMORY. The reward infrastructure is shipped — but with no real outcome signal, the learned router is worse than the heuristic.
✓ reward infra SHIPPED (forge-core aggregator v2a + policy). ✗ offline, the learned-router LOSES to heuristic (0.9075 vs 0.929) — no real outcome signal. #1 keystone gap.
MDESIGNmeta · RL infra landed
meta-faculty · designs the system itself
The self-designing faculty: the system proposing its own next improvements — RL auto-improver (retro-cluster → retro-propose → policy) plus FSM governance. Sits above the loop, not inside it.
✓ RL policy + retro-cluster / retro-propose daemons LANDED. Governs how the loop below evolves once real outcomes exist.
TWO NUANCESWhat the panel demanded we not bury
Two clarifications the review panel insisted on making explicit — collapsing either one produces a false picture.
① HDC/VSA substrate ≠ the community field
HDC / VSA is the fast-binding representation substrate — hypervectors, role-binding, SimHash bit-kernels; the 256-bit popcount kernel is a SPEED win (HDC-as-a-fidelity-lever was a verified NULL). The community / belief FIELD that occupies the INTUITION plane is a different object: degree-damped PPR + Fisher-Rao belief-prior (graph-propagation + information-geometry). The operator-facing "fused object" = the field computed over the HDC bit-substrate — but the field and the substrate are separable, not one thing. It sits at L4-recall / L5-composition and IS the C2 Intuition plane.
② B3 gates learning — but not everything
Everything above static MEMORY that claims to LEARN or ADAPT from real utility gates on B3 living data (INTUITION's temporal/outcome mechanism, CREDIT's real reward). But non-learning work does NOT wait for B3: productizing the Modulum client, wiring the PROVEN grounding stack, enabling the dormant recall planes, qrel/gold ingestion, DUR-1 prod-wiring, and basic serving integration can all proceed in parallel. "Gated on B3" is not "gated on everything."
THE BOARDShipped-dormant · research-validated · designed / NULL
The three-tier read, laid out flat. Left is real prod TS (but inert); middle is validated python that never entered the product; right is the honest design/NULL frontier.
Merged · tested · DORMANT prod TS · default-OFF
Durability codecbraid-MDS Cauchy-RS + SECDED · ran 34/34 · not wired
Provenance spinesnowflake-Merkle + NTT-chord + ultrametric · ran 26/26 · TR_WICK_PROVENANCE off
The enable machinePPR re-ranker + gold-gate + GOLD⊥GRAPH guard + qrel + content-resolver · no gold ingested
DUR-1 scrub daemonprod logic (tested 31) · not production-wired
Research-validated python · not in prod TS
Chaos-security latticeredteam-validated but "low-novelty"; NOT killed. Tail primitives (ultrametric, Merkle-weave) shipped separately — the chaos front-end + composition are research-only
VSA multi-slot binding + E8 cleanupresearch-only (E8 = a code-comment reference in prod)
Chaos-lattice + pinball-SFT kernels2 of the "6" addressing kernels — never ported to TS
Unified 3-tier grounding gate + bloomseal/cert primitives shipped separately; the assembled gate + bloom neg-gate = research/formal-spec only
The MODULUM CLIENTwick_modulum_client.py — the actual substrate connection. ZERO Modulum code in product TS.
Designed / NULL frontier
THINKING compositionfloor proven; composition-beats-selection NULL on real data
INTUITION temporal/outcomewarm co-access only (+0.040); blocked on B3 trajectory data
DECISIONsynthetic PoC only (calibrated abstention, toy)
EXECUTORonly the capture-half built; Modulum wiring = the tip, a build
CREDIT learned-routerloses to heuristic offline (0.9075 vs 0.929) — #1 gap
CRITICAL PATHRoadmap in dependency order — the tip, then the loop
All three verification agents converge here. EXECUTOR → Modulum wiring is the true tip — a build, not a flag. The serving-readiness gate must clear before loop-closure; "data collected → loop closes" is NOT automatic.
1
Productize the MODULUM CLIENT + EXECUTORTHE TIP · BUILD
Currently python-only. Connect the product to the live partial-recall substrate. Nothing above MEMORY genuinely learns until this runs.
2
Turn the built live-trajectory-sink loose on real usage
The capture-half already exists (emit-only). Real usage → B3 living data: timestamped find/fetch/expand + OUTCOME (~1000+ recurring events).
3
B3 unblocks INTUITION + CREDIT
Enable the HDC-community field on real outcomes (INTUITION); real reward lets the credit-learned router replace the heuristic (CREDIT).
4
Wire the PROVEN grounding stack into live runnersPARALLEL · no B3
PERCEPTION grounding I1–I4 is already proven — just unwired. Parallelizable; does not wait on B3.
5
CREDIT policy + PERCEPTION candidates → DECISION
Grounded candidates + learned reward finally make the DECISION faculty real (graduates it out of synthetic PoC).
6
SERVING-READINESS GATEGATE · before loop-closure
Before activation: EXECUTOR circuit-breaker + fallback for Modulum degradation / confabulation (a volatile ~60–70% external API — define fail-open/closed), flag-rollout plan, observability, gold/qrel policy, and rollback criteria.
7
Close the loop
utility-bind → seal → MEMORY → new edges update the INTUITION field. The loop closes; the store compounds.
8
Enable the dormant recall planes + PPR re-ranker on real gold
Needs qrel / gold ingestion — a separate operator decision. Flips the default-OFF planes against living-data gold.
DIRECT ANSWERSTwo questions, answered without spin
Q — Chaos-security lattice: is it a real prod component?
Research-validated, NOT a unified prod component. It is redteam-validated python, judged "low-novelty" (elementary chaos + standard LCP ultrametric + reused Merkle) — but not killed. Its tail sub-primitives (ultrametric, Merkle-weave) ARE in prod separately; the chaos front-end + composition are research-only. Short answer: sub-primitives shipped; the assembled lattice is research.
Q — What is "hyper-dimensional community computing"?
The HDC substrate × the community field = the C2 Intuition plane.HDC/VSA supplies the fast-binding hypervector representation (speed, not fidelity); the community/belief field (degree-damped PPR + Fisher-Rao) is what computes over it and occupies the INTUITION plane. The fused object is the field-over-substrate — but remember nuance ①: they are separable, not one object.
The honest through-line: WICK's moat is real and already built — an unbounded verifiable STORE that never loses data and proves it. But the store is dormant, the Modulum client that would feed it is not in the product, and every faculty above MEMORY is designed, synthetic, or NULL until real outcomes flow. The single build that unlocks everything is EXECUTOR → Modulum wiring — the tip. Nothing here is killed; kills aren't permanent if a real regime is found (NO-DISCARD).