how it stacks*
Eight layers per V2.1 taxonomy. Click any product card to see its full composition in the inspector. Toggle a persona to dim non-touched products. Pick a value stream to pulse-walk it.
module composition*
Module-centric view. 7 modules, each with its components, skills called, and products that instantiate it. Click a product chip to jump back to Hierarchy with that product selected.
what interacts with what*
EDR-style graph. Every entity as a node, every relationship as an edge. Click a node to filter to its connections; the inspector shows direction + edge type.
instantiates + contains + reads-writes. Edges curve rightward for external-call (to connectors), leftward for uses-skill. Click any node to highlight only its connections. TBD nodes have dashed amber borders.
data schema*
Six conceptual ERDs + two structural diagrams. Module instance pattern is the central one. Full schema-design questions at Sprint 9 §11 (12 open Werner gates).
IA + UX review*
10 review sections. Per-layer findings plus cross-cutting concerns. Findings are actionable recommendations — the IA failures we'd hit if we shipped the V2.1 product without intentional surface design.
engineering layer*
Four GitHub repos. One customer machine. DATA feeds INTEL, INTEL drives COMMS, COMMS reaches the customer. The PRODUCT surface ships 1 Oct 2026. Assessed 2026-05-27; topology remains current.
The four repos cleanly sort into four layers. Customer signals flow up through the stack; documents and outbound flow back down.
Lead pool, batch management, DataSoap quality flags, Aircall sync + reconciler, daily/source/agent reporting. Source of truth for who can be called and when. Operator-gated writes — no automated Aircall pushes.
Transcript-to-recommendation engine, content bank (17 documents), generation + 19-check QC, gap analysis, belief-state system, multi-channel campaign intelligence. The most feature-dense repo in the estate.
WhatsApp API gateway. Multi-session, webhook-driven, HMAC signing, REST API, n8n integration. Production-deployable. Integration code connecting OpenWA to LPM + Content-Unlocker is not yet wired.
The investor-facing subscription surface. Decision support for HNW UK individuals. Currently a README + skeleton directory — no application code. The strategically loaded gap: everything else in this estate is upstream of this slot.
Content-Unlocker → OpenWA WhatsApp send → investor
OpenWA inbound → Content-Unlocker belief-state update → next recommendation
Investor signals → Unlock-Alpha-Desktop dashboard (the missing piece)
Six engineering principles these repos do well. Transferrable to the OS and any new build.
LPM's Aircall reconciler runs only on explicit operator click — no schedulers, no auto-kicks. "Operators want per-batch consent before any rate-limit budget is spent." Any OS automation touching external APIs should inherit this gate.
LPM's README spells out which DDL changes require manual SQL Editor application. Each migration has a number, a one-line description, and an idempotent re-run note. The OS should adopt this style for vault-canon changes.
Content-Unlocker gates document recommendations on tracked investor beliefs. The concept extends to any recommendation engine: gate outputs on a tracked state, not just a profile.
Content-Unlocker's Gap Analysis identifies content holes before they cause a failed recommendation. Same shape applies to any content or skill estate: a missing-matrix surfaces gaps automatically before they become failures.
OpenWA swaps database, storage, and cache via config — SQLite/PostgreSQL, Local/S3, Memory/Redis. Every OS engine should declare its dependencies in a way that allows substitution without code change.
Content-Unlocker scans compliance-critical content atoms for contradictions. The vault equivalent: canon rules + citations + an audit job that flags drift. The daily-system-audit is a rougher version; ACU is the mature shape.
Assessed 2026-05-27. Not a fail — not yet a product. The set is more valuable assembled than as parts. The next four months of work is the assembly.
Source: Intelligence/research/2026-05-27-karpathy-github-projects-review-V1.md · Full Karpathy review at that path.