Most learning content is produced by humans, slowly, and can't prove it worked. This is the machine that does it differently.
One cubelet. Six faces. One quality standard. Before we explain how the pipeline works, we need to establish what it produces.
WHAT — The concept, defined precisely. What a practitioner must know to act on it.
WHY — The business or compliance driver. Why this matters in context.
HOW — The mechanism. How a practitioner implements or evaluates this.
WHERE — The application context. Where in a system, workflow, or regulation this applies.
WHEN — The trigger or timing. When this knowledge is activated in practice.
APPLY — The scenario. A practitioner-facing situation that tests judgment, not recall.
Every face must score 6 or higher. The aggregate must reach 42 out of 60. No face may score zero. The QA gate runs on Opus — enforced in code, not config.
The system doesn't guess what to teach. It reads the market. Assay harvests signals from regulatory documents, job postings, and practitioner discourse.
Each signal is citation-verified before it enters the substrate. Unverified signals are rejected — they never reach the Foundry.
Output: a MarketSignal with source, domain, verified_at, and relevance_score — the minimum facts the Foundry needs to structure knowledge.
Verified signals become structured substrate. The Foundry assembles four artifacts before any cubelet can be produced.
TAXONOMY.json — The domain tree. Every competency, its grain, its ID scheme.
ONTOLOGY.ttl — Semantic relationships between concepts across domains.
Vocabulary — Preferred labels, alternate labels, definitions, source citations.
Competency graph — Domain-specific ID patterns: govern-1.1, euact-art9, comp-ac-1.1. The ID is the competency.
All four artifacts write to R2 (Cloudflare object storage) — the primary store. Neo4j is a derived index rebuilt from R2. R2 can survive a Neo4j rebuild; the reverse is not true.
Five agents. Three model tiers. One gate. The pipeline converts substrate into quality-gated cubelets.
Research (Haiku) — Retrieves grounding context from the substrate. Fast and cheap at scale.
Writer (Sonnet) — Drafts all six faces against the grounding context.
Editor (Sonnet) — Refines tone, accuracy, and specificity. Enforces face-level standards.
QA (Opus) — Scores each face. If the aggregate falls below 42 or any face scores below 6, the cubelet is rejected and re-enters the pipeline at Writer — this time with a Sonnet upgrade.
Publisher (Haiku) — Writes the cubelet to R2 with a full quality_gate block: { passed: true, aggregate: N, scores: {WHAT, WHY, HOW, WHERE, WHEN, APPLY}, reviewed_by: "qa-agent-opus" }
Qualified cubelets reach learners across five apps. Each app is a different consumption mode. Passport handles cross-app SSO.
cubelet.ai (lms-app) — Primary learner interface. Rooms, canvas, simulation, and AI-guided practice.
design-app — Cubelet authoring and layout tooling for instructional designers.
scaffold — Pattern-judgment diagnostic. Measures what a practitioner can't yet do and fires gap signals back to the Factory.
tracker-app — Progress and performance tracking across learners and cohorts.
passport — Auth and entitlements. A session set by one app is readable by all others — the leading-dot cookie domain that makes cross-app SSO work.
The scaffold measures what learners can't do. The factory fills the gap. Mastery feeds back. The loop compounds over time.
Scaffold side: The diagnostic fires a HAS_GAP relationship in the graph — a signal that a competency is missing and needs production.
Factory side: The gap triggers cubelet production. When a learner masters the competency, a HAS_MASTERED relationship closes the loop.
Competency is the shared currency. It is the only concept the scaffold and the factory both understand natively. The system gets better at knowing what to produce as more practitioners use it.
Every stage is reachable by machine. The MCP layer is the integration boundary for technical partners.
course-factory-mcp (main) — Content retrieval: cubelet_get, cubelet_list, cubelet_search, catalog_cubelets, concept graph tools.
cmmc-consulting — CMMC L2 gap analysis and readiness reports. Entitlement-gated.
ai-gov — AI Governance reference and reporting: NIST AI RMF, EU AI Act, ISO 42001.
isaca — ISACA certification reference and exam reporting.
vulnops-mcp — Vulnerability operations. Dedicated worker only — not on the main surface.
assess-worker + factory-ops — Assessment delivery and domain registry / compile operations.
Entitlements resolve on every call. The bearer token proves the caller is trusted infrastructure. The KV store resolves which tools are gated for that caller.
The surface is open to technical partners. One MCP configuration block is all it takes to wire in.
Three entry points are available from day one — no custom integration required:
Request access and we'll provision a bearer token scoped to the tools your system needs.