Each section maps to a different artifact — the spec, the ADRs, the diagrams, the mockups, the conversational provenance, and the domain vocabulary.
44 numbered decisions across 12 sections. The single source of truth for V2's architecture — substrate, milestones, dispatch pattern, review system, KB integration, UI shape, build order.
Five load-bearing ADRs documenting the why behind every reversible commitment. Baseline, prompts-as-agents, LLM Wiki, the 0e0703d revert, and pipelined-sequential dispatch.
Visual maps for every load-bearing flow: the M1→M4→Publish state machine, linear review-proposal apply, the prompts-as-agents harness, and the LLM Wiki two-level structure.
Three-zone teacher workspace mockups — left rail (Agents / Curriculum / Knowledge Base), middle content area, right rail (Progress + Review tabs), chat dock with context band.
The grilling-session narrative that produced this design. 18 rounds of operator HITL turned a half-built prototype into a coherent vision — captured so the why survives compaction.
Every term in the design that has a precise definition — Workspace, Knowledge Base, Pipelined-sequential dispatch, Pure-workspace projection, LLM Wiki, and 16 others.
The older pipeline shape — 5 phase orchestrators, 12 HITL checkpoints, confidence-routed trust modes. Retained only as shared manifest metadata in library.yaml; historical detail lives in docs/pipeline/overview.md.
The 8-agent support manifest: Orchestrator, Planner, Generator, Evaluator, Formatter + 3 KB agents. Lives in library-course-generation.yaml and preserves persisted internal IDs while current V2 phase modules own new generation behavior.
The 8-agent baseline becomes markdown prompts, product skills, and bounded ToolLoopAgent sessions. Substrate moves to per-course git repos in Supabase Storage. KB becomes a per-teacher LLM Wiki. V2 has since cut over on the default branch.