Domain-Driven Redesign — decouple Phoenix from document structure #
Premise (from aicoding.leaflet.pub + the n=1 / Regenerative-Grain / Compile-to- Architecture essays): intent is the asset; document layout is an authoring accident — the "framework" of the input. The pipeline must derive a semantic domain graph from unstructured text (could be raw meeting notes) and let IUs emerge from domain cohesion, never from markdown headings. Structure survives only as a provenance anchor (a back-pointer for "why"), at most a weak prior — never an organizing principle or a hard boundary.
The three phases #
Phase 1 — Ingestion signal-classifier (this turn) #
Treat input as a flat stream of statements; classify each as signal (requirement / constraint / invariant / definition / decision) vs noise (chatter, greetings, process talk, questions, examples, meta). Noise is dropped from canonicalization (kept conservatively when unsure — the D-rate/eval trust loop catches the rest). This is "smart about what to ignore."
Phase 2 — section_path → provenance-only anchor (this turn: semantics; identity next) #
Phase 3 — Semantic IU clustering, mass-bounded (this turn) #
Replaced planIUs' (doc, section) buckets with domain clustering (src/iu-clusterer.ts):
Notes #
- The "fold refinement sections" heuristic (regex on heading names) is now deleted — domain clustering folds rules into their entity by meaning, not by heading.
- Risk: clustering is fuzzier than buckets → pushes correctness onto evals + the D-rate trust loop, which is exactly where the architecture says value lives, but raises the bar on canonicalization confidence and cold-start.