Mechanism Knowledge Design
Document status: architecture baseline. Mechanism knowledge constrains and explains recommendations; it does not create a second R&D run workflow.
Position
Mechanism knowledge stores engineering experience, process material, and source evidence as reviewable, versioned, scoped assets. It can explain a recommendation, exclude known unsafe combinations, or adjust candidate ranking; it cannot replace real production runs, quality outcomes, or an engineer decision.
Relationship to the Recommendation Loop
knowledge sources → claims and constraints → recommendation input snapshot
real runs and quality outcomes → optimization observation → next-recipe recommendation
engineer decision → actual-execution link → frozen outcome → evidence for knowledge updates
Each recommendation freezes the knowledge versions, source hashes, and applicability scope it used. An engineer's adoption, modification, or rejection reasons and the later quality outcome are the raw evidence used to strengthen or weaken knowledge.
Rules
- Each claim has a source, applicability scope, version, review status, and conflict treatment.
- Hard constraints only reduce the candidate space; soft constraints only affect ranking and cannot manufacture a quality outcome.
- When inputs do not match, scope is exceeded, knowledge is stale, or conflicts remain unresolved, the system degrades to data-driven recommendation or stops recommending.
- Mechanism explanation states its evidence level and never presents association as a definitive root cause.
- A recommendation retains its exact knowledge snapshot; later knowledge edits never rewrite it.
Knowledge Retrieval
Document knowledge reaches the analysis assistant through source ingestion and fragment extraction, human review, keyword and optional semantic indexing, hard site and applicability filtering, hybrid ranking, and fragment-level citation. Only reviewed sources and human-reviewed fragments are retrievable. Results retain the source record, page or sheet, source SHA, and content hash so an engineer can return to the original material and detect content changes.
The semantic index is rebuildable derived state. It does not change review status or become formal business evidence. Embeddings are disabled by default. When enabled, they reuse the protected model-service configuration and require an OpenAI-compatible /embeddings endpoint. If that service is unavailable or a query embedding fails, the system falls back to PostgreSQL keyword retrieval without relaxing authorization or review gates.
Current Boundary
The current path supports sources, fragments, claims, applicability, hard constraints, conflicts, versions, recommendation-use traceability, and scope-controlled hybrid retrieval for the analysis assistant. Real-run decisions and quality outcomes progressively support or weaken knowledge. A retrieval match does not establish a claim, and the repository does not claim proof of a particular mechanism or field benefit.