# Phase 5 Research: Ontology and Learning Materials

## Findings

The first useful ontology proof does not need heavy parsing. It needs a clean
boundary that proves uploaded material can become inspectable candidate
knowledge with provenance.

The MVP should:

- store material metadata and source text
- extract concept candidates from known backend interview concepts
- create prerequisite edges from a small deterministic rule set
- identify weak concepts when source support is thin
- never mark generated or inferred content as canonical

## Recommended Shape

- `internal/ontology` owns material ingestion, candidate storage, and snapshot.
- HTTP exposes JSON ingestion first.
- Evidence references use the existing workflow shared type.
- Gap records distinguish source-backed weakness from generated inference.

## Risks

- Overbuilding parsers too early would violate YAGNI.
- Treating keyword extraction as canonical knowledge would violate OpenSpec.
- A future parser can replace the analyzer behind the same service boundary.