Research corpus

research/ is a machine-readable evidence corpus, not a loose notes folder. AUDIT.md records the current accuracy/alignment review and the gated plan that turns claims into falsifiable work.

Evidence states

• seed — discovered, not fully reviewed;
• reviewed — source identity and relevance checked, but not every material claim is necessarily verified at its locator;
• verified — material claims are checked against primary locators and implementation claims have reproducible evidence;
• superseded and rejected — retained with typed relations so history is not silently rewritten.

Confidence and status are independent. A high-confidence reviewed record is not described as verified.

Required properties

Each research record has a stable ID, source identity, retrieval date, tags, confidence, claim links, typed relations and repository links. Verified records additionally require Summary, Evidence, Mechanism, NUIF relevance and Open questions sections. Exact sections, pages, versions or commits support non-obvious source claims; synthesis is separated from source statements.

Layout

• items/ — one durable record per source or synthesized subject;
• index.yaml — architectural claims and topic-to-record map;
• questions.yaml — open and decided research questions;
• experiments/index.yaml — reproducible investigations and their executable artifacts;
• coverage.yaml — coverage status per architectural front;
• schema/ — record schemas;
• AUDIT.md and roadmap.md — current audit and gated research process.

Important relationships must be structural. tools/research/validate.sh checks record schema, identifiers, claims, relations, topics, questions, experiments, coverage links and artifact paths. Network source-health checks are periodic rather than part of offline conformance.

Canonical source.