sourceright AGENTS.md

Sourceright Codex instructions

The live Sourceright repository has advanced beyond the previous overlay plan. Treat the current repository as the source of truth.

Do not bulk-apply the old overlay ZIP. Do not overwrite existing Rust modules. Do not split the crate into a workspace unless explicitly asked after an audit. Prefer additive changes: schemas, docs, examples, benchmark fixtures, demo apps, plugin manifests, validation scripts, and policy files.

Preserve these architectural boundaries:

• references.csl.json is canonical clean bibliographic data.
• references.verification.json stores provider evidence, provenance, confidence, conflicts, and review state.
• review-queue.jsonl is derived operational review work.
• Provider data must never silently overwrite canonical CSL.
• Legal citations remain separate from academic CSL.
• Claim/source/provenance work should not assert claim truth.
• MCP write tools must wait until read-only server contracts, schemas, audit logs, and dry-run semantics are stable.

Before modifying files:

1. Inspect the current repo.
2. Run existing checks where possible.
3. Use any older overlay only as design reference.
4. Produce a slice-by-slice implementation plan.

After each slice, run as many as apply:

cargo fmt --check
cargo clippy --all-targets -- -D warnings
cargo test
cargo check --locked

If new validation scripts are added, run them too.

Market-readiness pack guidance:

• Use .codex-plan/ material as planning guidance only, not as source files to bulk-apply over the repo.
• Prefer small, reviewable implementation slices for benchmark maturity, schemas, docs, rustdoc/API docs, public demos, plugin/provider status matrices, case-study fixtures, read-only MCP examples, and launch checklists.
• When applicable, also run:
  • cargo run --bin sourceright -- bench
  • cargo run --bin sourceright -- report --json examples/workspace
  • schema/demo validation scripts
  • the docs build command
• Keep benchmark claims separate from product claims.
• Accurate market wording includes “technical preview,” “pilot-ready,” “fixture-backed regression benchmark,” and “deterministic benchmark scaffold.”
• Avoid claims such as “production-ready institutional platform,” “SOTA benchmarked performance,” “legal filing compliance system,” or “AI detector” unless the repo evidence proves them.

Do not commit automatically. Summarize changed files, checks run, failures, uncertainty, and deferred work.

---