Objective: Build the SR8 proof lane so benchmark runs are local-first, reproducible, and useful to operators.

Scope:
- Phase 1. Add an evaluation harness for compile, validate, lint, and transform
- Phase 2. Add a benchmark corpus across founder, research, repo, procedure, weak, and self-hosting tracks
- Phase 3. Add regression comparison and self-hosting proof cases

Exclusions:
- Hosted SaaS benchmark portal

Constraints:
- Keep evaluation local-first and reproducible
- Do not hide weak confidence or contradictory intent

Context Package:
- WF11 gap closure receipt
- Architecture notes
- Benchmark rubric definitions

Authority Context:
- founder sponsor

Dependencies:
- Current compile, transform, and lint rails

Assumptions:
- Rule-first extraction remains the default path

Success Criteria:
- Benchmark runs emit scored JSON and markdown reports
- Self-hosting cases prove compile, transform, and recompile flows

Output Contract:
- Canonical artifact
- Self-hosting proof pack
