What problem does it solve?
It answers whether a spec-driven implementation actually fulfills its PRD, case by case, by decomposing every acceptance criterion into binary MET/UNMET checks backed by file:line evidence and rolling them up into a single comparable, reproducible grade.
Core Features & Use Cases
- Binary checklist scoring: Decomposes each acceptance criterion into atomic implementation (I) and test (T) checks, each requiring file:line evidence, then computes AC, story, and final scores with fixed 0.6/0.4 and P0/P1/P2 weights.
- Framework vs harness separation: Grades requirement extraction (Elicitation E: recall, precision, justified) and scope adherence (S) separately from test coverage, robustness (R), engineering gates (G), and test distribution (D).
- Reproducibility controls: Freezes the AC baseline checklist, runs k=3 self-consistency passes with majority voting, and requires script-computed roll-ups so the same PRD and commit yield the same grade.
- Use Case: Benchmark multiple spec-driven development frameworks against the same PRD by freezing one baseline checklist, then grading each implementation branch to produce comparable Final scores and ranked gap lists.
Quick Start
Ask the agent to run spec-driven-eval to grade the implementation on the current branch against the PRD, reusing the frozen baseline checklist and producing a scored report with evidence.
