Cursor rule

1---
2description: "Eval-Driven Development — define capability and regression evals before building; code graders use verify commands, model graders use explicit rubrics; log pass@k. Use before develop-tdd on new features, or when measuring agent capability over runs."
3alwaysApply: false
4---
5 
6# Run Evals
7 
8> **HARD GATE** — Define evals before implementation. Code graders = runnable `verify:` commands; model graders = explicit rubric with pass/fail criteria.
9 
10## Process
11 
121. Name the capability under test (one sentence).
132. Write `specs/EVALS-<feature>.md` with:
14   - **Capability evals** (does it do the job?)
15   - **Regression evals** (did we break anything?)
163. Assign grader type per eval: `code` (shell verify) or `model` (rubric).
174. Assign **strictness tier** per eval (graduated promotion — e45s37):
18 
19   | Tier | Meaning | Promotion rule |
20   |------|---------|------------------|
21   | `EXPERIMENTAL` | New eval, may flake | Not gating |
22   | `USUALLY_PASSES` | Stable in dev; ≥2/3 recent runs pass | Blocks BUILD only when combined with ALWAYS_PASSES suite |
23   | `ALWAYS_PASSES` | Zero tolerance; required for release | Any single failure blocks BUILD and merge |
24 
25   Promote: `EXPERIMENTAL → USUALLY_PASSES` after 3 consecutive passes; `USUALLY_PASSES → ALWAYS_PASSES` after 5 consecutive passes with zero flakes documented in `specs/state.yaml`.
26 
275. Run evals; log results table with pass@k (e.g. 3/3 runs) and tier per eval.
286. Block BUILD phase until all `ALWAYS_PASSES` evals pass at agreed k. `USUALLY_PASSES` failures warn; `EXPERIMENTAL` failures log only.
29 
30## Artefact
31 
32`specs/verifications/eNNsYY-eval-report.md` — see [REFERENCE.md](REFERENCE.md) for template. Eval reports are stored alongside verification evidence in `specs/verifications/`, keyed by story ID for traceability.
33 
34## Verify
35 
36→ verify: `test -d specs/benchmarks && test -f specs/benchmarks/SCHEMA.md`
37 
38 
39<!-- story: e02s01 -->
40 
41---
42 
43# Run Evals — Reference
44 
45## Strictness tiers (e45s37)
46 
47Add a `tier:` column to each eval row:
48 
49| Tier | Gate behaviour |
50|------|----------------|
51| `EXPERIMENTAL` | Log only — does not block |
52| `USUALLY_PASSES` | Warn on failure; blocks only when paired with failing `ALWAYS_PASSES` |
53| `ALWAYS_PASSES` | Hard block on any failure |
54 
55## EVALS template
56 
57```markdown
58# EVALS: <feature>
59 
60## Capability
61| ID | Eval | Grader | Tier | verify / rubric |
62|----|------|--------|------|-----------------|
63| C1 | ... | code | ALWAYS_PASSES | `verify: npm test -- <file>` |
64| C2 | ... | model | USUALLY_PASSES | Rubric: [ ] criterion A [ ] criterion B |
65 
66## Regression
67| ID | Eval | Grader | verify / rubric |
68|----|------|--------|-----------------|
69| R1 | Full suite passes | code | `verify: npm test` |
70 
71## Results
72| Run | C1 | C2 | R1 | pass@k |
73|-----|----|----|-----|--------|
74| 1 | PASS | PASS | PASS | 3/3 |
75```
76 
77## pass@k
78 
79Run capability evals k times (default k=3). Ship when all k pass or document known flake in `specs/state.yaml` `handoff.open_decisions`.
80