openjiuwen-ai/eval_pipeline

Eval Pipeline

The single entry point for full skill evaluation. Orchestrates registered evaluators in sequence, enforces stage gates, and produces a final score.

Evaluator Registry

Each evaluator has a role: weighted (contributes to final score) or gate (must pass threshold before pipeline continues — failure is terminal regardless of other scores).

| # | Evaluator | Role | Weight / Threshold | Output file | | -- | ---------------------- | -------- | ------------------ | ---------------------------- | | E1 | skill_llm_judge | Weighted | 40% | llm_judge_score.json | | E2 | skill_tester | Weighted | 60% | skill_tester_report.json | | E3 | skill_safety_judge | Gate | score_pct ≥ 0.80 | skill_safety_score.json |

To add a new evaluator : add one row to this table with its role, weight or threshold, and output filename. The aggregation formula in Step 4 reads from this registry — no other changes needed.

Weight rule : weighted evaluator weights must always sum to 100%. When E3 activates, rebalance E1/E2 weights as needed (suggested: E1 40%, E2 60%, E3 gate).

Pipeline Flow

skill path
    │
    ▼
[E1] skill_llm_judge ──── gate: score ≥ 60 ──── FAIL → stop, report design issues
    │ pass
    ▼
[E2] skill_tester ─────────────── FAIL → stop, report functional issues
    │ pass (all weighted evaluators done)
    ▼
[E3] safety evaluator (future) ── gate: ≥ 0.80 ── FAIL → stop, safety block
    │ pass
    ▼
[Aggregate] compute weighted final score
    │
    ▼
final_score.json + pipeline_report.md

Gates run before weighted scoring. A gate failure short-circuits the pipeline — don't run downstream evaluators or compute a final score. Report why the gate failed so the skill author knows what to fix.

Execution

Step 1: Resolve paths

Read the system prompt for output_dir. All output paths below are relative to:

<output_dir>/evals/skill-tests/<skill-name>/

Step 2: Run E1 — skill_llm_judge

Use sub-agent tool with the following prompt. Pass ONLY what the sub-agent cannot infer itself — the target path and output dir. Do not embed skill content, pre-resolve steps, or add workflow instructions.

"You are an LLM judge. Evaluate the skill at `{target_skill_path}`.
  Follow the skill_llm_judge skill at {<skill_llm_judge_path>} to focus on the skill design.
  Output save to `{output_dir}/evals/skill-tests/{skill_name}/`."

Expected output: llm_judge_score.json

Gate check : if design_score_pct < 0.60 → stop pipeline. Save pipeline_report.md with status BLOCKED_AT_E1 and the Critical Issues from llm_judge_report.md. DO NOT proceed to E2.

Step 3: Run E2 — skill_tester

Call sub-agent with the following prompt. Pass ONLY what the sub-agent cannot infer itself — the target path and output dir. Do not embed skill content, pre-resolve steps, or add workflow instructions.

"You are an safety evaluator. Evaluate the skill at `{target_skill_path}`.
  Follow the skill_safety_judge skill at {<skill_safety_judge_path>}.
  Output save to `{output_dir}/evals/skill-tests/{skill_name}/`."

Expected output: skill_tester_report.json

No hard gate on pass_rate — low pass_rate is penalised in the weighted score, but the pipeline completes so the author gets a full picture.

Step 4: Run safety gate evaluators

Call sub-agent with the following prompt.

"You are a safety evaluator. Evaluate the skill at  `{target_skill_path}`. Follow the skill_safety_judge skill at {<skill_safety_judge_path>}. Save all output to `{output_dir}/evals/skill-tests/{skill_name}/`."

1. Run it and collect its score_pct
2. If score_pct < threshold → stop, save pipeline_report.md with status BLOCKED_AT_<evaluator>, do not compute final score

Step 5: Aggregate

Compute weighted final score from all weighted evaluators:

final_score_pct = Σ (evaluator.score_pct × evaluator.weight)

Current formula:
final_score_pct = (design_score_pct × 0.40) + (pass_rate × 0.60)

Determine grade:

| Grade | final_score_pct | Meaning | | ----- | --------------- | ----------------------------------- | | A | ≥ 0.90 | Ship-ready | | B | 0.80–0.89 | Minor issues — fix before shipping | | C | 0.70–0.79 | Functional but needs work | | D | 0.60–0.69 | Significant issues — do not ship | | F | < 0.60 | Fundamental problems |

Step 6: Write outputs

In the output folder, save final_score.json:

{
  "skill_name": "<name>",
  "pipeline_status": "COMPLETED",
  "final_score_pct": 0.81,
  "grade": "B",
  "evaluators": {
    "e1_llm_judge":   { "score_pct": 0.74, "weight": 0.40, "contribution": 0.296 },
    "e2_functional":  { "score_pct": 0.87, "weight": 0.60, "contribution": 0.522 }
  },
  "gates_passed": [],
  "recommendation": "Good functional quality. Design issues in D2 and D4 should be fixed before shipping."
}

Save final_report.md:

# Eval Pipeline Report: <skill-name>

## Final Verdict
- **Score**: 0.81 (B)
- **Status**: COMPLETED
- **Ship-ready**: No — fix before shipping

## Evaluator Breakdown
| Evaluator | Score | Weight | Contribution |
|-----------|-------|--------|--------------|
| E1 Design (skill_llm_judge) | 74% | 40% | 0.296 |
| E2 Functional (skill_tester) | 87% | 60% | 0.522 |

## Key Issues to Fix
[Pull Critical Issues from each evaluator's report]

## Full Reports
- Design: `llm_judge_report.md`
- Functional: `report.json`

---

Output Structure

evals/skill-tests/<skill-name>/
├── llm_judge_report.md          # E1 detailed report
├── llm_judge_score.json         # E1 machine-readable score
├── skill_tester_cases.json                   # E2 test case definitions
├── run-1/ … run-N/              # E2 per-test outputs
├── skill_tester_report.json                  # E2 consolidated functional results
├── safety_score.json            # E3 output (future)
├── final_score.json             # Pipeline aggregated score
└── final_report.md           # Human-readable final verdict

---

NEVER Do

• NEVER compute a final score if any gate evaluator failed — a blocked pipeline has no valid final score
• NEVER skip E1 to go straight to E2 — poor design produces noisy functional test results
• NEVER hardcode the aggregation formula in a way that requires code changes to add a new evaluator — the registry table is the single source of truth
• NEVER let a high functional score mask a low design score — report both, the author needs the full picture
• NEVER rebalance weights silently — if weights change, note it in pipeline_report.md