eval-report
Use when the user has run multiple evaluation skills and wants a comprehensive analysis — maturity assessment, cross-skill signals, trends, prioritized actions, and an executive summary. Also use when the user mentions eval health check, evaluation audit, ship readiness, evaluation maturity, or "how good is my evaluation system itself." This is a read-only analysis skill.
agentscope-ai · 已有 AI 审核记录
安装与使用
运行命令前先阅读来源与依赖说明。复制命令不会自动安装或执行。
npx skills add agentscope-ai/OpenJudge --skill eval-report技能说明
name: eval-report description: > Use when the user has run multiple evaluation skills and wants a comprehensive analysis — maturity assessment, cross-skill signals, trends, prioritized actions, and an executive summary. Also use when the user mentions eval health check, evaluation audit, ship readiness, evaluation maturity, or "how good is my evaluation system itself." This is a read-only analysis skill.
Eval Report
Synthesize everything from your evaluation journey into a comprehensive report. This skill is read-only — it analyzes what exists, doesn't create new graders or datasets.
When to Activate
- You've run 2+ evaluation skills and want the big picture
- You need to report evaluation status to non-technical stakeholders
- You're making a ship/no-ship decision and need evidence
- The evaluation system has been running for a while — time for a health check
Checklist
You MUST create a task for each item and complete them in order:
- Inventory scan — catalog everything in eval-design.md + runs/ history
- Maturity assessment — 5 dimensions × 4 levels
- Cross-skill signal synthesis — consistent findings + contradictions
- Weakness diagnosis — failure concentration, correlations, stratum gaps
- Root cause classification — system / metric / data / unclear
- Prioritized recommendations — P0/P1/P2 actions with impact estimates
- Executive summary — ship readiness + top 3 risks + next actions
Step 1: Inventory Scan
Read eval-design.md and all runs/ directories. Build a timeline:
Timeline:
2026-04-15 01-eval-design → 5 failure modes → 3 dimensions from 200 traces
2026-04-18 02-metric-design → 4 graders configured (2 LLM + 1 rule + 1 executable)
2026-04-25 (evaluation run) → 90-sample stratified dataset scored
2026-05-01 03-align-human → 2 judges Phase 3, 1 Phase 2, 1 Phase 1 (TPR/TNR + kappa)
2026-05-10 07-redteam → safety audit not yet run
Report key metrics:
- Total skills run, total principles, total labels
- Calibrated judges: X of Y (with TPR/TNR range)
- Last activity date per skill
Step 2: Maturity Assessment
Rate the evaluation system across 5 dimensions:
| Dimension | L1 (Initial) | L2 (Developing) | L3 (Established) | L4 (Optimizing) |
|---|---|---|---|---|
| Failure Discovery | No systematic analysis | Failure modes identified | Coverage validated with stratification | Continuous triage from production |
| Judge Quality | v0 uncalibrated only | Some calibrated (TPR/TNR measured) | All calibrated with CI | Calibrated + aligned with humans |
| Label Coverage | < 50 labels | 50-200 labels | 200+ stratified labels | Coverage audit passed, drift monitored |
| Safety Coverage | No redteaming | Ad-hoc redteam run | Systematic redteam with policy doc | Continuous redteam with sign-off |
| Human Alignment | No alignment data | Kappa measured for some judges | Kappa ≥ 0.8 for all judges | Human spot-check only, quarterly audit |
Scoring rule: The overall maturity level is the minimum across dimensions (weakest link principle). If 4 dimensions are L3 but Safety is L1, the system is L1.
Step 3: Cross-Skill Signal Synthesis
Consistent Signals (high confidence)
Find themes confirmed by multiple skills. Example:
- "Factuality is the top risk" — evidence chain:
- 01-eval-design: #1 failure mode (38% prevalence in traces)
- 02-metric-design: weighted as the highest-impact dimension
- 03-align-human: TPR=0.92 TNR=0.88 (confirmed measurable)
Contradictions (needs investigation)
Find where skills disagree. These are the most valuable findings:
- "02-metric-design weighted hallucination as a top signal, but 03-align-human shows the hallucination judge has TPR=0.74" → Possible explanations: the judge prompt captures surface patterns, not real hallucination. Or the judge prompt needs refinement, or the labels are noisy.
Coverage Gaps (blind spots)
What hasn't been touched by any skill?
- "01-eval-design coverage shows multilingual input_type n=0, no workflow has addressed non-English queries"
Step 4: Weakness Diagnosis
Failure Concentration
Which principle/grader has the lowest pass rate? Where are failures clustering?
Failure Correlation
Compute Jaccard similarity between principle pairs — when sample A fails on principle X, does it also fail on principle Y? Highly correlated pairs (Jaccard > 0.5) likely share a root cause.
Per-Stratum Weakness
Which difficulty stratum performs worst across all principles? If boundary stratum TPR < 0.7 for 3 of 4 principles, boundary discrimination is a systemic weakness.
Step 5: Root Cause Classification
For each weakness area, classify the root cause:
| Type | Definition | Key indicator |
|---|---|---|
| system_problem | The application itself performs poorly | Low pass rate + high judge-human agreement |
| metric_problem | The judge/eval is flawed | Low pass rate + low judge-human agreement |
| data_problem | The eval dataset isn't representative | 01-eval-design coverage shows thin strata OR label drift detected |
| unclear | Not enough evidence | Conflicting signals, need more data |
This classification is critical — fixing a metric problem by changing the system (or vice versa) wastes effort.
Step 6: Prioritized Recommendations
Generate P0/P1/P2 actions. Each must include: priority, concrete action, current state, target state, expected impact, and the skill to use.
🔴 P0 | Calibrate hallucination judge
Current: TPR=0.74 (below 0.8 threshold)
Target: TPR >= 0.8
Impact: Judge becomes
阅读完整来源 ↗