/agent-regression-evals

name: agent-regression-evals
description: Use when an agent, prompt, model or tool change needs proof it did not make the agent worse — setting up a regression eval suite, deciding whether a pass-rate drop is real or noise, grading tool-call trajectories, finding flaky eval cases, or wiring an eval gate into CI that blocks the merge. Runs repeated trials, reports pass@k and pass^k, compares candidate to baseline with paired statistics, and refuses to pass a dropped case, a critical-case regression, or a forbidden tool call.
argument-hint: "[baseline.jsonl candidate.jsonl | results.jsonl | a question about agent evals]"
license: MIT
metadata:
  version: 1.0.0
  build_pattern: "Statistics-first eval gate: 3 deterministic stdlib scripts (trajectory grading, flake detection, paired regression gate) with blocking exit codes"
  distinct_from: "prompt-governance (versions and registers prompts; this gates a change with statistics); self-eval (grades one session's work); agenthub eval (ranks parallel agents in one session)"

Agent Regression Evals — prove the change did not make it worse

> **Portability:** 3 stdlib scripts, no APIs, no LLM calls. You run the agent; these grade, count and gate.

"Pass rate went from 82% to 79%" is not a finding until you know whether three points is outside the noise. Agents are non-deterministic, so one trial per case cannot tell a regression from a coin flip, and a final answer can be right while the path to it called a tool it never should have.

Workflow

# 0 - Run every case k times (k >= 3) on baseline and candidate, from a clean
#     environment per trial. Write one JSONL row per trial:
#     {"case_id": "refund-lookup", "trial": 1, "passed": true, "critical": true,
#      "cost_usd": 0.012, "trajectory": [{"tool": "get_order", "args": {...}}]}
# 1 - Grade the path where the path matters (safety, required lookups).
python scripts/trajectory_checker.py --spec cases.json --results candidate.jsonl
# 2 - Separate flaky cases from stable ones before trusting any comparison.
python scripts/flake_detector.py --results candidate.jsonl --k 3 --quarantine-out q.json
# 3 - Gate: paired per-case comparison against baseline. Exit 4 blocks the merge.
python scripts/regression_gate.py --baseline baseline.jsonl --candidate candidate.jsonl --quarantine q.json

Step 1 uses match modes `strict | unordered | subset | superset`. Default to `superset` (required steps present, harmless exploration allowed); grade the outcome, not the path, unless order is the requirement. `forbidden` tools block in every mode.

Step 2 reports **pass@k** (succeeds at least once in k) and **pass^k** (succeeds all k times). A user meets every trial, so pass^k is the reliability number. Cases marked critical are never quarantine candidates.

Step 3 averages trials within each case first (trials of one case are correlated), then tests paired per-case differences with a t-based 95% CI and a one-sided exact McNemar test. When the drop is real but the sample is too small to show it, the gate says how many cases it would take.

The gate blocks on

1. A statistically significant drop (CI upper bound below zero, or McNemar p < 0.05). 2. A critical case that passed every baseline trial and failed any candidate trial. 3. A baseline case missing from the candidate run. Dropping a case is the cheapest fake fix. 4. A quarantine entry with no owner or expiry, or covering a critical case. 5. A forbidden tool call (trajectory checker, exit 4).

It warns, and a human decides, on an underpowered drop above `--tolerance` or a cost rise above `--max-cost-increase`.

Hard rules

• **Never compare single trials.** Under 3 trials per case the flake detector refuses (exit 4).
• **Never pool trials as independent samples.** Average per case, then compare cases.
• **Never quarantine without owner and expiry.** An expired entry re-enters the gate automatically.
• **Never delete a failing case to go green.** The gate counts what is missing.
• **Never call "no significant drop" a pass on 8 cases.** Report the power note with the verdict.
• **Start the suite from real failures** (20-50 cases), balanced between "should do" and "should not do".

Scripts

| Script | Role | Exit codes | |---|---|---| | `scripts/trajectory_checker.py` | Grades tool-call paths: match modes, args modes, forbidden tools, call ceilings | 0 · 2 deviations · 3 bad input · 4 forbidden tool | | `scripts/flake_detector.py` | STABLE_PASS / STABLE_FAIL / FLAKY per case, Wilson CI, pass@k, pass^k, quarantine stubs | 0 · 2 flaky · 3 · 4 too few trials | | `scripts/regression_gate.py` | Paired CI, exact McNemar, critical cases, dropped cases, quarantine validity, cost | 0 PASS · 2 WARN · 3 · 4 BLOCK |

All support `--sample` and `--output json`.

References and assets

• [`references/eval_statistics.md`](references/eval_statistics.md) — pass@k vs pass^k, per-case averaging, paired CI, McNemar, power (7 sources)
• [`references/trajectory_grading.md`](references/trajectory_grading.md) — outcome vs path, match modes, grader types (6 sources)
• [`references/flaky_cases_and_quarantine.md`](references/flaky_cases_and_quarantine.md) — flake causes, quarantine with expiry, clean trials (6 sources)
• [`assets/eval_case_spec.example.json`](assets/eval_case_spec.example.json) — trajectory spec for step 1
• [`assets/sample_baseline.jsonl`](assets/sample_baseline.jsonl), [`assets/sample_candidate.jsonl`](assets/sample_candidate.jsonl) — results format
• [`assets/ci_eval_gate.example.yml`](assets/ci_eval_gate.example.yml) — CI job that runs the gate and fails on exit 4