Skip to content
cccskills
FreePublished on GitHub

compare-trials

Analyze graded `trial_result` JSONL outputs from the eval pipeline, including reliability and custom slicing.

The English summary is being prepared. Showing the original description.

See how to install

Files included (3)

  • SKILL.md1.2 KB
  • references/bootstrap.ts1.8 KB
  • references/compare.ts5.2 KB

SKILL.md (original)

Review the instructions the agent will receive before you install.

Compare Trials

Use this skill for analysis over graded JSONL outputs, typically after:

agent-eval-harness eval '{"mode":"grade",...}' > graded.jsonl

Default path

For normal baseline-vs-challenger comparisons, use:

agent-eval-harness eval '{"mode":"compare",...}'

Use custom scripts only when suite-specific analysis is required.

Custom analysis use cases

  • cost analysis
  • token usage analysis
  • slices by task metadata (category, difficulty, source)
  • custom regression gates
  • per-agent metadata summaries

Script guidance

When writing a custom script:

  1. Parse JSONL rows as trial_result.
  2. Validate graded rows (pass and score non-null).
  3. Group by stable identity fields (taskId, trialIndex, runId).
  4. Separate:
    • standalone run metrics
    • comparable-overlap metrics
  5. Report uncertainty where possible (bootstrap confidence intervals).
  6. Use exactPassAtK naming for combinatorics-based pass@k metrics (do not label exact values as estimatedPassAtK).

Reference helpers remain in references/ for bootstrap/statistics utilities.

Reviews

No reviews yet. Tell others what you thought.

Comments

No login needed. Share how it worked for you, ask a question, or note where it did not run.

No comments yet. Be the first to write one.

Comments are published right away. Do not include personal information such as email addresses or phone numbers. Up to 2 URLs.

0 / 1,000 characters

This form is protected by reCAPTCHA, and the Google Privacy Policy and Terms of Service apply.

More skills from this repository

Interview the user relentlessly about a plan or design until reaching shared understanding, resolving each branch of the decision tree. Use when user wants to stress-test a plan, get grilled on their design, or mentions "grill me".

The English summary is being prepared. Showing the original description.

behavioral-sh/agent-eval-harness7Updated Jun 18, 2026

tdd

Free

Use test-driven development for behavior-changing feature or fix work, and whenever the user mentions TDD, test-first, red-green-refactor, tracer bullets, integration tests, or public-interface behavior tests. Skip for docs-only, path-only rename, formatting-only, or purely mechanical chores unless explicitly requested.

The English summary is being prepared. Showing the original description.

behavioral-sh/agent-eval-harness7Updated Jun 18, 2026

Build command-only adapters for `eval` run mode using strict stdin/stdout JSON contracts.

The English summary is being prepared. Showing the original description.

behavioral-sh/agent-eval-harness7Updated Jun 18, 2026

End-to-end eval-suite orchestration with the `eval` command: run -> grade -> compare -> calibrate, using strict JSON contracts and schema discovery.

The English summary is being prepared. Showing the original description.

behavioral-sh/agent-eval-harness7Updated Jun 18, 2026

See all skills by behavioral-sh

Report a problem with this skill