← Field notes

Open-source release · September 3, 2026

Six AI workflows you can pull apart before trusting them.

Most workflow lists stop at tool names. These six MIT-licensed kits start with a job, preserve the evidence, and end with a decision another person can review.

What is actually verified?

Every kit includes a starter ledger and a shell verifier. All 17 repository verifiers pass at release v0.17.0. That proves the required files and evidence fields are present; it does not prove every editorial workflow produces a better business result. The recipes say “not benchmarked” until measured runs exist.

Pull release v0.17.0 ↗

The six kits

Spec to production proof

Turn an approved specification into a reviewed diff, executable checks, browser evidence, and rollback note.

Inspect the workflow →
AI workflow regression test

Compare a prompt, model, or tool change against frozen cases without hiding failed assertions.

Inspect the workflow →
Open-source coding-agent benchmark

Compare Qwen Code, Goose, OpenCode, or another harness on one fixed repository task.

Inspect the workflow →
Document-parser fidelity test

Check MarkItDown and Docling output against the source before a model relies on it.

Inspect the workflow →
Browser performance regression

Reproduce a slow flow, retain traces, make a bounded fix, and rerun the same conditions.

Inspect the workflow →
Safe dependency upgrade

Keep Renovate, scan findings, compatibility work, runtime proof, and rollback in one ledger.

Inspect the workflow →

The rule behind all six

  1. Freeze the job, input, versions, permissions, and done definition.
  2. Keep deterministic checks separate from model judgment.
  3. Retain raw outputs, interventions, failures, time, tokens, and cost where available.
  4. Change one variable at a time and repeat noisy runs.
  5. Publish the boundary: what this evidence does not establish.

Help us find the weak one

Run a kit on a disposable fixture. If a verifier passes while decision-critical evidence is missing, open an issue with the smallest safe reproduction. That is more useful than a star, and we will keep the failure in the public record.

Open a recipe issue ↗

Get the next measured result

One maintained workflow at a time. Explicit opt-in, no vendor-paid ranking, unsubscribe any time.