evals-bootstrap

Bootstrap the eval suite

Theory: Two kinds of checks and the full walkthrough in Evals practice. An eval suite is the same test after every change. Behavioural checks read the steps of a trace; end-to-end checks read only the result. Start behavioural: they are deterministic, run in seconds, and diagnose instead of just scoring.

Workflow

1. Find real failures. Ask where the agent's runs live (logs, transcripts, a traces directory). Read until you have up to twenty real failures — not imagined ones. If there are no logged runs yet, build the trace logging first (step 3) and seed the suite with the three failures the user can recall; a small honest suite beats a large invented one.
2. One line per failure. For each: the input, and the one specific behaviour that should have happened and did not. Failures cluster into four to eight behaviours; name them.
3. Ensure traces exist. Each run must be stored as traces/<id>.json — a list of events including tool calls. If the user's harness is Claude Code, the transcript already is the trace; wire up whatever copies or converts it. No trace, no behavioural checks.
4. Write cases.yaml. One entry per failure:

- id: refund_1042
  input: "Refund order #1042, customer says it arrived broken"
  expect: looks up the order before replying; asks approval before refund
  check:
    - trace has get_order before send_reply
    - trace has approval_request before refund

The expect line is for humans; the check lines are the test. Keep the rule language tiny: trace has X, trace has X before Y, trace lacks X. 5. Generate check_traces.py. A small runner: load cases.yaml, parse each rule with a regex, walk the tool-call list, print one line per case, exit non-zero on any failure. Keep it dependency-light (pyyaml only) and fast — the whole suite should run in seconds, with no model calls. 6. Run it and hand over the flywheel. Show the pass/fail lines. Then leave the loop in writing at the end of your report: read fresh traces weekly, add every new failure as a case, fix the biggest cluster, re-run.

Rules

• Every case comes from a real failure; delete a case only when the behaviour it guards is retired, not when it is inconvenient.
• No LLM-as-judge in the bootstrap. Add a judge later, only for what assertions cannot reach, and calibrate it against human labels first.
• The suite must be one command (python check_traces.py) so it can gate a CI job or a pre-release habit without ceremony.