Page 4 of 8~112 min topic

Eval gates in code

Measure whether the CI eval release gate works

Page 4 turns “it ran” into executable checks for the release eval gate in CI.

~14 min this pageEvaluation

1Learn the idea

Read

Make the metric executable

Translate the claim into assertions or a tiny eval harness. The metric to protect is: pass/fail exit code; metric deltas vs baseline; seed recorded. Always record the denominator (how many cases) beside any rate. A percentage without a denominator is marketing, not measurement.

Read

Run the checks

baseline=0.92; candidate=0.81; threshold=0.9
assert candidate < threshold
print({'exit_code':1,'delta':round(candidate-baseline,2)})

Expected evidence: gate fails candidate. A passing assertion proves only the behavior it names; broader usefulness still needs the chapter’s full limits.

Read

Say what the metric does not prove

Be explicit: beating the baseline (last known good release metrics) on this fixture does not prove behavior under flaky eval that flips without seed control, or soft-fail that never blocks merge. Label observations separately from conclusions so the next page inherits honest evidence about the CI eval release gate.

Read

Lab notebook: denominator discipline

Compute pass/fail exit code; metric deltas vs baseline; seed recorded with the denominator written beside the rate every time. For this chapter, the evaluation set is intentionally tiny; that is allowed only if you say so in the evidence. Compare against last known good release metrics before celebrating.

Add one negative case aimed at flaky eval that flips without seed control, or soft-fail that never blocks merge. A suite with only happy cases cannot protect the CI eval release gate when the characteristic failure appears in review.

Read

Worked judgment

If a check is expensive or flaky, shrink it until it is deterministic on golden JSONL + threshold.yaml. Flaky green builds teach the team to ignore gates. Record what this page does not prove so security-ops and mastery-ship inherit honest limits.

Read

Why this stage matters for the CI eval release gate

At the evaluation stage for eval-in-code, the job is narrower than finishing a product demo. You are creating one progressive evidence piece about golden JSONL + threshold.yaml that later pages inherit without redefining success. Keep that fixture small enough to inspect by hand, keep outputs copy-pasteable as text, and refuse to narrate this baseline as if it were a production SLA: last known good release metrics.

For this page specifically, success looks like metrics with explicit denominators and a negative case while still centering the user decision to fail a candidate release when golden-task metrics regress past a pinned threshold. If you cannot point to a file, command, or assertion that proves that for the CI eval release gate, stay on this page instead of advancing.

How-to: ship agent with eval gate · Glossary: eval set

Previous · Next

Go deeper

Before you start

Why this matters

Write one independent check that would catch a fake pass for this lab. Prefer a check tied to pass/fail exit code; metric deltas vs baseline; seed recorded over a check that only asserts “no exception.”

In the wild

See how this idea shows up as a product and a company — then come back to the lesson. Skills transfer across vendors.

Check your understanding

Page assessment

Answer from memory. Completion is saved from this evidence, not from opening the next page.

1. Is the metric computed with an explicit denominator?
2. Does a failing gold case actually fail the harness?
3. Did you separate observations from conclusions?
4. What remains unproved after these checks?

All responses are required.