Eval gates in code
Instrument the CI eval release gate
Page 6 adds signals that distinguish bad input from component failure in the release eval gate in CI.
1Learn the idea
Read
Emit stage signals
Instrument the release eval gate in CI so a run records enough structure to debug offline: counts, latency if relevant, pass/fail of pass/fail exit code, and a stable stage name. Redact secrets and raw credentials from every event.
Read
Emit and assert
import json
print(json.dumps({'stage':'eval-gate','seed':7,'baseline':0.92,'candidate':0.81,'exit_code':1}))
Expected evidence: eval gate telemetry. Prefer JSON or structured text you can grep in CI over prose logs for eval-in-code.
Read
Lock signals with a regression test
Turn one historical failure—especially flaky eval that flips without seed control—into a test that fails if the signal disappears for the CI eval release gate. Observability without a failing test is optional decoration; observability with a test is part of the eval-in-code artifact.
Read
Lab notebook: signal schema
Draft a three-field event for the CI eval release gate: stage, ok, and one domain field derived from pass/fail exit code; metric deltas vs baseline; seed recorded. Add fixture_id or docs_version when content can change. Explicitly list fields that must never appear (tokens, passwords, raw prompts) because editing golden expected answers to greenwash a bad candidate is in scope for this lab.
Wire one assertion that fails if the CI eval release gate event is missing after a run. Observability that cannot fail a test will not survive contact with a busy eval-in-code repository.
Read
Worked judgment
Imagine a teammate opens only your event stream after a bad deploy. Could they tell whether golden JSONL + threshold.yaml was wrong, whether flaky eval that flips without seed control, or soft-fail that never blocks merge returned, or whether editing golden expected answers to greenwash a bad candidate slipped through? If not, rename fields until those three stories are distinguishable.
Read
Why this stage matters for the CI eval release gate
At the testing and observability 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 a structured event schema locked by a test 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.
Go deeper
Before you start
Why this matters
Write the single log line or metric event that would tell you whether a bad result came from input vs implementation for the CI eval release gate. If your line could not tell them apart, redesign it before coding.
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.
Related lessons
Check your understanding
Page assessment
Answer from memory. Completion is saved from this evidence, not from opening the next page.
All responses are required.