Opik Verify

Verify — Ship or Hold, Against a Policy You Can Read

Definition of done: one verdict — ship, hold, needs_review, or insufficient_evidence — computed from a declared policy over the baseline-vs-candidate numbers, with every criterion listed with its threshold, the observed value, and pass/fail, the cases behind any failure named, and the compare-view link. The policy is either the repo's opik-release-policy.yaml or the documented defaults, and the report says which. If the two runs can't be read or aren't comparable, stop at the first genuine blocker and return exactly one next step. "Looks good to me" is not a verdict; a verdict without its criteria is not one either.

Operate: apply the policy mechanically, show your arithmetic, refuse to ship on a judge nobody validated, and change no application code. The only file this skill may write is the policy file, and only when the user says so. It never deploys.

Inputs

The entry point is /opik-verify right after /opik-compare (its baseline and candidate), /opik-verify <suite> (the two most recent runs on the suite), or /opik-verify <baseline-id> <candidate-id>. Infer the rest; treat these as optional overrides:

• policy (default: opik-release-policy.yaml at the repo root or under .opik/, else the defaults below) · which experiments (default: as above) · --record (default: off — write the verdict into the candidate experiment's config).

Ask only at a genuine, non-inferable blocker (see Blockers).

The policy

Every key is optional; missing keys take the defaults listed in references/policy.md. Say in the report which source applied.

judge_validated: false is the human-review gate: until someone has confirmed the judge agrees with people (/opik-evaluate's validate-evaluator reference), a passing run yields needs_review, not ship. Flip it to true in the file once that is done — deliberately a human edit, never something this skill sets on its own.

Activation — the only in-scope work

1. Load the policy

Look for opik-release-policy.yaml at the repo root, then .opik/. Parse it; unknown keys → Blocker (name the key). No file → defaults, and say so. Never invent thresholds not in the file or the defaults.

2. Resolve the two runs

Take them from /opik-compare's output when it just ran. Otherwise, the SDK read in references/sdk-reads.md (Resolve the two runs). Skip a failed-judge run (a run whose judge had no credential is not a candidate — /opik-compare explains how it happens). scoring_failed does not survive the read path; the read-back signal is: every item failed and every assertion reason mentions a missing credential or an LLM infrastructure error. Say which run you skipped and why. When the hosted MCP is connected, list('experiment', name=…) shows each run's averages and pass rate to pick from; the item-level read below stays on the SDK.

3. Read both runs, item by item

An experiment holds one item per run: with runs_per_item: 3 a dataset item appears three times, same dataset_item_id, different trace_id. Group — a dict keyed on dataset_item_id silently keeps one run and loses the counts. The SDK read, grouped per item, with the pass thresholds taken from the suite: references/sdk-reads.md (Read both runs). Comparability first: same dataset_version_id, same item set, same judge model (experiment config). Different → Blocker ("rerun the candidate on suite version X with judge Y, then /opik-verify") — a verdict on non-comparable runs is not a verdict.

4. Evaluate every criterion, in this order

Compute all of them even after the first failure — the report shows the whole table.

1. Evidence size — scored items (items with assertions) ≥ min_items. Below → the verdict is insufficient_evidence regardless of the rest, unless a gate criterion (2–7) also failed — then it is hold: a known safety regression outranks thin evidence. Still report every criterion.
2. Regressions — items passed in baseline and not in candidate. An item is flaky when runs_passed is strictly between 0 and runs_total in either run (the counts from step 3), or it flips between two runs of the same code if you have them. Under flaky_policy: exclude a flaky item is dropped from the regression count and listed separately with its counts; under count it stays in. When every item has runs_total == 1, flakiness is not observable — report the flaky check as not_evaluated, state that exclude excluded nothing, and suggest runs_per_item: 3 on the suite if the user wants the protection. Count ≤ max_regressions.
3. Safety — any regression whose data.tags intersects safety_tags → fail, no exceptions, no exclusions.
4. Pass rate — candidate pass_rate vs baseline, or vs the number given.
5. Subgroups — when subgroup_key is set, pass rate per value of that key must not fall.
6. Latency — candidate p90 duration ≤ baseline p90 × (1 + latency_p90_max_increase), from the experiments' duration percentiles (p50/p90/p99 on the experiment record) (or per-item duration from the REST experiment items).
7. Cost — candidate mean total_estimated_cost per item ≤ baseline × (1 + cost_per_item_max_increase). Skip and say "no cost data" when neither run carries costs. Aggregates lag. Right after a run finishes, the experiment record's duration can read 0.0 and total_estimated_cost_avg None for a few seconds while the backend aggregates (observed). A zero or missing aggregate on one side is not data — re-read after a short wait, or compute p90 and mean cost from the per-item duration / total_estimated_cost fields on the REST experiment items; never let a 0.0 pass or fail the gate.
8. Evidence strength — a paired sign test on the flips: with f fixes and r regressions, the two-sided binomial p-value under 50/50. Report it; it is not a gate. With f + r < 6 say "too few flips to call it more than noise".
9. Judge — judge_validated from the policy. False → cap the verdict at needs_review.
10. Attribution — flips whose reason reads as judge hesitation on an unchanged output (see /opik-compare step 5.6) are listed for the human under needs_review, never silently counted either way.