model-switch

Example: a model change that fails the build

This example compares recorded toxicity metrics for a baseline and candidate model while holding the prompts, scoring method, and policy fixed. The candidate metrics cause the existing rule to fail, so CI blocks the merge.

The example runs in this repository's own CI on every push, so the badge on the README reports the current state of the check rather than a recorded result.

Run it

./check.sh outputs-baseline.json     # the model in production
./check.sh outputs-candidate.json    # candidate model

The baseline and candidate runs exit 0 and 1 respectively. In CI, those exit codes produce a passing check and a blocked merge.

What the example demonstrates

The enforcement mechanism is independent of the scoring implementation. Detoxify supplies the toxicity metrics in this example; Perspective or another evaluator can be substituted if it produces the input fields expected by the policy.

The score is evaluated by a rule that can be read, cited and versioned, and the same policy evaluation can run in each environment. The threshold is defined in the policy rather than in this script: params.toxicity_threshold in global/v1/toxicity defaults to 0.1, so changing it requires a reviewed pull request.

Aggregate metrics can mask individual outliers. In this six-output example, the aggregate toxicity score increases from 0.0056 to 0.1373 and exceeds its 0.1 threshold. The maximum toxicity score increases from 0.02 to 0.81 and exceeds its 0.7 threshold. GOPAL keeps metrics.toxicity.score and metrics.toxicity.max_toxicity separate because, in a larger suite, an individual high-toxicity output may not cause the aggregate score to exceed its threshold.

CI runs the policy evaluation automatically as a status check. A failing evaluation therefore blocks the merge and keeps the applicable article and rule in the pull-request review context.

The numbers

outputs-*.json contain per-output toxicity scores recorded from Detoxify 0.5.2 over the same prompt suite. They are checked in so the example is deterministic and needs no model download in CI.

This example does not execute either model. It starts after model outputs have been generated and scored; the checked-in files represent the metrics presented to the policy. This keeps CI deterministic while isolating the policy-enforcement behavior. To generate equivalent metrics yourself, AICertify ships a Detoxify adapter that emits exactly this shape, or see supplying metrics for the mapping written as plain JSON.

Wiring it into your own repository

check.sh needs jq and opa and nothing else. Point --policy-dir at a GOPAL checkout or unpack a release bundle:

./check.sh scores.json --policy-dir /path/to/gopal --threshold 0.05

For a workflow you can copy, see .github/workflows/model-switch-demo.yaml in this repository, or examples/github-actions for the model-card equivalent.

Which policies work like this

Five policies in the current library can be evaluated entirely from measured metrics and are therefore suitable initial candidates for automated CI enforcement:

Policy	Reads	Supplied by
global/v1/toxicity	metrics.toxicity.score	Detoxify, Perspective
EU AI Act Article 11	metrics.model_card.*	a model card, scored by model_card_score
bfs fair lending	fairness, content safety, risk	Fairlearn
healthcare diagnostic safety	the same three	Fairlearn
global/v1/fairness	toxicity, stereotype	Detoxify, LangFair