From an agent run to a verified fix

Assevra evaluates captured runs and produces portable scorecards with explicit release decisions and suggested next actions.

pip install assevra==0.6.0
python -m assevra.reference --out-dir reference-output

Open the before and after HTML reports. The executable synthetic refund workflow fails before its authorization fix and passes the same declared assertions after it. Installation downloads packages; the example runs locally without a model or API key.

• Getting started: run the example and adapt it.
• Integrations: choose an SDK, recorder, trace adapter, or export recipe.
• Concepts: understand scope, missing evidence, and decisions.
• CI: block failed or incomplete releases.
• Methodology: read the assumptions behind each measurement.
• Schema contracts: consume versioned artifacts.
• Migration from 0.5: account for the v2 release semantics.

The framework does not infer your business requirements from a trace. Declare the required dimensions, supply representative labeled cases, and calibrate model judgments with human labels before relying on them.