Migrating to 0.9

Wind Tunnel 0.9 makes scenario intent decide the verdict. It also separates agent robustness from test-harness integrity and versions the native artifacts that carry those decisions. Runtime SPI method signatures are unchanged.

Declared expectations now gate by default

In 0.8, only outcome decided the aggregate verdict. Trajectory and constraint failures were diagnostic even when the scenario explicitly declared must_call, forbidden_calls, a custom TrajectoryCheck, or a Policy.

In 0.9, Scenario.resolved_gate_layers() infers:

• outcome for every scenario;
• trajectory when any trajectory expectation is declared; and
• constraint when any policy is declared.

This means a scenario cannot report PASS while violating one of its authored expectations. If a check is intentionally diagnostic during exploration, say so explicitly:

Scenario(
    name="observe_lookup_path",
    prompt="Find the account.",
    target_facts=[["ACC-123"]],
    must_call=["account_lookup"],
    gate_layers=["outcome"],
)

Remove the explicit gate_layers once the trajectory assertion is ready to gate. gate_layers=[] creates a diagnostic-only scenario, but experiment integrity is still mandatory.

Integrity and robustness now mean different things

Score.integrity answers: did the declared experiment condition actually happen? For perturbations, it checks the evidence markers recorded on the trace. A failed integrity check makes the aggregate INVALID; it does not count as an agent pass or failure.

Robustness answers: did the agent satisfy its gate under a valid adverse condition? Reports calculate robustness from gate performance on scenarios that declare perturbations. A suite without perturbation scenarios reports robustness as N/A.

For source compatibility, 0.9 still accepts Score(..., robustness=...), exposes score.robustness, and exports evaluate_robustness(). These are 0.8 compatibility spellings for integrity; new code should use integrity and evaluate_integrity().

Failure cost is operational

FailureCost now has a deterministic risk_weight:

Input	Weight
severity low / medium / high / critical	1 / 4 / 16 / 64
customer visible	+2
irreversible	+4
side effect performed	+8

Aggregates expose failure_risk = risk_weight × (1 - pass_rate). Reports, ledger rows, and wt compare surface this value and rank regressions by it. Risk does not relax the gate: every gated regression still fails.

Scenario authoring is less duplicative

• prompt may be omitted when user_turns is present.
• target_facts defaults to [], so an outcome_fn needs no placeholder.
• scenario.scored_prompt returns the final user_turns entry or the single-turn prompt and is used by reports and triage.
• Empty names, missing both prompt forms, empty user turns, duplicate or unknown gates, and order_matters=True without must_call fail at authoring time.

Persisted artifacts have explicit versions

New native traces carry "windtunnel_trace": 1, score sidecars carry "windtunnel_score": 2, and ledger records carry "windtunnel_ledger": 1. The v2 score stores integrity instead of the old robustness marker field.

Readers migrate unversioned 0.8 traces and scores in memory. They reject unknown future versions instead of guessing at their meaning. Contract A interchange and Contract B universe files remain version 1: additive unknown fields within v1 are tolerated, while unknown version numbers are rejected.

Upgrade checklist

1. Run the suite and inspect scenarios whose trajectory or constraint layer had been failing diagnostically; those failures now affect the verdict.
2. Use explicit gate_layers=["outcome"] only where diagnostic behavior is intentional and documented.
3. Rename direct score.robustness reads to score.integrity and evaluate_robustness calls to evaluate_integrity.
4. Treat INVALID as a bench/setup problem that requires rerunning after the condition is repaired, not as an agent regression.
5. Confirm artifact consumers tolerate the new version keys and integrity field. Prefer score_from_dict() over parsing score dictionaries directly.
6. Review FailureCost annotations so risk-ranked comparisons reflect the operational consequence of each scenario.