Count the evidence.
Never the assumption.

The AI delivery harness now fails closed when an adapter omits attribution, preserving honest success and cost denominators before the production-shaped comparison.

Inspect the attribution proof Release evidence

0Unattributed successes

3/3Focused harness checks

0 BBrowser output delta

0.16.18

AI ATTRIBUTION INTEGRITY

Require the trace.
Preserve the denominator.

FAIL CLOSED

No synthetic success

An exit-zero adapter without its raw trace is retained as an incomplete failure, never a zero-token success.

SUCCESS RATE

Keep failures visible

Missing attribution contributes zero successes to both paired variants instead of inflating either framework.

COST

Withhold unsupported claims

Per-success token and monetary summaries remain null when no attributable success exists.

RAW EVIDENCE

Retain the attempt

The runner still records source, artifacts, acceptance, commands, and the explicit incomplete attribution state.

SCOPE

Claim no ranking

The five-task production comparison remains active; this patch only repairs its evidence boundary.

OUTPUT

Ship zero extra bytes

The change touches benchmark, tests, and release records only, adding no compiler, runtime, API, or browser artifact.

UPGRADE

Keep AI evidence attributable.

$npm install @kudzujs/core@^0.16.18

Count the evidence.Never the assumption.

Require the trace.Preserve the denominator.