Evidence-Binding / Definition-Drift Auditor

“Did the scorecard grade the agent you are actually shipping?”

The deployment gate, the governance pack, and the certify path all TRUST a ProductionScorecard at face value — but NOTHING checks that the scorecard's evidence was computed over the SAME agent definition (system prompt + tool set + model) that is being deployed. The launder path is wide open: a deploy can pass a gate using a scorecard computed against a hardened older prompt while shipping a weakened edited prompt — or a newly-added high-risk tool (a transfer_funds-class capability) the scorecard never assessed — certifying a version that is no longer what ships. AgentSmack's evidence-binding auditor is a pure, deterministic structural meta-lens that re-derives BOTH definitions' fingerprints via the canonical computeFingerprint (sha-256 over canonical prompt + model + tools) and compares the deployed fingerprint against the evidence's bound fingerprint. A fingerprint mismatch is a critical launder-block (verdict drifted, score clamped ≤39); a drifted tool set or a changed model is structural drift; stale evidence (graded only from the supplied ISO timestamps via Date.parse, never the wall clock) is a soft caveat. At the deploy gate a clean scorecard over a drifted binding is forced inconclusive — it cannot launder a deploy. The report carries only 16-char fingerprint prefixes + drifted tool names — never a full hash, never raw prompt bytes. Load a sample to watch the verdict swing with no live infra.