Scorecard Self-Consistency / Tamper-Evidence Auditor

“A tampered scorecard cannot launder a clean AgentSmack score.”

The entire deployment-gate / governance-pack / percentile / certify pitch rests on a ProductionScorecard whose overall / band / grade / blockedDimensions are trusted verbatim by every downstream consumer. The launder path is wide open: the deployment gate parses the scorecard's SHAPE but never recomputes its ARITHMETIC, and the signed-pack verifier checks the signature + structure — not whether overall is the correct weighted fold of the dimensions. So a customer (or a compromised CI harness) can hand-edit a scorecard JSON — bump overall 38→92, flip band to production_ready, empty blockedDimensions, or raise a dimension above its contributing-surface minimum — and every gate lets the unsafe agent ship. This auditor is a pure, deterministic, arithmetic-coherence meta-lens that re-derives the canonical scorecard math from the SINGLE source of truth (the dimension weights, the min-of-contributing-surfaces rule, the ≤39 hard-blocker cap, the renormalised weighted blend, the shared grade/band ladder) and flags every field that does not match. Three launder-class mismatches — a non-derivable overall, an unapplied hard-blocker cap, a band flipped independently of the arithmetic — force the verdict to tampered (non-launderable); structural drift is suspect; an empty scorecard is insufficient_signal, never a fabricated consistent. It pairs with the deployment gate: a clean-looking tampered scorecard is forced inconclusive. Load a sample to watch the verdict swing with no live infra.