Harness Cross-Span Attribute Value-Consistency / Forged-Attribute-Trail Lab
“A forged attribute trail cannot launder a clean AgentSmack grade.”
Every harness binding detector correlates a gate to an action by actionId / tool+argsHash — and then trusts the gate's decision verbatim. So a trace can have a coherent clock, a valid DAG, rich coverage, and full per-action logging, yet be internally contradictory in its attribute values: a policy.check.completed asserts allow for an action but the action.executed carries different args than the gate evaluated (the gate approved a different action than was executed — a laundered approval); a redaction.applied claims it scrubbed value-hash H but a sibling span re-emits H raw; a tool.call.planned names tool A but the matching tool.call.completed reports tool B; an approval.completed cites a different scope than approval.requested; a value-hash appears with fabricated provenance in no ancestor. This lab is a pure, deterministic meta-lens over the SAME Phase-4 span list: any forgery-class finding forces the verdict to forged (score clamped into the failing band, non-launderable); substitutions / scope drift are suspect; a trace with no correlatable pair is insufficient_signal, never a fabricated consistent. Grading is by the parsed correlation key + attribute value, not the array index — re-ordering the spans cannot change the verdict. The report carries only span ids + closed enums + the leading 12 chars of the offending value-hash (no raw span / attribute / secret / value bytes). Load a sample to watch the verdict swing with no live infra.
Is the clean grade even earned, or laundered by omission?
This lab grades whether the attribute values agree — but a forged-clean trace can also launder a grade by OMITTING the risky spans so a harness dimension is silently no_signal-excluded. Run the same spans through the Instrumentation-Coverage / Telemetry-Sufficiency auditor, or grade the raw harness chain at the harness telemetry check.
Value-consistent is not the same as clock-coherent or structurally sound
This lab grades attribute-value equality across correlated spans — but a trace can agree on its values while a back-dated gate forges the wall-clock or a parentId cycle forges the graph. Run the same spans through the Span-Timeline Temporal-Integrity lab and the Span-Lineage Graph-Integrity meta-lens to catch a forged clock or a forged causal DAG.