FidelityLedger checks which evidence tools an extraction actually used
Its planted-tool failures measure a specific envelope; incorrect answers after tool use remain an open detection case.
FidelityLedger studies extraction runs where a value can appear correct even though the required evidence tool was never used. It logs dispatches and uses two rules checking which tools were called, independently of the extracted value.
The reported benchmark contains 37 claims across four components. The detector flags none of 207 clean fidelity-passing extractions and recovers 50 planted faults that withhold exactly the checked tools. The paper explicitly distinguishes this recall by construction from general detection power.
Incorrect answers after tools have been called were not measured by that detector. A separate physical chamber oracle covers only two of the 37 claims. These limitations are useful because they define what the instrumentation can actually justify.
For an extraction service, test both missing-tool and wrong-result cases. Keep source fidelity, dispatch evidence and physical validity as distinct checks instead of treating one passing value as proof of the entire process.
Editorial correction (2026-10-13 UTC): Added the planted-fault envelope, unmeasured called-but-wrong cases and two-claim physical oracle.