The black box is an operational liability
Talk to any team lead who has used an AI-assisted ISR tool in the field. They will tell you the same thing. The system says 'hostile contact at grid X.' The operator asks why. The system has no answer.
What camera feed produced that assessment? What was the detection confidence? Was there a corroborating RF signature? What behavioral pattern triggered the classification? What other entities were in the area and ruled out? The system does not know. Or at least it does not show its work.
An operator who cannot verify an AI assessment has two options: trust it blindly and risk acting on a false positive, or ignore it entirely and lose whatever value the AI was supposed to provide. Both outcomes are bad. Both are the direct result of AI systems that present conclusions without evidence.
How EdgeLance couples evidence to every output
Every threat assessment in EdgeLance is designed to carry the source data that produced it: camera frames, correlated sensor signals, badge/IFF context where available, acoustic data, detection confidence, entity history, and the decision window the assessment applies to.
The operator does not see 'hostile' on a map icon. They see: unrecognized face at camera 3 (confidence 0.91), no NFC IFF response after 30 seconds in zone B, RF signature matching a vehicle not on the approved list, movement pattern consistent with deliberate approach rather than routine transit. They can look at the camera clip. They can check the RF log. They can make their own call.
Challenge, override, and the mission record
Evidence coupling gives operators the ability to push back. If an AI assessment does not match what they see on the ground, they can trace the reasoning, find where it diverges from reality, and override with documented rationale.
That override becomes part of the mission record. The AI recommendation, the source evidence, the operator's challenge, and the final decision all live in the event timeline. After-action review can reconstruct not just what happened, but why the AI recommended one thing and the human chose another. That kind of data is what makes AI systems actually improve over time.
Calibrated trust is the only kind that works at 0300
Responsible AI depends on human control, source evidence, and the ability to inspect why a recommendation was made. Without evidence coupling, those principles are hard to operationalize.
Operators who can see and challenge AI evidence develop calibrated trust. They learn which scenarios produce reliable assessments and which produce noise. That calibration only comes from repeated exposure to the AI's reasoning, not just its conclusions.
Without evidence coupling, operators either over-trust (and act on false positives) or under-trust (and ignore the system they paid for). Evidence coupling is what turns an AI tool from something that demos well into something an NCO will actually rely on in the middle of the night.