Apex Labs

An event log in, evidence-linked findings out, every miss on the record. WITNESS audits a fixed-camera assembly procedure from its event log. It reports skipped steps, out-of-order steps and undocumented parts, and each finding either carries an evidence id or does not exist. The audit is deterministic and writes a report whose hash chain anyone can recompute from the output bytes. The intended user is a university CubeSat team, where integration and test runs on a Word checklist and a folder of phone photos. When a launch integrator asks which fastener was fitted, in what order, on which unit, the evidence cannot be reconstructed. The ten-box panel in the film stands in for that hardware: a cheap, controlled scene where a real mistake can be produced on purpose and the detector measured against it. We first tried producing the event log from real footage of a ten-box panel with geometry alone. That extractor missed and invented marks, then refused both clips rather than guess, and we froze it. A Meta segmentation model, sam-3.1, then proposed the boxes every five seconds, deterministic code named them, and the same texture rule judged them: on both clips the audit now agrees with the key, for 78 calls and USD 0.1950. The naming rule was calibrated on six frames from those same clips, so this is a demonstration, not independent validation. The limit: all 25 risk-gate cases are synthetic, and the only real footage is two clips of one scene. The film runs 152.2 s. Every real-footage number sits in docs/RISK_GATE.md and docs/REAL_MEDIA.md, beside the key it was measured against.