Survey QA Agent

Survey QA Agent is a Meta Muse Spark agent that reviews LiDAR-to-CAD survey drafts before delivery, gathers extra evidence when it needs it, and produces an auditable fix queue. The problem: CAD survey drafts need visual QA before they go to a client. Geometry checks can flag suspicious linework, but they cannot tell whether a line traces a real road feature or an artifact like grass, tree canopy, a parked vehicle, a barrier or a nearby building. Manual review is slow and hard to prioritize across a job. What it does: for each station, Muse Spark looks at the plan render and street photos. When the evidence is weak it asks for more (another photo, QA flags, or the next station), then returns SEND, CAVEAT or HOLD with a reason and the first fix a drafter should make, plus defect boxes on the image in the same response. Station decisions roll up into frame verdicts and a self-contained HTML report with a work queue grouped by defect type. A replay page shows each evidence request and decision. Every API call is reserved against a spend cap before it is sent. Results: on a held-out set of 62 stations a human reviewer marked HOLD, Muse (prompt v3) flagged 60 and held 55; geometry-only checks flagged 42 and held 10. A regression check of the current checklist on all 122 human-labelled stations flagged 112 of 112 defects but also flagged all 10 clean stations. Those stations shaped the checklist, so that is not a fresh test, and cutting false alarms is the next step.