ProofStep

ProofStep is a stateful visual practice agent that coaches a five-step physical placement exercise on a printable 3x3 board. It distinguishes PASS, CORRECTION and UNCERTAIN, explains the rule evidence, and advances only after independent geometric checks succeed. A wrong placement stays at the same step for correction and reobservation; an occluded or unreadable image requests a retake. Later steps recheck all previously placed pieces, and progress survives a browser refresh. Meta Muse Spark 1.3 observes image quality and board visibility. SAM 3.1 returns segmented object masks and boxes; the official Meta parser extracts mask centroids and extents. A deterministic homography and rule engine then check placement. The model cannot declare task completion. A separate local pixel scan verifies the empty board and registration, and does not claim a model was used. Try the app: https://proofstep.vercel.app/?lang=en The public demo clearly replays genuine hosted-Meta observations on synthetic inputs and recomputes the rules; it makes no new API calls. A separate bounded local live mode analyzes newly selected images with a private server-side key. The published dataset used 5 Muse + 11 SAM calls, and additional real browser correction/retry checks bring this task to 7 Muse + 13 SAM calls. Automated tests: 171 passed. Scope is explicit: manual calibration, synthetic validation and box/cell overlays are implemented. Physical-camera accuracy, automatic marker detection, multi-frame recovery and the proposed larger evaluation study are not yet validated. This is an AI-assisted hackathon project built with Codex and DeepSeek during the event implementation period.