Show Me Once: Verified Receiving Agent

Show Me Once — record a workflow once, run it verified forever Receiving teams reconcile every delivery against its purchase order: count the cartons, match the PO, flag shortages. It's repetitive, and a wrong count becomes a wrong payment. Show Me Once is built for ops teams with no engineers. How it works. An ops user records one reconciliation (match, shortage, missing count). Muse Spark 1.3 Vision turns the frames + narration into a reviewable procedure — steps, selectors, and rules, each citing the frames it came from. A person resolves every ambiguity and approves the exact procedure; nothing unreviewed can execute. A 10-turn Playwright agent then runs it on unseen deliveries across changed layouts: it files true counts, stops on persisted-state completion, and escalates to a supervisor queue instead of guessing. Two code-enforced safeguards back it up: an element-resolved quantity guard in the browser and receiving-evidence enforcement at the persistence boundary. Every run ends with an independent check — ok is true only for a completed run plus a passed persisted-record match. Measured results. Fresh match/shortage/missing cases × two layouts: 6/6 first attempts pass with ok: true. Guard intervention (0 needed live) is proven separately by element-guard, persistence-evidence, and injected-action integration tests — 129 tests green. SAM 3.1 was evaluated over 36 paired runs and honestly omitted: 0 usable regions, latency without benefit. All failures and fixes are preserved as evidence, including the original invented-count failure and its labeled retry. Limitations, stated plainly. Dock counts are simulated fixtures; both layouts share selectors (layout-robustness, not arbitrary-app adaptation); supervisor human-correction auth is demo-grade. Built with: Muse Spark 1.3 via the Meta Model API, Playwright, Python.