guiproof
Is this UI automation still true — today, for this version?
Continuously-verified, version-pinned GUI behavioral facts for agents — skip re-discovery every session.
A published agent skill (“Export is under the chevron”) is a snapshot that silently rots as the app changes — the button moves, a modal appears, a label goes inert. The failure is quiet: the agent exports the wrong report and nobody knows. There's no freshness signal and no version pinning, so serious deployments re-discover every app from scratch.
The value isn't the knowledge — it's the freshness signal on it. UI changes weekly, so a trustworthy, version-pinned freshness oracle is enormously valuable and hard to clone. Crowdsource cheap, noisy agent-run observations; own the re-verification engine that converts them into version-pinned, trust-weighted truth.
Decaying confidence
Each behavioral fact carries (app, version, fact, last-verified, confirmations, decay) — a freshness score, not a static tuple.
Re-verification engine
Batched sandbox runs re-confirm facts across tenants; cost-per-fresh-fact drops with scale.
Version fingerprinting
Maps observable signatures to app releases so facts pin to the exact version a tenant runs.
Neutral trust layer
A vendor-independent provenance schema agent runtimes read instead of re-discovering apps.