StepProof

StepProof solves the core crisis facing web and physical automation: autonomous AI agents are non-deterministic, expensive, vulnerable to prompt injection, and hallucinate success. When an enterprise web agent runs 10 times, it pays 10x token costs and can drift or fail catastrophically when a single DOM ID changes. StepProof introduces an architectural paradigm: "Models Perceive, Code Decides." Instead of leaving an unconstrained LLM in an infinite decision loop, StepProof watches a human demonstrate a workflow once (or learns from an agent's first successful trace) and compiles it into a portable, cryptographically signed Procedure artifact with multi-locator fallbacks ({slot} substitutions, role/CSS/XPath, and deterministic success checks). Key Pillars & Production Capabilities: 1. Deterministic Replay (0 Tokens, Pass^5 = 1.0): Runs 2 through N execute with 0 model calls, taking 0.10s per run compared to 12.6–26.4s and 152,139 tokens on open-ended agents—achieving 100% repeatability with zero token cost. 2. Self-Healing Visual Grounding (93.8% Recovery): When web layouts mutate and DOM selectors break, Meta Muse Spark visual grounding locates target coordinates on screenshots with SAM 3.1 open-vocabulary segmentation fallback. Across 16 mutated layouts and real screenshots, StepProof recovered 15/16 (93.8%) broken locators automatically, re-signing the procedure without manual intervention. 3. Code-First Safety Spine: Irreversible actions (submit, pay, restock, discard) hit deterministic code gates requiring human sign-off. Page text is treated as untrusted data—prompt injections (verified against hostile demo suites) cannot hijack control flow. 4. Tamper-Evident Proof & Audit Spine: Every run creates an append-only SHA-256 hash-chained log with before/after screenshots and token/frame telemetry. Offline verifiers can validate run integrity on an air-gapped machine with zero network calls. 5. Physical Procedure Verification & Robotics: Extends beyond the browser with SAM 3.1 video tracking to verify multi-step physical assemblies from video clips, paired with an MCP Safety Supervisor enforcing speed, workspace envelopes, and immediate e-stop watchdogs (<50ms). StepProof transforms volatile web agents into reliable, enterprise-grade, cryptographically verified automation.