Ivan Romanteyev (solo dev)

TrustFlow is a verifiable AI-agent workflow prototype for synthetic financial operations, built during the Meta Global AI Developer Hackathon. Autonomous agents can propose financial actions, but a model's claim of success does not establish whether the exact operation was authorized, executed only once, or independently verified after a timeout or crash. TrustFlow solves this through a structured workflow: Planner → deterministic PolicyEngine → Human Approval → Executor → Independent Verifier. High-risk operations require single-use human approval bound to an exact plan version and SHA-256 hash. Backend-owned operation identities and SQLite idempotency protect against duplicate synthetic transfers within tested boundaries. Independent ledger verification establishes whether committed transactions match approved payment intents. Following uncertain execution or a controlled post-COMMIT crash, TrustFlow reconciles the original operation and retries verification without replaying the transfer. The project was developed with extensive assistance from Meta Muse Code, alongside developer-directed engineering and independent Codex reviews. MetaResponsesProvider integrates Meta's Responses API with structured planning and semantic review using configurable Muse Spark and Muse Wonder models. One real Muse Spark planning inference has been verified. Full live execution and Wonder inference are still under validation. The repository includes reproducible offline demonstrations, safety tests, and independent acceptance evidence. TrustFlow demonstrates how probabilistic AI reasoning can be combined with deterministic authorization and independently verifiable execution.