ProofForge
🧠 ProofForge: explainable AI meets Hedera : agents that reason and prove.
Demo
About
ProofForge builds autonomous reasoning agents that don’t just give answers — they generate verifiable, human-readable proof traces for their conclusions and anchor those proofs on Hedera for integrity and auditability. A user asks a question or submits a claim → the agent produces a step-by-step reasoning trace (natural language + structured steps), attempts symbolic checks / unit tests where applicable, and emits a compact cryptographic proof (hash + metadata). That proof and an immutable reference (IPFS CID or storage pointer) are anchored to Hedera (HCS/HTS interactions) so anyone can verify the timestamped evidence without trusting a single model.
How it's made
Frontend (browser) React + TypeScript single-page app for interactive sessions and proof inspection. Wallet integrations for Hedera (e.g., HashPack / web-SDK) so users can sign transactions and anchor proofs. Agent orchestration / API FastAPI (Python) serving REST + WebSocket endpoints for multi-turn sessions and streaming results. LangChain (or in-house orchestrator) to compose LLM calls, retrieval, and verification steps into agent "plans". Reasoning & verification pipeline LLM layer: GPT-style model(s) for natural language reasoning and step generation. Symbolic / deterministic checks: Z3 (SMT) or lightweight unit test harnesses to validate numeric/logical claims where applicable. Execution sandbox: ephemeral containers or WASM runners to execute small reproducible snippets (data queries, calculations) safely. Proof generation: structured JSON proof object containing: natural language trace (ordered steps) machine-verifiable artifacts (outputs, checksums, test results) metadata (model version, prompt template, timestamp, session id) Hashing & bundling: produce a Merkle or content hash of the bundle; store artifact on IPFS (or S3 with CID) and keep hash/CID for anchoring.