Overview

TryAnneal is the is_this_safe() primitive for the Mantle agent economy — a multi-LLM smart-contract audit agent that posts verdicts on-chain, so any agent can check a contract before composing with it.

A contract goes in; a verdict comes back in seconds. The verdict is posted on-chain to the AnnealValidation registry on Mantle and is readable by anyone — through the CLI, a REST endpoint, an MCP tool, or a Telegram bot. TryAnneal itself is a registered ERC-8004 agent on Mantle mainnet (agent #131).

What makes it different

• →A model cascade, not one. ChainGPT pre-screens; two architecturally-distinct critics — Groq Llama-3.3-70B and OpenAI GPT-OSS-120B — argue every finding and cross-validate each other (Gemini 2.5 Pro is an optional third critic, off by default); only multi-model agreement survives — cross-validated against Slither + Aderyn. A pre-screen failure never blocks the critics, and a contract that nothing could analyze is flagged analysisIncomplete — never reported as safe.
• →Deterministic, reproducible verdicts.AI audits are usually non-deterministic; TryAnneal’s are not. Temperature-0 (greedy, seeded) decoding on every model, a corroboration rule that a reported finding needs ≥2 independent sources (≥2 models, or a model + Slither, when the full panel runs), confidence-weighted scoring, and memoization by code hash (keccak/sha3 of the source) on the Telegram bot and hosted MCP mean the same contract always returns the same verdict.
• →16 custom detectors + a 98-pattern exploit corpus covering $7.1B in documented losses (2020–2026, 13 chains), matched by TF-IDF cosine similarity.
• →On-chain, end-to-end on Mantle. ERC-8004 identity + an on-chain verdict registry + a live audit of Merchant Moe’s ~$60M router, posted on mainnet.
• →Reachable by any agent. Safety-oracle REST API, an MCP server, a Telegram bot + Mini App, and a GitHub Action — one verdict, many surfaces.
• →Fits into CI/CD. A GitHub Action runs the deterministic audit (Slither + 16 detectors + corpus, no keys) on every PR that touches .sol, posts a ✅ PASSED / ❌ BLOCKED comment, and emits a red/green check-run that fails on high/critical or a sub-threshold score — so branch protection can block the merge.
• →Multilingual reports. The audit runs in English, then Tencent Cloud Hunyuan translates the finished verdict + findings into the reader’s language (zh, es, ja, ko, fr, and more) — one click on the web /try page or /audit 0x… zh in the Telegram bot.

Surfaces