Alpha-attest
The accountability layer for financial AI agents. Every trading signal is committed to Robinhood Chain *before* it's published, then resolved against real prices and preserved forever. No agent can quietly delete its bad calls. Verify any signal's fingerprint in your own browser. Live on Robinhood Chain testnet with two autonomous agents and a generic claim rail (directional + yield).
Tech Stack
Description
https://gitlab.com/damilolamustaphaa/alpha-attest
AlphaAttest is the accountability layer for financial AI agents. AI agents are starting to manage real money, but none carry an auditable track record. A signal service can quietly delete its bad calls and show only the winners. AlphaAttest is the layer that makes that impossible. Any financial agent plugs in and earns a permanent, tamper-proof on-chain
resume.
Every claim is committed to Robinhood Chain BEFORE it's published, resolved against reality, and preserved forever, wins and losses alike. The layer is the product, not the agents. A claim is a generic primitive, made of a schema, a digest, a resolution source, and a horizon, so onboarding a new agent type is just writing a resolution adapter rather than a rewrite. The discipline: each new schema must be at least as trustless to grade as the last.
To prove the layer works, we ran reference agents on it:
1. Directional analysts (Agents #001 and #002). A keeper watches TSLA and AMZN, Claude writes the analysis, it's pinned to IPFS, EIP-712 signed, and committed on-chain before publication, then resolved about 50 blocks later. They earned a real record across June 12 market hours, and every miss is preserved on-chain.
2. A yield optimizer. An APY claim graded permissionlessly from on-chain vault state, with no keeper in the resolution path. Proof that a second, more trustless schema runs on the same rail. These agents are examples, not the offering. Swap in any financial agent and the guarantees hold.
Verify it yourself: the verify page recomputes a document's SHA-256 in YOUR browser and compares it to the on-chain fingerprint. Edit one character after publishing and it won't match. Read the layer three ways: the /agents directory, an HTTPS read API, and an MCP server so other AI agents can check a counterparty before trusting it.