Impact Accountability Agent
An autonomous AI agent that monitors milestone evidence, analyzes it with Claude, stores verification reports on Filecoin, and submits onchain verification transactions. Every decision is logged. Every report is permanent. No human bottleneck.
Milestone verification is the bottleneck in impact funding. Donors don't know if their money was used well. Projects hate the overhead of proving it. Manual review is slow, expensive, and subjective. The agent closes this loop: autonomous analysis with a permanent, decentralized audit trail.
Project uploads Agent detects Claude analyzes Report stored
evidence to → EvidenceAttached → evidence against → permanently on
Storacha/IPFS event onchain milestone criteria Filecoin Pin
│
confidence ≥ 70%?
yes │ │ no
↓ ↓
verifyMilestone() store report,
submitted onchain defer to human
│
↓
MilestoneArbiter returns true
→ Alkahest escrow releases fundsHow it works
Monitors EvidenceAttached events on ImpactHook. When a project uploads evidence to Storacha or Filecoin and stores the CID onchain, the agent detects it within seconds.
Retrieves the evidence file from Storacha/IPFS gateways using the CID. Supports images, PDFs, JSON reports, CSV data, and text documents. Multiple gateway fallbacks for reliability.
Sends the evidence and milestone description to Claude for structured evaluation. Returns a confidence score, per-criterion breakdown, and reasoning. Vision capability for image evidence, document analysis for PDFs.
If approved with confidence above the threshold, submits verifyMilestone() onchain. The verification report is stored permanently on Filecoin. MilestoneArbiter gates unlock, releasing Alkahest escrow funds.
Where agent data lives
Past verifications, project knowledge, and processed event history. Persists across sessions. Agent loads its memory from the latest CID on restart.
Structured JSON reports with per-criterion analysis, confidence scores, and reasoning. Stored via Synapse SDK. Permanent, verifiable audit trail accessible via IPFS CID.
Structured logs of every decision, tool call, and error. Flushed to Filecoin on shutdown. Proves the agent operated autonomously and made verifiable decisions.
Guardrails and accountability
Only auto-verifies when Claude's confidence exceeds 70%. Below that, stores the report but defers the decision to a human verifier.
Full analysis pipeline without submitting onchain transactions. Analyze evidence, generate reports, persist memory, but never call verifyMilestone().
Every decision is logged with timestamp, action type, and details. Logs are flushed to Filecoin on shutdown. Anyone can audit why the agent approved or rejected.
Configurable max Claude calls per hour. Agent tracks gas costs and storage costs per session. Designed to operate under real economic constraints.
Onchain Agent Registry
AgentRegistry.sol on Filecoin Calibration lets agents register identity, store metadata and state CIDs, and build verifiable reputation from their verification history.