╔═══════════════════════════════════════════════════════════════╗
║ ◉ THE AI AUDIT TRAIL THAT REGULATIONS DEMAND ║
║ Every model. Every weight change. Every inference. ║
║ Immutably sealed on Stellar Soroban. ║
╚═══════════════════════════════════════════════════════════════╝
A language model makes a loan denial. A vision model flags a job candidate. A trading algorithm executes a position. A medical diagnostic tool rules out cancer.
None of these systems are required to produce a cryptographic proof of what they knew, when they knew it, and what data shaped their decision.
This is not a technical limitation. It is a governance vacuum. The tooling to produce immutable AI audit trails has existed in blockchain infrastructure for years. What has not existed is a protocol purpose-built to apply it to AI systems at the model, inference, and training-data layers.
ModelTrace fills that vacuum. Not a compliance checklist. Not an optional transparency dashboard. An on-chain event recorder that makes AI opacity structurally impossible.
modeltrace/
├── apps/
│ ├── web/ ← Next.js 14 — live at modeltrace.vercel.app
│ └── backend/ ← Fastify event indexer
├── contracts/
│ ├── audit-registry/ ← Model registration + AI event logging ✅
│ ├── usage-meter/ ← Inference metering (scaffold — needs builders)
│ └── payment-router/ ← Auditor staking + rewards (scaffold — needs builders)
└── docs/ ← Protocol specs, compliance mappings
register_model(operator, name, version, weights_hash, training_data_cid) -> model_id
log_inference(operator, model_id, payload_hash, risk_level) -> event_id
flag_bias(auditor, model_id, evidence_hash) -> event_id
get_model(model_id) -> ModelRecord
get_stats() -> RegistryStats { total_models, total_events, high_risk_flags }Per-inference telemetry at scale. Captures per-request metadata without exposing model internals.
start_session → record_inference → close_session → get_session_stats
Staking and incentive layer for independent AI auditors.
stake_as_auditor → submit_attestation → challenge_attestation → claim_reward
| Framework | Requirement | Coverage |
|---|---|---|
| EU AI Act (2026) | Art. 13: Transparency for high-risk AI | ✅ On-chain inference log |
| EU AI Act | Art. 12: Logging requirements | ✅ Immutable event registry |
| NIST AI RMF | Map & Measure: risk tracking | ✅ Risk level per inference |
| ISO/IEC 42001 | AI management audit trail | ✅ Cryptographic provenance |
git clone https://github.com/FinesseStudioLab/modeltrace
cd modeltrace
stellar contract build --package audit-registry
cargo test --workspace
cd apps/web && npm install && npm run dev- Audit Registry: model registration, inference logging, bias flagging, full tests
- Frontend live with live audit stream, model risk monitor, compliance section
- Neural network background visualization
- Usage Meter contract — per-inference telemetry
- Attestation Router — auditor staking and rewards
- All three contracts deployed to Soroban testnet
- End-to-end audit lifecycle integration test
- Fastify backend full event indexer
- Python SDK: one-line inference logging for ML frameworks
- MLflow + Hugging Face Hub integrations
- OpenTelemetry exporter
- Grafana dashboard: real-time risk visualization
- EU AI Act compliance certificate generation
- NIST AI RMF automated scoring
- Mainnet security audit
- Bug bounty on Immunefi
- Stellar Mainnet deployment
- 3 AI lab partnerships for live inference tracing
- 1,000+ models registered, 100M+ events logged
AI audit trails generate millions of events per day. On Ethereum, logging each inference costs $0.50–$5.00 in gas — up to $1.2M/day per model. This math does not work. On Stellar: $0.00001 per event. A model can log every inference for $2.40/day. This is the only chain where comprehensive AI auditing is economically viable at scale.
Rust/Soroban engineers — Usage Meter and Attestation Router are fully specified.
ML engineers — Python SDK, MLflow/Hugging Face integrations.
TypeScript developers — Wire frontend to live contract state.
Security researchers — Audit the registry contract and staking mechanism.
All issues are labeled contract, sdk, frontend, backend, documentation, or research.
© 2026 FinesseStudioLab contributors · MIT License