LP-0130Finalevmprecompileaimininginference

LP-130: AI Mining and On-Chain Inference

Abstract

Implements the AI mining precompile for compute-to-earn verification on the Lux EVM. Miners submit work proofs attesting to GPU compute time, verified via TEE attestation and ML-DSA signatures. The precompile calculates rewards based on compute minutes, privacy level, and hardware trust score. Includes double-spend prevention via a spent-set and BLAKE3 work ID computation.

Specification

Address

0x0300 -- AI range.

Interface


0x01 VerifyMLDSA:     paramSet(1) + pubkey + signature + message -> bool(32)
0x02 CalculateReward: work_proof -> reward(32)
0x03 VerifyTEE:       tee_quote + nonce -> { verified(bool), trust_score(uint8) }
0x04 IsSpent:         work_id(32) -> bool(32)
0x05 ComputeWorkId:   work_proof -> work_id(32)  (BLAKE3 hash)
0x06 MarkSpent:       work_id(32) -> ()           (state write)

Work proof layout:


bytes  0-31:  device_id (32 bytes)
bytes 32-63:  nonce (32 bytes)
bytes 64-71:  timestamp (uint64, big-endian)
bytes 72-73:  privacy_level (uint16)
bytes 74-77:  compute_minutes (uint32)
bytes 78+:    tee_quote (optional, variable)

Privacy Levels

| Level | Multiplier | Description |

|-------|-----------|-------------|

| 1 (Public) | 0.25x | Open compute, no TEE |

| 2 (Private) | 0.50x | Software isolation |

| 3 (Confidential) | 1.00x | Hardware TEE (SGX/TDX) |

| 4 (Sovereign) | 1.50x | Full CC + local attestation |

Gas Schedule

| Operation | Gas |

|-----------|-----|

| VerifyMLDSA | 3,000 |

| CalculateReward | 1,000 |

| VerifyTEE | 5,000 |

| IsSpent | 100 |

| ComputeWorkId | 50 |

| MarkSpent | 5,000 |

GPU acceleration is used for batch ML-DSA verification (NTT-based) and batch attestation verification when available.

Security Considerations

1. Work proofs are bound to a specific device_id and timestamp. Replay protection via the spent-set.

2. TEE attestation is platform-agnostic (NVIDIA NVTrust, Intel TDX, ARM CCA). Verification is local.

3. ML-DSA signatures (FIPS 204) provide post-quantum authentication of work proofs.

4. Privacy level multipliers incentivize higher security tiers. The Sovereign tier requires both hardware CC and local-only attestation (no cloud dependencies).

5. The MarkSpent operation is a state write -- only callable by authorized mining contracts.

References

Source: github.com/luxfi/precompile/ai/
LP-035 (AI VM) -- the broader AI virtual machine specification
LP-127 (Remote Attestation) -- TEE verification details

Copyright

Licensed under the MIT License.