Claim · Mixed

A training run stayed within declared limits

On this page

Mechanisms

R2 Training-transcript verification (proof-of-learning)⚠primary
Transcript checks for rules on training compute, data and hyperparameters (Shavit; Choi et al.).
R2 Zero-knowledge proofs of training constraintsprimary
Proves training followed a committed specification and data; the frontier design adds compute-threshold attestations.
R1 Hardware-enabled guarantees (flexHEG) and guarantee processorsprimary
Verifiable claims about total training compute, and enforcement of compute thresholds ¹.
R3 TEE remote attestation for AI workloads ⚠supporting
PAL*M attests single-node training and fine-tuning operations ⁹. Distributed training is left open.
R2 Bandwidth limits and compartmentalizationsupporting
Bounds the size of model that can be trained efficiently across pods ².
R2 Confidential multi-party verificationsupporting
Zero-knowledge audits can prove properties of committed training data and weights ³.
R2 On-chip telemetry from timing, memory and performance counters ⚠supporting
Counters for FLOP, memory and interconnect traffic are proposed as meters for compute accounting ⁴ ⁵.
R2 Workload classification from telemetry and side channelssupporting
Can flag training on hardware declared for other uses; does not measure training size by itself.
R1 Hardware performance throttling and licensingsupporting
Licenses that authorize a fixed amount of work would bound compute per license period ² ³.
R1 Whole-workload recomputation (reproducible packets)supporting
Proposed for later R&D verification by treating training steps as packets ¹ ².

Implementations

R2 Kaizenprimary
Assessed use: proving gradient-descent training of small image models on committed dataProves gradient-descent training of committed weights on a committed dataset according to public training specifications; it does not account for training outside the proven computation ¹.
R2 VeriLoRAprimary
Assessed use: proving single-sample LoRA fine-tuning steps for 3–13-billion-parameter language modelsProves the declared computation for individual LoRA fine-tuning steps. The evidence is narrower than a full-run training or compute-budget claim ¹.
R3 Verde and RepOps (Gensyn)supporting
Assessed use: reproducing declared-model inference from receipts in Gensyn's information-market serviceAlso covers training and fine-tuning jobs delegated to several providers ¹.
R2 PAL*Msupporting
Assessed use: attesting declared model operations on a confidential CPU–GPU prototypeAttests declared training and fine-tuning operations and their measured inputs and outputs; it does not account for undeclared training elsewhere ¹.