K-Veritas
Guide

Compute cost & hardware

Bind the declared workload to physical hardware evidence, measured for the run's own process - not the whole machine.

Compute-cost attestation

Declare a model card so K-Veritas knows the scale of the work:

print("KVERITAS_MODEL params=25600000 arch=resnet50 precision=fp16")
print("KVERITAS_WORKLOAD dataset_size=1281167 epochs=90 batch_size=256")

At seal time a certificate checks the declared FLOPs against the hardware evidence using three physical bounds. A hard time or energy violation is a physical impossibility, so it is non-deniable - anyone can recompute the inequality from the inputs printed on the report.

Time

Declared FLOPs cannot exceed device peak throughput times GPU-active seconds.

Energy

Declared FLOPs cannot exceed measured joules divided by the minimum energy per FLOP.

Memory

Declared model weights should fit within the observed GPU memory.

Verdicts are PASS, REVIEW, FABRICATION-IMPOSSIBLE, or N/A. It proves the work ran at the declared scale; it does not prove the result is correct.

Hardware consistency (HMCA)

The Hardware-Metric Consistency Analyzer runs at seal time and flags metrics that do not match observed hardware activity - a metric reported with no measurable computation, or GPU claims on a CPU-only machine. It produces a score and verdict, both in the report.

Per-process measurement

K-Veritas measures only the process tree that kveritas run launched - CPU time and memory from the process tree, GPU memory and utilization filtered to its PIDs, and GPU power scaled by its share of utilization. If a browser or another job is running at the same time, it does not inflate the run's numbers.

Per-process attribution is Linux-only today (it uses /proc and nvidia-smi). On macOS and Windows the sampler falls back to system-wide readings until per-process support lands there.