kveritas Tool Verifies Computational Experiment Results with Tamper-Evident Proofs

Introduction to kveritas

kveritas is a newly released tool designed to provide tamper-evident verification for computational experiments. It creates a cryptographically signed PDF that links a published result to the specific code, hardware, and time used to produce it. This allows reviewers to verify claims without needing to rerun the entire experiment themselves.

Core Functionality and Features

The tool operates by running a command under its supervision, capturing a signed timeline of content-addressed snapshots. It supports any programming language and has zero runtime dependencies, functioning as a single static binary. Key features include cross-platform verification, sealing, proof generation, and management of benchmark artifacts and provenance. While file/subprocess activity mapping and per-process hardware attribution are currently Linux-only, other platforms fall back to system-wide readings.

Telemetry and Integrity Checks

During execution, kveritas records per-process telemetry such as CPU, memory, context switches, page faults, CPU frequency, and I/O at approximately 10 Hz. For GPU usage, it also tracks utilization, memory, power, and temperature. At the sealing stage, a Homogeneous Multi-Channel Analysis (HMCA) algorithm assesses whether these telemetry channels co-fluctuate consistently, indicating a genuine run versus a fabricated or replayed trace. The verdict can be PASS, WARN, FAIL, or N/A, with hard violations resulting in a "FABRICATION-IMPOSSIBLE" bound into the signature.

Model Card Verification

When a run declares a model card, kveritas produces a certificate that checks the declared FLOPs against the physical capabilities of the hardware. This mechanism helps to identify discrepancies, such as a large model declared on an idle CPU, ensuring that computational claims are consistent with the resources used. The time bound sums all devices, whether CPU or GPU, to ensure honest runs pass while detecting impossible claims.

Usage and Accessibility

Users can install kveritas via a prebuilt binary or build it from source using Go 1.22+. The workflow involves initializing a session, running commands under kveritas, sealing the results into a PDF report, and then verifying the report. The tool also auto-detects common metrics from frameworks like Keras and scikit-learn, simplifying its use for many computational tasks. The generated signed PDF can be verified offline or through a web verifier that renders telemetry channels over time.