Jan 2026
GPUs Do Not Promise Determinism. Here Is the Hardware Reason.
Why running the same CUDA code twice can produce different floating-point results, and what you can do about it.
Verifiable, offline AI systems.
We focus on traceable model runs and offline deployment.
Clear provenance, documented configs, and repeatable setups.
Design goals
Offline by design
Designed to run without outbound network calls or telemetry.
Repeatable runs
We aim for reproducible setups with documented tolerances and signed artifacts.
Data minimization
We design for minimal collection and keep evidence local where possible.
Verifiable is not truth
We do not promise the model is right. We aim to show what model ran, with what config, on what input. That's the part you can verify.
Provenance
Artifacts should trace back to their origin. Model weights, adapters, and runtime are identified where possible.
Manifests
Structured declarations of what should run. Machine-readable. Diffable.
Signed Configs
Configurations can be signed so tampering is detectable.
Hash-Chained Logs
Log entries can reference the previous; deletion or modification becomes detectable.
Energy is a constraint
In transfer-heavy workloads, data movement dominates energy cost. Unified memory architectures can reduce this cost by eliminating copies between CPU and GPU memory. We measure this with Joules per token.
Methodology
We document a measurement methodology for Joules/token benchmarking on Apple silicon.
macOS powermetrics API • 10-run averaging • thermal normalization • documented tolerances
Recent Research Notes
View all →Jan 2026
Verifiability Is Not Truth
Verification proves what happened, not that the output is correct. This distinction matters for compliance and trust.
Nov 2025
Joules Per Token: A Practical Metric
Why we measure inference efficiency in Joules per token, and how to do it repeatably on macOS.
Stay informed
Get notified when we publish new research or open access to our tools.
No spam, ever. We only email when we have something worth sharing.