Plain Theory Labs

Plain Theory Labs is an open, independent assessment framework for AI compute infrastructure. We measure what GPU clusters are doing, score those observations against a documented methodology, and deliver reproducible findings with disclosed assumptions.

What PTL produces

A PTL assessment produces a PTL Score — a number between 0.0 and 1.0 — derived from analytical engines running independently against the provided data. The score determines the assessment tier. The tier is shorthand. The score carries the detail.

Nine engines

Each engine addresses a distinct dimension of infrastructure performance:

Data collection. PROFILE characterizes your cluster — scheduler type, GPU fleet, telemetry sources — and routes data to downstream engines. CLAW is the intake agent. It installs as an OpenClaw plugin (openclaw install ptl-claw) and automates data collection inside NemoClaw environments.

Efficiency measurement. ACE (Adaptive Compute Efficiency Engine) measures GPU utilization from Slurm, Kubernetes, or DCGM telemetry. PACE measures scheduler efficiency — how well your cluster allocates resources across competing workloads.

Facility and hardware. COOL measures cooling system performance relative to a PUE benchmark. FLUX grades your carbon accounting methodology. CORE evaluates hardware-workload fit, fleet age, and embodied carbon.

Scoring and recommendations. GRADE aggregates engine scores into a composite PTL Score and produces the assessment report. ATLAS generates a ranked action plan — the specific changes most likely to improve your score in the next assessment.

Public dataset examples

These results were computed from publicly available GPU cluster datasets. Every number is intended to be reproducible from source data using open converters in the ptl-engines repository.

Organization	ACE Score	Telemetry	Jobs	Reference
MIT Supercloud	0.339	DCGM	73,367	Samsi et al., SC22
Alibaba Helios 2020	0.214	Hardware sensor	361,498	Weng et al., NSDI 2022
Microsoft Philly 2017	0.502	nvidia-smi	74,020	Jeon et al., USENIX ATC 2019

ACE reports two efficiency metrics for each dataset:

• GPU-hours weighted efficiency (GRADE primary) — of all GPU-hours allocated, what fraction did useful work. This is the primary GRADE input.
• Per-job mean utilization — average per-job GPU utilization, equal weight per job. This is what most facilities self-report internally.

GRADE uses the GPU-hours weighted metric because it reflects infrastructure efficiency in physical terms: wasted GPU-hours equal wasted energy. The public examples are comparable with published EE HPC WG benchmarks for GPU environments.

Methodology worked examples using public datasets are available in ptl-methodology/VALIDATION.md with full assumptions documented. These are illustrations of the methodology, not independent endorsements of any facility.

Longitudinal assessment

A PTL assessment is designed to be repeatable. Year one is a baseline. Year three is a trend. Organizations that improve from DEVELOPING to CAPABLE to OPTIMIZED across assessments have a documented record of what changed and how the score moved. PTL maintains the assessment method. You own the findings.

Get started

Read how assessment works or start a pilot.

Repositories

Every number PTL produces is reproducible from source. The engines, methodology, and this documentation site are all public.

Repository	Description	License
ptl-engines	All nine analytical engines, public fixtures, dataset converters, 1,426 passing tests	MIT
ptl-methodology	Scoring formulas, coefficients, tiers	CC BY 4.0
ptl-website	This documentation site	MIT
ptl-context	Engineering session logs and internal notes — not methodology or scoring logic	Private

All methodology is public and citable. Source code is open under MIT license. The organization is at github.com/plain-theory-labs.