GRADE — Granular Rating for AI Datacenter Efficiency

Granular Rating for AI Datacenter Efficiency (GRADE) is the aggregation and reporting engine. GRADE ingests all upstream engine outputs, applies published coefficients, computes the composite PTL Score, assigns an assessment tier, and produces the assessment report.

What GRADE produces

PTL Score — composite 0.0–1.0
Assessment tier — FRONTIER, OPTIMIZED, CAPABLE, DEVELOPING, BASELINE, or PENDING
HTML assessment report — score-first layout with engine findings, ATLAS recommendations
Assessment record — tamper-evident report ID with SHA256 hash

Composite formula

GRADE computes the PTL Score as a weighted average of engine scores using published coefficients:

PTL_Score = Σ(engine_score_i × weight_i) / Σ(weight_i)

Current coefficients (v2.0):
  ACE  × 0.35
  PACE × 0.25
  COOL × 0.20
  CORE × 0.12
  FLUX × 0.08
  ────────────
  Sum  = 1.00

When an engine is excluded from the assessment, weights are normalized proportionally:

normalized_weight_i = weight_i / Σ(active_weights)

Example: If COOL and FLUX are excluded:

active_sum = 0.35 + 0.25 + 0.12 = 0.72

ACE  normalized = 0.35 / 0.72 = 0.486
PACE normalized = 0.25 / 0.72 = 0.347
CORE normalized = 0.12 / 0.72 = 0.167

Worked example

Input (ACE only, MIT Supercloud — HPCA22 public dataset):

Engine	Score	Metric	Weight (rescaled)
ACE	0.339	gpu_efficiency_rate (GPU-hours weighted)	1.00

Calculation:

PTL_Score = 0.339 × 1.00 = 0.339
Tier: BASELINE (ACE only, first measurement)

Result: ACE Score 0.339 — BASELINE. 73,367 Slurm jobs from the MIT Supercloud HPCA22 public dataset. 33.9% of allocated GPU-hours did useful work (GPU-hours weighted). Per-job mean utilization is 0.257 — also in the findings, but not the GRADE primary metric.

Five-engine methodology illustration (hypothetical, showing formula):

Engine	Example Score	Weight
ACE	0.740	0.35
PACE	0.853	0.25
COOL	1.000	0.20
CORE	0.713	0.12
FLUX	1.000	0.08

PTL_Score =
  (0.740 × 0.35) + (0.853 × 0.25) + (1.000 × 0.20) +
  (0.713 × 0.12) + (1.000 × 0.08)
= 0.83781 → 0.838 — Optimized tier

This illustrates how GRADE aggregates all five engines. The specific numbers above are hypothetical inputs used to demonstrate the formula.

Composite scoring

GRADE weights engine scores using published coefficients and aggregates them into the composite. Engines for which data was not provided are excluded from the composite — they do not count as zero. Partial assessments are labeled by the engines included, so a two-engine assessment has a legitimate score on those two dimensions.

Engine weights are published in Coefficients.

Data quality tracking

GRADE tracks the input quality of each engine score and reports a composite confidence level. Not all engine inputs are equally verifiable.

Engine	Input type
ACE	Trace-derived — computed from Slurm sacct or DCGM telemetry
PACE	Trace-derived — computed from job wait-time metrics
COOL	Operator-reported — facility provides annual PUE
CORE	Operator-reported — facility provides hardware inventory
FLUX	Operator-reported — facility reports carbon accounting method

The composite confidence label:

Label	Meaning
HIGH	All present engines are trace-derived or directly measured
MEDIUM	Mix of trace-derived and operator-reported engines
LOW	All present engines are operator-reported

When ACE and PACE are both run, the composite confidence is MEDIUM at minimum — those two engines anchor the score in measured data. A five-engine score with all operator-reported inputs (COOL + CORE + FLUX only, no ACE or PACE) carries LOW confidence.

This is displayed in the assessment report and included in the assessment record JSON.

Tier assignment

PTL Score	Tier	Engines required
≥ 0.85	FRONTIER	All five engines
≥ 0.70	OPTIMIZED	Four or more engines
≥ 0.60	CAPABLE	Three or more engines
≥ 0.45	DEVELOPING	Two or more engines incl. ACE
ACE only	BASELINE	First measurement
None	PENDING	No engines complete

Assessment record

Every GRADE run produces an assessment record with a unique report ID in the format:

PTL-YYYYMMDD-ORGSLUG-TIER

For example: PTL-20260317-MITSUPERCLOUD-DEVELOPING

The report ID is paired with a SHA256 hash of assessment fields. PTL can review a report ID against its record when an organization chooses to share it.

Report structure

The GRADE report contains six sections:

Executive summary — tier, composite score, plain-language summary, top ATLAS recommendations, path to next tier
Engine findings — one card per engine: what was measured, the finding, the score, and what the score means
Key findings — five numbered findings with specific data callouts in monospace
Path forward — the composite gap to next tier, per-engine improvement potential sorted by maximum composite gain
Recommended next steps — ATLAS specific action text
Methodology — engine weights, tier thresholds, scoring methodology, confidence levels, disclosed assumptions

What GRADE does not do

GRADE does not soften findings. If ACE scores 0.339, the report says 0.339 and describes what that means in operational terms. PTL’s value is the specificity of the measurement, not qualitative adjustments that obscure how the number was produced.

CLI usage

# Generate one assessment report from a grade_input JSON file
grade certify \
  --input grade_input.json \
  --output grade_output.json \
  --report assessment_report.html

# With ATLAS recommendations section
grade certify \
  --input grade_input.json \
  --output grade_output.json \
  --report assessment_report.html \
  --atlas atlas_output.json

# Run all synthetic organizations and print tiers
grade demo

# Validate a ptl_output_v1.json export against the PTL schema
grade validate --input grade_output.json

Example grade_input.json:

{
  "organization": "NERSC Perlmutter",
  "period": "2026",
  "ace_report":  "ace_output.json",
  "pace_report": "pace_output.json",
  "cool_report": "cool_output.json",
  "core_report": "core_output.json",
  "flux_report": "flux_output.json"
}

Engines are excluded simply by omitting their report path — absent paths are not counted as zero. Partial assessments are labeled by the engines included.