Archived v6 diagnostic. These results predate the current model. They have not been rerun for the September 16 audit changes and do not validate current site scores. Read the model review.

Reproducible local comparison · 2026-09-14

How do reported data-center locations score?

The public inventory yields 67 campus anchors. In 43 local matched comparisons, their median score difference from alternatives is -1.1 points.

This is a diagnostic of the model, not proof that it predicts development. No scoring weights were fitted to these projects. Appearing in Cleanview earns no points. Campuses share geography and infrastructure, so the number of records overstates the amount of independent evidence.

A consistent project scenario

Every location is evaluated as 100 MW peak facility demand, 90% steady utilization, PUE 1.25, WUE 0.5 liters/IT-kWh, target 2031, minimum 115 kV and a 25 km study radius. Facility demand already includes overhead; PUE is not applied again to the energy bill. Reported project MW has uncertain IT-versus-facility basis and is kept separate. This does not reconstruct any campus's actual capacity, electricity contract or building design.

Prices are historical wholesale nodal values from 2025, not retail tariffs. Environmental snapshot: 2827bcdb6114. Selecting 2031 does not forecast utility energization timing or future prices.

Results by reported status

Local comparisons allow 50 km separation; regional comparisons allow 200 km. Both match developed-land intensity, voltage and grid-access characteristics. Operating, construction and planned cohorts remain separate. A planned project is not a successful outcome.

Before and after the methodology correction

The same matched locations and project assumptions are used for both methods. The recorded v6 model treats generator-queue activity and nearby generation as context rather than evidence of available load service or contracted supply. Removing unsupported proxies changes the score's known-evidence normalization; a higher score is not proof of improved economics or prediction accuracy.

What drives the local comparison?

Component scores are shown before weighting. Unknown inputs are excluded, with the number of known pairs shown. Load-service readiness remains unscored until suitable service-capacity and timing evidence is available.

Geographic concentration

The 67 campus anchors share 24 environmental samples and 41 price nodes. A 5 km screening sample does not establish parcel conditions; adjacent urban sites can share evidence or inherit nearby non-urban land cover.

Source quality and comparison rules

The source snapshot has 86 records, including 84 public map points; 76 pass status, coordinate and California land checks. 10 are excluded with reasons in the downloadable data, including 7 whose published point disagrees with its reported county. Coordinates are not guessed or corrected to improve results. Published coordinates are not surveyed parcel locations, and this inventory is not a census of all California data centers.

Buildings belonging to an identified campus contribute one anchor, chosen by maturity then stable source ID. Mixed-status memberships remain in the data. Alternatives are mapped substations at least 1 km from every source project, deduplicated geographically. Their absence from Cleanview does not establish that they are vacant, developable or without other data centers.

Matching requires the same voltage band and broad NLCD developed-intensity class, substation-distance difference at most 2 km, line-distance difference at most 3 km, previous-method non-queue coverage difference at most five percentage points, and previous-method grid-score difference at most eight points. Overall, operations and market scores never select matches. Each alternative is used once per comparison scope. There are 24 unmatched local anchors and 23 unmatched regional anchors; criteria are not relaxed to hide them.

What is still missing?

Utility will-serve commitments and energization dates, retail tariffs and demand charges, carrier access and latency, parcel and connection costs, local approvals, actual cooling and water supply, backup-generation permits, and contracted clean electricity remain material gaps. Middle-mile routes and nearby generating plants do not establish these services.

2025 prices and later public evidence cannot validate predictions made before older campuses were built. A future predictive test must freeze dated inputs and reserve whole metropolitan areas or utilities before tuning weights. These results should guide which missing evidence to add, not teach the score to copy the observed map.

Data and reproducibility

This comparison made zero external requests. The runner reads the same prepared files used by the app. Extract the previous evaluator archive and run node siting/scripts/validate-datacenter-siting.mjs --baseline=path/to/baseline-v5/model.js. The JSON contains every score, component, match, source exclusion and consumed-file hash.

Recorded module hashes describe files at evaluation completion. Unscored inventory context and interface additions were made during the first run; numerical scoring semantics remained unchanged. Linked-campus scores are checked against the final browser runtime separately. The previous evaluator was frozen before changes.