Reported capacity and operating capacity are different facts
Behind the Watt is the open evidence database I built to track behind-the-meter power at data centers. Its genesis dataset begins with three verified operating facilities totaling 1.03 GW.
That is not a market total. It is the first set of facilities for which I can publish the receipts. A separate public inventory lists 143.3 GW of reported gas projects. I keep those numbers separate because announced, permitted, under construction, and operating are different states.
The database is a pet project, but the method is meant for any workflow where a claim has to survive contact with a source document.
AI extracts the claim, then the system tries to disprove it
The system watches public registries, archives original documents, and uses AI to extract claims. It does not publish the extraction as truth.
- The quote must be found in the source document.
- The number must be relocated on the page where it appears.
- A second model cross-checks the claim.
- A human approves publication.
This reverses the usual trust request. The reader does not have to believe the model. The reader can inspect the document and repeat the check.
The most useful result came from aligning dates
The sharpest result so far is a time-to-power value of minus seven months: one facility was operating before its air-permit application was filed.
The value is not an AI opinion. It is a relationship between dated source records. AI helped find and structure the claims, but the conclusion remains inspectable.
This is where AI becomes useful for finance, legal, analytics, and construction teams. The system can read volume, but the deliverable should expose the exact evidence behind a decision.
Open formats make the method testable
The data is available under CC BY 4.0 as CSV, JSON, OpenAPI, and read-only MCP. The point is not only to publish a number. It is to make the evidence usable by people and machines.
Behind the Watt has no funnel behind it. I built it to test the method in public and to keep my own AI claims honest.
Use AI for the volume. Make the source do the persuading.