An AI first-pass reviewer for PJM interconnection submissions

To plug a new data center into the grid, the developer files a multi-hundred-page application packet with the regional grid operator (in the mid-Atlantic, that is PJM Interconnection, the nonprofit that runs the wholesale electricity market across 13 states). Weeks later, a reviewer often emails back a short deficiency letter saying the packet is missing a piece of data, contains a self-contradictory number, or labels the grid-connection point ambiguously. The project then drops to the back of the review queue and loses a full cycle, which can be months. Most of those errors are catchable on the developer's side before submission. Almost none of them are caught today.

The Cliff submission validator is a tool that pre-checks those filing packets. It reads the developer's application against the grid operator's rulebook (the 180-page tariff) and its detailed study manuals before the developer hits send. It returns a readiness score, a ranked list of findings at four severities, the specific rulebook section that backs each finding, and a recommended fix (sometimes an auto-patch). Roughly ten seconds of language-model inference replaces four weeks of waiting in the grid operator's review queue.

This is only possible because frontier language models can finally read a 180-page tariff, a 60-page study manual, and a 40-field application form together in one context window and reason about how they interact. Pre-2023 rule engines could catch obviously missing fields. They could not tell you that the curve describing how your generator supplies grid-stabilizing reactive power is inconsistent with the fuel type you claimed, or that the in-service date you requested is impossible given your slot in the study queue. That is the difference.

Start on the sample submission to see what the output looks like, or bring your own at /submission-validator/run.