Campaign — One letter sent by 10 or more people, copied or lightly reworded. One of 130 submissions in its group; the sender added words of their own. See the letter, its submissions and topics.
Carries the letter's score — A copy of a family's letter; it holds the score and answerability level of the letter it sent.
Each dimension is scored 0–3; the eight sum to the substance score out of 24.
A0 none: Counted, not answered. This rating is the one its shared letter earned.
These results are meant to be shared.
How it works. A large language model (an open-weight Qwen3.8 model) reads each substantive comment and copies, word for word, the passages that carry its analytical gap, evidence, request, alternative and connection to the place; it returns no scores. Code grades those passages 0–3 on eight dimensions — evidence, legal grounding, engagement with the environmental analysis, analytical gaps, alternatives, specific requests, position, and local knowledge — with legal grounding taken from a citation scan, not the model. For the strongest tenth on each side, the copied passages are checked against the comment and shown as exhibits. A comment counts as “substantive” when it clears a floor filter — it must show at least one substantive signal (a named entity, a specific citation, a stated analytical deficiency, or first-person testimony), which screens out one-line and purely conclusory submissions before any model scoring.
The full method. Every step, why it is done that way, its known limits and a glossary: Methodologies →
The data. The full canonical dataset — every comment on Docket FS-2025-0001 — is available from Regulations.gov: https://www.regulations.gov/bulkdownload.
Roadless.org — analysis of the public comment docket.
Media inquiries welcome — nicholas@wanderingnature.com.
© 2026 Roadless.org, Wandering Nature, LLC. All rights reserved.