Growth GarageA proposal for Calforests
Back to the overview

The rule making assistant

Rule making archive assistant.

The finished assistant will digitize the CDF-Cal Fire memos and the Board of Forestry rule making files. This page is a proof of concept. It shows we can pull the text out of the different kinds of documents that were provided, quote the page, and hand a question to a person when the record does not reach.

Dataset

CDF-Cal fire memos informing timber harvest and board of forestry rule making files

Demo corpus
8 files in this test
Board of Forestry and Fire Protection
1979 to 2018
Pages
1,441

Working demo

Ask the record

This test answers from the files we digitized for the proof of concept and cites the file and page. The finished build covers the full set of CDF-Cal Fire memos and Board of Forestry rule making files. The last question falls in a gap in this test on purpose. It shows what happens when the record does not reach.

Archive assistant8 files in this test · … passages
Nothing asked yet

Pick one of the questions below or type your own. Answers come from the files in this test and cite the file and page they came from.

Method

How it answers

Reads the discussion

The answers sit in the comment letters, the Board's replies and the meeting minutes, not in the regulation text. That is ordinary prose and it recovers cleanly from the scans.

Keeps replies attached

The record runs as a suggestion followed by the Board's response. Split those apart and an answer reports a rejected proposal as policy. They stay together and each is labelled.

Hands over when unsure

This test runs from 1979 to 2018 and does not include every year. A question about a year that is not in the test is handed to a person rather than answered from the nearest page.

Pilot

What we checked

1,441pages recovered from the scans, each answer traceable to one
2,735passages indexed and labelled by record type
8files in this test, chosen because the documents were not all the same kind
43pages flagged for review rather than left silent

Cost

One setup fee, then a monthly range. The monthly is not a single figure, because it depends on how capable a model we run and how far the answers and their reasoning need to go.

Setup

Building the assistant on the dataset above, with the scanned record, a citation back to the file and page, and a handoff when the record does not cover the question.

$15,000

Running cost

A lighter model sits at the low end. Deeper answers and stronger reasoning sit toward the high end. The figure is chosen when we settle how far the assistant needs to go, not before.

$150 to $450per month

Back to the overview