Toggle Daily · Methodology
Toggle Daily is built by two kinds of machine work. Software with no AI involved does the collecting and the measuring. That software was written by me, a human, and like any human-written software it can have edge cases where an error flows through. Errors and corrections describes how those are handled. An AI language model writes the text that presents the results. This page describes the boundary between the two and the checks that run on the model's writing.
The software follows the day's topics, gathers each event's coverage, identifies the information each outlet reports about the event, attributes each piece of information to the outlet that reported it, and compares the accounts across outlets for agreement and disagreement. It also computes the three shares of reported, contested, and developing described in Reported, contested, developing. None of this output is generated by a model.
A language model is the kind of software commonly called AI. The language model writes the event summaries, the panels, and the audiocast scripts, working from the day's verified information and nothing else. It does not assign, adjust, or override any number. The voices on the audiocasts are synthetic text-to-speech, and every audio page states this.
Toggle is model agnostic. We use the latest and most capable models available, today Claude Fable and GPT 5.6 Sol, and set one to check the work of another. The model matters less than the algorithms built to enforce the rules on what it writes.
Everything the model writes must clear a set of automated checks before release. The checks are code. When one fails, the release stops until the text is rewritten.
Quotations are verified word for word. Any quotation of five or more words must appear, character for character, in the day's source material. A quotation that has been trimmed, smoothed, or invented stops the release.
Attribution is verified event by event. A quotation or claim from one event's sources may not appear in another event's text. A speaker named in the text must be the person the source quotes.
Labels are verified against the data. Text that describes an event as corroborated when the computed shares say it is still developing is rejected.
The writing must be original. Quotation marks are reserved for words a person actually said. Journalist writing is never reproduced. That covers an outlet's sentences, its headlines, and its titles. What an outlet reports is stated in Toggle's own words, with the outlet named and the original article linked. A check compares the unquoted text against the day's source material and rejects any passage that reproduces a source's wording.
The checks make errors rare. They do not make them impossible. Errors and corrections describes how the ones that get through are handled.