Toggle Daily · Methodology
Toggle is not a news site. We take real events reported across multiple news wire services and outlets and distill them, leaving out the opinions and the editorializing. Human-designed algorithms scan the day's news topics and pick the five to ten events with the most activity, then find the facts most cited and most contested across each event. A language model, the AI, compiles a report of the topic, separating what is corroborated from what is contested. A further set of algorithms checks that report, including a check that facts from one event are not contaminating another event on a similar topic: a quote from an official in one event can never appear in the report of a different one.
We are also 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 we built to enforce our rules.
We only provide information on real events, and we try our best to attribute everything, every time. The Reported, Contested, and Developing shares on every event are computed by algorithms from what the outlets themselves published: information carried by independent accounts counts as corroborated, information the outlets disagree on counts as contested, and information with only one account behind it stays developing. No model assigns, adjusts, or overrides those numbers.
Everything the model writes goes through automated checks before release, and they are code, not judgment calls. Quotations of five or more words must appear character for character in the day's source material. A fact from one event's sources stays in that event's text. An outlet named in an event must be one of that event's sources. Text that calls an event corroborated when the data says it is still developing is rejected: the data wins over the words. When any check fails, the release stops until the text is regenerated.
Yes, in two layers. The first is a multi-gate set of algorithms every build must clear before assembly: an event-quality check on the source material, a word-for-word quote check, an attribution check that ties every fact to its own event and its own speaker, an editorial check that the on-screen labels match the underlying data, and an audit that confirms recently added features made it into the build. One failure stops the release. The second layer is me: I read the final copy of every build before I upload it. The gates run the same way every day, however busy the day gets. The final read is mine.
By our own accounting, Toggle is currently around eight to ten percent AI, and we want to be below 5%. AI is a great way to get started but our goal has always been to use AI to create MVPs or minimum viable products. From there we whittle away at the AI models and replace them with our own tools to create the final goods and services. Getting there costs money: voice-over work alone costs more than I can dedicate to a project with no timeline to profitability, so the audiocast voices are synthetic and every audio page says so. The rest of the pipeline is neither AI nor plain code; it is well-thought-out algorithms, designed by a person, that do the collecting, the sorting, and the checking. The model's job is narrow: turn a checked list of facts into readable sentences, then submit them to the algorithms that verify what it wrote.
No. Every voice is synthetic text-to-speech, and every audio page says so above the players. If you would like to volunteer to voice some of our audiocasts, reach out. We would like to hear from you.
No. Toggle is an information service. The reporting belongs to the outlets that did it, and they are named and linked on every event. We publish the information around an event: the news outlets dictate what the facts are, and we keep track. When the information differs from one account to another, that is contested. When the accounts line up, that is corroborated. We let corroboration and contest surface on their own. When an event lacks enough information, it is considered developing. We do not break news, we do not editorialize, and we do not rank outlets in what you read.
We correct it, note the correction with a date, and list it on the corrections page, which carries the running tally of every error that got through.
From a fixed list of named news organizations, gathered fresh each day. Unlisted outlets can corroborate a fact but never lead the reporting on an event and are never quoted. Every event page lists a sample of its sources with direct links. Read the original reporting yourself; we encourage it.