Corrections

What readers told us seemed unfair, and what we did about it. Trust from both sides is the whole product, so this page is public.

Privacy and data

PHILOS is built so that a school can adopt it without handing over its students. This page says exactly what we keep, where, for how long, and how to get it out.

What we store

• Without an account: nothing about you on our servers. A random device number is used only to count the daily free story; it is not linked to a name, email, or phone.
• With an account: your sign-in identity (email or phone, or a Google-provided identifier), your saved stories, topics, roots, categories, and settings, and your membership status.
• In a course: which course you joined, which assignments you opened, which lenses you read, and any exercise you submitted for grading. Instructors see students as "Student 1, 2, 3" unless a school's roster integration is configured.

What never leaves

The text sent to our model provider for analysis is the news article, or text you paste, and nothing else. No name, email, phone, course, grade, or account identifier is ever included in any request to the model. Your reading choices are not used to train any model.

Where it lives

Accounts and course data: Supabase (United States). App and cache: Render (United States). Payments: Stripe; we never see card numbers. Analysis: Anthropic, which processes article text under its API terms without training on it.

How long

Account and course data are kept while the account exists. Analyzed stories are cached for seven days and contain no personal data. Reader reports on the Corrections page are kept for the life of the program. When you delete your account, everything tied to it is removed immediately and permanently; backups roll off within 30 days.

Your controls

From Settings you can download everything we hold about you as a file, or delete your account and all of it. Instructors can export a course or delete it, which removes every student record in it. Students can leave a course at any time.

Sub-processors

• Anthropic (analysis of article text)
• Supabase (accounts and course data)
• Render (hosting and cache)
• Stripe (payments)
• Google (sign-in, only if you choose it)
• Twilio (sign-in codes by SMS, only if you choose phone)

We will list any change to this set here before it takes effect.

Questions or requests

Write to privacy@philos.news. For schools, a signed data-processing agreement and a FERPA addendum are available on request.

How PHILOS works

PHILOS exists to reinvigorate civil discourse. Every screen is designed so a reader on either side can check the work, and this page is the work.

1. What the model reads

For a published story, PHILOS fetches the article text directly. For a pasted email or document, it reads what you pasted. Nothing about you is included: no name, account, course, or reading history. The same text produces the same analysis for every reader.

2. Facts

A statement counts as a fact only if it is verifiable and the article attributes it to a source, document, record, or named person. Numbers keep their attribution. Claims presented as fact without a source, or that are disputed, are listed separately as contested. Context a fair report would include but this one omits is listed as well. There is no target count; one item or none is a correct answer.

3. Framing and the spectrum

A framing highlight marks a phrase whose wording favors one side: word choice, characterization, what is foregrounded. Each highlight names the side it favors and, in one line, what the phrase does to a reader. Highlights are notes, not accusations; emotion and vivid writing are not by themselves framing.

The spectrum is a score from −100 to +100 for the article's framing as a whole: 0 is balanced; the sign points toward whichever side the wording favors; magnitude reflects how much of the piece does so. It describes this article, not the outlet. The outlet's general rating is shown separately and is, today, the model's judgment within the reader's country; we intend to replace it with a licensed rating dataset and will say so here when we do.

4. The two sides

When a story is ordinary domestic politics, the sides are Left and Right. Otherwise PHILOS names the two principal sides as they name themselves. Each side's reading is written to a fixed standard: the strongest honest case a thoughtful holder of that view would make, the same length and seriousness for both, no strawmen, and no fallacies even where the original used them. The rule we test against: would a fair-minded person on that side accept this as their view?

5. Fallacies

Reasoning is flagged only when an actual argument is present and clearly matches one of seventeen named patterns: ad hominem, straw man, false dilemma, slippery slope, hasty generalization, cherry picking, false equivalence, post hoc, appeal to fear or consequences, appeal to popularity, appeal to authority outside its field, whataboutism, loaded question, begging the question, no true Scotsman, motte-and-bailey, and the genetic fallacy. Quoted speakers own their own fallacies; the outlet is not blamed for them. Strong language alone is never a fallacy. When nothing qualifies, the feature stays out of sight.

Known limits. Whether an inference is fallacious often depends on context the article does not supply; the model can flag a compressed but valid argument, or miss one dressed in neutral language. Every flag shows the passage and the reason so a reader can disagree.

6. History and the bridge

Timelines are researched with live web search and cite sources. Each moment records how each side responded at the time. The bridge names what both sides already agree on, what the disagreement is really about (a dispute over facts, a tradeoff between shared values, or different priorities), what each side sees that the other underweights, and questions worth asking in good faith.

6b. Exercises and the rubric

Say it back asks a reader to restate the other side’s position so well that its holders would accept it. It is graded 1–4 on four criteria: fairness (would a holder accept it as their own view), completeness (are the side’s strongest reasons present), charity (stated in the holders’ own terms with sincere motives), and accuracy (the claims attributed are actually theirs). Verdict: fair needs fairness 4 and nothing below 3; close needs fairness 3 and nothing below 2; otherwise strawman. Grading is against what the side actually holds, never against whether the grader agrees. Reflections written after a history root are graded on evidence, perspective, honest change, and clarity. The same rubric appears in every classroom worksheet, and instructors can read every submission and its feedback.

7. Where people decide and where the model decides

People wrote the rules above and the exact instructions the model follows, and people change them. The model applies them to each story; no person reviews each analysis before it appears. People review the reports readers send, decide what to change, and publish the decision on the Corrections page. An advisory council balanced across the political spectrum reviews the rules and the corrections.

7b. How well does the model agree with people?

8. Limitations

• The model is a careful reader, not an oracle. It can misjudge tone, miss context, or be wrong about which side a phrase favors.
• "Facts" means sourced and verifiable within the article, not confirmed true by us; the In context and Primary sources tools exist so readers can check.
• Outlet ratings are an estimate until a licensed dataset replaces them.
• Analyses are cached for a week; a correction to the rules reaches a cached story only when it expires or is cleared.

9. When we get it wrong

Every lens has a link to tell us it seems unfair. Every report is read. What we changed, and what we didn’t and why, is published on the Corrections page. How your data is handled is on the Privacy and data page.