FAQ & Challenges

Project Overton is a tool that measures where written content sits on seven independent questions of political-economic ideology. It does not measure bias (which implies a correct center), quality, or accuracy. It measures position — and shows you the evidence behind every score.

Who owns: What is privately owned, and what is held in common?

Who provides: Can people get what society requires of them, such as schooling and health care, without already having the money to buy it, and may everyone use that route?

Who pays for harm: When something a business or a person does might harm other people, who carries the burden: the one acting, or the people exposed?

Who decides: How much say do the people a decision binds, everyone it places a duty on or protects, have in making it?

Who counts: May some people count for less than others, by birth or by legal status, and can a person leave the rank they are given?

Who has a say: Does government answer to everyone equally, or to a few?

Who sets the terms: When people make a deal, such as a job, a lease, a loan or a sale, who sets its terms, including what it costs to get out of it: the stronger side alone, the two sides with a check on each other, or a public rule?

Each question is its own axis, scored independently from -4 to +4. When an article does not take a question up, that question is marked No Signal.

Because "left" and "right" collapse seven different questions into one word. An article can lean left on one question and right on another, and either side of politics can land at either end of a question. A conservative article defending parents' say over their children's schooling leans left on Who decides, and a progressive article defending an emergency order that no one it binds can refuse leans right. One axis cannot capture that. Seven questions can.

Everyday fights such as guns, immigration, housing and wages are not axes of their own. They are topics, and an article about one of them is read on whichever of the seven questions it takes up.

-4 is one end of a question and +4 is the other. 0 is the centre: an arrangement of its own, which both ends have complaints about, and not simply whatever is normal today. Each of the nine rungs has a name and a description of the arrangement it stands for. The definitions come from political philosophy, not from current media norms. An article that does not take a question up is marked No Signal on it, never 0.

No. Bias implies a correct center. We do not claim one exists. We measure position on defined axes. An article at +2 on Who owns is not "biased right" — it is at a specific, documented position on a specific dimension, with evidence you can examine.

Through structured analytical lenses that examine the article at multiple levels, from surface-level framing down to what is missing entirely. Each lens records observations tied to specific evidence in the text. For each axis, the tool then names the arrangement the article treats as proper and matches it to the rung of the rubric whose definition it fits. That rung is the score.

So that the score belongs to the article and not to chance. A language model does not give the same answer every time it reads the same text. Each article is read three times, independently. When the readings agree closely on a question (two place it on the same rung, and the third places it on that rung, one step away, or at No Signal), that rung is the score. When they are further apart and at least two of them found a position, the question is shown as a range, and the page explains what each reading found.

Until October 2026, several scoring models converted the observations into numbers, each by a different method. They disagreed with one another, and the same article could score differently from one run to the next, so they were retired in favor of the three readings described above. An analysis scored before that date says so on its own page.

Yes. A large language model reads the article through structured analytical lenses, records what it observes, and states for each axis which arrangement the article treats as proper and which rung of the rubric that arrangement fits. Each article is read three independent times, and a question gets a single score only when the readings agree closely. Every sentence on the card is checked against the rung it explains before it appears.

Yes. Every score comes with an evidence chain: the score, the rubric level it matched, the exhibits from the article (described in our own words, with short exact quotations where the wording itself is the evidence), and the observations that produced them. If you disagree with a score, the chain tells you exactly where your disagreement sits.

Every score has an evidence chain you can follow back to the article text. The tool does not claim objectivity. It claims auditability — every link in the chain is available for you to examine. If you disagree, you can trace the chain to find the specific observation or rubric match where your interpretation diverges.

Any reader, human or machine, can introduce variance. The mitigations are a rubric specific enough that the choice of reader matters less than the rubric itself, three independent readings that must agree closely before a question gets a single score, and checks on every sentence and quotation before publication. Inter-rater reliability testing with human scorers is planned; it has not yet been conducted.

Anchors are real-world reference points: specific policy actions whose positions have been determined from governance records. Attlee's NHS, the Tennessee Valley Authority, and Britain's water privatisation are examples. Every anchor is sourced from primary documents — legislation, executive orders, budget data — not from summaries or impressions.

The anchors were scored on the five axes the rubric used until October 2026. They are out of the menu until they are re-scored on the seven questions, and until then article scores are not compared against them.

We do not select — we filter. For each presidential administration since 1900, we score every piece of legislation that created or eliminated a durable institutional change lasting 10 or more years. An agency, a program, a legal framework, a right. If it meets that threshold, it goes in regardless of where we think it will score. The criterion is durability, not ideology.

International anchors cover arrangements that no US administration has reached.

The library covers every major policy domain across US presidents since FDR, plus key international governance actions. Each anchor is a specific policy action scored from primary source documents — legislation, executive orders, court opinions — and independently verified. The library is closed while it is re-scored on the seven questions.

No. All scores are unvetted drafts. They become more confident through the challenge process — if someone challenges a placement with evidence and the placement survives, it earns verified status. Scores can also be revised when new evidence emerges. Both the original and revised scores are preserved with full documentation.

The anchor is revised. Both the original and revised scores are preserved. Nothing is ever silently updated. The version history is permanent.

You are right. This tool does not measure everything. It measures one thing — how power is distributed or consolidated — across seven questions, in written media coverage. That is a defined scope, not a theory of everything.

A piece of music criticism has aesthetic dimensions the tool does not touch. A sports column has drama and narrative it does not score. Not everything is about power, and the tool does not claim it is. When an article does not engage these questions, the tool says so. Low salience is a valid finding. "No signal" is a valid output.

That said, the pieces that seem furthest from power — the tech profile, the lifestyle piece, the education feature — are often where the most embedded assumptions live. They do not think of themselves as political. That is exactly what makes their assumptions worth surfacing.

The tool does not investigate topics. It analyzes how articles about any topic are framed — whose perspective organizes the story, what is assumed without being argued, what alternatives are absent, and where the framing sits on seven defined questions.

Climate change is a topic, not an axis. An article about it is read on whichever of the seven questions it takes up. It can engage Who owns (who owns energy infrastructure: public utilities or private energy companies) and Who pays for harm (whether the costs of pollution and a warming climate fall on the businesses that cause them or on the people exposed).

The tool examines whether a climate article treats fossil fuel companies as "the energy sector" without naming public utilities as an alternative. Whether it frames transition costs as a burden on consumers without asking who profited from the status quo. Whether it proposes market solutions without considering public investment or community-owned renewables. Each of these is a measurable observation on a defined axis.

Racism is not absent from the framework. It is built into Who counts, which asks: May some people count for less than others, by birth or by legal status, and can a person leave the rank they are given? That includes racial rank. It also surfaces on the other questions when racial dynamics are structurally relevant.

The tool examines whether a piece about crime treats criminal behavior as individual moral failure without naming structural relationships between criminalization and race/class. Whether a piece about education ignores funding formulas tied to property values that perpetuate segregation. Whether vocabulary carries racial coding. It surfaces racism not as a cultural attitude but as a structural feature of specific policy domains, measurable in how articles frame those domains.

Each question reads the arrangement an article endorses, never the party behind it. An article on the right praising employee-owned companies leans left on Who owns, and an article on the left praising a founder's fortune as the reward of genius leans right. The tool flags left-coded framing in Jacobin with the same rigor it flags right-coded framing in the WSJ.

Each question is defined by its two ends, each put in the words of its own supporters, and by a centre that both ends have complaints about. On Who owns, the ends are “Nobody owns a business” and “Everything privately owned.” If someone disagrees with a placement, they can challenge the definition of the question or the specific evidence.

We agree. That is why we do not use one number. Every article is read on seven independent questions, and a piece can be left on one and right on another. Each question also tracks how centrally the piece engages with it, and whether the position is argued explicitly or assumed implicitly. The score card shows each question separately. The evidence chain shows the reasoning. No single number appears anywhere.

Yes. We are building a formal challenge submission system where anyone can contest a placement with specific evidence. If you cite real governance records or scholarship that argues for a different placement, it triggers a review. If your evidence is stronger, the score changes. Until the formal system is live, submit challenges by email to challenge@projectoverton.org with the article or anchor name, the question, your proposed score, and your evidence.

A free account is browse-only. You can read all analyzed articles on the homepage, access all published reports, and explore the full methodology documentation. You cannot run your own analyses on the free tier. Analyzing new articles requires a Pro subscription.

Pro ($15/month) is the full product. It includes 100 analyses per month, full evidence chains on every analysis, evidence files with primary source links, and exports. The Anchor Library is closed while it is re-scored on the seven questions. Three-band profiles (speech, platform, governance) are coming soon and will be included with Pro when available.

Platform ($50/month) is for researchers, journalists, and analysts who want to build and analyze their own datasets using the same tools the project uses internally. It includes 500 analyses per month, batch article upload, all 10 analysis types including Tilt/Shift and corpus-level analysis, a query interface across scored datasets, custom evidence files, and API access. Pro users consume individual analyses. Platform users produce datasets at scale and then run analytical tools on those datasets.

Enterprise is for institutions that need multi-seat access, custom measurement axes, custom analytical weighting, or institutional workspaces. It is not a separate tier on the pricing page. Contact us for details.