MOD-10

Atlas

ConceptNews

One objective world, made visible. Facts stripped of voice, placed on a living globe, at every scale.

An idea with research behind it, documented before anything is built. Moduloa’s active build is Manufacturing.

The problem

Today the same event runs under two entirely different headlines, because most outlets answer to a leaning before they answer to the reader, and the reader inherits the bias along with the news. What is missing is not more coverage. It is a neutral layer: the facts, separated from the voice that carries them.

Not an outlet: an atlas

An outlet has a voice; that is what an outlet is for. Atlas deliberately has none. It is a record. An AI research agent reads many articles on the same event (across outlets, languages, and leanings), keeps what independent accounts agree on, marks what only one side claims as exactly that, and writes one objective account. Then it places that account on the globe.

The interface

The globe

01 · Two pins per storyEvery account carries its source and its subject. A journalist publishing from Oslo about a border conflict between Thailand and Cambodia appears twice on the globe: where the report came from, and where the thing is happening. The distance between the two pins is information in itself.

02 · The risk layerAI-generated graphics drawn from the digested reporting: where forces are massed, where a front actually stands, what is under whose control. An active front becomes one readable, current picture of risk instead of a thousand contradicting headlines about it.

03 · Every scaleThe machinery is not reserved for wars. A grandmother in a small Norwegian town baking cookies to sell is an event with a place: on the globe she is visible to anyone nearby, who can walk over and buy them. The smallest news gets the same treatment as the largest.

The field as it is

This is a crowded field, and right now it is shrinking

Nothing above describes an empty market. Hundreds of organisations already do the fact layer by hand, a standards body codifies the method, and a machine-readable output format exists. The opening is not absence: it is that the incumbents are losing money and distribution at once.

437Fact-checking projects counted worldwide, 1 June 2026

10 vs 30+Launched versus stopped posting across 2025, on the census’s own two counts

01 · Someone already counts themThe Duke Reporters’ Lab census counted 437 active projects as of 1 June 2026, across 116 countries and 70 languages, down from 441 at the end of 2025. Across 2025 the lab counted ten launches against more than thirty that stopped posting; the running total moves by far less than that, because the census keeps adding projects it had not previously counted (its own caveat, not ours).

02 · The method is already codifiedThe IFCN Code of Principles sets five commitments, the first being this page’s entire claim: “fact-check claims using the same standard for every fact check,” without concentrating on one side. Then sources “in enough detail that readers can replicate their work,” transparent funding, a published methodology, scrupulous corrections. Someone wrote this down first.

03 · The format exists; its main consumer just leftClaimReview is a schema.org type for “a fact-checking review of claims made (or reported) in some creative work.” Google’s structured-data documentation now opens with “We’re phasing out support for ClaimReview markup in Google Search,” leaving it live only in the Fact Check Explorer tooling. Full Fact’s Andrew Dudfield told Poynter fact-checks “now risk being just another result.”

04 · The constraint is money, not the ideaMeta announced in January 2025 that it would end its US third-party fact-checking programme, arguing in the announcement that “a program intended to inform too often became a tool to censor,” with Community Notes phasing in “over the next couple of months”. The announcement is what that document supports, and we do not cite it for a completion date. Three-quarters of the hundred-plus fact-checkers the IFCN surveyed report being financially vulnerable or in crisis, and the same census records Logically Facts ceasing publication after losing its Meta and TikTok revenue.

05 · Europe’s instrument is voluntaryThe disinformation code was folded into the Digital Services Act framework, with its commitments auditable from 1 July 2025. It is worth naming precisely rather than treating as a counterweight: it binds signatories who choose to sign, and signing is a choice each platform makes separately. A voluntary code and a withdrawn markup standard are not opposite forces of equal size.

The hardest shelf here

Finding the claim is the easy half: settling it is what nobody has solved

The literature splits this into claim detection and claim validation. Detection is the tractable half. Validation (what is true, from evidence, on a claim nobody has adjudicated yet) is where the best system in the one shared task measured below still misses more than a third.

01 · The benchmark numberThe 2024 AVeriTeC shared task counted a claim verified “if and only if both the verdict is correct and retrieved evidence is considered to meet a certain quality threshold.” Across 21 submissions the winning system scored 63%, so more than a third of claims came out misjudged or unsupported, on claims humans had already settled. One shared task on one dataset is one measurement, not a law.

02 · Earlier benchmarks flattered the machinesThat dataset exists because previous ones had “substantial limitations, such as relying on artificial claims, lacking annotations for evidence and intermediate reasoning, or including evidence published after the claim” (the authors’ words). The last is temporal leakage: scoring a model after the world published the answer. Its replacement holds 4,568 real claims from 50 organisations.

03 · The serious builders aim lower on purposeThe survey literature splits the task into detection and validation, and the standard framing for deployment, Nakov and colleagues, is technology that “support[s] the human expert” rather than issuing the verdict. Both surveys are from 2021 and measure nothing themselves; they are cited for the framing, not for a number. Atlas as written is the more ambitious claim, and it should be read as the less proven one.

04 · Rating the raters fails in documented waysBias charts are the nearest shipped product, and their limits are conceded by their authors: AllSides told Poynter “our five-point scale is inherently limited,” and Poynter’s Kelly McBride warned overreliance “is going to probably give some consumers a false level of faith.” Where fact-checkers overlap, though, the evidence cuts against the cynical read: a PLOS One comparison found “near consensus about whether the statement was true or false,” with agreement falling away toward half-truths.

05 · Selection is the measurable biasNot “no viewpoint”. That is not available. Which claims get picked is itself a position, and unlike tone it can be counted: in that same comparison, 116 of 513 statements PolitiFact checked, 22.6%, never appeared in The Washington Post’s database. Whatever “unbiased” can honestly mean here has to include publishing the selection rule, because the selection is where the measurable bias lives.

What would have to be true before this starts

A test, not a promise. Atlas stays a concept until a pipeline can take claims already adjudicated by IFCN signatories, judge them blind, and publish the agreement rate and every disagreement, clearing 63% on that joint verdict-and-evidence measure, with no temporal leakage and a published selection rule.

How it stays honest

Method, not promises. Every account publishes its source list, what was corroborated and by how many independent reports, and what remains single-source. Bias is not defeated by claiming neutrality; it is defeated by showing the work. The same open-by-default rule as everything else on this site, pointed at information.

Honest status

Drafted and researched, nothing built. The full globe assumes AI that can digest the world’s reporting in real time, closer to AGI than to today’s tools, and this page says so plainly. The fact layer (many articles in, one objective account out, sources shown) is buildable far sooner, and is where the work would start.

Documentation, decisions, and build logs accumulate here. Everything is public.