⊕ The Political Atlas · projection layer · established August 2026
Navigators use dead reckoning when there is nothing to sight against: you work from known heading and speed, you plot a position, and you accept that error accumulates. That is exactly what this section does. One documented source, six axes, every score published with its reasoning and open to challenge.
The problem this addresses
A forecast gets published, a politician gives a two-hour interview, a new movement puts out a programme — and within a day the commentary has decided it is authoritarian, or socialist, or fascist, or a threat to democracy. Those are all claims about position. None of them name an axis, state a scale, or offer a way to be wrong.
The Atlas exists because a single left-to-right line has to carry six independent questions at once, and compressing six numbers into one does not blur the information, it destroys it. That argument applies with more force, not less, to things that have not happened yet. When there is no governing record to point at, the temptation to substitute a moral verdict for a measurement is at its strongest.
So this section does the harder version. It takes one source that already has an audience, applies the same anchored codebook used on all 319 Atlas organisations, publishes every rationale beside a timestamp in the original recording, and invites you to score the same evidence before you see what we scored.
The failure counter is deliberate. A reading where the six axes cannot represent the subject is a finding about the instrument, and it gets published at the same weight as any other result. If that number stays at zero, something is wrong with our honesty rather than right with our tool.
House rules
| Rule | What it means in practice | Why it is here |
|---|---|---|
| No reading leads with a distance | A composite figure may appear in a table. It never appears in a headline, a summary, a social post or an image. Every reading leads with which axis carries the change. | A stripped number is one screenshot away from becoming "the Atlas rates X worse than Stalin." Composition cannot be stripped, because composition is an argument. |
| Right of reply, offered before publication | Where a reading scores a living person or an active organisation, the subject is sent the scores and rationales in advance and offered unedited space beside the reading. Status is shown on every page. | It is the single convention separating analysis from a hit piece in the eyes of any editor who might pick this up. It also produces the best story when a subject engages. |
| Every score deep-links to the tape | Each rationale carries a timestamp that cues the embedded source to the exact second. Written sources get a paragraph-level anchor. | An auditable trail is the difference between a reading and an opinion. It also means a hostile reader can check us in fifteen seconds, which is the point. |
| Predictions are dated and scored in public | Each reading commits to falsifiable expectations with review dates. On the date, the ledger records correct, wrong or unresolvable. | Being publicly wrong on schedule buys more credibility than being vaguely right forever. Every review date is also a fresh reason to return. |
| Null readings are published | Where an axis cannot represent the subject, we say so and explain the failure rather than forcing a number. | A site that advertises how often its own tool breaks is doing something no engagement-farm would ever do. |
| The selection rule is published and the balance is tallied | Subjects are chosen by the stated rule below, and the running tally of what has been read is kept on this page whether or not it looks good. | Selection bias, not scoring bias, is how an instrument loses its claim to symmetry. Individual readings can each be careful while the choice of subjects is not. |
The readings
Daniel Kokotajlo's AI 2040: Plan A proposes five different answers to what happens in 2029. Each answer is a different kind of state. Two of the six axes return no valid reading, and the reason they fail is the most interesting thing in the piece.
Strip the technology from Plan A and what remains is a programme: a universal dividend, a binding supranational treaty, radical transparency, and a deterrent held in reserve. Coded as though it were any other organisation seeking power, it lands somewhere very specific — and not where its critics or its supporters assume.
Both readings run off the same recording. That is deliberate: it demonstrates that one source supports two entirely different questions — what kind of world does this scenario produce, and what kind of political actor is the person proposing it.
Selection
A reading is commissioned when a source meets all four tests. The rule is fixed in advance so that the choice of subject is not itself an argument.
| Broad position of subject | Readings | Subjects |
|---|---|---|
| Left of centre | 0 | — |
| Right of centre | 0 | — |
| Cross-cutting or not placeable on a left–right line | 2 | AI Futures Project; the Plan A scenario space |
| Governments and state actors | 0 | — |
Two readings is far too small a sample for this table to mean anything, and both are of the same source. It is published anyway, from the first reading onward, because a balance ledger that only appears once the balance looks good is worthless. Judge it at ten readings, and hold us to it.
Disagreeing
The Atlas method page already concedes the thing most rating projects hide: it has a single coder, and inter-coder reliability cannot be self-reported. Dead Reckoning turns that admission into the main feature. Every reading opens its scoring to you first — same evidence, same anchors, score withheld — and then shows you the gap.
Where you disagree by more than two points, the Atlas asks you to file a challenge with your reasoning. Those challenges are the reliability data the method page says is missing. The mechanism that makes this section worth arguing about on social media and the mechanism that closes a real hole in the science are the same mechanism, which is a rare and fortunate thing.
No reader-agreement distribution is published anywhere on this site yet, because none has been collected at a size worth reporting. The coding instrument keeps your scores in your browser and says so. When the pool passes one hundred independent coders, the distribution publishes with its n visible. A number invented to look impressive would destroy the only asset this project has.