⊕ Dead Reckoning · method and limits · v1.0
The Atlas earns its standing from one rule: every score carries a rationale and at least one citable artefact. Readings cannot satisfy that rule, because the thing being scored has not governed or has not happened. This page sets out what they must do instead.
The distinction
| Atlas entry | Dead Reckoning reading | |
|---|---|---|
| What is scored | An organisation's stated positions and its governing record | A published forecast, or a programme by an actor with no governing record |
| Evidence required | A rationale and at least one citable artefact per cell | A rationale and a locator into the primary source per cell. No artefact of the position in practice exists |
| Enters the dataset | Yes — downloadable, comparable, audited | Never. Excluded from the dataset, the distance matrix, the comparison tool and the audit |
| Confidence | Declared per entry; may be high | Low by definition. Every reading is marked contested |
| Can be challenged | Yes, through the challenge form | Yes, and additionally through the blind coding instrument on every reading |
| Right of reply | Standing offer to any organisation coded | Actively offered before publication for any living person or active organisation |
| Scored against reality | Revised as the record develops | Committed to dated predictions, logged correct or wrong in public |
Readings live on a different URL branch, use a different visual identity, and carry a standing disclaimer on every page. That is deliberate. The dataset's value rests entirely on the artefact rule, and a projection layer that blurred into it would cost more than it earned. If this section ever begins to outrank the dataset in search or attention, that is a problem to fix rather than a success to celebrate.
Evidentiary rules
Persona readings score living people and active organisations, and carry a materially higher bar.
Right of reply
Any living person or active organisation scored in a persona reading is sent the complete coordinates and rationales before publication, together with an offer of unedited space on the reading itself. The offer has no word limit and no editorial conditions. Disputing individual cells, rejecting the exercise entirely, and declining to engage all publish verbatim.
| Marker | Meaning |
|---|---|
| Prepared, not yet sent | Reading drafted; the offer has been written but not despatched. The reading may be published in this state only where the subject is a scenario rather than an actor. |
| Sent — awaiting response | Offer despatched on a recorded date. Fourteen days are allowed before publication proceeds. |
| Replied — published in full | A response was received and appears unedited on the reading. |
| Declined or no response | Recorded neutrally, with the date of the offer. No inference is drawn from silence. |
It is the single convention separating analysis from a hit piece in the judgement of any editor deciding whether to cover this. It also produces the best outcome available: a subject who engages generates a genuine story, and a subject who declines has still been treated fairly on the record. The cost is fourteen days.
Corrections
Where a factual error is identified — by a subject, a reader or us — the reading is corrected and a dated correction notice appears on it. The original claim is described, not deleted. Where a score changes rather than a fact, the previous coordinate stays visible alongside the new one with the reasoning for the revision.
Where a challenge through the Atlas challenge form concerns a reading rather than a dataset entry, it is routed here and answered here. Challenges that change a score are named as such.
Readings carry a version number and a publication date. Any substantive revision increments the version and adds a line to the revision log at the foot of the reading.
Reliability
The Atlas method page concedes that the project has a single coder and that inter-coder reliability cannot be self-reported. That is an unusually honest admission and it is also a genuine weakness. Every reading in this section therefore opens a subset of its cells to the reader first: same evidence, same anchors, score withheld until the reader commits a number.
Three things follow. The reader sees the gap between their coding and ours, which is more informative than either number alone. Disagreements above two points are routed to the challenge form, where they become recorded reasoning rather than a comment. And in aggregate, the pool of independent codings is the inter-coder reliability data the method page says is missing.
No distribution of reader scores is published anywhere on this site until the pool passes one hundred independent coders for a given cell, and when it publishes the sample size appears with it. Codings currently remain in the reader's browser and the interface states this. A fabricated agreement statistic would destroy the only asset this project has, and the temptation to produce one grows with the traffic, so the rule is written down here in advance.
Instrument
Reading 001 produced two cells the six axes could not represent, and both failures share a root. Rule of membership cannot represent a line drawn between species rather than within one. Organisation of power cannot distinguish a government that can lose from one that merely never does.
Both are the same missing measurement: the Atlas has no coordinate for who or what actually decides. From Athenian democracy through to every party contesting an election this year, that could be safely assumed. The decider was human; the only questions were which humans and by what procedure. The six axes encode that assumption everywhere and nowhere explicitly, which is why they fail silently rather than loudly.
| Score | What it means | Anchor cases |
|---|---|---|
| −10 | All binding decisions are deliberated and taken by humans. No non-human system holds delegated authority over any consequential matter. | Athenian assembly, the Icelandic Althing |
| −05 | Non-human systems advise and administer. Humans decide, and routinely decide against the advice. | Actuarial sentencing guidance, present-day central bank forecasting models |
| 00 | Mixed. Substantial classes of decision are effectively delegated to automated systems; others are firmly reserved to humans. | Algorithmic benefits administration, automated market surveillance |
| +05 | Non-human systems decide by default across most of government. Human override is lawful, available, and rarely exercised. | Plan A's 2035–2040 pause regime |
| +10 | Non-human systems decide. The human role is advisory or ceremonial, whether or not a formal right of override survives on paper. | Plan A's post-2040 end state; Plan D's aligned outcome |
Adding this axis would do something the existing six cannot: it separates the two outcomes that AI-governance argument keeps conflating. A benevolent aligned superintelligence and a hostile one sit at the same AG coordinate. What differs between them is whose interests the system serves — which is exactly what the other six axes measure. That is a clean division of labour, and it makes the framework's answer precise. Superintelligence does not create a new position on the political map so much as add a dimension the map did not previously need.
The Atlas audit already reports that two of its axes correlate at 0.86, and that dropping any single axis leaves the distance matrix correlating above 0.95 with the full model. A seventh axis has to clear a higher bar than being interesting. The honest test is whether AG separates any pair among the existing 319 organisations that the current six place together. Until someone demonstrates that it does, AG stays in this projection layer, where it belongs. Challenges arguing either way are welcome and will be published.