Methodology
How this register is built
The value of this dataset depends entirely on its integrity. This page states what counts as evidence, what does not, how sources are verified, how chronology is handled, and how the framework can be shown to be wrong.
What counts as evidence
An observation is recorded only where a market development can be connected to an observable change in economic structure, such as pricing, activity, interaction, capability scarcity, governance, value capture or revenue durability.
Recorded
- A company changes pricing because agents reduce seat dependence.
- A client demands lower fees because AI reduced the hours required.
- A sponsor builds shared AI deployment infrastructure across a portfolio.
- A regulator treats the control layer around a model as the unit of concern.
Not recorded
- A company launches an AI assistant.
- A firm adopts a model.
- An investor backs an AI startup.
- Any announcement of intent with no observable economic consequence.
The distinction is deliberate and strict. Adoption is not evidence about economic structure, and a register full of adoption stories would be worthless for the questions this project exists to answer.
Source hierarchy
| Tier | Kind | Includes |
|---|---|---|
| 1 | Primary | Company filings, earnings calls, investor presentations, company announcements, regulatory and government publications, central banks, official pricing pages and product documentation, academic papers, original datasets. |
| 2 | High-quality independent reporting | Financial Times, Reuters, Bloomberg, Wall Street Journal, The Economist and major specialist financial publications. |
| 3 | Institutional research | Recognised research institutions, professional-services research, ratings and analyst houses, and university research centres. |
| 4 | Industry publication | Trade press and specialist industry reporting. |
| 5 | Commentary | Blogs, social posts and commentary. Tier 5 should trigger research for a stronger underlying source and should never become the public evidence itself. |
Source verification
No record is published without a source that has been opened. Verification means resolving the canonical URL, retrieving the page, and confirming it supports the stated fact. Seeing the item in a set of search results is not enough. A search, redirect or tracking URL is never stored as the canonical source.
Links are re-checked on a schedule. Where a link breaks or the page materially changes, the record is flagged and never quietly repointed at a different article, and the original URL and its verification history are retained.
An automated check can establish that a link resolves; it cannot establish that the page supports the claim. Those are recorded as two separate fields, and the second is only ever set by a person.
Observed fact and interpretation
Every record separates what the source supports from what Autonomy Economics makes of it. The two are stored in different fields and displayed differently. No source cited on this site has endorsed the framework, and none is presented as having done so.
Evidence strength
| Level | Meaning |
|---|---|
| 1 | Anecdotal |
| 2 | Emerging |
| 3 | Moderate |
| 4 | Strong |
| 5 | Directly observable |
Level 5 means the predicted mechanism itself is occurring. It does not mean the theory is proven. Any record at level 4 or 5 must state the strongest alternative interpretation of its own evidence before it can be published.
Chronology
Each record carries two dates, one for when the market development happened and one for when the reporting appeared. Autonomy Economics was posted on 30 April 2026. That posting date is the pivot, and the date the paper was written is not, because what makes an observation a test of a prediction is that the prediction was publicly on the record when the observation was made. Evidence predating publication is recorded as prior art and is never described as post-publication validation. Of the 32 observations currently published, 28 were observed after publication.
For surveys, datasets and retrospective reporting, chronology follows the underlying observation period, and the date an article or report was published does not decide it. A survey conducted before publication does not become post-publication evidence because its report appeared later. Where a measurement period crosses the publication date and cannot be separated, it is treated conservatively and is not counted as a clean post-publication test.
Correlation and causation
Nothing on this site claims that autonomous capability caused an observed economic change. Each record states the mechanism it is consistent with, the strongest competing explanation, and the future observation that would separate the two. Where an observation has several plausible causes it is classified as ambiguous and stays there until something distinguishes them.
How contrary evidence is classified
| Classification | Meaning |
|---|---|
| Contradictory | The observation is inconsistent with a stated direction of the framework at the time it is observed. This records a current counter-signal; it does not by itself mean that a multi-year falsification horizon in Section 11B has been completed. This is not a quota. A record is never made contradictory simply to balance the register. |
| Boundary condition | The predicted mechanism is constrained by timing, context, human control, measurement, regulation or another observable condition. It narrows where or how quickly the proposition applies without showing it to be false. |
| Ambiguous | The same observation supports more than one material reading and the available evidence cannot distinguish between them so far. |
Failure to support a prediction is not the same as contradicting it. This distinction is particularly important during a short observation window, when persistence conditions may not have had time to fail.
How status is assessed
Proposition status is derived from the published register and never set by hand. Supportive observations contribute in proportion to their strength. A moderate observation counts once and a directly observable one counts three times. Contradictory observations subtract on the same scale. Observations recorded before publication of the framework count at half weight, because they cannot corroborate a prediction that was not on the record at the time.
The public labels are deliberately descriptive. They state how much registered evidence is currently consistent with a proposition. They do not say how likely the proposition is to be true. They are not confidence scores, grades or peer-review verdicts. The underlying calculation still uses a stable internal ladder so changes can be tracked consistently over time, but the site does not present labels such as strong orvery strong as conclusions about the theory.
A predicted mechanism may be labelled directly observed only as an explicit editorial assessment. That judgement is shown as editorial and carries a stated reason.
Current editorial assessments
Two capital-coordinated deployment vehicles were announced and staffed within a week of each other (AE-2026-001, AE-2026-002). The predicted structure is not being inferred from second-order effects; it exists and is named. Whether it produces the predicted portfolio-level returns remains open, and that is tracked separately under Revenue Durability.
No proposition is scored numerically in public. The Evidence Index becomes a published number only after the weighting has been tested against a substantially larger register.
Assessment history
When an assessment changes, the previous one is preserved and never overwritten. 9 assessments have been recorded to date. Someone reading this register in several years should be able to see whether its reading of the evidence evolved honestly, including where it weakened.
How predictions can be falsified
The April 2026 paper sets the canonical empirical tests in Section 11B, including the trigger of credible commercial deployment and the one to two, two to three and three to five year test windows. The monitor also carries proposition-specific operational search conditions so contrary evidence can be found now. Those operational conditions are labelled as monitor tests; they are not presented as verbatim pre-published wording. Both are shown on the challenges page.
The monitor runs separate confirmation, falsification and boundary discovery lanes. The lane in which a source was found does not move a proposition's status. Only observations that have been reviewed, verified and published enter the status calculation.
Retrospective falsification audits
The daily monitor is a continuing discovery process. It makes no claim to have found every relevant event on the public web. Periodically, the register is therefore reviewed retrospectively over a fixed historical window. The audit starts from the paper's canonical framing and Section 11B test, then uses an operational counterevidence query to search deliberately for the world in which that proposition would be wrong, narrower or slower than stated. Evidence is recorded as it appears; the longer horizon determines when persistence becomes a stronger framework-level test.
A retrospective audit searches economic variables as well as AI vocabulary. Examples include billable hours and worked rates, seat pricing and renewals, model spend concentration, human approval at payment, governance costs, service margins and portfolio-level operating performance. Sources include primary pricing and financial disclosures, empirical datasets, academic research, buyer and customer surveys, specialist industry data, regulators and independent reporting.
Reposts and syndicated coverage do not become separate observations. Multiple reports of the same underlying dataset are treated as one event with additional sources. Independent datasets may be recorded separately when they measure different populations, periods or economic outcomes.
These reviews are systematic audits of retrievable evidence. They make no claim that the register is a complete census of everything published. The method is designed to make contrary evidence easier to find and harder to suppress, while keeping the threshold for contradiction unchanged.
How AI is used
Automated research assists with discovery, deduplication and a first-pass classification of candidates against the framework. It does not publish. Every record on this site was reviewed by a person, who checked the source, the separation of fact from interpretation, the chronology and the classification before it became public.
Citation
Source material is not reproduced. Each record states the fact in its own words, links to the original publisher, and retains the publisher, headline where available, publication date and the date the source was retrieved. Where a source is subscriber-only, the link goes to the publisher and never to a copy of the article.