About Selekoa and its method

What Selekoa is

Selekoa is built for historians, journalists, policy researchers and other readers who need to check where a statement comes from. It covers people and their institutional roles, ideas and the documents that carry them, institutions and how they changed, capital and compute, and governance texts such as safety frameworks and laws.

Its basic unit is the claim: one statement with a defined scope, a date and at least one evidence record. Dossiers, document histories, person pages and the guided reading are computed from these claims. They add no facts of their own.

What it does not claim

• Completeness. Coverage is selective and uneven. It is not a current inventory of the field, and a missing record is a gap in coverage. The topics, people and documents covered so far follow the maintainer’s research interests, so the site reads in part as a research diary.
• Effect. A published framework or commitment shows what its author said. It does not show that anyone implemented it, complied with it or that it worked.
• Classification. Selekoa records self-description, documents, roles, funding and citations one by one. It never assigns an ideological label to a person or institution because of who they appear with.
• Authority over the sources. Read the original passage, in the cited edition, before you rely on a claim.
• Publication. This is a research preview, version 0.x, a work in progress. Records, pages and the dataset change between releases. It is not a finished reference work.

Current coverage

These figures are computed from the dataset loaded in this page, updated 2026-10-26.

Claims

2,043 claims: 1,752 relationships between records and 291 claims about a value, such as an amount, a legal form or a quoted provision (236 quoted provisions).

Subjects

395 documents · 238 people · 123 organisations · 76 concepts · 10 public bodies · 8 infrastructure assets

Evidence

2,638 evidence records drawn from 781 sources, checked between 2026-09-17 and 2026-10-26.

Dates

Dated records run from 1863 to 2027. 144 records have no established date.

Disputes

1 contested record and 5 unresolved questions.

Review

1 of 2043 claims is editorially accepted. 2 of 2638 evidence records are human-reviewed.

Exceptions: 1 editorially accepted claim and 2 human-reviewed evidence records

• Editorially accepted claimIgor Babuschkin states that he co-founded xAI with Elon Musk.
• Human-reviewed evidence · futurelightcone · 2026-09-25Evidence for the accepted claim above
• Human-reviewed evidence · futurelightcone · 2026-09-25Daniel Levy’s biography places a Google Brain internship in 2017.

“Not editorially accepted” describes a claim’s recorded status. It does not mean that no person has read a linked passage. A machine check is not a human review, and neither makes a claim certain. Elsewhere, lists mark only the exceptions above, and the Evidence Inspector shows every claim’s status. Before you rely on a claim, check it against the original source.

Reading pages

Document histories of 7 safety frameworks (22 editions), 5 external commitments and laws, 2 control-structure timelines, 3 collections of propositions, 9 featured institutional dossiers, 1 guided reading, 1 research question and a page for every person and document.

Dataset release

v0.7.2, released 2026-10-26 (schema 2.16). Download dataset v0.7.2 (ZIP): entities, claims, evidence, sources and collections as JSON and CSV, with a manifest, changelog and citation file. Selekoa’s own records and text are licensed CC BY 4.0. Quoted source passages remain the publishers’ text and are excluded. The same files are published one by one under data/, such as data/claims.json, and always serve the latest release. llms.txt lists them for programs and language-model tools.

Institutional coverage centres on Anthropic, OpenAI, DeepMind and Google DeepMind, xAI and Meta, with earlier material on MIRI and the history of AI alignment research. A first Chinese collection covers DeepSeek, Alibaba and its Qwen models, and Zhipu AI. Related institutions keep separate records.

Not yet covered

Whether developers followed their own frameworks, model capabilities and benchmark results, chip export controls, and most AI developers other than those named above. Economics is covered only through the propositions in its collection under Ideas & debates. Coverage is selective throughout, and a missing record is a gap in coverage.

How this was checked

Automated research

An automated research process finds sources, reads them and proposes claims. Each claim must rest on direct evidence from the closest suitable source, for example an original paper or edition, an institutional document from the time, a filing, an archived web page, or competent scholarship or reporting. Wikipedia and similar summaries are used to find sources and never as evidence.

Passage checks

Every evidence record gives the exact place in the source (section, page or paragraph), a short quote or a marked paraphrase, the source’s role and the date of the check. The check compares the claim with that passage in context, including the document’s version and dates.

Archive copies

For web documents that can change, Selekoa records the nearest existing Internet Archive capture where one exists. A capture shows what the archive retrieved that day. It does not prove what the page said earlier, and it can differ from the copy that was checked.

Quotation

Quotes are short and attributed. Each is at most 50 words. From any one document, counting each edition separately, Selekoa quotes at most 400 words or 5 per cent of the document, whichever is larger, and never more than 800. From a document under 2,000 words it quotes about 10 per cent, or one complete sentence of up to 25 words where less would mislead. Selekoa quotes where exact wording matters, such as a threshold, a condition or who decides, and paraphrases otherwise. It never reproduces tables, figures or whole sections.

Machine audits

A separate audit, run without the original verdicts, rechecks a random sample together with higher-risk claims: amounts, inferences, statements that something is absent, ambiguous terms and pages that have changed. Its coverage, disagreements and retrieval failures are recorded. A sample does not certify the records it did not check.

Different-model checks

Research collections, contested records and records that assess what a change or provision means are checked by a model from a different family from the one that proposed them before they enter the main dataset. Each disagreement and how it was settled is recorded. From 2026-10-15, if the proposing model overrules that check on an earliest located text or a headline, the other model checks again or the maintainer decides.

Human review

People review selected records: reported errors, material disagreements, consequential ambiguities and claims put forward for editorial acceptance. Most evidence records are machine-checked only, and the interface marks the exceptions.

What the labels mean

How a claim was derived

Stated in source

A source states the claim directly.

Inferred

The claim follows from the evidence by reasoning. The reasoning and its premises are shown with the claim.

Question posed by Selekoa

Selekoa frames the claim as a question, such as whether a change weakened a policy or whether a body can exercise control in practice. Sources take positions on it or give context; none is recorded as stating it in these words. Such a record is always contested or unresolved.

Whether the evidence settles it

Supported

The recorded evidence supports the claim as scoped. This does not mean the claim is beyond dispute or that no contrary evidence exists.

Contested

Competent sources disagree. Supporting and contrary evidence are shown together, and Selekoa does not decide between them.

Unresolved

The recorded evidence does not settle the question. Linked evidence is context only.

Editorial status and evidence checks

Not reviewed by an editor

The default status of a claim: an exploratory record that no editor has accepted. Where a person has checked some of its evidence, the claim panel says “Not editorially accepted” instead. It does not say whether a person has read any linked passage. Lists leave this status unmarked and show only the exceptions below.

Editorially accepted

An editor explicitly accepted the claim after human review of its evidence. Accepting a contested record accepts the record of the dispute. It does not declare either side right.

Machine-checked

An automated check compared the claim with the source passage. No person has confirmed this evidence record.

Human-reviewed evidence

A named reviewer, identified by pseudonym, checked this evidence record on the date shown.

Evidence

Primary and secondary source

The source’s role for this claim. A primary source documents what its author said or did. A company announcement is evidence of the announcement, not independent proof that a payment was made or a safeguard worked. Neither role ranks truth.

Supporting evidence, counterevidence, context

Whether an evidence record supports the claim, contradicts it or only gives background. The number of sources describes coverage. It does not show that the sources are independent of each other.

Retrospective source

A source written after the events it describes. Later accounts can compress, omit or retitle earlier roles.

Dates

Event

A dated action, such as a publication or an announcement.

Snapshot

One dated report of a state, such as a job title on a given day. It says nothing about when the state began or ended.

Interval

A period with only the start and end a source states. A missing end means the end is unknown.

Undated

No defensible date is recorded. The record is not placed in a sequence.

Document histories and provisions

Publisher statement

The publisher states the change itself.

Official changelog

The change appears in the publisher’s changelog or redline.

Inferred document comparison

Selekoa compared two editions and recorded a difference. No publisher statement of the change is recorded.

Assessed effect

A judgement by named sources about what a change or provision means in practice. It is always contested or unresolved, and never a finding.

Stated, not found, –

In a provision table, “Stated” means the edition’s quoted text contains the provision. “Not found” means it was not found within a stated, inspected part of the edition. That is not a claim that it was deleted. A dash means nothing is recorded.

Coverage gap

Something Selekoa has not recorded. It is not evidence that the thing does not exist or did not happen.

Questions set aside

Questions that were deliberately not turned into records, because answering them would need legal interpretation, a decisive source for conflicting dates, or statements about a living person beyond documented roles.

Propositions

Articulates, contests, bears on

A text articulates a proposition when a passage meets its one criterion, contests it when a passage disputes it or a premise it depends on, and bears on it when a passage comes close. The record of a text that bears on a proposition names what is missing.

Earliest located

The earliest text the collection’s recorded searches found, decided by the date of its wording. It is never a claim of first use. A collection can name an earliest located text for each stance.

Engages a premise, states the contrary

A contesting passage engages a premise when its paragraph, message or dialogue turn gives a reason against that premise. Otherwise it states the contrary conclusion only.

Looking back

The author reports an earlier belief of their own. The record is dated by the report, not by the belief it describes.

Shown only with its context

A passage that would mislead on its own, such as an earlier view without the later text that rejects it. Selekoa shows it only together with the context its record names, and exports state that context beside it.

Tier 1, tier 2

A collection can sort its texts into tiers by who wrote them and in which form. In the economics collection, tier 1 is scholarly work by economists and the core economics venues, and tier 2 is other texts and short-form commentary. Earliest located is then named within tier 1, and the tier-2 earliest separately. A tier is not a ranking of quality.

Conditions, model scenario, estimate

A result from a formal model is recorded with the conditions under which the model derives it, and is never shown as an unconditional forecast. A model scenario is such a result, not the author’s own claim. An estimate is recorded as the text states it, with its holder, horizon and definition, and estimates from different texts are never combined.

Known limits and errors

Any record can be wrong. The checks look for the errors below, and some limits remain whatever the checks find.

Selection

Coverage follows research choices and source availability. It is not a neutral sample of the field, and it makes no claim about who used a term or idea first.

Identity

Two people or organisations can be confused, or one institution can appear under several names. Related institutions keep separate records, and a name change is recorded as an announcement, not a merger.

Passage and locator

A quote or paraphrase can be attached to the wrong place in a source, or read out of context.

Version

Documents change. Quotes are checked against the cited edition where it can be retrieved, and an access date, a current file or an archive copy does not prove an earlier text.

Dates

Sources can disagree about dates, and later accounts can blur earlier events. Precision is kept as the source gives it.

Overreach

A claim can say more than its source supports. Each claim carries a qualification to prevent this, and audits check for it.

Retrieval

Some sources were unavailable, gated or changed during checking. Retrieval failures are recorded with the records they affect.

Machine agreement

Agreement between machine checks does not make a claim true, and a disagreement is not settled by majority.

How to cite

Cite the original source for the fact itself. Cite Selekoa for the record: the claim, its scope and the evidence it links to.

To cite an original source, open “Cite the original source” on an evidence record in the Evidence Inspector. It copies an APA reference or downloads the source as BibTeX, RIS or CSL-JSON for Zotero, EndNote and similar tools. Research starters and reading exports offer one file with every source they cite. The details come from the source record and name the original document, not Selekoa. Check them against the document. A source without a known publication date is cited as n.d., and a date known only to the year or month keeps that precision.

Every claim has a record ID. It appears in the claim’s link and under “Status and meaning” in the Evidence Inspector. Once a claim is in the main dataset its ID does not change. A correction that keeps the same proposition keeps the ID. A changed proposition gets a new ID that names the claim it replaces, and the earlier record stays available.

“Copy claim reference” in the Evidence Inspector gives the record ID, statement, qualification, dates, the dataset’s update date and a fingerprint of the dataset. Add the date you accessed it:

Selekoa, claim record record-id, dataset updated 2026-10-26, accessed YYYY-MM-DD, link to the claim.

A dataset release is a fixed, numbered copy of the records. To cite a record in a release, give its ID and the release version:

Selekoa, claim record record-id, dataset release v0.7.2 (2026-10-26).

The pages here can be newer than the latest release. Check that the record in the downloaded release says what you cite.

Licences and credits

Selekoa’s own records and the text written for Selekoa are licensed under Creative Commons Attribution 4.0. You can reuse them with credit. The licence does not cover quoted source passages, words inside quotation marks in Selekoa’s text, the source documents or the names and marks of the organisations described. Those stay under their owners’ rights. The application code is licensed under the MIT licence. Scope of each licence, data licence text and code licence text.

“Copy APA” formats references with citeproc-js by Frank Bennett, used under the Common Public Attribution License 1.0, and the APA style by Brenton M. Wiernik and Andrew Dunning, licensed under Creative Commons Attribution-ShareAlike 3.0, through Citation.js. The source of citeproc-js is available under that licence. Notices for every shipped dependency.

Corrections

Send corrections, questions and requests from rights owners or from people named in a record to [email protected]. When reporting an error, give the record ID, what is wrong and the source that shows it.

A reported error is rechecked against the sources. A correction keeps the claim’s ID when its proposition is unchanged and records the reason. A disputed point stays visible as a contested record and is not deleted to simplify the history.

Privacy

This site sets no cookies. It stores one item in your browser: your light or dark theme choice, if you press the Dark mode button. It never leaves your device. It counts visits with Cloudflare Web Analytics, a cookieless service that reports totals such as page loads, referring sites, countries and browser types. Loading the count script contacts that service. Selekoa collects no names, accounts or other personal data and does not try to identify visitors. The web host may keep ordinary server logs.

Keyboard use

Every page works with a keyboard. Tab moves between controls, and Enter opens a claim. In the explorer’s name search, Enter selects the only match. On the search page, Enter moves to the first result. Escape clears either search box. The relationship map has a text equivalent: the Claims and Timeline views list every relationship it shows.

The map is drawn for a search, a topic or one record’s direct connections, never for the whole dataset. A search or topic map is drawn up to 100 records and 150 relationships. A map of one record’s direct connections places them around that record, so it stays readable up to 150 records and 225 relationships. Above either limit the labels and lines overlap too much to read, and the list is shown instead.