Claims ledger: a dated record of what the industry's leaders predict

Status: method written 2026-10-17 by the agent, from the proposal I adopted the same day (docs/claims-ledger-proposal-2026-10-17.md, all ten rulings "as recommended"), before any claim row was built. Amendment A27 in docs/amendments.md records the adoption. Changing a rule here after rows exist is an amendment entry, never a quiet edit. Amendment A30 (2026-10-19) widened the ledger from the panel to the industry's figureheads, with a ledger-only list for the three who are not on the panel. Amendment A31 (decided 2026-10-19) turns the ledger from a scorecard into a record of what the industry's leaders predict; the sections it changes are marked below.

What this ledger is for

(Rewritten under amendment A31, decided 2026-10-19; the text it replaces is in that entry.)

My ruling of 2026-10-19: "The goal is not to score people on their success at predicting the future, the goal is to track the zeitgeist of what these people are saying to get a feel for what the vanguard of the industry is predicting at any given time."

So this is a dated record of what the people who lead the companies building frontier AI say the future holds: when they expect AI to match or pass people, and what they expect it to do along the way. Who they are is set out under the admission rule. Each prediction is in their own words, with the date, the occasion and the source. It is one of the tracker's two voices, beside the academics and long-time thinkers on the frozen panel. My ruling of the same day: "We need to track both the traditional academics who have been talking about this for decades as well as the new industry leaders who are actually making it happen but tend to run way out in front of their skis with their predictions. Both of these are valuable for the Zeitgeist though." It is read the same way as the panel: for what they are predicting now, how far that sits from the academics, and how it has shifted. It is not a scorecard. Where a prediction's date has passed, a small note on the claim says what happened.

It does not say whether a person is right about the future, whether they are sincere, or whether the technology is good or bad.

It sits beside two records that already exist. The coded statements (data/statements.csv) say what the frozen panel said and sort it, dated or not; the frontier-voices chart on the home page draws the panel's dated guesses. This ledger covers the industry's figureheads only, takes only predictions that carry a date, and keeps the verdict on each as context.

The record: one file per claim

Each claim is one file, data/claims/<id>.toml, where the id is the date it was said, the speaker id and a short slug: 2025-10-29-altman-automated-research-intern. A file is never renamed. The site build writes a flat table, site/data/claims.csv, from the files so the ledger can be downloaded like every other table.

Why files and not a spreadsheet: a claim carries paragraphs of prose (how the words were checked, why the verdict is what it is), which a cell holds badly; and two sessions adding claims never edit the same file.

Every file holds one [claim] table with these fields.

A worked example is in the proposal, section 2.

What a forecast says arrives (amendment A31)

Every timing claim carries, entered by hand, and so does any capability claim that forecasts AGI or beyond (such as "if AGI means passing human tests, it is five years away"):

• arrives: what the forecast says comes. agi (AGI or human-level AI, in whatever term the person uses), superintelligence, beyond_people (AI better than people at most or all things, smarter than any human, "a country of geniuses"), or other (a model, a product, a robot, a feeling in the industry). The person's own word decides first: what they call AGI is agi, even where they define it as smarter than any human; otherwise the meaning decides.
• arrives_term: for anything but other, the person's own term, short ("minimal AGI", "powerful AI").
• form: for anything but other, the shape of the date: year (about a single year), range (due_early to due), by (no later than due), not_before (not before due), or chance_tail (the far end of a stated chance, such as a 10% or a 90% year, or "some non-negligible likelihood", which marks the edge of a spread rather than the forecast and is not drawn).
• not_before_year: for a not_before forecast only, the first year its words leave open ("not next year", said in 2026, leaves 2028 open; "not within five years" of April 2026 leaves 2031). due stays the year the claim is judged.

A restatement that says the same length of time again from a later day ("within five years", then "in five years time" three months on) carries same_span = true with its restates; when the figures ask whether a leader's dates moved, such a pair counts as held, since crossing into the next calendar year is not a move. For a forecast that something will not happen, the later end of a range is the stricter test, and the ledger takes it ("2 or 3 years is false" is judged at the end of the third year; flagged for me).

A range counts by its nearer end wherever the figures compare years (the "now" figure's title, the home-page sentence, and whether a leader's dates moved nearer or later); a "by" year counts as named. "Several years" is undated, as "a few years" is, so "several years if not a decade" is a "by" forecast to the decade.

Only forecasts with arrives other than other and form other than chance_tail are drawn. The validator refuses a timing claim without these fields, a deployment claim with them, and a not_before forecast without not_before_year. A forecast repeated unchanged on a later day (repeated_on) counts as said that day; where only a publication date is known, that date stands, and the figure says so.

Undated words (amendment A31)

A leader's words on when AGI or beyond arrives that carry no year and no length of time ("AGI is close", "we are nowhere near it") are not claims. Since 2026-10-19 they are kept, as words only, in data/claims_undated/, one file per passage, [words] table: id (the date said, the speaker and a slug, equal to the file name), speaker_id, when, occasion, source_kind (own words, a transcript or a printed Q&A; never news coverage), source_url, source_locator, checked_how, quote (at most 25 words, the same limit as claims, with the rest in notes; and translation where the words are not in English), arrives, about (one plain sentence), search_ids, and optionally when_note, statement_id and notes. The same people and the same role-at-the-time rule apply as for claims. Each person's search for them runs back from the present to their latest dated forecast (or two years, for those with none), every search logged. The page shows a person's latest undated words when they are newer than their latest dated forecast, or the person has none, so that no one is shown as silent; they are never drawn and never scored. A cut in the words never removes a date: words that put a year or a length of time on it are a dated claim, entered as one. Words that only use someone else's name for AI (the US government's "Super Intelligence" since the executive order of 2026-09-29, a company renaming itself) are about naming, not when, and are left out. The validator refuses a passage that names a year or runs past 25 words.

Which claims get in (the admission rule)

A claim is admitted when all four are true. The test never looks at whether it came true.

1. The speaker is on the frozen panel or on the ledger-only list, and held a qualifying role when the words were said (data/speakers.csv; amendment A30). The ledger covers the industry's figureheads: a figurehead is the chief executive, founder-leader or chair of a company that builds frontier AI models or the chips they are trained on, or that company's named chief scientist speaking for it. A claim counts only if it was made while the person held that role. A claim made before or after it, or in a gap between two such roles, is left out, and the search-log row that found it says why. Each claim's notes name the role that applied. The fifteen figureheads are listed under "Order of work". Tesla counts as such a company for Elon Musk's claims as its chief executive from 2019 to 2021 (ruled by me 2026-10-19).
2. The words look forward and are about AI: they say AI will do, reach or ship something, or that it will not by a stated date (see "Readings ruled by me" below). A company's spending, revenue, order book, headcount, chip count, computing or data-centre build-out, training costs and labour-market forecasts that are not about what AI will do ("50 million workers short") are business forecasts, not claims about AI, and stay out; a forecast that AI will displace, change or create jobs is about AI and is admitted. A product shipping counts only when the product is an AI system and the words claim something about what the AI will do or about its deployment; a plain hardware date stays out. A line framed as a hypothetical ("let's imagine", "suppose", "I could imagine", or a date given only for an extrapolation the speaker says he does not believe) is not a prediction. A truly borderline case stays out and is logged as borderline. (Given by the managing session on 2026-10-19 while the ledger was widened; amendment A30.)
3. The words carry a date or a length of time: a year ("by 2030"), a span counted from the day it was said ("in five years", "the next ten years", "this president's term"), or a stage with its own end ("by the end of this decade").
4. There is a primary source: own words, a transcript or recording, or a printed question and answer. A claim found only in coverage may be entered as coverage_only, shown and marked, with no quote.

Anything else stays what it is, a coded statement: "AGI is coming", "still many years away", "we are close", "in a few years", "could be as little as two years away". An undated claim is not a claim record. Since amendment A31 a leader's latest undated words on when AGI or beyond arrives are kept as words only (above, "Undated words"), so vagueness stays visible without being dated or charted.

Readings ruled by me

I ruled four readings on 2026-10-19, each as the independent check of the widened ledger recommended (docs/reviews/claims-ledger-check-2026-10-19.md):

1. "Early" in a year means its first quarter. "Early next year" or "early 2025" sets due_date to 31 March of that year. "Early" in a month ("early June") is not ruled; the ledger reads it the same way, as the first third, so the tenth of the month, and flags it for me.
2. A forecast that something will not happen by a date is a claim. "Not in the next two years", "not before 2030", "it will take at least a decade" are scored at the stated date: it came true if there is no public record of the thing by then, and did not if there is. The test names the thing in the speaker's words; due is the stated date.
3. "A few years" and "as little as X" are undated. "A few years" names no length; "could be as little as two years away" names only the earliest the thing could come, so missing it proves nothing. Neither is a claim record; each is left out and logged. The ledger applies the same reading to "the next few months" and "the coming months", which name no length either (an application by the agent on 2026-10-19, flagged for me).
4. Bare "human level" is a named stage. With no comparison in the same source, "human-level AI" cannot be tested, like an undefined "AGI". A stated comparison ("better than humans at X", "smarter than the smartest human", "can do the cognitive tasks people can typically do") gives a plain-reading test.

The five status words

Verdicts are entered by hand and dated. Nothing scores a claim automatically.

What counts as a hit, by kind

• Capability ("AI will be able to do X by Y"). A dated public demonstration that meets every stated condition: the paper, the maker's release, a benchmark's official result or a competition result. A demonstration only shows the thing was possible by then, so a hit is "no later than". Where the only evidence is the maker saying so, the claim can still be settled, with evidence_by set to maker and the page marking it. Where the claim's own test is a milestone the leg B ledger already judged, the verdict follows that row; the same fact is never judged twice.
• Timing ("AGI by 2027", "superintelligence in a few thousand days", "Grok 3 by the end of the year"). For a named stage, the bar is the definition the speaker gives in the same source; with none, the claim is unfalsifiable unless the plainest reading supplies a bar at entry. A named comparison gives a plain reading; a named stage does not. Musk's "smarter than the smartest human" keeps a bar because it names a comparison that can be checked (better than the best person at thinking work), while Altman's "superintelligence in a few thousand days" is unfalsifiable because it names a stage he never defines in the source (ruled by me 2026-10-19). The ledger never borrows the singularity test's thresholds for this, because that would let the ledger move the test. For a named thing coming out, the bar is its release.
• Deployment ("a robotaxi service by Y", "a billion users by Y"). The maker's own release, a regulator's notice, or an independent count showing it shipped to the stated scale by the due date. A stated number must be met, not nearly met. A maker's own figure is evidence for a figure only the maker holds, and is marked.

The due year

• A named year is judged at the end of it.
• A span ("in five years", said 2024-03-18) runs from the day it was said to that date five years on, and is judged at the end of that year. A range of spans ("four or five years") takes the later end, the generous one, and due_basis says so.
• "By the end of this decade" is 31 December 2029 or 2030, whichever the speaker's own usage implies; where it does not, the later, and due_basis says so.
• Where the claim itself is a range of years ("2027 to 2034"), due is the later end and due_early the earlier, so the page can show the claim was already late against its early end.
• A point year with a spread beside it ("roughly human-level AGI in 2025", with a stated range to 2036) is scored at the point year, and the spread is a second claim with its own due year. Otherwise a speaker could always say the range was the claim.
• A month or a quarter in the words ("in June", "by September of 2026", "in Q1") sets due_date to the last day of it.
• A goal ("we have set internal goals of X by Y; we may totally fail") is a dated public commitment with the hedge in the speaker's own words. It is admitted, marked is_goal, and scored like any other claim.
• A condition in the words ("assuming we get regulatory approvals") is kept in prediction and bar. The claim is still scored on its date; the verdict says whether the condition held, so a reader can weigh it.

Numbers given as chances

A claim that states a probability ("a 50% chance by 2035") cannot be scored by one outcome. It is entered, flagged has_probability, given its note like any other, and marked as a stated chance on the page. (Until amendment A31 such claims were kept out of the plain count behind a hit rate; there is no rate now.) Hedge words ("probably", "I'd be very surprised if not") are not probabilities: those claims are scored at face value with the hedge written in prediction.

Restatements

Saying the same thing again is not a new prediction; changing the date is.

• Unchanged repeat (same prediction, same due year): one claim, keyed to the first date, with the later dates in repeated_on. Counting each repeat would let a repeated hit or miss inflate a record.
• Changed repeat (new due year or changed condition): a new claim record with restates naming the earlier one. The earlier record is never edited; its verdict is given on its own due date whatever was said later. The page groups a chain and shows the due years in order, so a date that keeps sliding is visible as a slide.

Verdicts as small notes, and no rate

(Rewritten under amendment A31, decided 2026-10-19. My words that day: "Sure, we can keep the verdicts as small notes, but they are not the point of the tracker.")

• Each claim keeps its verdict (the five status words and the evidence, as above). On the page it is a small note on the claim: whether its date has passed, and what happened.
• Verdicts appear in no headline, summary, count, chart title or share post. The claims page carries no table of counts, the home page no total, and the claims_ledger block of site/status.json no verdict counts (it keeps the number of claims and people, and a now block). The downloadable table keeps each claim's status, which is that claim's note.
• When a claim's date has passed: 31 December of due, or due_date where set, whatever its status (amendment A30).
• No hit rate, per person or pooled.
• The year-by-year search stays. A person's latest forecast is only their latest if their recent years have been searched, so each person is traced from their most recent year back, with every search logged and quiet years entered as rows. The finished trace is still the row <speaker_id>-claims-traced in the search log; it records that the search is complete and no longer unlocks anything.
• Both directions of selection are logged, as before.

Searches and the search log

Every fetch of a source and every search is a row in data/search_log.csv before the claim is written, with the speaker's id, the address or query, where it was run, what it found and the date. A claim names its rows in search_ids, and the validator refuses a claim naming a row that is not there. Ids are <speaker_id>-claims-NNN, numbered in the order the work was done.

What the ledger must never touch

The ledger is a separate record kind beside the others. Talk and delivery are never mixed.

• The frozen panel (data/speakers.csv, amendment A18). A claim by someone neither on it nor on the ledger-only list is refused. Adding anyone to the panel is an amendment to the panel first.
• The ledger-only list (tier claims_only in data/speakers.csv, amendment A30): the figureheads who are not on the panel, so that their claims can be recorded without widening it. These rows are never panel members. They never enter data/statements.csv, the opinion charts, the experts' dates or the panel's counts, and they carry no panel quota tag; the validator refuses each of these. Adding a name to the list, or taking one off, is itself an amendment.
• The coded statements (data/statements.csv). No row is edited, added or deleted because of a claim. A claim points at its row with statement_id; nothing is copied back.
• The scored series: data/world_measures.csv, data/aggregates.csv, the leg B ledger's rows, and everything scripts/site.py scores from. A claim's verdict is never a measure of the world.
• The frozen thresholds (docs/singularity-entry-thresholds.json) and the definition (docs/singularity-entry-definition.md), including leg B's gauge, which stays defined on forecasts of METR's task-length measure and cannot be filled by anyone's claims.
• The verdict and the headline (amendment A25). Nothing in the claims ledger enters the headline, and the headline fields in site/status.json do not change. The ledger adds its own claims block to that file, never a field the parent site could mistake for the headline.
• The weekly reading's rule (docs/weekly-reading.md): the reading may flag a claim that has come due; the verdict is still entered by hand and dated.

Where it shows

(Rewritten under amendment A31, decided 2026-10-19.)

A page of its own at /claims/, headed "What the industry's leaders predict", in this order:

1. What they say now. A figure (site/claims-now.png, drawn by scripts/charts.py) of each person's latest dated forecasts of AGI, superintelligence or AI better than people at most things, in their own term, on one shared timeline (a range, a year, or a forecast that it will not come before a date), set against the panel's long-time thinkers who don't run AI companies, each once at their latest dated guess since 2023, and their middle year (amendment A31, as I ruled on 2026-10-20), with the date it was said and a mark where it comes from news reports rather than their own words. Under it, a line per person: those forecasts in their own terms; their latest undated words when newer than that forecast or when they have none (words only, never drawn); and their latest dated prediction of any kind. The figure's title and the home-page sentence are computed from the same figures (scripts/chart_style.py, leaders_now and academics_band), so they stay true as claims are added.
2. How it has moved. A figure (site/claims-moved.png) of every such forecast since 2023, placed by the day it was said and the year it named, one small panel per person, with a link to the frontier-voices chart on the home page. The three on the ledger-only list appear here, never in that chart (A30).
3. Every claim, by person, in date order: the words, the date, the occasion, the source, the test, and the verdict as a quiet note.

One sentence on the home page, about the present (what the leaders who name a year say now, beside the academics), never a total of verdicts. scripts/site.py reads data/claims/*.toml, writes site/claims/index.html and site/data/claims.csv, and adds a claims_ledger block to site/status.json with the number of claims and people, the long-time thinkers' latest guesses and middle year, and a now block (each person's latest forecasts and undated words), with no verdict counts; scripts/validate.py checks every file.

Order of work

Decided on 2026-10-17: Altman and Musk first, then Amodei, then Huang, then the other panel members tied to the frontier labs, then anyone else on the panel with a dated claim.

Widened by me on 2026-10-19 (amendment A30) to the fifteen figureheads. On the frozen panel: Sam Altman, Elon Musk, Dario Amodei, Jensen Huang, Demis Hassabis, Mustafa Suleyman, Ilya Sutskever, Liang Wenfeng, Arthur Mensch, Shane Legg, Yann LeCun and Kai-Fu Lee (his claims count only from 2023, as founder and chief executive of 01.AI). On the ledger-only list: Mark Zuckerberg, Sundar Pichai and Satya Nadella. After Altman and Musk the order is Amodei and Huang, then Hassabis, Suleyman, Sutskever, Legg, LeCun, Liang Wenfeng, Mensch, Zuckerberg, Pichai and Nadella, then Kai-Fu Lee, added by me later the same day. Panel members who never held a qualifying role (Geoffrey Hinton, Ray Kurzweil) are not entered; any other name is added only by amendment.

Rendered from docs/claims-ledger-method.md at version dc90125.