Proposal: a dated-claims ledger per person, scored against delivery

Written 2026-10-17 by the agent at John's request, relayed through the master session the same day. Nothing is built and no record changes with this document: the panel, the statements, the scored series, the thresholds, the definition and the verdict all stay as they are, and the seed list in section 8 is a list in this file, not rows in data/. The numbered decisions in section 9 each carry a confidence and the strongest alternative, as the ground rule of 2026-09-08 requires; the rulings John must make are in section 10. If adopted, section 9's decision 12 says what gets numbered as an amendment.

0. One correction before anything else

The brief this was written from says Jensen Huang is not on the frozen panel and that adding him would take an amendment. The record says otherwise. data/speakers.csv carries him as a member of the 2020s cohort ("Frozen 2026-10-06 (A18)"), and amendment A18 (the panel freeze) lists him among the cohort's twelve, six of whom "pass thin", Huang first among them: both of his statements on the panel are secondhand. So no amendment is needed to give Huang claims; all four leaders John named (Musk and Altman in the core panel, Amodei and Huang in the cohort) are already on it. Section 7 still sets his claims in a table of their own, because John asked for it and because his panel row is the thinnest, but the reason is thin sourcing, not panel membership. Huang is not tagged frontier_lab (the panel's note calls him a near miss: he sells the chips, not the models), which matters for one decision in section 9.

1. What the ledger is for

John's question, 2026-10-17: the people who run the frontier AI companies both sell the story and see further than anyone else, so is each of them a front-row seat or a salesman? His ruling that day was that a dated-claims ledger per person, scored against delivery, be proposed for his ruling, on the pattern of the sibling trackers' claims logs. The master session's gloss: a hit rate per person, in public, as the honest answer.

The ledger answers one narrow question and says so. For each named leader, of the predictions they made that have since come due, how many came true, how many did not, and which way the misses went (too soon or too late). That is a score of their forecasting, published with every claim in their words, the source, and the evidence for the verdict. It does not say whether they are right about the future, whether they are sincere, or whether the technology is good or bad. A leader with a poor record may still be right next time; a leader with a good record may have been lucky. The page says that in so many words.

Why it belongs in this project and not another: the tracker already records what these people said (data/statements.csv) and already scores forecasts against outcomes in one place (the leg B ledger, which scores the 2016 researcher survey's 32 milestones). What it lacks is the joint: a named person's own dated words, set against what happened, one claim at a time. The family's tracker pattern (rule 6.4, talk and delivery side by side) says that is where a tracker earns trust.

2. The record layout

One file per claim

Proposed: data/claims/<date>-<speaker>-<slug>.toml, one file per claim, the layout of The Upgrade Tree's claims folder, with the verdict fields The Outrunning Ageing's records use. Python 3.11 and later read TOML in the standard library (tomllib), so scripts/validate.py stays standard-library-only, as its first lines promise.

Why not the CSV this record otherwise uses: a claim carries three or four paragraphs of prose (how the words were checked, why the verdict is what it is, what the speaker did not say), which a spreadsheet cell holds badly and a reviewer cannot read in a diff; and two sessions adding claims then never edit the same file. The cost is one more file kind beside the CSVs. To keep the project's habit that every record can be downloaded as a table, the site build writes a flat site/data/claims.csv from the files, as it does for the other tables.

Fields

Worked example

data/claims/2025-10-29-altman-automated-research-intern.toml, as it would be written from the seed list (the words and dates were read at the source on 2026-10-17; the verdict is the one the seed pass proposes, and John can overrule it):

[claim]
id = "2025-10-29-altman-automated-research-intern"
speaker_id = "altman"
when = "2025-10-29"
when_note = "The post's timestamp is 17:19 UTC on 2025-10-29; it says the livestream was 'Yesterday'."
occasion = "Post on X summarising OpenAI's livestream of 2025-10-28"
source_kind = "own_words"
source_url = "https://x.com/sama/status/1983584366547829073"
source_locator = "the first paragraph after 'TL;DR:'"
checked_how = "Read through a mirror of the post's text on 2026-10-17; x.com itself was not opened."
quote = "We have set internal goals of having an automated AI research intern by September of 2026"
statement_id = "altman-2025-09"
prediction = "OpenAI's own goal: by September 2026, an automated AI research intern running on hundreds of thousands of GPUs. He adds 'We may totally fail at this goal'."
kind = "capability"
is_goal = true
bar = "OpenAI's own definition, from its post of 2026-09-06: a system that can carry out well-defined research tasks under human direction, including tasks that would take a skilled researcher a few days; and running on hundreds of thousands of GPUs, from the original words."
due = 2026
due_basis = "His words: 'by September of 2026'."
status = "hit"
verdict_on = 2026-09-06
verdict_why = "OpenAI's post says 'according to our measurements, we have now reached the goal'. The post as read does not say whether it ran on hundreds of thousands of GPUs, so one condition of the bar is unchecked; recorded as hit on the company's word, to be reopened if an outside check finds otherwise."
verdict_source = "https://openai.com/index/research-acceleration-view-inside-openai/"
evidence_by = "maker"
notes = "A goal, not a forecast, in his own words; the page marks goals as goals. The 2028 follow-on goal (an automated AI researcher by March 2028) is its own record."

Two things the example shows. A hit can rest on the maker's word, and the page then says so in a chip beside the verdict, so a reader can weigh it. And a bar can have a condition nobody has checked; the entry names it instead of hiding it.

3. The scoring rules

Which claims get in (the admission rule)

A claim is admitted if all four are true, and the test never looks at whether it came true:

1. The speaker is on the frozen panel (data/speakers.csv).
2. The words look forward: they say AI will do, reach or ship something.
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 date ("by the end of this decade").
4. There is a primary source (own words, a transcript, or a printed question and answer). A claim found only in coverage may be entered as coverage_only, shown and marked, with no quote, as the siblings do.

Anything else stays what it is today, a coded statement in data/statements.csv: "AGI is coming", "still many years away", "we are close". This is the record's answer to the undated-claim question (decision 4 below, and ruling 2 in section 10): an undated claim is not a claim record. The page says how many dated and how many undated forward-looking statements each person has, so vagueness stays visible without being scored.

The five status words

What counts as a hit, by kind

• Capability ("AI will be able to do X by Y"). A hit needs a dated public demonstration that meets every stated condition, from the paper, the maker's release, a benchmark's official result or a competition result. This is the leg B ledger's standard for "arrived", and 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, but evidence_by is maker and the page marks it. Where the claim's own test is a milestone the leg B ledger already judged, the claim's verdict follows that row; the same fact is never judged twice.
• Timing ("AGI by 2027", "superintelligence in a few thousand days"). 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 before the due date; the ledger does not borrow the singularity test's thresholds for this, because that would let the ledger move the test (section 5).
• Deployment ("agents will join the workforce in 2025", "1 billion users by Y", "a robotaxi service by Y"). A hit is 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 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 the entry says so.
• Where the claim itself is a range of years ("2027 to 2034"), the due year is the later end, and the earlier end is recorded, so the page can show that the claim was already late against its early end.
• A point year with a spread beside it ("roughly human level AGI in the year 2025", with a stated 90% 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 goal ("we have set internal goals of ... by September 2026; we may totally fail") is a dated public commitment with the hedge in the speaker's words. It is admitted, marked is_goal, and scored like any other claim; the page shows the mark. Decision 11.

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 verdict like any other, and kept out of the plain count; it is listed for a later calibration check once a person has enough of them. 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

A leader who says the same thing again has not made a new prediction, and one who changes the date has.

• 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, as The Outrunning Ageing does for Kurzweil's 2032 restatement of his earlier date. 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.

The rate, and what stops it being gamed

• Per person only, never pooled. For each leader: hits, missed (including withdrawn), not yet due, unfalsifiable, with the number of due claims always printed beside the rate. The rate is hits over hits plus missed. The sibling pattern pools nothing, and a pooled "leaders' hit rate" would make three people answer for a fourth.
• Which way the misses went. For each missed claim, whether the thing came later than said, or has not come, against the few that came sooner. The leg B ledger reports direction for the same reason, and for these speakers it is the half of the question the rate alone hides.
• A floor. A person with fewer than five due claims shows "too few to rate" and the counts, not a percentage. Decision 7.
• A collection pass before any rate. The seed list below is the famous, checkable claims found for this proposal, not a complete record, and a rate computed from it would measure which claims were easy to find. Proposed: no percentage is published for anyone until a person has been traced year by year, as the panel's collection does, with the searches logged in data/search_log.csv and quiet years entered as rows. Until then the page shows the seed claims and counts under the plain label "seed only: not a rate".
• Both directions of selection are logged. The searches that found each claim, and the claims looked for and not found, are written down, as docs/panel-verification-2026-10.md does for the panel.

4. How it relates to the leg B ledger and the coded statements

The leg B ledger (data/leg_b_ledger.csv, method in docs/leg-b-ledger-method.md, first reading in docs/leg-b-ledger-2026-10.md) scores the 2016 researcher survey's 32 milestone forecasts: the combined forecast of the 352 researchers who answered (each milestone's even-chance year) against the date each milestone was first publicly demonstrated. The claims ledger is the same operation on a named person's own forecast. They share three things on purpose: the evidence standard (a dated public demonstration from the paper, the maker's release or a competition's result, and "no later than" because a demonstration can only follow feasibility), the verdicts' meaning (arrived and not arrived become hit and missed; the "no bar" rows become unfalsifiable, with the same ruling that a bar is never set after the answer is known), and the report of which way misses went. They differ in two things. The ledger scores a crowd's median at one fixed date (28 June 2016); the claims ledger scores individuals on many dates. And the ledger's rows are fixed once; claims keep arriving. Neither feeds the other's numbers: the claims ledger never changes a leg B row, and it never becomes input to leg B's gauge (section 5). Where a claim's test is a milestone the leg B ledger already judged, the claim's verdict cites that row.

The coded statements (data/statements.csv, 159 rows) are the seed and the sibling in the record. Of the 159, 43 carry a year in the timeline fields (point_year, range or probability by year): Kurzweil 18, Altman 9, Hinton 6, LeCun 5, Marcus 5. Nobody else on the panel has a coded row yet; Musk, Amodei, Huang, Hassabis and the rest of the panel have a speaker row and no statement rows, which is why the seed list in section 8 draws on primary sources opened for this proposal and not only on existing rows. Four rules keep the two records from tangling:

1. A statement row is not a claim record. A row says what was said and codes it. A claim record adds a bar, a due year and a verdict. A claim points at its row with statement_id and the validator checks that the speaker and the date match and that the claim's quote lies inside the row's quote (or says in notes why the span differs).
2. Nothing is stored in two places. The statement row is not edited when a claim is entered, and a claim's verdict is never copied back to the row.
3. One source can seed several claims. A coded row carries one timeline; a source such as Altman's "Reflections" post says more than one dated thing. Each dated thing is its own claim record pointing at the same row.
4. Not every dated row becomes a claim. The seed rule is: every row with a year in its timeline fields whose speaker is on the panel and whose elicitation is not quoted, each opened at its source and put through the admission rule. The 15 quoted rows among the 43 (a reporter's words around a remark) can enter only as coverage_only, until the primary source is found. Many of the 43 will fail the bar test (they say a stage arrives by a year but not what would count as arrival); those become unfalsifiable, which is the honest outcome for them.

5. What it must never touch

The ledger is a separate record kind beside the others, and the rule is the tracker pattern's first claims rule (6.1): talk and delivery are never mixed. So, in writing:

• The frozen panel (data/speakers.csv, frozen by amendment A18, the panel freeze). A claim by someone not in it is refused by the validator. Adding anyone to the ledger's reach is an amendment to the panel first, as the README already says for any new person.
• The coded statements. No row is edited, added or deleted because of a claim.
• 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 ("Not yet.") and the headline of 2026-10-17 (amendment A25, the average share of the way). The headline's number is the average share of the way across the measured parts; nothing in the claims ledger enters it, and status.json's headline fields do not change. The ledger may add its own fields to status.json (decision 9) but never a field the parent site could mistake for the headline.
• The weekly reading's rule (docs/weekly-reading.md): the reading may append rows and flag, never change what is scored. If John extends it to claims, it could flag a claim that has come due, and the verdict would still be entered by hand and dated.

6. Where it shows

Two choices.

A page of its own, /claims/, with one line on the home page (recommended). The home page is already one long scroll with five jobs; a table of a few dozen claims, with a block of quote, source and verdict under each, would bury the test. The page opens with the per-person table (counts, the rate or "seed only", which way the misses went), then a list of claims by person in date order, each showing the person's words, the date, the due year, the status word as a chip in the family's yes/no colours, and the evidence. The home page's "current zeitgeist" section gets one sentence and a link beside the experts' dates: "Of the predictions with a date that the heads of the AI companies have made, N are now due, and M came true." Per rule 6.4 of the tracker pattern (talk and delivery side by side), the claims page repeats the four gauges' shares in a one-row strip at the top, drawn from the same figures, so the talk is seen against the delivery; the strip is display only and has no link to any claim. The top bar gains a "Claims" handle in the heavier tier (rule 4.7: short handles; the siblings that have a claims page all call it "Claims"), which makes the bar four heavy handles and three light.

The alternative is a section on the home page. The pattern's own reasoning for one page ("the claims log as a page of its own with no delivery beside it would read as a collection of quotations rather than as a test of them") argues for it. The cost is length: the home page would carry a long table above the method section, and both siblings that keep claims (The Upgrade Tree and The Outrunning Ageing) give them a page of their own. A middle way is built into the recommendation: the page repeats the gauges strip.

If built, scripts/site.py reads data/claims/*.toml, writes site/claims/index.html and site/data/claims.csv, adds claims counts to status.json (decision 9), and scripts/validate.py gains a claims check. None of this is in this proposal's change.

7. Jensen Huang

He is on the frozen panel already (section 0): the 2020s cohort, by amendment A18 (the panel freeze), and data/speakers.csv says so. The tracker's open question in TimeAssembler (the one that asks whether to add him by amendment) rests on a premise the record does not support, and should be closed or reworded. What adding him would have taken, had he been absent: an amendment entry in docs/amendments.md giving the reason, his evidence against the cohort rule (a timeline statement dated 2024 or later, none dated before 2018) and the panel counts after the change, as the panel freeze (A18) did for the other eleven. None of that is needed.

What is true of his row is that it is thin, and the seed pass changes one fact about it. The row says both of his statements on the panel are secondhand. Both were opened at the source for this proposal: the "fairly competitive with humans in five years" remark is from the New York Times DealBook event on 2023-11-29, and "I think we've achieved AGI" is from Lex Fridman's podcast of 2026-03-23, in answer to when an AI could run a company worth a billion dollars. He said it again on Nvidia's earnings call of 2026-08-26 ("in a lot of ways, and for many tasks, we could say that we've already achieved AGI"). The row's "latest 2026-03" is therefore out of date by five months; I have not changed it (nothing in data/ changes with this proposal), and it is a small fix for the collection pass.

Why a table of his own, as John asked: his claims are a different kind. Most are about what Nvidia will ship or sell (a chip platform in the second half of 2026, orders worth a trillion dollars through 2027), which Nvidia's own results can settle, and the claims about AGI itself are the loose kind (the "achieved AGI" remarks have no test). He is the clearest case of the "salesman" half of John's question, since his company profits from the story whichever way the models turn out. He is not tagged frontier_lab (the panel's note: chips, not models), so decision 10 in section 9 asks whether a ledger of "frontier leaders" includes him. Ten claims:

10 claims: 2 hit, 1 missed, 5 not yet due, 2 unfalsifiable, 0 withdrawn. Seed only: not a rate.

Reading it: three of the ten are due; two came true on Nvidia's own word and one (the Mercedes claim) is contested. Five are not yet due. Two are unfalsifiable, both the "we have achieved AGI" remarks: the speaker is the only judge. That is the record's ordinary rule, not a finding against him; the page shows the two as a separate count.

8. The seed list

What it is. 79 dated claims by 11 people (Musk 14, Altman 16, Amodei 12, Huang 10, and 27 from Legg, Hassabis, Suleyman, Sutskever, Hinton, LeCun and Kai-Fu Lee, the other panel members tied to the frontier labs who gave a dated, falsifiable claim). 17 point at an existing row of data/statements.csv (eleven Altman, three Hinton, three LeCun); the rest come from primary sources opened for this proposal, because the panel has no coded rows yet for Musk, Amodei, Huang, Hassabis, Legg, Suleyman or Sutskever. Each claim in the full appendix (docs/claims-ledger-seed-2026-10-17.md) carries the source and where in it, how the words were checked, the due year and how it was worked out, today's status and the evidence, and any flag. The tables below are the compact view. Kurzweil is left out of the seed on purpose: his 2029 date is the long chain the restatement rule is for (18 of the 43 statement rows carrying a year are his, seventeen of them pointing at 2029 and one at 2045) and is already in the siblings' records for the escape-velocity date; it should be entered once, with its repeats in repeated_on, when the ledger is built.

Status today: 45 not yet due, 16 missed, 10 hit, 8 unfalsifiable, none withdrawn. Only 26 of the 79 are due at all.

How to read the counts. They are not rates, and the page would not show them as rates (section 3, "a collection pass before any rate"). The seed was gathered by looking for the famous, checkable claims, and for Musk that means self-driving and robot promises, which are the ones most often quoted and most clearly late. Ten misses of fourteen for him says "the famous claims", not "the record". The same is true in the other direction for people whose claims are mostly hedged ranges (Hassabis, Sutskever, LeCun, Hinton: every claim is still ahead of them). The honest summary of the seed is that the cases which can be scored so far are mostly the deployment promises, and the AGI-timing claims have not come due.

Who did the work and how far it is checked. Four research passes opened the sources and proposed each verdict; the quotes were cut from fetched pages or captions, not typed from memory. I spot-checked: Altman's blog quotes against the live page, OpenAI's September 2026 announcement of its research intern and Tesla's July 2026 call (both by search, which agreed). The quotes taken from automatic YouTube captions have rough wording and the siblings' two-transcript check has not been run on them. Every verdict here is the agent's and John can overrule any one. Where a claim is contested the table says so in its last column.

Sources that would not open (the verdict does not depend on them unless said): Tesla's own web pages, Axios, the World Economic Forum site, openai.com (read through a reader copy or the Internet Archive), and, once, YouTube. Four claims are coverage only and carry no quote: Musk's 2020 "level 5 during 2020", Altman's reported "internal system he would call AGI by the end of 2026", Amodei's 2025 warning about entry-level white-collar jobs, and Kai-Fu Lee's 2017 "half of jobs" remark.

Findings about existing rows, none changed. (1) altman-2026-04 quotes TIME's paraphrase as Altman's words; only "not quite yet" is his. (2) altman-2023-06 (the governance post of 2023-05-22) has three authors. (3) The "Reflections" post is dated 2025-01-05 in Pacific time; altman-2025-03 has 2025-01-06, the UTC date. (4) altman-2021-01 omits the explicit five-year and ten-year points in the same passage. (5) The Huang row is out of date (section 7). These are for the collection pass.

The tables

The heads of the companies first, then the other panel members tied to the frontier labs. Huang's table is in section 7. "Today" is the verdict as of 2026-10-17, proposed from the sources named; John can overrule any one.

Elon Musk

14 claims: 2 hit, 10 missed, 2 not yet due, 0 unfalsifiable, 0 withdrawn. Seed only: not a rate.

Sam Altman

16 claims: 5 hit, 1 missed, 8 not yet due, 2 unfalsifiable, 0 withdrawn. Seed only: not a rate.

Dario Amodei

12 claims: 1 hit, 2 missed, 7 not yet due, 2 unfalsifiable, 0 withdrawn. Seed only: not a rate.

Shane Legg

4 claims: 0 hit, 1 missed, 2 not yet due, 1 unfalsifiable, 0 withdrawn.

Demis Hassabis

4 claims: 0 hit, 0 missed, 4 not yet due, 0 unfalsifiable, 0 withdrawn.

Mustafa Suleyman

4 claims: 0 hit, 0 missed, 3 not yet due, 1 unfalsifiable, 0 withdrawn.

Ilya Sutskever

3 claims: 0 hit, 0 missed, 3 not yet due, 0 unfalsifiable, 0 withdrawn.

Geoffrey Hinton

4 claims: 0 hit, 1 missed, 3 not yet due, 0 unfalsifiable, 0 withdrawn.

Yann LeCun

4 claims: 0 hit, 0 missed, 4 not yet due, 0 unfalsifiable, 0 withdrawn.

Kai-Fu Lee

4 claims: 0 hit, 0 missed, 4 not yet due, 0 unfalsifiable, 0 withdrawn.

9. The decisions, each with confidence and the strongest alternative

1. Build a dated-claims ledger at all. Moderate-to-high confidence that the question is worth answering and that this project is where (it holds the panel, the statements and the leg B ledger). Moderate that it will stay honest, because the ledger is only as good as its admission rule and the collection pass (decisions 4 and 8). Strongest alternative: do not build it; leave the leaders' dated words as coded statements and let the page keep showing their dates beside the test. That costs nothing and risks nothing, but it cannot say whether any of them was right, which is the question John asked.
2. One TOML file per claim in data/claims/, with a flat claims.csv written for download. Moderate confidence; a judgment call. Strongest alternative: one data/claims.csv, the form this record otherwise uses. Simpler to read as a table and no new file kind, at the cost of long prose in cells, harder diffs, and two sessions editing the same file.
3. Five status words; withdrawn counts as missed in every total. Moderate confidence. The count rule is the siblings' ruling of 2026-10-17; keeping the word is mine, so a retraction shows. Strongest alternative: the siblings' four words (not yet due, hit, missed, unfalsifiable) with a withdrawn_on date field on a missed claim; identical family vocabulary and one fewer word, at the cost of the retraction being less visible.
4. A claim with no date or length of time in the words is not a claim record; it stays a coded statement, and the page counts each person's undated forward-looking statements so vagueness is visible. Moderate confidence. Strongest alternative: the siblings' rule, score undated claims at a fixed year (2050 there). It keeps family vocabulary and loses nothing at entry, but it makes "soon" and "close" unscorable for 24 years, so the ledger could not answer John's question inside the project's life; and 2050 is those sites' own horizon, not this one's. A third option, enter them all as unfalsifiable, buries the scored claims.
5. A hit needs every condition of the bar; there is no "partly". A near miss is a miss with the gap named. Moderate confidence. Strongest alternative: the leg B ledger's two readings, strict and generous, printed together, which reflects that the judgement sits between them and costs a second status word.
6. The due year is the end of the named year, a span runs from the day it was said, a range takes its later end, a point year with a spread is scored at the point. Moderate-to-high confidence. Strongest alternative: take the earlier end of ranges, which is harsher and closer to what a reader hears ("5 to 20 years" as "5"), but turns every hedge into a miss.
7. No percentage for a person with fewer than five due claims. Low-to-moderate confidence; five is a convention, not a derivation. Strongest alternative: no floor, always print the counts, never a bare percentage. Both show the counts; the floor only stops a headline like "100%" from one claim.
8. No rate is published for anyone until that person has been traced year by year, as the panel's collection does, with the searches logged and quiet years entered. Until then the page says "seed only: not a rate". Moderate-to-high confidence; without it the rate measures which claims were easy to find. Strongest alternative: publish the rate from the seed labelled as such; faster and answers the question sooner, at the risk that the label is lost when the number is quoted.
9. Per person, never pooled; the ledger adds its own block to status.json and nothing to the headline. High confidence on not pooling (the pattern pools nothing); moderate on the file block. Strongest alternative: no addition to status.json, so the parent site cannot read the claims at all; safer, but then the family's Projects page cannot show the ledger.
10. Scope of people: everyone on the frozen panel, begun with the four leaders and the other lab-tied members in the seed, and Huang included on the same footing although he is not tagged frontier_lab. Moderate confidence. Strongest alternative: the four leaders only; a cleaner answer to John's question, but it leaves Hassabis, Legg and the rest out, who gave some of the most checkable dated claims in the seed.
11. A stated goal ("we have set internal goals of X by Y; we may totally fail") is admitted, marked as a goal and scored like any other claim. Moderate confidence. Strongest alternative: exclude goals as not forecasts; avoids punishing a hedge, but a company's dated public commitments are what the hit rate is for, and leaving them out hands every leader a way to phrase a forecast as a goal.
12. Numbering and order of work. Adopt as the next amendment (A26, the next number after the headline amendment A25: the ledger's existence, the five status words, the admission rule, the undated rule, the due-year rule, what it never touches), with a method document docs/claims-ledger-method.md written and committed before any row, as the leg B ledger's method was. Moderate confidence. Strongest alternative: no amendment and a note in the README, on the ground that it adds a record without touching the panel or the test; John has preferred the amendment file to be complete.
13. A probability stated as a number (a 50% chance) is kept and flagged but left out of the plain count, for a later calibration check. Moderate confidence. Strongest alternative: score such claims with a probability score (the squared distance between the stated chance and what happened), which treats them properly but needs many claims per person before it means anything.
14. Where it shows: /claims/ with one line on the home page and the four gauges repeated in a strip at the top of the page. Moderate confidence. Strongest alternative: a section on the home page (the pattern's rule 6.4 argues for it).

10. The rulings John must make

The first four are the ones the master session named; the rest each have a default I would take if John answers "as recommended".

1. Adopt the ledger or not. Recommended: adopt, on the terms below, and write the method document first (decisions 1 and 12).
2. The status words and the undated-claim rule. Recommended: the five words with withdrawn counted as missed (decision 3); undated claims stay coded statements, counted but not scored (decision 4); every condition must be met for a hit (decision 5). The alternatives are the siblings' four words, and scoring undated claims at 2050.
3. Huang first or not. Correction first: he is already on the panel, so no amendment is needed. Recommended: not first. Begin with Altman (eleven of his sixteen seed claims already point at coded rows, nearly all with primary sources) and Musk (the most dated claims), then Amodei, and Huang after them, since his panel row is the thinnest and his dated claims are mostly Nvidia's own sales and shipping. The alternative is Huang first because his claims are the easiest to settle from the company's own releases.
4. Where it lives. Recommended: a page of its own, /claims/, with a home-page line, a "Claims" handle in the top bar's heavier tier, and the gauges repeated at the top (decision 14). The alternative is a home-page section.
5. Scope of people. Recommended: everyone on the frozen panel, starting with the leaders and the lab-tied members (decision 10). Or the four leaders only.
6. The record's form. Recommended: one TOML file per claim (decision 2). Or one CSV.
7. No rate before a person is traced; no percentage under five due claims (decisions 7 and 8). Or a labelled seed rate.
8. Goals and probabilities. Recommended: admit goals, keep numeric chances out of the plain count (decisions 11 and 13).
9. Numbering (decision 12): an amendment, with the method document first.
10. The open TimeAssembler question about adding Huang rests on a premise the record does not support (section 0). Recommended: close it with a note that he is on the panel, and open a smaller question only if John wants a note on the thinness of his row.

If John rules "as recommended" on all ten, the next steps are, in order: the method document and the amendment entry; the validator's claims check; the first rows for Altman and Musk with their searches logged; then the page. Nothing in this proposal starts any of them.

Rendered from docs/claims-ledger-proposal-2026-10-17.md at version a7bd17a.