Independent check of the widened claims ledger (pull request 23), 2026-10-19

Update, third round (commit 4bc00cc): merge once the five small fixes in "Third round" below are made.

• One fix changes a count: Hassabis's 2011 "came true", which does not meet his own definition. It must become "did not come true", or go to John.
• The other four are wording. None needs another check from this session.

Third round (commit 4bc00cc): the named fixes, confirmed at the sources

The builder made the second pass's section A fixes, and John ruled items A3 and A5, in commits 4953661 and 4bc00cc.

• What was checked. Each fix, John's two rulings, and the builder's two extra changes, the search-log label and the word "create". The files changed were read again, with two helpers re-reading the sources.
• The page now reads 325 claims, 108 due, 37 came true, 54 did not. That is exactly what the four status changes give:
  • Pichai −1;
  • Suleyman +1;
  • Hassabis's Astra +1 (John's ruling);
  • Hassabis's 2011 claim +1, which does not hold (below).
• Checks. validate.py (325 claims, 4,922 search-log rows), check_news.py and test_check_news.py pass. The site builds, the panel is still 37, and the ledger-only list does not leak.

The five fixes before merging:

1. A4: set Hassabis's 2011 claim to "did not come true", with a dated note, or put it to John. The page then reads 36 came true and 55 did not.
2. A2: add the paper's "actively monitor ... code changes" sentence to Suleyman's reason, and say why the test is still met.
3. A14: correct whose account posted the WIRED repost, and the remaining 2024-11-10.
4. A15: correct the Vectara figures in Lee's hallucination reason.
5. The last label, in Zuckerberg's 20th claims search.

Fixes 2 to 5 change no verdict and need no further check from this session. Fix 1 changes the counts the news post's dated note will quote, so that note should be written from the page after it.

The third-round notes for each item are in claims-ledger-check-2026-10-19-per-claim-third.md.

Second pass (commit cf8fec5)

Verdict after the second pass (commit cf8fec5): merge after the named fixes in its section A.

• The first pass failed the claim files and is kept below as found. In it, 12 of 40 sampled files failed and about ten "came true" verdicts were in doubt.
• What changed since. John ruled the open readings, and the builder passed every file again.
• The re-pass worked.
  • In a fresh random sample, 3 of 40 files fail (first pass: 12 of 40).
  • Of the 35 claims now marked "came true", 2 fail.
  • Of the 33 first-pass failures, 27 are fixed, 4 were taken out properly, and 2 still need work.
  • The rules, the code, the search log, the ledger-only list and the page all pass.
• What is left is specific:
  • two verdicts that are wrong by their own tests;
  • a handful of tests that drift from the speaker's words, which must be fixed before publishing, because a test cannot change afterwards;
  • one quote whose speaker is uncertain;
  • one missed claim;
  • and Kai-Fu Lee's files, which were new this round and fail at 8 of 33.
• None needs another full re-pass. The managing session can merge once the "must fix" list in section A is done, without a third check, if the builder's commit for it is read by the managing session.

Checked by a session that did not build the work. Nothing was merged, pushed to the pull request's branch or published. The per-claim notes for both passes are beside this file: claims-ledger-check-2026-10-19-per-claim.md for the first pass and claims-ledger-check-2026-10-19-per-claim-second.md for the second.

The branch was read at cf8fec5: 325 claims by 15 people, 109 due, 35 came true, 56 did not.

What was checked: 138 files, plus the 4 taken out

• The 33 files that failed the first pass. 29 still exist and were read again; for the 4 taken out, the search-log row that records each was found, and each says "taken out before merging" with its reason.
• Every claim marked "came true": 35.
• Every forecast that something will not happen: 20. These are scored under John's second reading, and the bare "human level" in them falls under his fourth.
• All 33 of Kai-Fu Lee's files.
• A fresh random sample of 40 new files. It was drawn with Python's random.seed(2026101902) from the 172 new files in none of the groups above and not read in the first pass.

The groups overlap, so 138 files were read at their sources.

One count is my own call, not the helper's. The helper marked Suleyman's 2025 self-improvement experiment a pass with a flag. I count it as a failure, because its verdict contradicts its own test (item A2).

A. Must fix before merging

Verdicts that are wrong by their own tests or records (change a count):

1. 2025-04-24-pichai-waymo-atlanta-later-this-summer (the builder's own call f). It is entered as a separate record, a hit, for narrowing Atlanta to "later this summer". But on the same earnings call Pichai also says "first paid rides in Atlanta, via Uber, later this year" (official transcript, page 19). So the date was not clearly narrowed, and one Atlanta launch now scores two hits. Take it out as borderline, and add 2025-04-24 as a repeat date on the February Austin-and-Atlanta record. Pichai: 10 came true becomes 9.
2. 2024-10-suleyman-teams-experiment-with-self-improvement-in-2025 (the builder's call d).
   • It is scored "did not come true" because the Darwin Gödel Machine paper mentions "human oversight".
   • But the file's own test reads his words "doesn't have a human directly overseeing it" as no person approving each change. Its own reason grants that the paper meets that; the paper's safety line is about oversight of the whole experiment.
   • By its test it came true on 2025-05-29, and the test cannot be changed now. Suleyman: 3 came true becomes 4, and 6 did not becomes 5.
   • Then check 2024-04-suleyman-self-improving-ai-five-to-ten-years ("not before 2029") does not later read the same paper the other way.
3. 2024-05-14-hassabis-astra-capabilities-gemini-app-later-this-year. It is scored "did not come true", but its test asks for the capabilities "available to the public in a Google product such as the Gemini app". From 2024-12-11 Google AI Studio's "Stream Realtime" let anyone talk to Gemini over a live camera or shared screen. That rests on outside coverage and Google's developer forum; Google's own Gemini 2.0 post confirms the live video interface underneath it. Rescore it as came true, or put the "is a developer tool a Google product" question to John. Either way, the reason must deal with it.
4. 2011-01-hassabis-transfer-learning-concepts-five-years. It is marked "cannot be tested", but the same 2011 talk defines conceptual knowledge (22:32) and transfer learning (35:10). By the rule that the speaker's own definition in the same source is the test, write the test from them and score it (due 2016).
5. 2024-amodei-new-model-most-intelligent-by-airing. The re-pass made it "did not", saying the lead "rests only on Anthropic's word". That is wrong, and so was my first pass, which cited only the Chatbot Arena leaderboard (second place). An outside benchmark of the same week, LiveBench (arXiv 2406.19314, version 1, 2024-06-27), put Claude 3.5 Sonnet first overall. The outside record is split. Rewrite the reason to name both, log the LiveBench search, and put the status to John. "Did not" is defensible under "leads on the main published comparisons", but it should be his call with both facts in front of him.

Tests that do not hold the speaker's words (a test is fixed at entry, so these cannot wait):

6. 2026-01-lecun-smarter-than-humans-not-within-two-years: the test adds "at thinking work across the board" to his "smarter than humans". The same addition was taken out of his London file in the re-pass, so the two now disagree.

7. 2025-07-30-amodei-same-cost-much-more-capable-two-years: the test asks only for "scoring clearly higher on public tests". He said "much more capable of doing work much more autonomously much more broadly".

8. 2024-10-kaifulee-agi-better-at-almost-everything-five-to-ten-years: the test narrows his "almost everything we do" to "thinking work". Its timestamp also stops about 20 seconds before the quote.

9. 2024-10-kaifulee-robot-cannot-clean-a-room-within-three-years. The test changes his words twice:

   • it widens "this robot" (the home humanoid robots from Tesla and Figure the host had just named) to any household robot;
   • it raises "cannot even begin to do" to full cleaning across homes.

   As written, Physical Intelligence's π0.5 robot, shown on 2025-04-22 cleaning in homes it had never seen, may already defeat it. Write the test in his words.

10. 2025-10-30-kaifulee-company-run-by-ai-agents-five-to-ten-years: "we won't get there for another perhaps 5 or 10 years" is the shape of John's second reading (his own example is "it will take at least a decade"). Score it as a forecast that something will not happen, as Lee's GPT-5 file does with the same construction.

Kai-Fu Lee: a speaker, a missing claim, and the record:

11. 2026-09-23-kaifulee-embodied-agi-five-to-ten-more-years: could not be verified. In the Endgame podcast's captions a speaker-change mark falls at the start of this line (1:40:51), so it may be the host, Gita Wirjawan, speaking. Listen to 1:40:48 to 1:41:05. If it is the host, take the file out and correct its search-log row (Lee's sixth claims search, kaifulee-claims-006). If it cannot be settled, take it out as borderline.
12. A missing claim. In the dialogue of 2026-02-04 (the Fudan University repost, Lee's 60th claims search, kaifulee-claims-060), section 04, Lee says that within two to three years an AI will appear that does 90% of human tasks better than 90% of people, and that this will certainly happen.
    • It is dated, about AI, and testable by his own comparison.
    • It is the middle step between his October 2025 "about five years" and September 2026 "maybe two years", and neither the ledger nor the search log has it.
    • Enter it (due 2029, early end 2028, restating the 2025-10-02 claim), and point the September 2026 file's restates at it.
    • Log two lines of the same dialogue as left out: the conditional "if ... we can multiply the value humans create a hundredfold within five years", and the undated peak of AI hardware.
13. 2024-11-10-kaifulee-gpt5-not-within-six-months and 2024-11-10-kaifulee-consumer-ai-apps-explode-first-half-2025: a printed Chinese text of the full conversation exists. Recode China AI links it, and Tencent Tech's WeChat account posted it on 2024-11-05. So these are not coverage only, and the words were said by 2024-11-05, not on 2024-11-10 (the podcast page). The file names cannot change and the validator ties when to them. So say it in when_note and the source fields, and upgrade the source kind to the printed interview. The GPT-5 verdict stands.
14. 2024-11-26-kaifulee-ai-apps-proliferate-within-two-years and 2024-11-26-kaifulee-flood-of-ai-apps-in-2025. The notes say the WIRED column is signed as 01.AI's chief executive. Its byline calls him chief executive of Sinovation Ventures, and it never names 01.AI. He held the 01.AI role that day, so the claims still count; correct the stated fact.
15. 2024-03-kaifulee-hallucination-solved-in-eighteen-months. The verdict stands, but its reason misreports Vectara's table: it says frontier models make things up 3 to 6.5% of the time, but those figures are small models. The frontier models run about 6.5 to 14%. Correct the figures, and prefer OpenAI's post of 2025-09-05 ("Hallucinations remain a fundamental challenge"), which is dated before the due date.
16. 2024-02-20-suleyman-hallucinations-largely-eliminated-next-year. The verdict stands (the European Broadcasting Union study), but the reason gives the GPT-5 system card's error rate as "11.6% to 22%" and leaves out its 4.8% for gpt-5-thinking, the figure most favourable to him. State the full range and why the miss still stands.

Fixes that change no verdict but were asked for and not fully made:

17. The Grok 2 and Microsoft 365 Copilot dates.
    • Musk's Grok 2 post (2024-07-01-musk-grok-2-in-august) was made on 2024-06-30 in his own time.
    • Nadella's shareholder letter (2023-10-nadella-m365-copilot-generally-available-this-year) is signed 2023-10-16.
    • Both files explain this in when_note, but when cannot change while the validator ties it to the file name and the method forbids renaming.
    • Accept the notes for now; whether the validator should allow it is a question for John.

B. The search log, the page and the counts

• The search log.

  • Every search id named by every claim exists, with no duplicates and none borrowed from another speaker.
  • Kai-Fu Lee has 77 new rows.
  • No row says "withdrawn" any more.
  • The four files taken out after the first pass are each logged with their reason.
  • One wording fault (not blocking). In 22 rows a new reason is appended after an old label, "under the strict reading of the admission rule's second condition (the managing session's ruling: only claims about AI):". The reason after it is right: undated, "as little as", a hypothetical, or a borrowed span. But the label in front of it is wrong.
  • The rows include Amodei's 14th, 17th, 28th and 49th claims searches, and several of Suleyman's and Zuckerberg's. Drop the label where the reason is not a business forecast.

• The ledger-only list still does not leak.

  • The home page says "the 37 people on the expert panel".
  • The experts' dates come from 27 experts.
  • The frontier-voices chart reads no row by Zuckerberg, Pichai or Nadella.
  • data/statements.csv is unchanged by the pull request.
  • Kai-Fu Lee is on the panel already (tier core), so his statements belonging there is right.

• The page.

  • The summary reads "325 dated predictions by 15 people. 109 are due ...: 35 came true, 56 did not and 18 cannot be tested".
  • The counts table's "cannot be tested" column now gives the due ones in brackets, so every row adds up (Altman: due 6 = 2 + 1 + 3).
  • The coined label "ledger-only list" is gone from the page.
  • At 375 pixels wide nothing scrolls the page sideways. At 1,280 the table still scrolls inside itself, as before.

• The checks. validate.py passes (325 claims, 4,906 search-log rows), and so do check_news.py (3 posts) and test_check_news.py (8 tests).

• The method and amendment A30 now carry, with the date and John's words:

  • the four readings;
  • the managing session's rulings;
  • the rule that a test uses the speaker's words;
  • the change to when a claim is due, as A30 item 8, with the first thirty's 17 due;
  • Musk's Tesla years;
  • Kai-Fu Lee.

  The stale Kurzweil line is gone.

• Every first-pass item in "Wrong words, dates or records" is fixed except item 17 above. That includes Pichai's YouTube-chapters quote, now in his words.

C. The news post

The live post still says "twenty are due" in its summary, its after: line and its table, with Altman at 9. After the merge the page will say 17 for the first thirty, and 6 for Altman. Amendment A30 item 8 records why.

Main's correction already moved "five came true" to four, and the page agrees: Altman 2, Musk 2. The dated note the managing session adds at the merge must also give 17 due and say that the counting of "due" changed. Otherwise the post's table disagrees with the page it links to.

The home-page sentence ("325 predictions ... 109 are now due, and 35 came true") will move with items A1 to A5, and only from the merged files.

D. The builder's seven calls, which John has not ruled

E. Open for John after the merge (no change needed first)

1. Seasons. "This summer" is taken to end at the September equinox, applied the same way to hits and misses. Two hits rest on it: Pichai's Ask Photos (2024-09-05) and Docs Live (2026-09-03).
2. "A few thousand days". Altman's superintelligence line is still a dated claim (cannot be tested), though it names no length any more than "a few years" does. Under reading 3 it would be an undated statement. No due count changes, and the live post uses it as its example of a claim that cannot be tested.
3. Floors. "At least" floors are taken at the later end in one Suleyman file (no scaling slowdown for "at least two to three years", due 2027) and at the earlier end in another (self-improvement "a good five to 10 years", due 2029). One rule is needed. Forecasts that something will not happen "within N years" are judged at the end of the year, slightly later than the stated date; the tests carry the exact date, so nothing is wrong today.
4. Roles.
   • LeCun's chair of AMI Labs, a company that has released no model, decides every LeCun claim from 2026.
   • Legg counts as "Chief AGI Scientist", a co-founder in Google DeepMind's leadership.
   • Zuckerberg counts from 2013.
   • Lee: by his own account in February 2026, 01.AI now works only for businesses and leaves large base models to the big firms. His claims since then stay in (the ruling has no end date), but the role notes should say so.
5. "PhD level" is read as an undefined stage, like bare "human level"; it changes no count.
6. Restatements and repeats.
   • Whether a repeat that adds a definition becomes its own scored record: Legg's 2028 cannot be tested in its first source but is defined in 2023 and 2025.
   • Whether Huang's Jaguar Land Rover and Hyperion 8 misses count the same 2025 event twice.
7. The validator and dates, as in item A17.
8. Links inside the claim records' own text, as for the first thirty.

The other smaller points (caption mishearings, a verdict date set at the due date, a verify field left on the research-intern file, a podcast's channel name) are in the second-pass per-claim notes.

First pass (commit 2b66e6a), kept as found

Verdict then: do not merge yet. The rules, the code and the ledger-only list were sound; the claim files were not. The rest of this part is as written then, and its "Kai-Fu Lee" placeholder is answered in the second pass.

1. Claims checked against their sources

What was checked: 90 files.

• A random sample of 40 of the 278 new files, drawn with Python's random.seed(20261019) from the 233 that are neither "came true" nor coverage-only (the other two groups were checked in full).
• Every claim marked "came true": 45, the 40 new ones and the 5 among the original thirty.
• Every coverage-only claim: 6, one of which is also a "came true" (LeCun's "not next year"), so 5 more.

For each one the source was opened (captions through yt-dlp, pages directly or through an archive copy), and the quote, the date, the role on that date, the admission rule, the due date, the verdict and its record, and the maker-or-outside mark were checked.

Result: 57 pass, 33 fail, none left unverified. Every quote but one is in the speaker's own words where the file says. The one exception is listed under "Wrong words, dates or records" below.

Not every failure changes a number. They fall into four kinds.

"Came true" verdicts that do not hold as written (change a count, or need John)

1. 2025-10-29-altman-automated-research-intern-by-september-2026 (one of the live thirty). The test includes "running on hundreds of thousands of GPUs". OpenAI's post of 2026-09-06 says nothing about that, and the file admits the condition is unchecked. Under the ruling that a hit needs every condition, it is not a hit. The test was also taken from OpenAI's own announcement of success, not from Altman's words. This changes the live news post's "five came true".
2. 2024-04-12-amodei-asl3-this-year-or-next. Anthropic switched on its Level 3 protections in May 2025 "as a precaution", saying it had not determined the model crossed the line. In the same interview Amodei defines the level as a real rise in bioweapons risk. By his own definition the record shows no crossing: a miss, or a call for John.
3. 2024-amodei-new-model-most-intelligent-by-airing. "Most intelligent in the world" rests only on Anthropic's word. The main outside leaderboard of the time (LMSYS Chatbot Arena) put Claude 3.5 Sonnet second, behind GPT-4o. A miss, or a call for John.
4. 2023-12-06-hassabis-gemini-ultra-developers-enterprise-early-next-year. The record cited (Google Cloud, 2024-02-15) offers Ultra only "via allowlist", the same by-invitation access the promise said would come before the wider roll-out. No record of open developer access by 2024-03-31 was found.
5. 2018-05-10-huang-self-driving-cars-2020-to-2021. In the sentence just before, Huang sets Level 2 against "a Level 5 self-driving car". The test ignores that same-source usage and asks only for Honda's Level 3 traffic-jam car (about 100 leased in Japan). On his own usage it is likely a miss.
6. 2023-09-27-zuckerberg-glasses-multimodal-ai-early-next-year, 2024-10-30-zuckerberg-smaller-llama-4-early-next-year, and 2024-12-05-suleyman-copilot-vision-general-availability-early-2025. All three read "early next year" as the first half and are hits only on that reading (releases on 2024-04-23, 2025-04-05 and 2025-06-12). The Hassabis and Huang files read the same words as the first quarter, where that produced a miss (Huang's Mercedes date) or a hit anyway. A generous reading that appears only where it makes a hit looks like choosing the reading by outcome. See the open readings below.
7. 2024-07-23-zuckerberg-meta-ai-most-used-by-end-of-year. Meta's own counts show its size. "More people than any other assistant" rests only on Meta's statement about its rivals, which is not a figure only Meta holds.
8. 2025-01-24-zuckerberg-meta-ai-leading-assistant-billion-people. The billion is shown; "leading" is not. On the same call of 2025-07-30 Zuckerberg named becoming "the leading personal AI" as a goal still ahead, and OpenAI reported about 800 million weekly ChatGPT users that October. A miss.
9. 2023-10-17-legg-agi-not-next-year-or-year-after and 2024-04-09-lecun-not-surpass-human-intelligence-next-year. Both are hits for a forecast that something will not happen. Whether such words are admitted at all is open point 4, and a "did not happen" can only be shown by a later statement, not by "a dated record of the event itself". They should not go out as hits before John rules on point 4.
10. 2024-06-02-huang-mercedes-self-driving-production-2025 (the hit may stand). The record cited (electrive, 2025-06-04) never mentions Nvidia; only Nvidia's own blog of 2026-01-05 says the car runs its software. The mark should be "maker", not "outside", and the blog should be the record. electrive also says Nvidia's system reaches the US in 2026 and only "potentially" Europe.

"Did not" or "cannot be tested" verdicts that do not hold as written

11. 2023-01-25-lecun-teenagers-train-image-models-within-two-years (changed from "came true" to "did not" before the pull request). The test adds conditions not in his words: home equipment, built from scratch, adapting a model not counting. The dated sentence says only "any teenager in the parents basement is going to be able to do this". Hugging Face's guide of 2023-01-26 fine-tunes Stable Diffusion on one consumer graphics card. Read plainly, it arguably came true. Leave it out as borderline or put the test to John.
12. 2024-02-20-suleyman-hallucinations-largely-eliminated-next-year (coverage-only). "As accurate as any content that you can find on the web" is a named comparison, so under John's ruling it gives a plain-reading test, not "cannot be tested". Scored, it is very likely a miss.
13. 2011-12-31-legg-proto-agi-within-eight-years. Marked "cannot be tested" on his remark that it "might be impossible to falsify", but his reason there is military secrecy. The post defines proto-AGI as a list of abilities an outsider could check, so it can be scored, probably as a miss.
14. 2023-07-14-suleyman-modern-turing-test-as-little-as-two-years. "Could be as little as two years away" is an earliest date, not a deadline, and his test is well defined. "Cannot be tested" is the wrong status: leave it out as borderline, or score it.
15. 2022-05-11-pichai-youtube-auto-chapters-80-million-goal. The miss rests only on finding no record. The record cited ("tens of millions" of videos) does not rule out 80 million, so "has not come" claims more than the record shows. (Its quote is also wrong; see below.)
16. 2025-05-20-pichai-gemini-25-to-search-this-week. The file's own source shows Gemini 2.5 in Search by July 2025, so the direction is "came later", not "has not come". The record only repeats Google's announcement, so the mark is "maker".

Admitted when the rules say leave out, or with a test that misreads the words

17. 2024-11-11-amodei-lex-2026-or-2027-straight-line. The date is conditional on a straight line he says he does not fully believe ("if you believe the straight line extrapolation"). It is closer to a hypothetical: leave it out.
18. 2025-05-23-hassabis-agent-fleets-four-or-five-years. "I could imagine that in the next four or five years", posed as a question for economists. A possibility, not a forecast: leave it out.
19. 2026-01-14-mensch-not-superhuman-in-every-domain-within-two-years. A forecast that something will not happen, with its two-year span borrowed from the sentence before. Leave it out, or hold it for point 4.
20. 2025-05-22-amodei-one-person-billion-dollar-company-2026. The test never requires AI to do the work; a one-employee holding company worth a billion would pass. Add that condition (from the question he was answering) or leave it out as borderline.
21. 2026-09-12-amodei-pace-swarm-could-take-over-internet-six-to-twelve-months. The test weakens "taking over the entire internet with a persistent botnet" to "at large scale". It must use his words. A test is fixed at entry, so fix it before publishing.
22. 2024-04-suleyman-self-improving-ai-five-to-ten-years. His words ("still a good five to 10 years before") are a floor, "not before about 2029", but the test is "by the end of 2034". Work arriving early, such as the Darwin Gödel Machine paper the ledger uses elsewhere, could then score as a hit, the opposite of what he said.
23. 2021-08-19-musk-tesla-bot-prototype-next-year (one of the live thirty; John has now ruled that the Tesla years count). The test is only that a physical robot prototype is shown, with no AI condition: a plain hardware date under the managing session's ruling. Give the test an AI condition or put it to John. It is one of the live five "came true".

Wrong words, dates or records (no count changes unless noted)

24. 2022-05-11-pichai-youtube-auto-chapters-80-million-goal. The quote drops words without marking it: "the number of videos with from eight million". The captions read "the number of videos with auto generated chapters from eight million". Restore the words (start at "a goal to 10x" to stay under 25 words).
25. 2021-03-16-altman-five-years-legal-documents-medical-advice (live thirty; the hit stands). The medical half was settled by OpenAI's ChatGPT Health page of 2026-01-07, not by the 2023 GPT-4 report cited. The verdict date, record and direction ("on time", not "sooner") need changing, and the quote is in the opening section, not "near the end".
26. 2024-07-01-musk-grok-2-in-august (live thirty). The post went out on the evening of 2024-06-30 in the speaker's own time, so when is 2024-06-30. The file name stays.
27. 2023-10-nadella-m365-copilot-generally-available-this-year. The shareholder letter is signed "October 16, 2023", so when is 2023-10-16, not just the month.
28. 2023-07-25-nadella-security-copilot-early-access-this-fall. The record should be Microsoft's own post of 2023-10-19 (not a news article relaying it), with that verdict date and the mark "maker". "Paid", which Nadella said, should go back into the test.
29. 2023-10-24-pichai-gemini-models-throughout-2024. The test needs more than one model released in 2024, but the one record cited shows one. Cite the 2024-02-08 Ultra launch or the May 2024 Flash preview; it also calls Gemini 1.5 Pro a February public release, while Google called it a limited preview.
30. 2024-06-25-suleyman-real-time-interruptible-voice-by-end-of-year (the hit stands). The record cited never says the user can interrupt the AI. Google's Gemini Live post of 2024-08-13 ("You can even interrupt mid-response") does.
31. 2024-10-suleyman-permanent-ai-memory-done-in-2025 (the hit probably stands). In context "we" is Microsoft's Copilot, not the field. Test it on Copilot's 2025 memory feature, marked as the maker's word.
32. 2024-12-05-suleyman-copilot-vision-all-apps-always-on-2025 (the miss stands). The reason describes mid-2025; Microsoft's post of 2025-10-16 says Vision went worldwide with full desktop sharing. What failed is the always-on part; rewrite the reason.

Smaller notes that change nothing (the darioamodei.com posts carry only a month; Pichai took office as Google's head on 2015-10-02, not the announcement date of 2015-08-10; caption mishearings in two Huang quotes; LeCun's AMI Labs title given as "executive" chairman where AFP says "non-executive") are in the per-claim file.

What the sample says about the files not checked

Of the 40 sampled files, 12 fail, and 7 of those 12 would change a status or take a file out. That rate, applied to the roughly 190 new files nobody has re-checked, means dozens more problems of the same kinds. That is why the verdict is a re-pass, not only the fixes above.

2. The search log

• Every search id named by every claim exists in data/search_log.csv; there are no duplicate ids, and each claim names only rows for its own speaker.
• 577 new rows, all dated 2026-10-19, none with an empty finding. Nothing earlier in the log was changed or removed.
• Every logged drop has a reason. About 200 rows record something left out. Each was read for a reason: business forecast, hardware date, borderline, hypothetical, not the speaker's words, role at the time, no length named, no test, a hope.
• The lead's 31 later drops (Amodei 18, Zuckerberg 11, Hassabis 1, LeCun 1) each appear in the row that first recorded them, with "later left out" and the reason, and none of their files remains.
• One wording note: those rows say a file "was withdrawn before merging". "Withdrawn" is one of the ledger's five status words (a speaker's retraction). "Taken out before merging" would not collide with it.

3. The ledger-only list does not leak

The site was built on the branch (.venv/bin/python scripts/site.py: "claims page: 308 claims by 14 people").

• The home page's panel count is 37: "This rests on the 37 people on the expert panel".
• The experts' dates still come from 27 experts.
• The frontier-voices chart reads 144 rows from 27 speakers, none of them Zuckerberg, Pichai or Nadella (checked by calling frontier_rows()). The tier filter is now the shared PANEL_TIERS.
• data/statements.csv is unchanged by the pull request (807 rows, no row by any of the three). The validator now refuses one if added.
• The three names appear in built files only on the claims page, the claims and search-log tables, speakers.csv and the docs, plus files that already held them on main and that this pull request did not touch. Those are the wide list of candidates, the Reddit pilot's codes, and one Google self-report in world_measures.csv.
• The claims page says plainly that the three are not on the list of experts. Their column of coded statements reads "none recorded: not on the list of experts".

Pass.

4. The change to how "due" is counted

Is the new counting right? Yes. A "cannot be tested" claim used to count as due from the day it was entered; now it counts only once its own date has passed. That matches the method's status table, which defines a claim's date as 31 December of its due year or its due_date.

• The page's summary now reads: "106 are due, their dates passed: 45 came true, 46 did not and 15 cannot be tested. Of those still ahead, 60 cannot be tested either".
• The per-person counts match the pull request's table.
• No claim is past its date and still "not yet due", and no claim got a verdict before its date.

Is it written down?

• Not in the method. The method nowhere says when a "cannot be tested" claim counts as due.
• Not in amendment A30. A30 does not mention the change at all. Its "Applies to" list covers site.py only for the tier filters.
• Fix: add one sentence to the method (under "The rate") and a numbered point to A30 (or its own amendment). It changes a published number.

Does it change the live numbers? Yes, and the news post would contradict the page.

The dated note the managing session adds to the post must say that the counting of "due" changed, giving 17 due for the original thirty, and must carry any change to "five came true". Otherwise the post's table disagrees with the page it links to. The new home-page sentence ("Of the 308 predictions with a date recorded so far from 14 of the people who run the AI companies, 106 are now due, and 45 came true") matches the page. Its 45 will move with the fixes in section 1.

5. The ten points left for John

John ruled points 1 and 2 on 2026-10-19:

• Point 1: Musk's Tesla-years claims count.
• Point 2: Kai-Fu Lee is added, for claims made while he led 01.AI.

For the other eight, the question is only whether any must change before the merge.

6. Checks and the page

• python3 scripts/validate.py: OK (speakers 53, statements 807, search log 4,744 rows, claims 308).
• python3 scripts/check_news.py: OK, 3 posts.
• python3 scripts/test_check_news.py: 8 tests, OK.
• /claims/ at 1,280 pixels wide. It reads correctly and nothing scrolls the page sideways. The counts table (1,318 pixels wide in a 760-pixel column) scrolls inside itself, so the rate and "which way the misses went" columns are off-screen until scrolled. That is the site-wide table style, and the same as on main.
• /claims/ at 375 pixels wide. Nothing scrolls the page sideways, and the claims read well.
• The counts table's "due" column no longer adds up from the columns beside it. "Cannot be tested" counts all such claims, so Amodei shows due 8 = 4 + 3 + 1 but "cannot be tested" 13. Showing that column as "13 (1 due)", as the pull request's own table does, would let a reader add it up.
• Plain language. The page names "(the ledger-only list)" in parentheses, a coined label. "On this page only" already says it.
• The method contradicts itself on Kurzweil. It now says panel members who never held a qualifying role, Kurzweil among them, are not entered. The next line, carried over from main, still says Kurzweil's 2029 date "is entered once ... when his turn comes". Delete that line.

7. Named fixes, in order

1. Settle the open readings John has not ruled. One rule each for: "early" (and "early June"); forecasts that something will not happen; "a few years"; earliest dates such as "as little as" and "at least"; "I could imagine"; and "human level" as a comparison or a stage. Then re-pass all 308 files against them.
2. Correct items 1 to 32 above, each a separate dated change to its file, with a dated note wherever a test changes.
3. Write the due-counting change into the method and into A30, and delete the stale Kurzweil line.
4. Add the role notes to the thirty original files and the two Legg files.
5. Check a fresh sample of the files after the re-pass; one helper per person is enough.
6. Write the news post's dated note from the merged counts: due for the thirty is 17, not 20, and say so; carry any change to "five came true"; and give the new totals.

Kai-Fu Lee's claims are checked in the second pass above.

Rendered from docs/reviews/claims-ledger-check-2026-10-19.md at version a7bd17a.