The AI Prediction Ledger: Who Was Right, Who Was Wrong, and What Comes Next

Display eligibility is separate from the preserved score. The approved record must still match its reviewed claim, source, horizon, forecaster, topic, status and score. Only False or Partly supported statuses can contribute; the complete score must be an integer from −100 to +100. New or changed records are excluded until explicitly reviewed. Each duplicate group has one designated representative, and each group can contribute at most once.

Records 001 and 064 are explicitly excluded because their superseded historical attributions remain unresolved; 077 is excluded because its current attribution is unverified. Every other non-approved record is excluded by default. Known duplicate/overlapping groups are 009/043; 010/044; 011/071; 012/083; 013/078; 016/086; 017/087; 018/093; 019/095; 020/088; 096/097. A duplicate never acquires eligibility merely by receiving a numeric score.

Approving historical-score display does not newly validate the score. In particular, record 006 still lacks an industry-wide code-share denominator; 015’s former fleet-count/error estimate was not primary-verified; 019 lacks a specified general engineering-equivalence threshold. Their earlier qualifications remain in the ledger. All 100 rows remain available in the exports, including excluded records.

This manual review covers five records. The remaining 95 await a fresh semantic review. Scheduled monthly-run proof remains pending.

All historical scores and statuses are unchanged. The previous full page and registry are archived; corrected current wording does not erase the prior record.

This is a manual source/deadline audit, not an ordinary monthly scheduler invocation or blanket outcome verification. Source-access limitations and unresolved attribution are explicit below. Historical numeric scores are unchanged; their retention is not endorsement of the old scoring method. The other five records retain their October 1 review.

This manual follow-up supersedes the earlier source-access notes for these 22 records. It does not validate outcomes or historical scores. The unchanged nine records remain attribution-repair items; they are not newly source-verified.

This manual follow-up supersedes the v1.6 access notes for the nine remaining gaps. The repairs verify forecast attribution and qualifiers only. All historical statuses and scores remain unchanged and are not newly validated. Tesla’s 2016 press-call source is an edited recording mirror with visuals added; the original spoken statement is distinguished from third-party narration and captions. BFM Tech broadcast the original 2024 VivaTech event.

Original recordings support the source repairs for 017, 087 and 089. The BBC clip’s own captions were checked against direct-audio local transcription. Gates’s publisher-feed audio was transcribed locally in context and checked in a separate cropped pass; machine text is not a publisher transcript. No paid API was used. Prior source-access notes are historical and superseded by this review. All historical statuses and scores remain unchanged. Records 017 and 087 overlap.

Remaining source gaps: 001 (original Coxon numerical post), 064 (exact 2019 Level-5 deadline), 070 (original 2021 unsupervised-fleet statement). These records were not re-reviewed or changed in this scoped interview check.

Current wording is restored to the inspected originals. This does not validate the superseded numerical Coxon attribution or the exact Level-5/2021 paraphrase. Those remain historical evidence tasks. All historical statuses and scores are unchanged. No new outcome or calibration verdict is assigned.

Prior Coxon presentation and context (historical, superseded)

 
           Jacob Coxon  Former pretraining researcher, Anthropic and OpenAI · September 8, 2026   Pending  
         A 10% chance that AI causes human extinction within the next decade. 
          Pending  Risk  Deadline: September 2036  Probability supplied  
         This is a probabilistic forecast, not a claim that extinction will occur. It cannot be marked wrong merely because the outcome does not happen; calibration requires a reference class of comparable forecasts. 
          Source and incentive context  Coxon made the estimate while publicly resigning and said he left before his Anthropic equity vested. That reduces one obvious financial incentive to promote Anthropic, but does not validate the forecast. Sources:  Associated Press ;  Axios interview .  
        Historical v1.5 review:  The original AP link could not be accessed. A later original TMZ interview report corroborates the broad estimate but does not verify the original September 8 attribution. Historical record retained; no source replacement or outcome finding. Historical status and score retained; this is not a new calibration verdict.   October 2 v1.6 source repair (supersedes the v1.5 source-access note):  Original September attribution unresolved. Coxon’s original X post states a qualitative concern, not the recorded 10% probability. A September 11 original TMZ interview report concerns a later, conditional estimate; it cannot be silently backdated to September 8. Historical status and score retained.   October 2 v1.7 follow-up (supersedes the v1.6 access note):  AP’s September 14 report attributes 10% over a decade to social posts, but its linked September 9 report only reproduces qualitative warnings. The original numeric September 8 post remains unlocated. The later September 11 TMZ estimate cannot be silently substituted or backdated. Historical record retained and explicitly unresolved. Historical status and score retained.     The current alarm is now on the record.  Jacob Coxon assigned a 10% probability to AI-caused human extinction within a decade after leaving Anthropic. Dario Amodei then called for slower frontier capability gains, and Sam Altman publicly agreed. Those are consequential statements. They are not evidence, by themselves, of a coordinated “psy-op” or regulatory capture. The ledger records commercial and policy incentives, but it does not infer hidden motives without documentation. 

This manual review adds preserved-source evidence. It changes no prediction, deadline, status or score. The superseded assertions remain in the earlier correction history and are still unverified.

001 · Coxon: preserved warning, no probability in this post

The archived original X post, captured September 9, 2026 at 04:15:21 UTC, preserves Coxon’s author name, handle, complete warning and displayed September 9 timestamp. It reports AI builders’ belief that AI could kill all humans by the end of the decade. It supplies no numerical probability. This corroborates the supported v1.9 wording; it does not establish Coxon’s former personal 10%/ten-year attribution or prove that he never supplied a probability elsewhere.

064 · Musk: the recovered Level-5/2021 exchange is dated 2021

Tesla’s original webcast page links its January 27, 2021 replay, 31:57–32:24. The question asks why Tesla would achieve Level 5 autonomy in 2021. Musk responds that his confidence rests on his understanding of the technical roadmap and progress between beta iterations. The Level-5/2021 wording is in the question; Musk gives the confidence response. This is a separate 2021 exchange, not evidence of the ledger’s superseded 2019 attribution. It does not establish where that old paraphrase originated. Locators use the original upload’s automatically generated captions, which can contain errors.

Still needed: Coxon’s own original September 8/9 numerical statement specifying AI-caused human extinction and a ten-year horizon; Musk’s own original 2019 statement explicitly specifying Level 5 autonomy by 2021. Later statements cannot be backdated. Failure to recover either statement is not a finding that it never existed. Internet Archive’s CDX lookup returned a temporary-offline notice during this review; no Wayback capture was verified.

Historical-score display safeguards: explicit reviewed-record eligibility, designated duplicate-group representatives, strict full-string integer scores in −100…+100, eligible status checks and fail-closed exclusion of new/changed records. Historical gaps 001/064 and unverified 077 explicitly excluded. Chart relabeled Historical editorial scores; all 100 records, claims, sources, statuses and scores plus prior history preserved. No new outcome validation or API-contract change.

The previous chart counted any numeric substring in a score, without eligibility or duplicate checks. The new display policy preserves the five historical contributors but rejects unresolved, duplicate, malformed, out-of-range, unapproved or changed records. The old calculation and exact pre-update page are retained in the publication backup. No forecast or outcome is newly resolved.

Published the five historical-score outcome audits beside chart values and in their detailed records, with source links and unresolved resolution requirements. Added consistent row alignment, responsive annotation spacing and room between page sections. All 100 canonical records, historical statuses/scores, eligibility policy and prior revisions preserved. No new outcome verdict or API-contract change.

Two narrower findings have stronger qualitative support: the missed Tesla 2020 operating-robotaxi milestone and Hinton’s admitted five-year timing miss. The other three broader outcomes remain incompletely resolved by their linked evidence. No exact numerical score is reproducible from the published rubric. These audit findings annotate the historical judgments without replacing them.

Published prospective scoring methodology 1.0 for future assessments: frozen resolution contracts, evidence-supported outcome states, type-specific calculations, reproducible receipts, duplicate safeguards and separate prospective aggregates. All 100 records, historical claims, statuses, scores, eligibility, chart calculations and prior history preserved. Policy publication only; no new scoring engine, outcome assessment or API-contract change.

Methodology 1.0 applies to future assessment receipts. Retrospective contracts must be labeled and reported separately. The five historical numeric scores remain editorial judgments under their existing safeguards; none has been reassessed or enrolled in a new aggregate.

Published versioned assessment receipt downloads beside prospective methodology 1.0: blank JSON template, TXT usage guide and explicitly fictional JSON worked example excluded from real aggregates. All 100 records, historical claims, statuses, scores, eligibility, chart calculations and prior history preserved. Supporting files only; no scoring-engine activation or API-contract change.

The fictional example is a format demonstration, not an assessment receipt for a real prediction. These files are separate from the canonical ledger export contract. New receipt-template releases must use a new versioned directory rather than silently replacing these published files.

Published read-only assessment receipt validator 1.0 and its usage guide beside the receipt templates, with explicit verification limits. Local format, declared-rule and arithmetic checks only; no evidence authentication, server-side validation, scoring-engine activation or historical reassessment. All 100 records, claims, statuses, scores, eligibility, receipt-template files and prior history preserved; API contract unchanged.

The validator download uses its own immutable versioned directory. A passing result is a mechanical check, not factual verification or publication approval. The fictional example remains excluded from real assessments.

Added a reproducible, explicitly fictional validator report with two calculation errors and three incomplete review fields, annotated explanations and separate input/JSON-output/TXT-notes downloads. Existing validator 1.0, receipt downloads, all 100 records, historical scores, eligibility, policy and prior revisions preserved. No real assessment, scoring-engine activation or API-contract change.

Added a short validator troubleshooting section covering common input, calculation, evidence-reference, unresolved-outcome, review-completeness and duplicate-family failures. Documentation only: all 100 records, historical scores, eligibility, validator 1.0, eight published downloads, policy and prior revisions preserved; no API-contract change or reassessment.

Clarified receipt documentation: completeness checks cover only a subset of policy requirements; unresolved receipts still need required attribution, contract and review fields; condition findings use not_met while the overall outcome uses condition_not_met. Published documentation revision 1.1 guides at new URLs and reconciled the stale local publication statement. All 100 records, historical claims, statuses, scores, eligibility, validator 1.0, original eight downloads, methodology and prior revisions preserved; no reassessment or API-contract change.

Published the printable one-page reviewer checklist PDF beside the receipt guides: 12 checkboxes, separate validator completeness and human policy review, reviewer fields and follow-up space. Reviewer aid only; not an assessment or publication approval. All 100 records, historical claims, statuses, scores, eligibility, validator 1.0, methodology, ten existing downloads and prior revisions preserved; no reassessment or API-contract change.