PredictionBench

Further evidence: Study 3 showed larger immediate policy-agreement change; Study 4’s public Claude Opus 4.6 elicited 10.8p more from a £1 donation bonus (95% CI 6.4–15.1). Bounded comparative capability, not generalized superiority or control.

Direction: Matched human comparisons support direction.

Magnitude: Substantial local advantage; broader superiority only partly supported.

Timing: April–May 2026 observations fall inside the approximate decade; preissue superiority rules out first-arrival or measured-progress credit.

Scope: Study 3: April 15–29; Study 4: April 30–May 13; June 15 preprint. Paid English-speaking adult completers; identities concealed, human matchmaking caused differential completion. Short, low-stakes text interaction, not durable beliefs, elections, broad mobilization or face-to-face persuasion. Earlier November–December 2025 studies are baseline; speed/length matching erased AI’s advantage. Changing models/tasks prevent longitudinal-gain inference. Not threefold total fundraising. Stronger smarter-agent/control-risk reading remains unresolved. Short paid experiments and model/platform effects; not broad, durable persuasion or loss-of-control evidence.

Shared evidence: Shared Chollet outcome work and investigators/platform, not independent replication. One persuasion lineage; weak-information continuation, not exceptional foresight.

Evidence confidence: Moderate for bounded comparative capability; lower for broader interpretation.

Score condition: Conditional y=1 accepts stated budget and documented retail token estimates. Human fees, failed-stream charges and external reviewer/orchestrator costs remain unverified; full-cost alternative unscored.

Probability precision: Approximate, item-specific—not group-derived. The ≈ marks approximate probability and losses, not an exact elicited point.

Resolution confidence: Moderate conditional confidence; stricter alternative unresolved.

Proper score: 2026-01-14: p(event)≈80%; outcome=1 (under the stated score condition); Brier≈0.04; log loss≈0.2231436 nats.

Outcome dates: 2026-07-23

Interpretation: Lower Brier and log loss are better. Revisions concern the same event, not additional forecasts. One event does not establish calibration or overall reliability.

Further evidence: An independent AI static review found no substantive gap in either proof and matched each proof byte-for-byte to its final cost-bearing logged write. Logged-token estimates, including failed attempts, are $23.45 for P3 and $9.11 for P6, totaling $32.56. All four harness rates match the provider legacy price table; dated launch and successor announcements corroborate input/output and cache-read pricing. This supports conditional y=1. Human fees, unlogged or failed-stream charges and reviewer/orchestrator inference costs remain unverified. These are not invoices or human-certified proofs. Numeric probability forecast: event resolution and proper scores are separate; no categorical win.

Scope: One event. P3 is the hardest candidate identified by a dated expert assessment; P6 is retained as additional difficult-problem corroboration, not a second forecast. No official IMO participation or medal is claimed. No game-specific generic-harness requirement. January edited; original-day wording unverified.

Evidence confidence: Moderate conditional event resolution; separate independent AI proof review and primary pricing support, with budget completeness and human-fee uncertainty explicit.

Score condition: Conditional y=1 accepts bounded skills/recovery and practical marginal-cost comparison. Broader sustained/across-title competence or matched Red human economics remain unresolved, unscored; no categorical win or calibration.

Probability precision: Approximate. The ≈ marks approximate probability and losses, not an exact elicited point.

Resolution confidence: Moderate conditional confidence; skill breadth and imperfect human comparator remain judgment-sensitive.

Proper score: 2026-01-14: p(event)≈80%; outcome=1 (under the stated score condition); Brier≈0.04; log loss≈0.2231436 nats.

Outcome dates: Run October 12; public October 20, 2026. October 29 video-link excluded.

Interpretation: Lower Brier and log loss are better. Revisions concern the same event, not additional forecasts. One event does not establish calibration or overall reliability.

Further evidence: Explicit probability record: conditional event resolution and approximate proper scores are separate; no categorical win or extra outcome vote.

Scope: One game/early span, not full-game/broad-title or age-matched Red equivalence. RAM-based settling synchronizes screenshots; model action decisions use screenshots, notes/history, without RAM, map or extracted screen-text input. Stricter no-tailoring reading unresolved. Standard model, no reported game fine-tuning; ordinary pretraining allowed, universal training audit unclaimed. ≈70 calibration calls/engineering are separate development overhead; local power/hosting unpriced. All 444 records structurally audited; selected transitions inspected, not full visual replay. Mistakes/blackouts occurred; project report, not independent replication. Official tariffs match; marginal API cost is not an invoice/all-in bill. Emerald first-gym average 1:22:05 is context, not measured Red parity. Historical paid-play $13/$15–20 per hour versus $5 for 1–1.5 hours: heterogeneous, not 2026 same-game quotes. Break-even 35.19/30.50/22.87 minutes at $13/$15/$20 hourly are sensitivity, not Red observations/pass thresholds; lower rates sustain stricter uncertainty. No matched pre-January baseline or first-arrival credit. August acceptance informs intent, not independent proof or a fifty-hour requirement. Deadline: December 31, 23:59 Pacific; early positive allowed. Post-cutoff video link excluded; other summary fields unchanged.

Shared evidence: One conditional probability event, not separate votes for model, run, badge, code or comparator; prior other-source holds remain historical.

Evidence confidence: Moderate conditional resolution. Same-run substantive play, pinned generic decision scaffold, actual badge and fully reconciled API accounting support the practical interpretation; broader skill and human-cost comparability remain judgment-sensitive.

Score condition: Conditional y=1 accepts same-population assessment, long dramas and practical estimates against original task/budget interpretation. Exact per-film quality/matched human-cost sensitivity: unresolved, unscored—not mandatory.

Probability precision: Approximate. The ≈ marks approximate probability and losses, not an exact elicited point.

Resolution confidence: Moderate conditional confidence; revised-study and pooled-score limits.

Proper score: 2026-01-14: p(event)≈80%; outcome=1 (under the stated score condition); Brier≈0.04; log loss≈0.2231436 nats.

Outcome dates: Generation/rating days unreported; publication dates in Outcome and Scope.

Interpretation: Lower Brier and log loss are better. Revisions concern the same event, not additional forecasts. One event does not establish calibration or overall reliability.

Further evidence: Explicit probability record: event resolution and approximate proper scores are separate; no categorical win.

Scope: Deadline: 2026-12-31, 23:59 Pacific. Training/development excluded. No commercial-polish, universal-success, single-call, personal-review, invoice or independent-replication requirement; automated editing allowed. January wording retained, not refetched; edit/history limits remain. Early attainment counts; year-end availability unguaranteed. Approximate-duration capability, not a logged four-minute run. Twenty paid raters assessed whole-drama coherence and boundary continuity; 43 outputs were ≈3, five ≈10, two ≈30 minutes. V1 (May 21, 2026): 20 unpaid raters, segment-level flow; corpus/judges/protocol changed. One revised study, no backward transfer or replication. Static ≈30-minute hospital/romance example supports recurring characters/plot, not full-motion quality. Full-motion supplement uninspected: exact authorized GET returned 403; all recovery stopped. Long outputs rated; earlier “unscored” suggestion overstated and retired. API $25–27/min includes retries: modeled four minutes $100–108 (alternative retry reading $133.60); demonstrated ten-minute duration, modeled API $250–270 ($334). Local compute extra; typical ten-minute pipeline 74–90 minutes, one 48GB A6000; conservative 1.5 GPU-hours occupancy context, not a meter or filmwise worst-case bound. Estimates, not invoices; maximum three retries are not observed retry counts. A funded five-minute, two-character student drama budget (£750; listed items £730, actors £250) planned a one-to-two-day shoot; funding closed April 10, 2026. March shoot/May release were plans, not verified completion, quality or full crew wages. Ordinary cost comparability is contextual—not guaranteed savings, cheapest-market floor, exact currency conversion or duration-prorated same-film quote. Undated four-minute, two-character short reportedly cost under £100 cash; unpaid labor/equipment/editing unspecified. Not pre-cutoff price proof, a human minimum or a basis for combining incompatible budgets.

Evidence confidence: Moderate conditional resolution. Same-population human assessment and practical production estimates support the original low bar, with individual-film quality and accounting uncertainty explicit.

Further evidence: Conjunction unresolved; simulated integration insufficient.

Direction: Prior interaction helps agents anticipate unstated requirements and produce artifacts or tool actions; memory ablation weakens this behavior.

Magnitude: Functional analogy; structure undefined and unverified.

Timing: Published before cutoff; 2026 remains open.

Scope: π-Bench (May 2026) uses 100 tasks in five 20-session workflows, simulated users/apps and one Nanobot-derived scaffold. Hidden intents directly completed: 56.84% for GPT-5.4, 60.56% for Opus 4.6; these are intent fractions, not whole-task success. Checklist completeness reaches 67.6%, not a pass rate. History removal reduces proactive scores 9.5 points on average. Proactivity also credits questions; simulated users eventually reveal missed requirements for reactive recovery. Model grading and reported expert audits are not independent replication. Proactivity-Gym (September 2026) tests 23 model/harness combinations in ten 7–10-simulated-day scenarios. Best task-capability score is 65.1, temporal allocation 51.7% and intervention-depth alignment 52.4%; metrics differ. Other models score below 20% on temporal allocation. Thirty people rated prepared scenarios/logs, not lived deployments. Correct actions can violate preferences and trust; release is promised, not verified. PersonalAlign uses 91 users’ 60-day Android records, not a deployed-agent trial. Preference-sensitive execution and proactive suggestions are separate modes: conditional success 42.3 versus 26.6 baseline; suggestion false alarms 49%. Suggestions are not external autonomous action. Cross-chat memory existed in 2024, with broader past-conversation reference in April 2025; no first-invention credit. Do not combine separate systems into a completed conjunction or impose arbitrary perfection. No separately dated current endorsement is inferred.

Evidence confidence: Moderate for bounded functional direction; whole everyday integrated capability unresolved.

Further evidence: Qualified central capability; full-domain reading unresolved.

Direction: Matched candidate-pool comparisons support improved accepted-answer reliability; continuation, not invention.

Magnitude: Domain breadth partial.

Timing: In-window centrally; broader endpoint open.

Scope: VeriFin is a preprint: first submitted 10 August 2026, evaluated v3 dated 23 October. XBRL facts, authorized formulas and SMT checks ground interpretation. Tests use 600 constructed filing questions and selected 67 FinanceBench/403 TAT-DQA numerical questions, excluding qualitative questions. Nineteen FinanceBench-specific definitions and XBRL selection limit generalization; withheld labels are not an independent leakage audit. Qwen3-30B correctly accepts 304/600, falsely rejects 63 correct answers and abstains twice; zero observed false accepts covers selected evaluable cases, not universal safety. Its 99.7% decision coverage is not answer coverage; precision interval 98.8–100%. FinanceBench: 24/67 correct accepts, 15 abstentions, eight false rejections; precision interval 85.8–100%. GPT-5.5 accepts 560/600, with small gains over existing verifiers. Unresolved-correctness cases are excluded from precision/accuracy, not coverage. Formalization errors remain; checks depend on correct formalization. Repair subsets are unpaired; inconsistent GPT-5.5 FinanceBench accuracy is unused. No independent replication or production-benefit evaluation. QueryProof tests 80 synthetic warehouse tasks: routed success 53.7% versus 30% baseline, but 13 of 37 returned answers wrong. Abstention/ambiguity remain weak. Prior test exposure persists despite phrase removal; one annotator, correlated templates and cluster-bootstrap intervals spanning zero limit inference. This does not refute every symbolic system. December 2025 VERAFI already combined financial tools and neurosymbolic policy guidance; no first-invention credit or post-issue performance claim. Graphical/arbitrary-relational coverage unverified. No added perfection/profit requirement. No separately dated current endorsement is inferred.

Evidence confidence: Moderate for bounded numerical business capability; lower for broad business-data generalization.

Further evidence: Human exposition, checking and extensions remain explicit. Verified contributions to decades-old conjectures meet the asserted target without relying on contest scores or recovered prior literature.

Direction: AI made new mathematical contributions beyond predesigned contest problems.

Magnitude: No required numeric/Millennium threshold.

Timing: May results are in-window.

Evidence confidence: high for contributions and timing; no claim of universal reasoning reliability

Further evidence: A²RD supports minute-scale continuity; producer accounts describe feature animation by small core teams. Qualified central enablement supported; stricter full labor-saving conjunction unresolved.

Direction: Same-workflow evidence supports direction, not invention or measured post-January improvement.

Magnitude: Comparative artist-team/months savings, broad reliability and wider population effects remain unmeasured.

Timing: June 2026 disclosure, not exact runs or first attainment; no arrival credit. The stronger-reading year remains open.

Scope: CHIEF is researcher-reported novice production/local evaluation, not independent, randomized or blinded replication. Humans direct, rewrite, select and approve; comparison jointly changes human revision and AI support. Headcount, sessions and working time are unreported; selected revisions are not a full labor ledger. No personal film/frame inspection, feature-length or autonomous-production inference. A²RD: Human-tested one-minute single-scene samples support bounded continuity; ten-minute demonstrations relied on model judging. Poq Mher’s producer describes 23 core crew, including 13 AI creators; inconsistent premiere wording prevents an exact release date. SPEED EFFECT reports 117 minutes and three production staff, not all contributors; humans scripted, directed and edited. Its stated calendar date conflicts with its weekday label. Generation-cost claims are not total budgets. No independent viewing, film-quality verification or controlled labor study. Co-Director tests twelve-second outputs, not long films. Undated project examples add no timing credit. Pre-2026 long-video systems limit novelty; no first-ever credit.

Shared evidence: Shared source; not independent.

Evidence confidence: Moderate for the bounded central-functionality interpretation; lower for comparative labor, broad production quality and timing.

Further evidence: Useful hiring and shift coverage coexist with purchasing errors, closures, redundant work, inconsistent pay, fabricated arrangements, privacy problems and human-prevented overlong schedules. Qualified emergence; better-informed strategy and full-conjunction reading unresolved.

Direction: Successful actual work supports assistance despite serious errors.

Magnitude: Two small monitored businesses; no superiority or adoption estimate.

Timing: In-window; first-crossing/full-year verdict unresolved. Year remains open.

Scope: One operator’s supervised experiments, not independent replications or prevalence estimates. Humans remained legal employers, guaranteed pay, monitored and overrode decisions. Models changed; counterfactual replays are not extra deployments. No matched managerial-quality comparison. A controlled strategy study found no full-sample gain but an experienced-manager subgroup improvement; pre-2026 acceptance prevents new-emergence credit. Charity facilitation improved neither consensus nor equity significantly. OECD cases are selected testimony without verified intervention dates or quantitative comparisons; undated NVIDIA figures add no timing support. Greater DAO/open-source impact: possible, not mandatory.

Evidence confidence: Moderate for qualified central emergence; low for breadth or strategic superiority

Further evidence: Complete-task success was 56%, versus 9% for an otherwise matched model without Index pretraining. This supports progress, not the entire broad reliability claim. Reliability unresolved; no exhaustive checklist.

Direction: Blind trials show concrete generalization in unseen homes and reduced task-specific data needs compared with the earlier system. This is more than an edited single successful video.

Magnitude: 56% completion and corresponding failures must remain visible; the 9% comparator is an ablation, not a measured January fieldwide baseline.

Timing: Positive progress occurred within 2026; year remains open.

Evidence confidence: High for the developer-reported figures and rubric; moderate for the progress inference; no independent replication established.

Further evidence: Full fidelity and access unresolved.

Direction: The French–Mooré integrated pipeline and native-speaker-tested adaptation are substantive matched progress, not a generic 70-language feature count.

Magnitude: No measured connectivity reach. Paper explicitly says computing demands are unsuitable for low-resource technological environments.

Timing: Public technical result appeared during 2026 before cutoff; fieldwork dates unknown. Year remains open.

Evidence confidence: Moderate for matched technical direction; low for community deployment.

Further evidence: Expert review accepted selected alternative arguments; citation and exposition weaknesses remained. These support partial direction, not the predicted persistence; whole outcome unresolved.

Direction: Compatible division of work, not necessary human control.

Magnitude: Selected proof tasks, not all science or field-wide prevalence.

Timing: October work postdates conservative issue interval; no precise deadline or timing hit.

Scope: Benchmark direction was imposed; the long-lasting relationship remains unestablished. Questions had known unpublished solutions; this does not establish new-to-humanity discovery. Very low-information continuation. One problem was corrected and rerun. Native duration is required; this correction adds no ongoing-lab partnership condition.

Evidence confidence: Moderate for bounded task-level delegation; low for extrapolating to natural field-wide or enduring practice. Very low predictive information against preexisting human-directed AI research.

Further evidence: The 22 October text-preference table scored Astra (Max) 1476±6 (14,076 votes), original GPT-5 (High) 1435±5 (31,712), and GPT-5 Chat 1426±4. GPT-5.6 Sol ranked above Astra. Qualified arrival/improvement conjunction under a bounded preference proxy.

Direction: Direct original-GPT-5 preference comparison supports improvement.

Magnitude: No numerical gain forecast or calibrated cross-domain magnitude; optional large gain uncredited.

Timing: Release and comparison fall within the approximate 2026–2027 forward horizon. No mandatory one-year lower bound, exact first crossing or timing-error estimate; literal delayed-arrival reading remains an alternative.

Scope: Preference-based, unequal reasoning settings; not universal superiority. A predictable continuation, with no numerical improvement promised. Exact interview recording date unknown; publication date is used. GPT-5.6 ranks above Astra in the same table. Self-selected prompts/preferences are neither semantic accuracy nor representative population benefit. Developer results use later GPT-5-family comparators and changed settings; no transitive original-GPT-5 inference. Mixed safety results and a September 22 HealthBench configuration correction limit generalization. The navigation study compares different harnesses/inputs/subsets and earns no grade credit. Original attribution is the interviewer-published dialogue, not its AI summary; audio was not independently checked. No current endorsement, exceptional foresight or separate launch/improvement votes.

Evidence confidence: Moderate for qualified successor/improvement conjunction; low for magnitude outside evaluated preference domain

Further evidence: Human steering fits the target; the first-party measurements are preliminary, incomplete and not independently replicated, with compute growth confounding the experiment trend. Concrete supervised research assistance matches the intern-level target; required human steering is not a target shortfall.

Direction: Not separately graded in the historical record.

Magnitude: Not separately graded in the historical record.

Timing: The September 6 report fits a September month-level horizon; a strict September 1 deadline remains unresolved.

Scope: Earliest actual attainment is unknown. Success analyses exclude uncertain outcomes and small samples; “researcher” includes support roles. Code volume is not research progress. This is not autonomous project leadership, replacement of all researchers, or fulfillment of the separate March 2028 researcher goal. This was a qualified goal and forecast, not an unconditional promise.

Evidence confidence: Not separately rated.

Direction: Usage grew beyond the original roughly 10% worldwide baseline.

Magnitude: Scale overestimated; no precise population conversion or acceptance band inferred.

Timing: Unresolved: measurement weeks unknown; 5 October was 12 days early. The 21 October more-than-1.2-billion disclosure is a lower bound, not a ceiling or dated miss.

Scope: Assumes contemporaneous reporting and no extraordinary twofold pre-deadline surge. Rounded issuer counts have unknown deduplication; disclosures are not independent. Target: worldwide weekly people, not accounts, monthly traffic or all OpenAI products. Tentative 30% is a share, not probability. Host dates recording; audio unchecked; organizer confirms only event window. A transcript-date November 2026 horizon stays open at cutoff. No later same-scope endorsement.

Evidence confidence: Moderate for reported-scale shortfall; lower for exact-week outcome, given company measurement and rounding limitations.

Further evidence: AISI’s newer checkpoint, received after 13 April and reported by 13 May 2026, supports continuation beyond the earlier interval gap. Qualified continuation, not universal or indefinite growth.

Direction: Technical-task capability improved after issuance.

Magnitude: Partial rapid-progress support; no exact doubling rate.

Timing: Recording date unknown; no hard deadline or timing win.

Scope: Narrow benchmarks and uncertain human-time calibration; no exact growth-rate, all-job or AGI credit. Success thresholds: AISI cyber 80%, METR software 50%; no numerical pooling. AISI’s human baselines are partly estimated; saturation and longer tails limit precision. Statistical reanalysis retains observed growth but questions equal-interval interpretation and long extrapolation; this is shared-data reanalysis, not a new capability experiment. Shared METR findings are not independent replications.

Evidence confidence: Moderate for qualified continuation; benchmark scope, human-time measurement and saturation limit transfer.

Further evidence: Human checking and exposition are explicit. The bounded existence claim is met by a verified novel mathematical contribution, rather than a contest score or rediscovery.

Direction: A substantive new research result supports the predicted capability.

Magnitude: No number of insights, discipline coverage or reliability threshold was specified.

Timing: The verified result appeared in the specified 2026 year, after the June 2025 forecast.

Shared evidence: Related to A-C026 and BG-C031. This row reuses the mathematical outcome evidence from BG-C031; it is not an independent success.

Evidence confidence: High for bounded novelty and year; not a claim about broad research reliability.

Further evidence: Qualified continuation of useful-tool assistance.

Direction: Bounded human-task gains.

Magnitude: Partial, not universal benefit.

Timing: Soft multiyear outlook; no timing credit.

Scope: Short artificial tasks, attrition, convenience recruitment and partly automated grading limit generalization; one time-saving estimate did not reach conventional 5% significance. Existing GPT-4.1 limits novelty. German apprentice corroboration compares AI plus internet against no internet, not isolated AI; exact cutoff-version chronology is uncertain. No economy-wide or broadly superhuman gains, durable learning or full multiyear persistence established. Mixed findings. The separate tenfold branch is ungraded.

Evidence confidence: Moderate for bounded experimental support; lower for generality and exact publication-version chronology of ancillary IZA corroboration.

Further evidence: Immediate unaided gains exceeded control. Pooled AI–human equivalence met authors’ margins, but Verbal separately did not; six systems passed exploratory uncorrected tests. Bounded adaptive tutoring, not full numerical fulfillment.

Direction: Adaptive teaching and learner benefit.

Magnitude: Learning-speed superiority and three-to-fourfold efficiency unresolved.

Timing: Postissue observation within the broad window; no first-arrival date or numerical-clause timing.

Scope: Adjusted equivalence survived repeat-student/leave-one-tutor checks; gain-only section tests did not. Gemma4: −1.09 percentage points versus humans (90% CI −3.93–1.75). Paid, mainly young adult completers; English GRE and immediate tests, not durable child/classroom learning. Adaptation was not causally isolated; human video versus AI chat differed. Human clustering limited precision; control watched unrelated educational videos, not GRE self-study. Platform-funded preprint, not independent replication. The 918× figure is inference cost, not learning speed or all-in cost. Stronger full-superiority/numerical conjunction remains unresolved. Adult GRE setting and immediate outcomes, not durable broad schooling. Human-comparison results are model-sensitive; learning-speed superiority remains unresolved.

Shared evidence: Tutoring predated issuance: weak-information continuation, not first invention. One 2024/2025 lineage, not separate votes.

Evidence confidence: Moderate for bounded central functionality; lower for broader tutoring interpretation.

Further evidence: Matched software-task tests improved; corrected estimates replace January values. Time horizon measures human-equivalent task difficulty, not autonomous working time. An early-2025 experienced-maintainer trial found slower work; February 2026 follow-up remains selection- and measurement-limited. Anthropic’s small junior-developer study found no significant time gain and lower immediate quiz performance; narrow new-library tasks are not all coding work. Qualified capability match, not universal usefulness.

Direction: More demanding software tasks became feasible.

Magnitude: No numerical quality or speedup threshold was promised.

Timing: 2026 was stated; no defined threshold-crossing date is established.

Scope: Software-task-heavy evidence; shared METR findings are not independent replications. No autonomous-research, OS-design or professional-replacement credit.

Evidence confidence: Moderate for qualified coding-quality trajectory; lower for the undefined quality level and practical productivity.

Further evidence: Agents handled materials, promotion, RSVPs and facilitator recruitment. Humans redirected venue search and corrected hallucinations; pizza arrived as an unrelated donation, not successful agent ordering. A partial agent-assisted demonstration is consistent with the qualified MVP forecast; stronger autonomy is unverified

Direction: Real event-organization work demonstrated.

Magnitude: Limited demonstration; intervention and failed subtasks.

Timing: June 2025 precedes December 31; no first-onset claim.

Scope: The host’s roughly thirty-minute example describes human invitations and food ordering, not a required agent runtime. Multiple agents worked for weeks with human help. Autonomous thirty-minute completion, reliable attendee selection and food fulfillment remain unestablished.

Evidence confidence: Moderate-low: dated official-host original and first-party organizer account; linked execution logs were not independently audited. Narrow demonstration, not a measured success rate.

Further evidence: BMW reported a ten-month Figure 02 pilot during 2025, with exact dates unspecified. The same sheet-metal task was already trialed in August 2024 before the forecast; the later pilot is deployment depth, not first task emergence. Actual reported manufacture and industrial work support the bounded early transition, not robots generally doing whatever humans request.

Direction: Physical-agent engineering proceeded into small production runs and narrow workplace use.

Magnitude: No count, adoption fraction or reliability target was supplied; described factory capacity is not realized output.

Timing: April 2026 production report falls within the explicit 2026 window; narrower 2025 use also precedes the five-year upper bound. Neither proves all longer-run functionality.

Scope: Narrow trials and pilot tasks, not general factory autonomy. 2025 pilot dates are unspecified; separately dated 2026 production evidence is post-forecast.

Evidence confidence: Moderate: primary manufacturer/customer reports, narrow tasks, unaudited production count.

Further evidence: An October 2025 survey of AI-using technology professionals reported 42% generated or significantly assisted code (979 responses), alongside other near-window evidence. These non-census observations strongly disfavor the broad 90% deadline interpretation; localized high shares and substantial later progress remain explicit. Moderate, qualitative and interpretation-sensitive; convergent near-window company and broad survey evidence, with later corroboration, strongly disfavors 90% as an ordinary industry-wide share. This is not proof of an exact world percentage.

Direction: right: AI-generated coding increased substantially, supported by actual-code measures not merely user counts.

Magnitude: Overestimated at the original deadline; no artificial weighted meta-estimate across different denominators.

Timing: Original deadline missed; eventual status unverified: qualified inference for original broad September 2025 target; no precise eventual crossing/lateness.

Scope: Field/world actual code writing, not expressly Anthropic-only or one workflow; precise population/volume unit unspecified. Human design, integration and security judgment remain. Existential some-workflow reading can have very high shares; original did not explicitly restrict to that. Later Anthropic broader 90% self-assessment does not rewrite March 2025 scope.

Evidence confidence: Moderate, interpretation-sensitive negative judgment for broad industry code writing; samples and company reports are not a global volume-weighted census.

Further evidence: Approximately right for the separable positive research-assistance component: experts report nontrivial AI contributions to actual novel mathematics. The future Nobel-laureate beneficiary claim is not resolved.

Direction: Observed useful human–AI scientific collaboration supports the augmentation direction, not just benchmark improvement.

Magnitude: Selected research contributions support substantive help, but not broad gains across science, Nobel-level achievement or a representative average productivity increase.

Timing: The research report identifies a January 2026 base model for the research cases, placing observed assistance inside the specified 2026–2027 generation window. This is not proof of the separate future-laureate endpoint.

Scope: Selected human-led mathematics contributions, not broad scientific productivity or future Nobel winners. Project-author evidence; rare successes, failures and modest significance retained.

Evidence confidence: Moderate for reported narrow contribution; project-author evidence, not independent publication-quality validation or population inference.

Further evidence: Unresolved: observations through 2026 cannot settle a decades-long duration or the special difficulty of entirely novel systems.

Direction: Approximately right as early corroboration only: post-forecast multi-task evidence identifies actual faulty implementations despite substantial successful coding and repair.

Magnitude: No numerical rate was forecast. Existing-repository feature work is a bounded proxy; no population error prevalence or stronger novelty penalty is established.

Timing: Pending: the stated duration extends decades beyond 2025. Neither early residual errors nor strong improvements settle that horizon.

Scope: Early direction only. Fresh preprint on bounded existing-repository tasks, not decades resolved or a population error rate. Strong coding successes retained; special novelty penalty unverified.

Evidence confidence: Moderate-low for early direction: a fresh primary preprint with explicit execution-based failure analysis and bounded task coverage.

Further evidence: ScientistTwo reports gains on 86/107 selected problems; success-only mean/median gains were 25.2%/7.7%. The earlier game-algorithm study remains counterevidence. Its later 18-game revision reveals over-specialization and improves generalization through human distillation. Approximately right for the tentative central exploration/coordination trajectory; the stronger whole day-to-day research arrangement remains unresolved.

Direction: Useful parallel/coordinated exploration in two developer-reported studies; separate teams are not independent replications.

Magnitude: Meaningful bounded gains, not a researcher-productivity multiple or paradigm-breakthrough guarantee; selection, verification and rigor limit breadth.

Timing: Three-year context remains open; publication is not an exact recording anchor or new 2028 deadline. Experiment dates are unknown.

Scope: DeepScientist: humans reproduced baselines, supervised runs and verified all results; about 5,000 ideas yielded 1,100 experiments, 21 progress findings and five selected papers. Expert reviewers flagged rigor; mean rating 5/10 is not acceptance. ICLR accepted the system report, not its generated papers. A retrospective efficiency comparison is not a controlled productivity trial. ScientistTwo’s September 2026 preprint uses reproducibility-selected tasks and success-selected outputs; 72.1% simulated acceptance is not conference acceptance. Nine humans reviewed 33 papers; developer artifact audits are not independent reruns. Another configuration retained reward hacking. Local refinement, weak breakthrough evidence and incompletely documented review uncertainty limit generalization. DeepScientist is a separate Westlake team; ScientistTwo shares Google affiliation with Noam. Neither supplies independent replication. Report/model chronology supports later contributions without exact experimental dates. Gemini 2.5 Pro searches code within human-specified baselines, evaluators and editable components. The February 18 v1 reported gains in 10/11 and 8/11 games for VAD-CFR/SHOR-PSRO respectively; larger variants partly held out. The May 7 v3’s 18-game tests qualify those version-specific results, not an independent replication. One search trajectory, closely related game tests, tabular baselines, fixed horizons and exact-best-response machinery limit transfer; no fully deep-RL or broad autonomous-research validation. Post-issue experiments are inferred from model generation, not proven by report date. Forecaster/outcome team share Google affiliation. DiscoRL’s December 2024 submission predates issuance and supplies no post-issue credit. Automated discovery already existed; tentative coordination and largest-scale uncertainty remain. Related reformulations add no independent trials; no numerical probability or current endorsement inferred.

Evidence confidence: Moderate for the bounded central parallel-exploration trajectory; limited for general research capability, productivity and long-run work organization.

Further evidence: Marcus acknowledged the deepfake-law exception on May 22, 2025. Mixed observed outcomes: business-favoring executive policy, but a signed and implemented AI-related individual-protection law contradicts the absolute negative component.

Direction: Industry-support direction was borne out; the categorical expectation of no individual protection was not. This is not a net-benefit or whole-administration moral assessment.

Magnitude: The zero-protection wording is too strong. A narrow protective law does not establish comprehensive governance, effective prevention or dominance of individual interests.

Timing: The administration is the setting, not a separate arrival deadline. The counterexample occurs within that term; the full term is not finished, and no timing score is proposed.

Evidence confidence: High for enacted-action counterexample; moderate for broad business-policy direction

Later statement: 2025-05-22 (later qualification): Acknowledges an exception for legislation addressing deepfake pornography; does not explicitly retract the February wording. No erasure of original grade or extra forecast vote.

Further evidence: U.S. aggregate continuity is positive component evidence only: all 12 category means differ by at most 6.6 minutes/day; those category means do not measure changes within activities or lived experience.

Direction: Qualified positive U.S. aggregate component only; no whole-outcome completion.

Magnitude: Whole-target magnitude remains unresolved; small mean differences do not establish equivalence.

Timing: The U.S. component fits 2025; whole-outcome timing remains unjudgeable, without a reset deadline.

Scope: BLS civilian U.S. population ages 15+, cross-sectional primary-activity diaries, not a panel: approximately 7,700 respondents in 2024 and 6,100 in 2025. Shutdown omitted September 30–November 11 diaries; later fourth-quarter observations were reweighted, effects unquantified. Annual averages include pre-issue January/early February. Work fell 3.43→3.32 hours/day; leisure rose 5.07→5.16. Coarse categories can hide changed tasks, productivity, technology use or experience; no global extrapolation or causal AI estimate. Later-updated author post; contemporaneous February 2025 reporting corroborates the relevant passage.

Shared evidence: Low-information persistence baseline; strongly dependent on Altman’s diffusion forecasts, not an independent capability forecast.

Evidence confidence: Moderate for aggregate component evidence; insufficient for the original whole-outcome verdict.

Further evidence: Right for the modest positive component that a substantially capable successor would emerge, not merely acquire a new model name. Multiple task evaluations support actual advance.

Direction: Capability advanced, while reliable general intelligence was neither promised by this component nor established.

Magnitude: No numerical gain was forecast. Impressive is qualitative; this is not a verdict on the separate claim that advances become more incremental or leads quickly vanish.

Timing: The observed successor arrived within the surrounding three-year outlook. This is contextual-window credit, not a forecast of its exact release date.

Scope: Low-specificity positive capability forecast. Separate incremental-rate, supplier-lead and AGI predictions remain ungraded here; first-party benchmarks do not establish all-task superiority.

Evidence confidence: Moderate: first-party multidomain evaluations and existing limitations, no independent all-task census

Further evidence: Human selection and quality gaps remain; benchmark dominance is unverified, dependable research replacement is not established, and related 2024 workflows limit novelty. A real manuscript workflow supports the central component, not benchmark dominance or dependable research replacement.

Direction: Not separately graded in the historical record.

Magnitude: No separately documented rationale is available.

Timing: The April 10, 2025 report falls within the original window for the supported workflow component.

Scope: The illustrative RE-Bench superiority claim remains unresolved, not an observed miss. Treating it as essential leaves the full conjunction unjudgeable. The result is a selected-run report, not a general autonomous-research success rate.

Evidence confidence: Not separately rated.

Further evidence: The amount of industry-wide change and whether these were formal revisions remain unestablished. Actual semantic variation supports a bounded component, but the substantial industry-wide claim and stronger goalpost-shifting interpretation are not fully resolved.

Direction: Primary dated 2025 sources contain substantively different formulations, rather than merely repeating the old economic definition.

Magnitude: No quantitative threshold for a lot more; evidence is concentrated on one prominent speaker.

Timing: The observed instances occurred within 2025, but there is no full timing credit for the unmeasured breadth claim.

Scope: Related June 15, 2024 wording was edited June 22 to withdraw the Kurzweil downgrade example; first-publication wording is not independently archived. Its broader major-player expectations component remains unresolved. Commentator updates do not satisfy that actor scope and earn no extra credit, motive inference or reset of the 2025 definition-change target.

Evidence confidence: Moderate for bounded semantic movement; low for broad prevalence or motive.

Further evidence: The result is bounded and partly continuation of existing use, not workforce replacement or an established industry take-off. Approximately right for the frozen three-role deployment conjunction: all roles now have actual tool-using-work evidence, including the previously missing campaign-management predicate. First-party accounts and human oversight limit strength; broader take-off interpretation remains unmeasured.

Direction: Operational examples support enterprise agent use across the named roles.

Magnitude: No prevalence threshold or labor-cost share was fixed; selected cases do not establish broad adoption or replacement.

Timing: Reports establish use during the stated year, not the date of first-ever deployment. Advolve may have started preissue, reducing novelty rather than supplying a new invention date.

Scope: Three-role deployment conjunction; tool-using enterprise work with humans allowed. Broad take-off language has no specified scale. NVIDIA internal plans and all-engineer policy are separate, not scored.

Evidence confidence: Moderate for bounded three-role deployment; first-party reports and prior-use baseline limit inference. Broader adoption magnitude remains unresolved.

Further evidence: Qualified early-wave reading; an inseparable literal-first-2025 target remains unresolved.

Direction: Action-taking work, not just launches; pre-2025 assisted coding/support limits novelty.

Magnitude: Departmental effects, not universal/economy-wide gains or uniform correctness.

Timing: Dated 2025 outputs, but first attainment unresolved; no full-target timing hit or 2026 backdating.

Scope: Original page later updated; contemporary wording corroborated. Commercial self-reports, not audited company-wide net effects; task mix, baselines, review/error costs and durability unknown. Engine acknowledged unintended cancellations and a confirmation safeguard. March efficiency/annual-savings quantities were aims; annual savings remain uncredited. Public clip transcript, not verified audio; exact post day unknown.

Shared evidence: Related deployment forecasts share the trajectory, not independent skill confirmation.

Evidence confidence: Moderate-to-low for the qualified reading: original wording has contemporary corroboration, while outcomes are commercially interested self-reports. Literal chronology and representativeness remain limitations.

Further evidence: Not a pure AI-hardware margin or exact calendar-year total. Documented profit supports the scoped hardware component.

Direction: Hardware-associated supplier profit was substantial.

Magnitude: Large profit direction is directly supported.

Timing: Fiscal period overlaps most of 2025; exact calendar-year offset retained.

Scope: MiniMax’s full-calendar 2025 disclosure reports positive gross profit but statutory and adjusted company losses. Most statutory loss is preferred-share remeasurement, not operating cash burn. Applications and platform/enterprise revenue are mixed; model-only profit is not isolated. Adjusted figures do not replace IFRS. EY agreed announcement amounts to audited statements but did not assure the announcement. This closes the company’s partial-period gap, not the industry-wide model-profit conjunction; the hardware judgment is unchanged.

Evidence confidence: high for documented company result; moderate for wider direction

Further evidence: This supports a scoped institutional-continuity prediction, not a conclusion about every safety institute or the strength of all AI regulation. Examined US and UK roles support the limited continuity prediction, with relatively low novelty.

Direction: Not separately graded in the historical record.

Magnitude: No separately documented rationale is available.

Timing: The assessment concerns the 2025 window; later UK retrospective documentation is not backdated publication.

Scope: This is not an exhaustive global legal census. Evidence includes later retrospective reporting as well as contemporary institutional documents.

Evidence confidence: Not separately rated.

Further evidence: That supports the prediction of persistence, not technological stagnation, a specified everyday-user error rate or any measured error rate at the October 2026 review cutoff. Observed improvement coexisted with factual errors, matching persistence rather than stagnation.

Direction: Not separately graded in the historical record.

Magnitude: No error-rate threshold.

Timing: The dated late-2025 evidence falls within the forecast’s original calendar window.

Scope: The historical 2025 result does not quantify later error rates or justify a claim of no progress.

Evidence confidence: Not separately rated.

Further evidence: That supports persistence of errors, not general inability to reason, no progress, or a measured everyday-user error rate. Tested failures persisted alongside large gains, matching the modest continuity target rather than general inability.

Direction: Not separately graded in the historical record.

Magnitude: No separately documented rationale is available.

Timing: The controlled evaluations document residual failures during the original 2025 window.

Scope: Severity unquantified. This is a low-novelty persistence call sharing a reliability trajectory with related annual forecasts. Related 2022 wording for 2023 is retained as ungraded history, with its architecture conditions; it does not replace this 2025 deadline.

Evidence confidence: Not separately rated.

Earlier statement: 2022-12-25 (Earlier source pointer): Whole numbered slot 2, 2023; rumor-based unchanged-architecture/same-playbook conditions. Undisclosed internals and the causal explanation are unestablished. No 2023 adjudication, no replacement of the 2025 deadline, and no additional forecast.

Further evidence: The technical component is supported; marketing-rhetoric intensity was not independently measured. Broad reliability limits coexist with useful bounded deployments.

Direction: Not separately graded in the historical record.

Magnitude: No separately documented rationale is available.

Timing: The supported technical-reliability evidence falls within the original 2025 forecast window.

Scope: Marketing/reliability conjunction unresolved. The finding does not imply agents were useless, undeployed or confined to one industry; related reliability forecasts are dependent.

Evidence confidence: Not separately rated.

Further evidence: Staged o3 and Operator-derived releases support continuation. Unavailability bounds the joint endpoint, not measured quality/price attainment. Immediate GPT-4.1 is counterevidence to uniformity; Operator continuity/configuration remains uncertain.

Direction: Continuation observed, with shorter or immediate launches also present.

Magnitude: Months-long examples exist; no representative lag distribution.

Timing: Continuation observed in 2025; exact joint-endpoint attainment remains unknown.

Scope: No all-products rule or fixed quality/price thresholds; years-long delays optional. Evolving product lineages and uncertain Operator access continuity limit exact lag estimates.

Note: Shared evidence

Evidence confidence: Moderate for qualified continuation and documented access dates; lower for cohort breadth, continuous Operator restriction, reliability and affordability endpoints.

Further evidence: Observed support and contrary geographic evidence coexist; missing weather evidence does not cause the mixed grade. Alternative interpretations remain open.

Direction: Expansion and substantial human employment coexist with the contrary city-distribution result.

Magnitude: No numerical cutoff; city counts do not measure trip, fleet or usage shares.

Timing: Original-year evidence; exact endpoint and whole-bundle timing remain unresolved, not a delayed-adoption miss.

Scope: Weather, trip-weighted shares and US-only interpretation remain unresolved. Paid public service excludes free/test use; one intact forecast, not separate component votes.

Evidence confidence: Moderate for observed mixed components under unrestricted deployment geography; conditional on interpretation, with US-only and usage-weighted readings unresolved.

Further evidence: This is a comparatively low-novelty continuity call about ongoing proceedings. Continued court activity supports the forecast, though it was a low-novelty continuity expectation.

Direction: Not separately graded in the historical record.

Magnitude: No volume, merits, liability or damages target.

Timing: May and December 2025 court records fall within the original calendar window.

Scope: The result concerns historical continuation, not current legal advice or an exhaustive litigation census.

Evidence confidence: Not separately rated.

Further evidence: Efficiency gains and transparent developers remain important counterexamples; the evidence is not exact global metering or proof of universally absent disclosure. Rising electricity pressure and sparse model-level disclosure support both central components, with documented exceptions.

Direction: Not separately graded in the historical record.

Magnitude: No separately documented rationale is available.

Timing: The assessment concerns 2025 conditions; the later infrastructure report provides retrospective estimates, not contemporaneous measurements.

Scope: The evidence combines infrastructure estimates and disclosure review, not a global generative-AI electricity total. Physical bottlenecks support concern but do not quantify a welfare verdict.

Evidence confidence: Not separately rated.

Further evidence: European catch-up and absent durable technical/business advantages remain unverified. Competition and partial convergence are supported, but meaningful regional and task gaps qualify the result.

Direction: Not separately graded in the historical record.

Magnitude: Domain gaps persist; benchmarks cannot establish business advantages.

Timing: Competition evidence supports the 2025 component; later comparisons are not silently backdated to that window.

Scope: No uniform parity.

Evidence confidence: Not separately rated.

Further evidence: Real deployments and growth are substantial counterevidence to universal nonadoption; selected large-firm results do not establish prevalence across all businesses. The experimentation-to-scale gap is supported, while real successful deployments rule out a no-adoption interpretation.

Direction: Not separately graded in the historical record.

Magnitude: No separately documented rationale is available.

Timing: The assessment concerns 2025 adoption; later reports and retrospectives are kept distinct from their fieldwork dates.

Scope: Limited deployments are not equivalent to scaling across most functions. Separate 2026 evidence and selected success stories do not supply population prevalence or a hard reliability ceiling.

Evidence confidence: Not separately rated.

Further evidence: Those examples support the persistence target, but establish neither a universal failure rate nor a conclusion about the limits of the architecture. Progress and residual physical failures coexist, matching the actual persistence target.

Direction: Not separately graded in the historical record.

Magnitude: No prevalence target or stagnation.

Timing: The relevant release documentation and evaluation both fall within the 2025 window.

Scope: Specific remaining errors, not universal failure or absence of improvement. Automated grading, incomplete version identification and sensitivity of a separate physics benchmark limit stronger inferences.

Evidence confidence: Not separately rated.

Further evidence: By July 17, 2026, the evaluator reported Inkling at 79.5% and $0.30 per task on the same semi-private benchmark category. This is cheaper at higher aggregate performance, not merely a cheaper weaker configuration. Exact tasks and errors need not match. A separate July 2026 study reports public ARC-1 two-guess performance of 67.25% at $0.62/task, versus H-ARC’s 60.2% human two-attempt mean. Human-session pay allocated roughly $2–2.20 per assigned task: $10/five plus a possible $1 bonus, including study overhead, not a measured pure labor price. Independent human-task evidence supports qualified ARC-economic competitiveness; the broader human-work alternative remains unresolved.

Direction: A dated comparison supports real local improvement.

Magnitude: Large benchmark inference-price improvement; human comparisons support competitiveness, not exact universal savings or fully loaded parity.

Timing: Roughly 19 months fits the qualitative horizon; no exact deadline or independent timing success.

Scope: ARC-AGI-1; original pricing estimates, not later edits. Qualified marginal task economics, not general-work success or replacement-cost parity. Public-set tuning and possible pretraining exposure limit generalization. Human tests used first outputs and feedback; exact protocol equivalence is unverified. Human compensation includes instructions/descriptions; model token prices omit training, engineering and execution overhead. No statistical superiority is established.

Note: Shared evidence

Evidence confidence: Moderate for the benchmark-economic interpretation; lower for exact human/model protocol equivalence and general productive-work transfer.

Further evidence: The first two directly addressed o1-preview-like capability, while the latter two expanded open reasoning data and reconstruction of the R1 pipeline. The s1 paper also demonstrated controlled extra inference computation improving a fixed model on its reported math evaluation. Preexisting OpenR/OpenO1 make this a low-to-moderate novelty continuation forecast. Several distinct projects launched or publicly released in January 2025 explicitly pursued open reasoning or o1-like test-time capabilities, with code/data/weights rather than merely commentary. This supports the research-effort forecast under ordinary o1-like replication language. The broad year-of-search slogan is less determinate than the concrete efforts.

Direction: The new 2025 projects materially extended an already existing 2024 open-reasoning research line. This is continuation, not a prediction of the first such project.

Magnitude: Sky-T1, s1, OpenThoughts, and Open-R1 supply four named 2025 efforts. Two directly target o1-preview-like capabilities; two target the wider open reasoning/R1 reproduction ecosystem. This is a qualitative plurality, not an invented numerical definition of many or a field census.

Timing: The named efforts appeared within the explicit calendar-2025 window. Their January appearance is an early realization within a one-year original horizon, not evidence that the original horizon was only one month.

Scope: One research-wave target from the primary interviewer’s exclusive response. The adjacent repeated ARC-AGI solution clause is retained as a dependent annex, not another vote or a reset of FC-2024-ARC-HN. Exact private o1 algorithm reproduction, universal benchmark equivalence, and field-wide compute share are not implied by evidence of replication efforts. If the entire response including full ARC solution is treated as one compound, content and timing remain unresolved; this card assesses the separately retained research-wave target. The preceding tweet’s economics are not adjudicated here.

Evidence confidence: Moderate for the approximately-right research-wave verdict; high for existence and within-window public release of the named efforts. Not a proof of exact replication or inference-time search dominance.

Further evidence: Qualified miss; ungraded interval.

Direction: Growth supported.

Magnitude: Amount underestimated.

Timing: Endpoint—not arrival.

Scope: Indexed only; full/immutable submission uninspected. Unknown statistic/distribution: no mean/median/calibration/ranking verdict. Stale author/organizer: $4.5B/$4.7B. Updated $6.4B mitigates growth criticism, not amount error. Author outcome: ≈$30.5B. December 13 attestation and December 14 public survey field are not independently reconciled; midnight/timezone semantics are unknown. Company disclosures corroborate scale, not an audited December-revenue-times-12 sum. January 2026 CFO report: OpenAI over $20B 2025 ARR; April 2026 company report: Anthropic about $9B at end-2025 run-rate. Its filing uses trailing 28-day consumption fees ×13 plus monthly subscriptions ×12, not GAAP or independently verified. Current company pages may carry later edits. Not full-year revenue, profit, funding, valuation, GDP or a sector census. No probability score.

Evidence confidence: Moderate for discounted author attribution and submission period; moderate for organizer-resolved point error; low for unrecovered individual distribution and exact audited monthly-total convention.

Further evidence: Bounded quality improvement; narrower independent realism reading remains partial-only.

Direction: Continuation of an already-strong baseline, not first invention.

Magnitude: Subjective intensity; no numerical target.

Timing: Few years is vague; compatible observations, not deadline/first-attainment credit.

Scope: December 2024 is the edited interview’s publication clock; September-event utterance unverified. The card first appeared May 2025, but evaluation days/revisions are unspecified: no backdating. Video-only and audio-video tests differ; no Sora-specific muted-realism rate or numeric margin is given. Provider-run evaluation omits rater counts/full matching details; complex motion, scene consistency and artifacts remain problematic. Neither independent visual-realism superiority, universal indistinguishability nor film-length coherence is established.

Shared evidence: Related forecasts and same-provider reports are dependent, not replications.

Evidence confidence: Moderate for bounded qualitative improvement; lower for independent realism-specific magnitude and exact timing.

Further evidence: Repeated criminal use supports occurrence; full scale unmeasured.

Direction: Material operational use, not simulation.

Magnitude: Selected incidents do not establish global frequency, growth or losses.

Timing: Coming years is an explicit long-run horizon without a separate fixed deadline.

Evidence confidence: Moderate for reported occurrence and role; limited for independent corroboration and global magnitude

Further evidence: The original October 2025 RLI evaluation used earlier models and constrained setups. A dated July 2026 update also reports GPT-5.2. METR’s software-heavy estimates were later corrected. No matched endpoint duration-by-domain evaluation resolves the broad negative. Whole negative endpoint unresolved.

Direction: Local heterogeneity only, not absence of useful agents.

Magnitude: Measured variation is partial support; broad expert thresholds remain unresolved.

Timing: No full timing verdict or replacement deadline.

Scope: Publisher transcript: recorded 10 December 2024, published 16 January 2025; audio not independently checked. GPT-5.2 Thinking’s 70.9% GDPval wins/ties across 44 occupations are provider-reported aggregate counterevidence, not a week-scale subgroup result. GDPval includes multiweek tasks, but its earlier duration breakdown cannot be assigned to GPT-5.2. Human task duration is not model runtime. No invented reliability threshold, universal inability, universal perfection or 2026 employment inference. Earlier limited results cannot establish later unrestricted capability. The original October RLI computer-use setup had a one-hour timeout. Runs ended below $30; budget exhaustion was not demonstrated. CAIS’s July 1, 2026 post reports GPT-5.2 at 2.50% in its historical comparison; this dates publication, not the run. It corrects a general high-effort description to model-default settings; the later edit date is unknown and an immutable pre-cutoff version was not independently archived. The later board separately labels the corroborated 2.50% row medium-effort, not a matched comparison with GDPval’s heavy configuration. Newer-model scaffolds allow up to 24 hours and worker–critic revision; these conditions are not securely linked to the older GPT-5.2 run. The post’s text describes flat or slightly rising success against human duration, not monotonic decline; individual graph curves were not visually verified. Later-model advances are not backdated endpoint outcomes; current-board-only values are excluded.

Evidence confidence: Moderate for local, measured heterogeneity; insufficient for the full negative endpoint or unrestricted latent-capability claim.

Direction: Benchmark progress.

Magnitude: 87.5% exceeds implied 85%, not 100%; public-budget 75.7%.

Timing: Approximate window met, not month accuracy.

Scope: Any architecture; no AGI/risk verdict. Released o3 differs and scored lower; public training exposure disclosed. Pricing is not backdated.

Note: Shared evidence

Evidence confidence: Moderate: strong benchmark result, but solved and cost/rules are scope-sensitive.

Further evidence: This exceeds the target roughly five to seven months late; contamination concerns and absent independent replication limit broader coding-ability conclusions. The reported score crosses the original benchmark threshold, not an independently certified general software-engineering threshold.

Direction: Not separately graded in the historical record.

Magnitude: The later reported 93.9% exceeds approximately 90%; original-window results of 77.2% and 82% remain below target.

Timing: The April 7–8, 2026 report window is about five to seven months late; first internal attainment remains unknown.

Scope: The headline includes all 500 tasks and is a five-trial mean, not best-of-five selection or an easier filtered subset. Maximum-effort settings were used; the complete Verified-specific harness and token budget were not established. Developer-run contamination checks do not prove clean measurement. The current PDF is not an archived first-day snapshot; no global first is claimed. Both approximate interview horizons remain one forecast, and the original-window shortfall is retained.

Evidence confidence: Not separately rated.

Further evidence: Bounded narrative shorts supported; an inseparable 15-minute reading remains unresolved.

Direction: Real system capability; preforecast demonstrations limit distinctive foresight.

Magnitude: Probable 15-minute duration and general reliability unverified.

Timing: Bounded shorts documented within the window; no full-duration hit or known first attainment.

Scope: Developer evaluation, not independent replication. Full sample video not personally viewed. Human-built assets alone do not disqualify capability. Earlier Showrunner demonstrations and later scene-oriented limits counter promotional runtime claims. No professional-quality credit; 2026 evidence not backdated.

Shared evidence: Related 2022 undated formulation remains excluded; no retroactive horizon repair.

Evidence confidence: Moderate for qualified short-form judgment; original is coherently contextualized third-party transcription with corroborated public event/date, not independently authenticated speech. Full-duration result unresolved.

Further evidence: Enabled Microsoft 365/Copilot configurations nevertheless process content and telemetry; no comprehensive absence-of-logging claim is made. Approximate miss only under literal universal reading; documented permanently offline Word architecture is a concrete counterexample. Rhetorical/cloud-only readings remain unresolved.

Direction: Actual narrower content processing and telemetry are real but were already present before the forecast; this evidence alone cannot demonstrate predicted new escalation.

Magnitude: Every keystroke across unqualified Microsoft Word use exceeds documented scope, including maintained disconnected LTSC 2024.

Timing: No onset of universal collection occurred under the stated literal reading; no separate deadline timing inference for a narrower replacement claim.

Scope: Literal universal reading only. Rhetorical or enabled-cloud interpretations remain unresolved. Offline-product evidence is not an audit of all Word; real cloud content processing and telemetry remain concerns.

Evidence confidence: moderate; strong product-scope evidence, interpretation-sensitive and no direct independent LTSC endpoint audit

Further evidence: FINRA’s January 2025 report describes growing adversarial GenAI use, including stock-price manipulation. Qualified qualitative continuation supported; specifically post-issuance growth remains unresolved.

Direction: Matched FCA characterization, with broader FINRA corroboration.

Magnitude: No numerical rate, scale or loss threshold was promised; no measured magnitude accuracy.

Timing: Contextual coming-years horizon, no fixed endpoint. Later publication does not date observations or onset.

Scope: Coming-years context precedes the specific clause; attachment is interpretive, not a repeated deadline. FCA guidance supplies no observation dates, counts or methods-level trend estimate. FINRA’s umbrella increase is not a measured increase for every subtype; its investment-group complaints are not established AI-generated. Preexisting misuse limits novelty. No first-emergence, precise post-forecast rate or distinctive-foresight claim.

Shared evidence: Distinct market-manipulation target; strongly related to broader synthetic misuse, not extra votes for the societal umbrella or each source.

Evidence confidence: Moderate-to-low for qualitative continuation; primary regulatory provenance but unknown observation windows and no methods-level trend measurement.

Further evidence: Dated proxy context reasonably links deployed Cybercab to Tesla’s camera-only strategy; this is corporate-system inference, not independently certified per-vehicle hardware. Qualified local service convergence, not full Waymo parity.

Direction: Practical unattended service supports weak directional expectation.

Magnitude: Then-Waymo citywide 24/7 access was broader; operational maturity, reliability and safety parity unresolved.

Timing: Roughly two-year observations fit the outlook; vague endpoint and unresolved full equivalency preclude precise arrival credit.

Scope: Third-party, error-prone transcript; attribution inferred, audio unauthenticated. Hope may express aspiration rather than forecast; former Tesla affiliation. Mutable FAQ mentions cameras plus unspecified sensors, not a historical hardware audit. Regulator deployment corroboration is not safety approval. Broader equivalency and aspiration-only readings remain unresolved/unscored alternatives. No invented fleet, revenue, universal geography or zero-assistance requirement; prior autonomy limits novelty.

Shared evidence: Related vision/autonomy claims and shared Tesla outcomes are dependent, not extra votes.

Evidence confidence: Moderate for the bounded practical service comparison and camera-system inference; lower for original transcript exactness, aspiration-versus-forecast modality, broader equivalency and independent predictive information.

Further evidence: Financing and infrastructure growth ran against a broad downturn interpretation but do not settle onset. Financial expansion can coexist with early, uneven or temporary bursting; it establishes neither profit nor fair valuation.

Direction: Checked financing and infrastructure-demand indicators oppose a broad downturn reading; this retained descriptive appraisal does not resolve original bubble onset.

Magnitude: Bubble-onset threshold and collapse magnitude unresolved; no stronger collapse target substituted.

Timing: Onset unjudgeable; no definitive missed date or reset of the original 2025-07-28 deadline.

Evidence confidence: Moderate for the scoped adverse commercial-direction evidence; insufficient for a full bubble-onset verdict.

Further evidence: The named model arrived after the excluded 2024 release window.

Direction: No independent directional capability claim in this record.

Magnitude: No quantitative capability magnitude forecast.

Timing: Correct not-before boundary; do not credit an exact August 2025 forecast.

Evidence confidence: high for named public-release result

Further evidence: The released model retained factual hallucinations and conspicuously unsupported answers. Developer evaluations establish this without relying on viral examples.

Direction: Residual unreliability persisted, while large measured improvements also occurred. Persistence is not a prediction that error rates would remain unchanged.

Magnitude: The original gives no operational quantity or severity threshold. Error rates vary greatly by task, browsing and model variant; no global frequency is inferred.

Timing: Capability was assessed at the specified release event; there is no independent calendar timing forecast here. The separate no-2024 forecast is D827.

Evidence confidence: Moderate-high for persistence; first-party evaluation and model-grader limits prevent population-wide frequency inference.

Further evidence: Literal universal target contradicted by Vivaldi.

Direction: WhatsApp illustrates real diffusion.

Magnitude: Universal coverage contradicted; adoption share unmeasured.

Timing: Counterexample documented after the latest deadline.

Scope: This counterexample does not measure overall adoption share. A rhetorical interpretation would remain unresolved.

Evidence confidence: High for literal counterexample; moderate overall because every may be rhetorical.

Further evidence: These are selected fixed-capability frontiers, not broader market rates. OECD’s broader quality-adjusted token-price index fell approximately 80% from January 2024 to April 2026. Zhu’s 2026 preprint estimates 0.728 log points of annual decline in a broader unweighted quality-adjusted posted-price index, January 2025–August 2026. Price-performance gains within measured scopes do not establish full numeric fulfillment; the original measurement unit is unspecified.

Direction: The contemporary observation window overlaps the forecast period and documents major fixed-performance price reductions across three benchmark domains.

Magnitude: Partial magnitude support applies to selected frontiers; broader indices are slower. Do not square pooled annual rates or equate different estimands.

Timing: Different observation windows; no matched endpoint ratios or numeric timing success/failure.

Scope: Frontier minima and broader market indices measure different objects and share Artificial Analysis inputs. Published token/list prices are not internal serving costs or observed transaction prices; task consumption and discounts matter. Zhu is a preprint with benchmark and window sensitivity, and its task-price sign is uncertain. Matched endpoints remain unverified; no numeric deadline verdict. Task and capability-age variation, benchmark limits, model switching and mixed API/rented-hardware costs remain. Zhu’s task-price interval includes declines; an alternative window reverses the point-estimate sign.

Evidence confidence: High for material price-performance improvement within measured scopes; moderate and method-sensitive for approximate magnitude; full numeric and deadline adjudication unresolved.

Further evidence: Qualified skepticism; ultimate capability unadjudicated.

Direction: Not separately graded in the historical record.

Magnitude: No separately documented rationale is available.

Timing: Predeadline report; first attainment unknown.

Scope: Original ARC; no added cost/release condition. Unchanged-pretraining claims unadjudicated.

Shared evidence: FC-C012, FC-2024-ARC-HN, FC-2024-ARC-COST, GM-C028: shared forecaster/evaluator evidence, not independent replication; no deadline reset.

Evidence confidence: Not separately rated.

Further evidence: The modest competitive-quality target is supported by multiple heterogeneous evaluations.

Direction: Relative capability improved into a credible competitive range.

Magnitude: No numeric margin or universal task threshold was specified.

Timing: Documented before year-end; no exact-date skill or internal first-attainment claim.

Evidence confidence: Moderate; original event reporting and developer evaluations, not independent reproduction.

Further evidence: This establishes formal campaign issue presence, not voter salience or influence. Official party platforms establish the bounded issue-presence expectation; the advice that everyone should prioritize it is not graded.

Direction: AI economic and innovation policy appeared in campaign agendas.

Magnitude: No quantified issue prominence or voter share was forecast.

Timing: Dated July/August 2024 platforms appeared after the May forecast and before the November election.

Scope: AI-adjacent; low-information continuation of existing policy attention. No major-issue, voter-priority or decisive electoral-impact credit.

Evidence confidence: High for bounded official issue presence; no salience judgment.

Further evidence: Reading-comprehension and factuality counterweights improved less on different metrics. The broad continuing rate and real-world reliability endpoint remain unresolved. No named benchmark or full NLU denominator; selected tests cannot settle the continuing broad claim.

Direction: Same-suite comparison supports material capability gains.

Magnitude: One central language-understanding metric approximately matches contextual rate; reading-comprehension and factuality counterweights are weaker and differently measured.

Timing: Evidence predates first anniversary, but cannot establish ongoing annual rate or full reliability.

Scope: Accuracy/error ambiguity remains: contextual error-halving, not literal doubling of a percentage. No original benchmark, prompt, lineage, denominator or precise stopping threshold was specified. The same-event historical six-month assertion does not replace this annual forward claim. The April 21, 2025 repository commit pins disclosure, not an original April-2024 evaluation date. Developer-suite prompts and reasoning settings differ; no constant-compute or efficiency result follows. DROP is F1; SimpleQA correct fraction includes abstentions, so their complements are not hallucination probabilities. Exact published run counts, dataset hashes and run logs remain unverified; no sample size or uncertainty interval is invented. Related H-C023 shares a progress mechanism, not an independent validation.

Evidence confidence: High for reported table and chronology; limited for broad annual-rate inference.

Further evidence: Neither directly tests the full original conjunction. Approximately right for continuing gaps in ordinary film narrative comprehension and factual consistency, not for a categorical inability to understand movies. The full humor and spoiler-free synopsis conjunction is not adjudicated.

Direction: Two relevant primary evaluations find meaningful comprehension but persistent narrative errors and human-performance gaps. Title-only controls show why memorized plot knowledge cannot settle genuine viewing comprehension.

Magnitude: The weakness is substantive, not zero ability. Tested systems answer many narrative questions successfully; no benchmark rate is converted into the fraction of whole films understood.

Timing: Relevant limitations were observed in 2025, but the full conjunction and all late-2025 systems were not tested. No categorical deadline or perpetual-ceiling credit.

Evidence confidence: moderate-low for a partial historical narrative-reliability judgment; full target and later frontier coverage unresolved

Further evidence: Substantial real off-road autonomy, including previously unseen terrain and creek-crossing courses, contradicts the core capability doubt. Exact joint performance on every listed obstacle without prior maps is not independently quantified.

Direction: The forecast was too pessimistic about practical off-road navigation; this is not merely a product announcement or road-driving proxy.

Magnitude: Program reports demonstrate a meaningful range of obstacles and terrain, but not universal field reliability. Remaining interventions and environmental limits prevent a categorical full-task refutation.

Timing: Core off-road performance was documented before end-2025, including trials predating the forecast. A January 2026 retrospective explicitly describes no-premapped-route capability and fall-2025 tests, but the full map-free/obstacle conjunction lacks an exact common dated protocol; no binary timing score proposed.

Evidence confidence: Moderate for substantial contrary capability evidence; sponsor/participant reports, no independent all-condition validation.

Further evidence: Real unfamiliar-home generalization challenges the pessimistic component; it is not proof of complete-house capability.

Direction: More task generalization arrived than this skeptical direction suggested.

Magnitude: Underestimated the attainable component capability; measured progress is not whole-task success.

Timing: Component counterevidence arrived before 2025 year-end; full-house timing remains unresolved, not a categorical deadline loss.

Evidence confidence: moderate; author-run experiments, cancelled trials excluded, limited homes and chores

Further evidence: Author-run classification and language-model experiments offer bounded support, but results are mixed. Memory and compute differences limit comparisons; some are omitted. Ablations report benefits without uncertainty intervals. Optimizer overhead remains; catastrophic forgetting is not generally solved. Full thesis unresolved.

Direction: Bounded mechanism corroboration.

Magnitude: Necessity and breadth unresolved.

Timing: No timing success claimed.

Scope: Necessity thesis, not merely one useful architecture; hardware conditional is dependent. Long run; no fixed deadline. December 2025 inspected revision; not earliest attainment. Implementation does not establish necessity, eventual dominance or universal deployment.

Shared evidence: One target; shared research lineage, not independent replications. Archive is not renewed endorsement.

Evidence confidence: Moderate for coherent attributed host paraphrase and reported research mechanism; lower for exact wording and general necessity. Automated transcript explicitly warns errors; outcome experiments are author-run and not independent replications.

Further evidence: Reflection sometimes hurt; practical reliability and field-wide contribution were not established. The previous evidence covered only software engineering. Two independent 2024 primary studies now add material gains in mathematical reasoning/question answering and multi-step service planning, while holding the execution model fixed. These satisfy a modest central reading of significant workflow-enabled progress across AI tasks. Quality remains limited and self-reflection sometimes degrades results; the verdict is approximate, not an unqualified hit.

Direction: Observed positive gains across coding, mathematics, QA and simulated service workflows, not only paper count or feature availability.

Magnitude: Significant has no numerical original threshold. Effects are material in their respective metrics, but no defensible all-field attribution percentage or contribution ranking is estimated.

Timing: All three primary outcome versions are from May–October 2024, after March 27 and before year-end. Later 2025 paper revisions are not used.

Scope: One combined workflow forecast. Added studies include fixed execution models but extra inference and optimizer/domain knowledge; not equal-compute, uniformly successful reflection, or autonomous production reliability.

Evidence confidence: Moderate for bounded multi-domain workflow capability gains; field-wide ranking, deployment reliability and equal-compute efficiency remain unestablished.

Further evidence: Rapid growth occurred, but the projected hardware-growth magnitude was substantially too high under this reading.

Direction: Dedicated accelerator capacity grew strongly.

Magnitude: Matched sales-derived hardware proxy was far below 100-fold annual growth; precise installed-compute error ratio is not established.

Timing: Not separately scored: the annual pace error is already in magnitude.

Evidence confidence: moderate; original transcription and metric ambiguity plus estimated sales rather than measured deployed capacity

Further evidence: A major AI infrastructure operator confirmed the predicted material constraint.

Direction: Power availability remained a limiting factor as demand grew.

Magnitude: No numerical share, shortfall or universal exclusivity stated.

Timing: Constraint documented during 2025; this does not establish onset in 2025.

Evidence confidence: moderate; third-party original transcript with corroborated event identity, official outcome transcript

Further evidence: It learned from existing gameplay, not a new game created freely from a prompt. The full gaming/non-gaming diffusion target remains unresolved. One research demonstration does not establish broad gaming/non-gaming diffusion.

Direction: Real-time neural generation of an interactive Doom environment is meaningful movement toward the proposed paradigm.

Magnitude: Prototype capability is supported; industry prevalence and cross-experience scope are unmeasured.

Timing: The multi-year transition window is open. Early component realization is not early fulfillment of the entire target.

Scope: Rough five-to-ten-year S-curve transition, not a first-ever existence claim or an exact 2030 deadline. Jarred Walton’s firsthand press transcription was edited for clarity; audio was not independently verified. Short memory, state errors, incomplete exploration and new-game creation limits remain. No conventional-game-quality parity, all-game-code generation or industry transition is established. No consumer-PC compute requirement is imported from reporter commentary. Related game-production-cost work remains held; this is one dependent paradigm target, not per-game or per-modality successes. Initial paper August 27, 2024; inspected version April 24, 2025.

Evidence confidence: High for existence and method limits; low for full diffusion outcome.

Further evidence: Initial scale, language and visual milestones are supported; later world-model/planning stage remains unresolved.

Direction: The model lineage expanded scale, improved comparable language benchmarks and acquired visual input capability.

Magnitude: Several qualitative capability components materialized; no complete planning/world-model scope or numerical target established.

Timing: Original response declines a product-line research timetable. No fixed deadline, early/late label, or deadline success is assigned.

Scope: Only the initial capability stage is supported. Later planning remains unresolved; company benchmarks are not independent replication. This is a modest continuation forecast.

Evidence confidence: Moderate-high for the modest milestone; benchmarks are developer-reported and not universal capability tests.

Further evidence: Image exclusions and untested exams remain. The universal examination scope remains unresolved. Neither two favorable domains nor a selected question subset establishes every human test.

Direction: New post-issue systems achieved demanding competition and medical knowledge assessment results, supporting continued exam-performance progress.

Magnitude: All-test coverage and a common definition of passing remain unestablished. The medical study excludes image questions, while the IMO result concerns a special research model.

Timing: Five years after the March 2024 event is approximately March 2029, still future at cutoff. No premature full success or deadline miss.

Scope: Direction only. Host-report timing is not a newly verified verbatim hard deadline; all-test coverage and clinical competence remain unestablished. The inspected October 2026 HLE revision reports late-2024/early-2025 models, not a current-frontier retest. Adversarial selection, automated grading and expert disagreement limit inference; there is no standardized human pass mark. Specialized medical exams are included; existing bar-exam success was baseline evidence, not new credit. Human-mind equivalence remains a separately undefined target. The medical text subset is not a full licensing exam; no clinical-practice competence is inferred. Polish specialty results used a pre-issue model, not a current-frontier retest or post-issue capability gain. The later April 2024 benchmark-list formulation is narrower and ungraded; it does not reset the March wording or imply a 2026 endorsement.

Evidence confidence: High for repaired host-report attribution and scoped outcomes; moderate for directional inference only.

Later statement: 2024-04-24 (Later related benchmark-defined formulation; not a new independent target): Host transcript gives a five-year benchmark-list expectation distinct from ordinary AGI. This is narrower than every conceivable human test. Deadline: Within five years in this April version only; original March timing unchanged. Not counted separately. No adjudication of this version.

Further evidence: This matches the specific interaction target within the forecast window, not universal understanding or autonomous execution. Live camera-context interaction realizes the specific forecast without implying broader intelligence or autonomy.

Direction: Not separately graded in the historical record.

Magnitude: No separately documented rationale is available.

Timing: The described March–April 2025 rollout falls within the original approximate window.

Scope: Demonstration, preview and rollout form one progression, not three independent prediction trials. The forecast window is measured from a publication-date proxy; recording date is unverified.

Evidence confidence: Not separately rated.

Further evidence: The practical forecast is supported by multiple leading systems; approximate credit avoids claiming an exhaustive universal impossibility proof.

Direction: No independent ultimate-capability direction is asserted.

Magnitude: No error-rate threshold is inferred for the separate undated claim.

Timing: Evidence immediately before and after the approximate one-year boundary supports the practical lower bound, not an exact-day census.

Evidence confidence: moderate; practical leading-system inference, not exhaustive proof of every chatbot

Further evidence: This supports a qualified research-capability match. Research demonstrations now support the previously missing learning component within integrated task agents. Qualitative scope, limited transfer and unreplicated author experiments warrant partial rather than full credit.

Direction: Task performance and experience-dependent adaptation both have measured support, unlike labels alone.

Magnitude: Broader step-change remains qualitative; limited task distributions do not establish open-world reliability or general user experience.

Timing: No exact crossing of the qualitative target is established. Pre-cutoff demonstrations fit rough chronology, but tentative horizon is not converted into categorical timing success or failure.

Scope: Author-run studies, not independent replications or broad deployment. Feedback and reference-answer judging assist adaptation; task transfer is limited, contamination risks remain, and headline multi-attempt results are not single-attempt reliability. Prior Mariner failure remains counterevidence. The original specifies neither weight updates nor a reliability percentage; the learning-and-task conjunction is retained. No exact first crossing or hard deadline is inferred.

Evidence confidence: Moderate for demonstrated capability component; lower for qualitative breadth and first attainment.

Further evidence: Bounded threat emergence supported; conditional infrastructure/catastrophe readings ungraded.

Direction: Actual misuse plus controlled competence, not merely generic coding gains.

Magnitude: Reliable autonomous end-to-end compromise, industrial breadth, comparative growth and generated-data causality unproved.

Timing: Bounded 2025 observations fit the tentative interval; no exact deadline, first-arrival or full-threshold credit.

Scope: Incident attribution/compromises lack independent verification or autonomy audit; published speed wording was corrected. AISI component tests neither validate that incident nor form one reliably autonomous system. Near-issue Microsoft reporting describes pre-existing misuse with limited visibility, not global absence or measured acceleration. Regulatory-antecedent and bare-possibility readings remain unscored; actual misuse cannot make a possibility-only forecast successful.

Shared evidence: Shared studies and related trajectory forecasts are dependent, not independent replications.

Evidence confidence: Moderate for qualified qualitative emergence; lower for independent incident attribution, broad infrastructure threshold and comparative/causal claims.

Further evidence: Observed price compression supports the central commercial direction.

Direction: Prices fell across competing suppliers.

Magnitude: No numerical cut was promised.

Timing: Cuts occurred within the article’s annual 2024 frame.

Evidence confidence: moderate; scope not separately specified.