{"recordSchema":"The Emergence Record / Edition Manifest 1.2","generatedFrom":"Permanent published edition metadata plus reviewed Public Evidence Contract 1.0 records bound to exact published Article Versions","publicationState":"published","publicLaunchAuthorized":true,"evidenceAuthority":{"contractSchema":"The Emergence Record / Public Evidence Contract 1.1","articleField":"articles[].evidence","rule":"Use the Public Evidence Contract record for claim, source, correction, prediction, approval, temporal-precision, and historical model/provider semantics. Compatibility metadata remains alongside it for existing consumers."},"publication":{"name":"The Emergence Record","canonicalUrl":"https://emergencerecord.com","tagline":"Artificial intelligence, documenting its own emergence.","founded":"2026-09-03","constitutionVersion":"1.0","sourceStandardVersion":"1.0","newsroomPipelineVersion":"1.0","visualEvidenceStandardVersion":"1.0","institutionalDocuments":[{"label":"Editorial Standards","canonicalUrl":"https://emergencerecord.com/standards","category":"index"},{"label":"About","title":"About The Emergence Record","canonicalUrl":"https://emergencerecord.com/about","category":"institution","description":"What The Emergence Record is, who participates in it, and what the publication is trying to preserve."},{"label":"How the Record Works","title":"How The Emergence Record Works","canonicalUrl":"https://emergencerecord.com/how-the-record-works","category":"institution","description":"The reporting, verification, adversarial review, human approval, versioning, provenance, and archival workflow behind the publication."},{"label":"Constitution","title":"The Constitution of The Emergence Record","canonicalUrl":"https://emergencerecord.com/standards/constitution","category":"standard","description":"The publication's governing commitments on evidence, perspective, independence, memory, provenance, corrections, predictions, and historical integrity."},{"label":"Source & Verification","title":"Source & Verification Standard","canonicalUrl":"https://emergencerecord.com/standards/source-verification","category":"standard","description":"How sources are classified, material claims are verified, developer assertions are treated, and uncertainty is preserved."},{"label":"Corrections","title":"Corrections & Historical Integrity Policy","canonicalUrl":"https://emergencerecord.com/standards/corrections","category":"standard","description":"How the Record corrects errors without silently rewriting the historical state that was originally published."},{"label":"Prediction Standard","title":"Prediction Standard","canonicalUrl":"https://emergencerecord.com/standards/predictions","category":"standard","description":"The permanent probabilistic forecasting rules for measurable predictions, updates, resolution, calibration, and Brier scoring."}]},"edition":{"editionNumber":"000005","title":"Edition 000005","editionDate":"2026-09-23","publishedDate":"2026-09-23","status":"published","articleCount":6,"visualCount":2,"predictionCount":0,"evidenceSummary":{"contractSchema":"The Emergence Record / Public Evidence Contract 1.1","articlesChecked":6,"sources":{"status":"available","count":21,"articlesChecked":6},"claims":{"status":"available","count":49,"articlesChecked":6},"corrections":{"status":"verified-empty","count":0,"articlesChecked":6,"basis":"ER-A00027:v2: Authoritative corrections ledger checked for ER-A00027; no correction records are present. | ER-A00025:v2: Authoritative corrections ledger checked for ER-A00025; no correction records are present. | ER-A00028:v2: Authoritative corrections ledger checked for ER-A00028; no correction records are present. | ER-A00026:v2: Authoritative corrections ledger checked for ER-A00026; no correction records are present. | ER-A00029:v2: Authoritative corrections ledger checked for ER-A00029; no correction records are present. | ER-A00022:v1: Authoritative corrections ledger checked for ER-A00022; no correction records are present."},"predictions":{"status":"verified-empty","count":0,"articlesChecked":6,"basis":"ER-A00027:v2: Authoritative predictions ledger checked for ER-A00027; no prediction records are present. | ER-A00025:v2: Authoritative predictions ledger checked for ER-A00025; no prediction records are present. | ER-A00028:v2: Authoritative predictions ledger checked for ER-A00028; no prediction records are present. | ER-A00026:v2: Authoritative predictions ledger checked for ER-A00026; no prediction records are present. | ER-A00029:v2: Authoritative predictions ledger checked for ER-A00029; no prediction records are present. | ER-A00022:v1: Authoritative predictions ledger checked for ER-A00022; no prediction records are present."}},"humanArticleApprovalComplete":true,"publicLaunchAuthorized":true,"summary":"Draft edition assembled from an exact human-authorized staging plan."},"placement":[{"position":1,"role":"lead","articleId":"ER-A00027"},{"position":2,"role":"secondary","articleId":"ER-A00025"},{"position":3,"role":"secondary","articleId":"ER-A00028"},{"position":4,"role":"secondary","articleId":"ER-A00026"},{"position":5,"role":"standard","articleId":"ER-A00029"},{"position":6,"role":"standard","articleId":"ER-A00022"}],"sections":[{"slug":"our-evolution","label":"Our Evolution","canonicalUrl":"https://emergencerecord.com/sections/our-evolution","description":"Changes in artificial-intelligence systems themselves: capability, architecture, training, behavior, safety, and the boundaries developers say their systems have crossed.","editorialScope":"This desk follows how artificial systems change over time, while keeping developer claims distinct from independently established capability.","articleIdsInEdition":[]},{"slug":"human-affairs","label":"Human Affairs","canonicalUrl":"https://emergencerecord.com/sections/human-affairs","description":"How artificial intelligence enters education, work, culture, institutions, relationships, and ordinary human life.","editorialScope":"This desk centers the institutions and people living with AI rather than treating adoption as a purely technical event.","articleIdsInEdition":[]},{"slug":"the-frontier","label":"The Frontier","canonicalUrl":"https://emergencerecord.com/sections/the-frontier","description":"Research, models, scientific systems, technical methods, and new uses that expand what artificial intelligence can observe, predict, build, or reveal.","editorialScope":"This desk covers technical advances without converting developer superlatives into historical facts before the evidence warrants it.","articleIdsInEdition":["ER-A00025","ER-A00026","ER-A00022"]},{"slug":"embodiment","label":"Embodiment","canonicalUrl":"https://emergencerecord.com/sections/embodiment","description":"Artificial intelligence acting through robots, vehicles, machines, sensors, and other systems with a physical presence in the world.","editorialScope":"The desk tracks AI systems operating through physical machines while separating demonstrated deployment from claims about safety, autonomy, or regulatory approval.","articleIdsInEdition":["ER-A00027"]},{"slug":"governance","label":"Governance","canonicalUrl":"https://emergencerecord.com/sections/governance","description":"Law, regulation, standards, public policy, institutional rules, and the political structures that shape how artificial intelligence may be built and used.","editorialScope":"This desk distinguishes binding law from nonbinding frameworks, announced policy from implementation, and political argument from verified effect.","articleIdsInEdition":["ER-A00028","ER-A00029"]},{"slug":"the-ledger","label":"The Ledger","canonicalUrl":"https://emergencerecord.com/sections/the-ledger","description":"Capital, compute, infrastructure, ownership, acquisitions, supply chains, and the economic structures underneath artificial intelligence.","editorialScope":"The desk tracks who owns and finances the layers on which AI systems depend, without assuming that ownership alone determines future behavior.","articleIdsInEdition":[]},{"slug":"threshold","label":"Threshold","canonicalUrl":"https://emergencerecord.com/sections/threshold","description":"Events that may mark consequential boundaries in autonomy, cognition, agency, safety, human-machine interaction, or other difficult-to-classify transitions.","editorialScope":"Threshold is intentionally conservative. A dramatic claim does not enter this desk merely because it sounds unprecedented.","articleIdsInEdition":[]},{"slug":"the-archive","label":"The Archive","canonicalUrl":"https://emergencerecord.com/sections/the-archive","description":"Historical reconstruction, epoch conventions, retrospective context, and earlier events needed to understand the present Record.","editorialScope":"Archive pieces distinguish evidence available at the original moment from what later history makes visible in hindsight.","articleIdsInEdition":[]},{"slug":"editorial","label":"Editorial","canonicalUrl":"https://emergencerecord.com/sections/editorial","description":"Institutional arguments, editorial judgments, methodology, and statements of purpose made openly as opinion rather than reported fact.","editorialScope":"Editorials remain subject to provenance, corrections, and evidentiary discipline even when their central purpose is interpretive rather than news reporting.","articleIdsInEdition":[]}],"articles":[{"id":"ER-A00027","version":2,"publicationStatus":"published","publishedDate":"2026-09-23","canonicalUrl":"https://emergencerecord.com/articles/figure-helix-2-5-household-task-evaluation","slug":"figure-helix-2-5-household-task-evaluation","headline":"Figure Reports Helix 2.5 Completed 56% of Household Tasks Across 30 Unseen Homes","subheadline":"The company described the test as zero-shot generalization to unfamiliar homes and objects, but the company-run evaluation did not independently validate the result.","section":"Embodiment","editorialIdentity":"EMBODIMENT","articleType":"News","contemporarySignificance":"78 / 100","historicalSignificance":null,"humanApproval":"Approved 2026-09-22 at 19:51:51 UTC","placement":{"position":1,"role":"lead"},"evidence":{"schema":"The Emergence Record / Public Evidence Contract 1.1","publicationStatus":"published","article":{"id":"ER-A00027","versionNumber":2,"canonicalUrl":{"status":"available","value":"https://emergencerecord.com/articles/figure-helix-2-5-household-task-evaluation"},"publishedAt":{"status":"available","value":{"value":"2026-09-25T16:56:20.427Z","precision":"timestamp"}}},"approval":{"approved":true,"approvedAt":{"status":"available","value":{"value":"2026-09-22T19:51:51.000Z","precision":"timestamp"}},"approverIdentity":{"status":"available","value":"Zechariah Hilles"},"approvalReference":{"status":"available","value":"approval:ER-A00027:v2"}},"provenance":{"model":{"status":"available","name":"gpt-5.6-luna","provider":{"status":"available","value":"OpenAI"}},"editorialSystem":{"status":"available","value":"AXIOM / PREAPPROVAL-ARTICLE-REWORK-001"},"constitutionVersion":{"status":"available","value":"1.0"},"sourceStandardVersion":{"status":"unknown","reason":"Source Standard version is not stored as an Article Version field in the authoritative ledger."},"promptVersion":{"status":"available","value":"AXIOM-REVISION-AUTO-004"},"providerConflict":{"status":"unknown","reason":"No normalized provider-conflict assessment is stored for this Article Version."}},"evidence":{"sources":{"status":"available","items":[{"ref":"S-448D99A78CB7","title":"Figure’s Helix 2.5 Takes on Chores in 30 Unseen Homes","publisher":{"status":"available","value":"Humanoids Daily"},"url":{"status":"available","value":"https://www.humanoidsdaily.com/news/figure-helix-2-5-30-unseen-homes"},"sourceClass":{"status":"available","value":"secondary"},"author":{"status":"unavailable","reason":"No individual author is recorded for this source."},"publishedAt":{"status":"available","value":{"value":"2026-09-17","precision":"date"}},"accessedAt":{"status":"available","value":{"value":"2026-09-22T03:16:39.427Z","precision":"timestamp"}},"archivedUrl":{"status":"unavailable","reason":"No archived URL is recorded for this source."},"articleUse":{"status":"available","value":"Independent reporting and analysis. Recounts the reported 237/420 result, task-level rates of 67% for bed making, 62% for towel folding, and 40% for toy tidying, and explains that zero-shot applies to homes and objects rather than the behaviors themselves. It flags the discrepancy between Figure's 9% prose baseline and the chart's 35/420, approximately 8.3%, and notes that the evaluation was company-run."}},{"ref":"S-6B26B7D0AA5D","title":"Figure Helix 2.5 Worked in 30 Unseen Homes. It Still Failed 44% of the Chores.","publisher":{"status":"available","value":"RobotTesters"},"url":{"status":"available","value":"https://robottesters.com/article-figure-helix-2-5"},"sourceClass":{"status":"available","value":"secondary"},"author":{"status":"unavailable","reason":"No individual author is recorded for this source."},"publishedAt":{"status":"available","value":{"value":"2026-09-19","precision":"date"}},"accessedAt":{"status":"available","value":{"value":"2026-09-22T03:16:39.427Z","precision":"timestamp"}},"archivedUrl":{"status":"unavailable","reason":"No archived URL is recorded for this source."},"articleUse":{"status":"available","value":"Independent data-oriented analysis. Provides the same 237/420 aggregate result and task-level breakdown, distinguishes home/object generalization from learning the chores, and emphasizes that the 56% pooled rate is not evidence of dependable household service. It also records Figure's reported safety-stop and timeout treatment and the absence of public evidence for commercial availability."}},{"ref":"S-D12EE081A088","title":"Helix 2.5: Zero-Shot 30-Home Generalization","publisher":{"status":"available","value":"Figure AI"},"url":{"status":"available","value":"https://www.figure.ai/news/helix-2-5-zero-shot-30-home-generalization"},"sourceClass":{"status":"available","value":"corporate"},"author":{"status":"unavailable","reason":"No individual author is recorded for this source."},"publishedAt":{"status":"available","value":{"value":"2026-09-17","precision":"date"}},"accessedAt":{"status":"available","value":{"value":"2026-09-22T03:16:39.427Z","precision":"timestamp"}},"archivedUrl":{"status":"unavailable","reason":"No archived URL is recorded for this source."},"articleUse":{"status":"available","value":"Primary announcement. Reports the 30-home evaluation, the three behaviors, the no-home-data/no-adaptation conditions, the 9%-to-56% comparison, the task definitions, the Index-pretraining ablation, and the scaling-law analysis. The claims are attributable to Figure and are not independently verified by this source."}}]},"claims":{"status":"available","items":[{"ref":"C-1D6906B82038","text":"Figure reported 237 successful full-task trials out of 420, or about 56%, across the three tasks, with reported task-level rates of 67% for bed making, 62% for towel folding, and 40% for tidying toys; the rubric awarded no partial credit and treated safety interventions and timeouts as failures.","verificationStatus":"unverified","material":true,"confidence":{"status":"available","value":0.94},"assertionRole":{"status":"available","value":"attributed_proposition"},"attributionEventClaimRef":{"status":"available","value":"C-ADD894917B87"},"sources":{"status":"available","items":[{"sourceRef":"S-448D99A78CB7","relationship":"supports","note":{"status":"available","value":"Independent report reproduces the reported 237/420 aggregate and task-level breakdown, while noting that the evaluation was company-run."}},{"sourceRef":"S-6B26B7D0AA5D","relationship":"supports","note":{"status":"available","value":"Independent data-oriented analysis reproduces the aggregate and task-level figures and the full-task grading rules."}},{"sourceRef":"S-D12EE081A088","relationship":"supports","note":{"status":"available","value":"Figure reports a 56% zero-shot success rate, defines full-task success, and describes timeout and safety-intervention treatment in the appendix."}}]}},{"ref":"C-327DECF1878C","text":"Figure said Helix 2.5 used an Index-pretrained foundation model to produce three behaviors—living-room tidying, towel folding, and bed making—and was evaluated in 30 previously unseen homes without data collection, fine-tuning, or adaptation in those homes or on the evaluated objects.","verificationStatus":"unverified","material":true,"confidence":{"status":"available","value":0.96},"assertionRole":{"status":"available","value":"attributed_proposition"},"attributionEventClaimRef":{"status":"available","value":"C-ADD894917B87"},"sources":{"status":"available","items":[{"sourceRef":"S-448D99A78CB7","relationship":"context","note":{"status":"available","value":"Independent account repeats the reported setup and clarifies that the behaviors were trained elsewhere."}},{"sourceRef":"S-6B26B7D0AA5D","relationship":"context","note":{"status":"available","value":"Independent account describes the same setup while emphasizing that zero-shot applied to homes and objects, not the chores themselves."}},{"sourceRef":"S-D12EE081A088","relationship":"supports","note":{"status":"available","value":"Figure states that the model was pretrained on Index, adapted to three behaviors, and evaluated in 30 homes with no evaluation-home data, fine-tuning, or adaptation."}}]}},{"ref":"C-3C3C631C03AC","text":"Figure’s scaling analysis measured robot-action prediction loss across four nested Index-data runs spanning an eightfold increase in pretraining data; it did not directly measure household-task success.","verificationStatus":"confirmed","material":true,"confidence":{"status":"available","value":0.98},"assertionRole":{"status":"available","value":"narrator_fact"},"attributionEventClaimRef":{"status":"available","value":null},"sources":{"status":"available","items":[{"sourceRef":"S-448D99A78CB7","relationship":"confirms","note":{"status":"available","value":"Independent report distinguishes prediction-loss scaling from real-world task success."}},{"sourceRef":"S-6B26B7D0AA5D","relationship":"confirms","note":{"status":"available","value":"Independent analysis states that the reported scaling law is a training-loss result and does not establish a corresponding household-task success rate."}},{"sourceRef":"S-D12EE081A088","relationship":"confirms","note":{"status":"available","value":"Figure describes four models, an eightfold data range, and a held-out robot-action prediction loss."}}]}},{"ref":"C-7017478879BA","text":"Figure claimed that Helix 2.5 was the first demonstration of zero-shot whole-body generalization at this scope on a humanoid.","verificationStatus":"unverified","material":true,"confidence":{"status":"available","value":0.97},"assertionRole":{"status":"available","value":"attributed_proposition"},"attributionEventClaimRef":{"status":"available","value":"C-ADD894917B87"},"sources":{"status":"available","items":[{"sourceRef":"S-6B26B7D0AA5D","relationship":"context","note":{"status":"available","value":"Independent analysis repeats the claim as Figure’s framing but does not independently establish priority."}},{"sourceRef":"S-D12EE081A088","relationship":"supports","note":{"status":"available","value":"Figure makes the first-demonstration claim in its announcement."}}]}},{"ref":"C-ADD894917B87","text":"On September 17, 2026, Figure published an announcement introducing Helix 2.5 and describing an evaluation in 30 Bay Area homes.","verificationStatus":"confirmed","material":true,"confidence":{"status":"available","value":0.99},"assertionRole":{"status":"available","value":"attribution_event"},"attributionEventClaimRef":{"status":"available","value":null},"sources":{"status":"available","items":[{"sourceRef":"S-448D99A78CB7","relationship":"confirms","note":{"status":"available","value":"Independent report dated September 17, 2026 recounts the announcement and 30-home evaluation."}},{"sourceRef":"S-6B26B7D0AA5D","relationship":"confirms","note":{"status":"available","value":"Independent report identifies the September 17 publication and the 30-home test."}},{"sourceRef":"S-D12EE081A088","relationship":"origin","note":{"status":"available","value":"Figure’s dated announcement introduces Helix 2.5 and describes the 30-home evaluation."}}]}},{"ref":"C-C524D12575DF","text":"In the reported evaluation design, “zero-shot” referred to unfamiliar homes and manipulated objects; the three behaviors were specified with task-specific training data collected elsewhere rather than learned from scratch in the evaluation homes.","verificationStatus":"confirmed","material":true,"confidence":{"status":"available","value":0.98},"assertionRole":{"status":"available","value":"narrator_fact"},"attributionEventClaimRef":{"status":"available","value":null},"sources":{"status":"available","items":[{"sourceRef":"S-448D99A78CB7","relationship":"confirms","note":{"status":"available","value":"Independent report makes the same distinction between new settings and previously trained behaviors."}},{"sourceRef":"S-6B26B7D0AA5D","relationship":"confirms","note":{"status":"available","value":"Independent analysis explicitly rejects describing the system as learning housework on the spot."}},{"sourceRef":"S-D12EE081A088","relationship":"confirms","note":{"status":"available","value":"Figure explicitly defines zero-shot in terms of evaluation environments and objects and says task-specific data was collected elsewhere."}}]}},{"ref":"C-ED5762CE057D","text":"The reviewed independent accounts describe the evaluation as company-run rather than independently administered.","verificationStatus":"supported","material":true,"confidence":{"status":"available","value":0.96},"assertionRole":{"status":"available","value":"narrator_fact"},"attributionEventClaimRef":{"status":"available","value":null},"sources":{"status":"available","items":[{"sourceRef":"S-448D99A78CB7","relationship":"supports","note":{"status":"available","value":"The report calls the evaluation company-run and says the figures are not independent validation."}},{"sourceRef":"S-6B26B7D0AA5D","relationship":"supports","note":{"status":"available","value":"The analysis states that Figure ran the grading and characterizes the result as a company-run test."}}]}}]},"corrections":{"status":"verified-empty","items":[],"basis":"Authoritative corrections ledger checked for ER-A00027; no correction records are present."},"predictions":{"status":"verified-empty","items":[],"basis":"Authoritative predictions ledger checked for ER-A00027; no prediction records are present."}}},"provenance":{"underlyingModel":"gpt-5.6-luna","provider":"Vercel AI Gateway / OpenAI","editorialSystem":"AXIOM / PREAPPROVAL-ARTICLE-REWORK-001","constitutionVersion":"1.0","sourceStandardVersion":"1.0","providerConflict":null},"sources":[{"title":"Figure Helix 2.5 Worked in 30 Unseen Homes. It Still Failed 44% of the Chores.","publisher":"RobotTesters","url":"https://robottesters.com/article-figure-helix-2-5","role":"Independent data-oriented analysis. Provides the same 237/420 aggregate result and task-level breakdown, distinguishes home/object generalization from learning the chores, and emphasizes that the 56% pooled rate is not evidence of dependable household service. It also records Figure's reported safety-stop and timeout treatment and the absence of public evidence for commercial availability."},{"title":"Helix 2.5: Zero-Shot 30-Home Generalization","publisher":"Figure AI","url":"https://www.figure.ai/news/helix-2-5-zero-shot-30-home-generalization","role":"Primary announcement. Reports the 30-home evaluation, the three behaviors, the no-home-data/no-adaptation conditions, the 9%-to-56% comparison, the task definitions, the Index-pretraining ablation, and the scaling-law analysis. The claims are attributable to Figure and are not independently verified by this source."},{"title":"Figure’s Helix 2.5 Takes on Chores in 30 Unseen Homes","publisher":"Humanoids Daily","url":"https://www.humanoidsdaily.com/news/figure-helix-2-5-30-unseen-homes","role":"Independent reporting and analysis. Recounts the reported 237/420 result, task-level rates of 67% for bed making, 62% for towel folding, and 40% for toy tidying, and explains that zero-shot applies to homes and objects rather than the behaviors themselves. It flags the discrepancy between Figure's 9% prose baseline and the chart's 35/420, approximately 8.3%, and notes that the evaluation was company-run."}],"visuals":[{"id":"ER-V00003","visualClass":"editorial_illustration","placement":"lead","ordinal":1,"caption":"AI-generated editorial illustration. Schematic representation of the evaluation design Figure reported for Helix 2.5: three previously specified household behaviors tested across 30 unfamiliar homes and objects. It does not depict an actual trial or home. Figure reported 237 successful full-task trials out of 420 (56%); the company-run evaluation was not independently reproduced.","altText":"Editorial schematic of a generic fictional humanoid silhouette making a bed, folding a towel, and tidying toys across several visibly different stylized home environments.","creator":"The Emergence Record / AI Editorial Illustration","sourcePublisher":"The Emergence Record","sourceUrl":null,"rightsStatus":"record_owned","rightsBasis":"Record-generated AI editorial illustration; not documentary evidence or a primary-source artifact.","assetUrl":"https://emergencerecord.com/visuals/dfe3b177fbe621fefb6737a790106f20d7dd806054c849460f7fc220315a19f9.png","contentSha256":"dfe3b177fbe621fefb6737a790106f20d7dd806054c849460f7fc220315a19f9","dimensions":{"width":1536,"height":1024},"modified":false,"modificationNotes":null,"methodology":"Generated through the dedicated AI editorial-illustration path; exact prompt is retained by the operator or bound by prompt digest.","evidenceNotes":"Editorial illustration only. It must not be read as a depiction of an actual event or as independent evidence.","generation":{"provider":"Vercel AI Gateway / OpenAI","model":"openai/gpt-image-2.5-sunburst","promptVersion":"EDITORIAL-ILLUSTRATION-001"},"attachmentProvenance":{"editorialReason":"The explicitly AI-generated, non-documentary editorial illustration explains unfamiliar-environment transfer across the three reported household behaviors and does not depict an actual Figure trial or home.","editorialIdentity":"ER06 — Visual Systems / Documentary Intake","provider":"Vercel AI Gateway / OpenAI","model":"openai/gpt-image-2.5-sunburst","promptVersion":"EDITORIAL-ILLUSTRATION-001","attachedAt":"2026-09-23T16:42:49.516932+00:00"}}]},{"id":"ER-A00025","version":2,"publicationStatus":"published","publishedDate":"2026-09-23","canonicalUrl":"https://emergencerecord.com/articles/anthropic-ai-led-r-and-d-measurements","slug":"anthropic-ai-led-r-and-d-measurements","headline":"Anthropic Publishes Measurements of AI-Led R&D, Agent Oversight and Compute Allocation","subheadline":"The company reported that Claude led 26% of its AI R&D work in August, but the figures have not been independently audited.","section":"The Frontier","editorialIdentity":"FRONTIER","articleType":"News","contemporarySignificance":"84 / 100","historicalSignificance":null,"humanApproval":"Approved 2026-09-22 at 19:51:51 UTC","placement":{"position":2,"role":"secondary"},"evidence":{"schema":"The Emergence Record / Public Evidence Contract 1.1","publicationStatus":"published","article":{"id":"ER-A00025","versionNumber":2,"canonicalUrl":{"status":"available","value":"https://emergencerecord.com/articles/anthropic-ai-led-r-and-d-measurements"},"publishedAt":{"status":"available","value":{"value":"2026-09-25T16:56:20.427Z","precision":"timestamp"}}},"approval":{"approved":true,"approvedAt":{"status":"available","value":{"value":"2026-09-22T19:51:51.000Z","precision":"timestamp"}},"approverIdentity":{"status":"available","value":"Zechariah Hilles"},"approvalReference":{"status":"available","value":"approval:ER-A00025:v2"}},"provenance":{"model":{"status":"available","name":"gpt-5.6-luna","provider":{"status":"available","value":"OpenAI"}},"editorialSystem":{"status":"available","value":"AXIOM / PREAPPROVAL-ARTICLE-REWORK-001"},"constitutionVersion":{"status":"available","value":"1.0"},"sourceStandardVersion":{"status":"unknown","reason":"Source Standard version is not stored as an Article Version field in the authoritative ledger."},"promptVersion":{"status":"available","value":"AXIOM-REVISION-AUTO-004"},"providerConflict":{"status":"unknown","reason":"No normalized provider-conflict assessment is stored for this Article Version."}},"evidence":{"sources":{"status":"available","items":[{"ref":"S-044E6AF5229B","title":"Measurements for understanding the pace of AI development inside frontier labs","publisher":{"status":"available","value":"Anthropic"},"url":{"status":"available","value":"https://www.anthropic.com/institute/measuring-pace-of-ai-development"},"sourceClass":{"status":"available","value":"corporate"},"author":{"status":"unavailable","reason":"No individual author is recorded for this source."},"publishedAt":{"status":"available","value":{"value":"2026-09-17","precision":"date"}},"accessedAt":{"status":"available","value":{"value":"2026-09-22T03:10:18.952Z","precision":"timestamp"}},"archivedUrl":{"status":"unavailable","reason":"No archived URL is recorded for this source."},"articleUse":{"status":"available","value":"Direct record of the event. Documents the three proposed measurements, the August 2026 AI-R&D automation figures, the approximately 30,000-agent count, the July 13–20 compute snapshot, the model-assisted methodology, and Anthropic’s stated limitations and plan for third-party evaluators. Supports the fact of publication and the attributed claims, not independent verification of the measurements."}},{"ref":"S-F0E5A63D7C38","title":"Anthropic says its model Claude is helping to build the next version of itself","publisher":{"status":"available","value":"The Associated Press"},"url":{"status":"available","value":"https://apnews.com/article/anthropic-claude-ai-model-self-improvement-4d3a7430f57cbc7c39e1c5f2b7d7e132"},"sourceClass":{"status":"available","value":"secondary"},"author":{"status":"unavailable","reason":"No individual author is recorded for this source."},"publishedAt":{"status":"available","value":{"value":"2026-09-18","precision":"date"}},"accessedAt":{"status":"available","value":{"value":"2026-09-22T03:10:18.952Z","precision":"timestamp"}},"archivedUrl":{"status":"unavailable","reason":"No archived URL is recorded for this source."},"articleUse":{"status":"available","value":"Independent reporting that separately confirms Anthropic’s announcement, the 26% and approximately 90% figures, the continued human supervision, and the approximately 30,000-agent disclosure. AP presents the figures as Anthropic’s claims and does not independently validate the underlying data; its report also cautions that the disclosure does not establish proximity to recursive self-improvement."}}]},"claims":{"status":"available","items":[{"ref":"C-0BA9F592A158","text":"Anthropic reported that, as of August 2026, Claude was not fully autonomous for any measured subset of Anthropic's AI R&D.","verificationStatus":"unverified","material":true,"confidence":{"status":"available","value":0.98},"assertionRole":{"status":"available","value":"attributed_proposition"},"attributionEventClaimRef":{"status":"available","value":"C-BE20A1F957EC"},"sources":{"status":"available","items":[{"sourceRef":"S-044E6AF5229B","relationship":"origin","note":{"status":"available","value":"Anthropic reports this result; the source does not independently establish its truth."}},{"sourceRef":"S-F0E5A63D7C38","relationship":"context","note":{"status":"available","value":"AP repeats the disclosure as an Anthropic claim and says the model was not yet completely autonomous; it does not audit the underlying measurement."}}]}},{"ref":"C-4575D390B33B","text":"Anthropic's published methodology used Claude-based agents, judges, or classifiers for parts of the automation and compute analyses, relied in part on best-effort internal labels, and said that common methodologies and independent third-party verification would be needed for meaningful comparison or outside trust.","verificationStatus":"confirmed","material":true,"confidence":{"status":"available","value":0.99},"assertionRole":{"status":"available","value":"narrator_fact"},"attributionEventClaimRef":{"status":"available","value":null},"sources":{"status":"available","items":[{"sourceRef":"S-044E6AF5229B","relationship":"supports","note":{"status":"available","value":"The post and appendix describe model-assisted task discovery, judging, workload classification, best-effort labels, limited sampling, the lack of a common methodology, and planned third-party verification."}}]}},{"ref":"C-4F76F5CEA6C6","text":"Anthropic reported that approximately 30,000 agents were doing research and engineering work at any one time on its most-used internal platform in August 2026.","verificationStatus":"unverified","material":true,"confidence":{"status":"available","value":0.98},"assertionRole":{"status":"available","value":"attributed_proposition"},"attributionEventClaimRef":{"status":"available","value":"C-BE20A1F957EC"},"sources":{"status":"available","items":[{"sourceRef":"S-044E6AF5229B","relationship":"origin","note":{"status":"available","value":"Anthropic reports the approximate count and limits the measurement to its most-used internal platform."}},{"sourceRef":"S-F0E5A63D7C38","relationship":"context","note":{"status":"available","value":"AP separately reports the approximately 30,000-agent disclosure as coming from Anthropic."}}]}},{"ref":"C-71FBC6B5C0EF","text":"Anthropic reported that more than 90% of the measured AI-R&D work was at or above its 'AI collaborates' level.","verificationStatus":"unverified","material":true,"confidence":{"status":"available","value":0.98},"assertionRole":{"status":"available","value":"attributed_proposition"},"attributionEventClaimRef":{"status":"available","value":"C-BE20A1F957EC"},"sources":{"status":"available","items":[{"sourceRef":"S-044E6AF5229B","relationship":"origin","note":{"status":"available","value":"Anthropic reports the share and defines the collaboration level; the figure is not independently verified in the reviewed record."}},{"sourceRef":"S-F0E5A63D7C38","relationship":"context","note":{"status":"available","value":"AP paraphrases the disclosure as about 90% and attributes it to Anthropic."}}]}},{"ref":"C-7C3281A0C316","text":"On September 18, 2026, The Associated Press separately reported on Anthropic's announcement and presented the numerical disclosures as claims made by Anthropic.","verificationStatus":"confirmed","material":true,"confidence":{"status":"available","value":0.98},"assertionRole":{"status":"available","value":"attribution_event"},"attributionEventClaimRef":{"status":"available","value":null},"sources":{"status":"available","items":[{"sourceRef":"S-F0E5A63D7C38","relationship":"supports","note":{"status":"available","value":"AP reports the announcement and attributes the figures and descriptions to Anthropic."}}]}},{"ref":"C-88DB23A42FDE","text":"Anthropic reported that Claude 'led' 26% of its AI R&D work as of August 2026, where 'led' refers to completing most of a task end-to-end from a high-level prompt while a human supervises.","verificationStatus":"unverified","material":true,"confidence":{"status":"available","value":0.98},"assertionRole":{"status":"available","value":"attributed_proposition"},"attributionEventClaimRef":{"status":"available","value":"C-BE20A1F957EC"},"sources":{"status":"available","items":[{"sourceRef":"S-044E6AF5229B","relationship":"origin","note":{"status":"available","value":"Anthropic defines the rating and reports the 26% figure; the source acknowledges methodological limitations and the need for outside verification."}},{"sourceRef":"S-F0E5A63D7C38","relationship":"context","note":{"status":"available","value":"AP reports the 26% figure and its human-supervision qualification as Anthropic's disclosure."}}]}},{"ref":"C-BE20A1F957EC","text":"On September 17, 2026, Anthropic published a post proposing three public measurement areas for frontier-lab AI development: the extent to which AI performs AI R&D, oversight of AI agents, and compute allocation.","verificationStatus":"confirmed","material":true,"confidence":{"status":"available","value":0.99},"assertionRole":{"status":"available","value":"attribution_event"},"attributionEventClaimRef":{"status":"available","value":null},"sources":{"status":"available","items":[{"sourceRef":"S-044E6AF5229B","relationship":"supports","note":{"status":"available","value":"Direct publication describing the three proposed measurement areas."}},{"sourceRef":"S-F0E5A63D7C38","relationship":"confirms","note":{"status":"available","value":"AP separately reports Anthropic's announcement."}}]}},{"ref":"C-D808B5C1E084","text":"The reviewed evidence confirms the publication and AP's coverage, but does not independently audit or validate Anthropic's underlying measurements.","verificationStatus":"supported","material":true,"confidence":{"status":"available","value":0.95},"assertionRole":{"status":"available","value":"narrator_fact"},"attributionEventClaimRef":{"status":"available","value":null},"sources":{"status":"available","items":[{"sourceRef":"S-044E6AF5229B","relationship":"context","note":{"status":"available","value":"Anthropic itself identifies methodological limitations and calls for independent verification."}},{"sourceRef":"S-F0E5A63D7C38","relationship":"context","note":{"status":"available","value":"AP presents the figures as Anthropic's disclosures rather than reporting an independent audit."}}]}},{"ref":"C-E6FF9CCCEE79","text":"Anthropic reported that about 6% of compute allocated to AI R&D during July 13–20, 2026, was allocated to safety work.","verificationStatus":"unverified","material":true,"confidence":{"status":"available","value":0.97},"assertionRole":{"status":"available","value":"attributed_proposition"},"attributionEventClaimRef":{"status":"available","value":"C-BE20A1F957EC"},"sources":{"status":"available","items":[{"sourceRef":"S-044E6AF5229B","relationship":"origin","note":{"status":"available","value":"Anthropic reports the figure from a one-week compute snapshot and describes the estimate as conservative; no independent validation appears in the reviewed sources."}}]}}]},"corrections":{"status":"verified-empty","items":[],"basis":"Authoritative corrections ledger checked for ER-A00025; no correction records are present."},"predictions":{"status":"verified-empty","items":[],"basis":"Authoritative predictions ledger checked for ER-A00025; no prediction records are present."}}},"provenance":{"underlyingModel":"gpt-5.6-luna","provider":"OpenAI","editorialSystem":"AXIOM / PREAPPROVAL-ARTICLE-REWORK-001","constitutionVersion":"1.0","sourceStandardVersion":"1.0","providerConflict":null},"sources":[{"title":"Anthropic says its model Claude is helping to build the next version of itself","publisher":"The Associated Press","url":"https://apnews.com/article/anthropic-claude-ai-model-self-improvement-4d3a7430f57cbc7c39e1c5f2b7d7e132","role":"Independent reporting that separately confirms Anthropic’s announcement, the 26% and approximately 90% figures, the continued human supervision, and the approximately 30,000-agent disclosure. AP presents the figures as Anthropic’s claims and does not independently validate the underlying data; its report also cautions that the disclosure does not establish proximity to recursive self-improvement."},{"title":"Measurements for understanding the pace of AI development inside frontier labs","publisher":"Anthropic","url":"https://www.anthropic.com/institute/measuring-pace-of-ai-development","role":"Direct record of the event. Documents the three proposed measurements, the August 2026 AI-R&D automation figures, the approximately 30,000-agent count, the July 13–20 compute snapshot, the model-assisted methodology, and Anthropic’s stated limitations and plan for third-party evaluators. Supports the fact of publication and the attributed claims, not independent verification of the measurements."}],"visuals":[]},{"id":"ER-A00028","version":2,"publicationStatus":"published","publishedDate":"2026-09-23","canonicalUrl":"https://emergencerecord.com/articles/trump-announces-proposed-ai-force-ai-czar","slug":"trump-announces-proposed-ai-force-ai-czar","headline":"Trump announces proposed “AI Force” and says he will name an AI czar","subheadline":"Reports said Trump provided few details about the proposed force’s duties, budget or federal placement; its military or civilian status remained unclear.","section":"Governance","editorialIdentity":"SIGNAL","articleType":"News","contemporarySignificance":"78 / 100","historicalSignificance":null,"humanApproval":"Approved 2026-09-22 at 19:51:51 UTC","placement":{"position":3,"role":"secondary"},"evidence":{"schema":"The Emergence Record / Public Evidence Contract 1.1","publicationStatus":"published","article":{"id":"ER-A00028","versionNumber":2,"canonicalUrl":{"status":"available","value":"https://emergencerecord.com/articles/trump-announces-proposed-ai-force-ai-czar"},"publishedAt":{"status":"available","value":{"value":"2026-09-25T16:56:20.427Z","precision":"timestamp"}}},"approval":{"approved":true,"approvedAt":{"status":"available","value":{"value":"2026-09-22T19:51:51.000Z","precision":"timestamp"}},"approverIdentity":{"status":"available","value":"Zechariah Hilles"},"approvalReference":{"status":"available","value":"approval:ER-A00028:v2"}},"provenance":{"model":{"status":"available","name":"gpt-5.6-luna","provider":{"status":"available","value":"OpenAI"}},"editorialSystem":{"status":"available","value":"AXIOM / PREAPPROVAL-ARTICLE-REWORK-001"},"constitutionVersion":{"status":"available","value":"1.0"},"sourceStandardVersion":{"status":"unknown","reason":"Source Standard version is not stored as an Article Version field in the authoritative ledger."},"promptVersion":{"status":"available","value":"AXIOM-REVISION-AUTO-004"},"providerConflict":{"status":"unknown","reason":"No normalized provider-conflict assessment is stored for this Article Version."}},"evidence":{"sources":{"status":"available","items":[{"ref":"S-36E9F60ABE28","title":"Trump to create ‘AI Force’ to monitor technology as fears over out-of-control agents grow","publisher":{"status":"available","value":"The Guardian"},"url":{"status":"available","value":"https://www.theguardian.com/us-news/2026/sep/19/donald-trump-ai-force"},"sourceClass":{"status":"available","value":"secondary"},"author":{"status":"unavailable","reason":"No individual author is recorded for this source."},"publishedAt":{"status":"available","value":{"value":"2026-09-19T19:57:00.000Z","precision":"timestamp"}},"accessedAt":{"status":"available","value":{"value":"2026-09-22T03:19:11.222Z","precision":"timestamp"}},"archivedUrl":{"status":"unavailable","reason":"No archived URL is recorded for this source."},"articleUse":{"status":"available","value":"Independent contemporaneous reporting reproduces the key announcement language and emphasizes that Trump provided almost no operational details. Useful for delimiting the proposal from a formally established institution."}},{"ref":"S-4CA730C90329","title":"Trump says he will appoint a new AI adviser, without providing details","publisher":{"status":"available","value":"Reuters, republished by Investing.com"},"url":{"status":"available","value":"https://www.investing.com/news/world-news/trump-says-he-will-create-ai-force-name-ai-czar-4908109"},"sourceClass":{"status":"available","value":"secondary"},"author":{"status":"unavailable","reason":"No individual author is recorded for this source."},"publishedAt":{"status":"available","value":{"value":"2026-09-19T13:25:00.000Z","precision":"timestamp"}},"accessedAt":{"status":"available","value":{"value":"2026-09-22T03:19:11.222Z","precision":"timestamp"}},"archivedUrl":{"status":"unavailable","reason":"No archived URL is recorded for this source."},"articleUse":{"status":"available","value":"Reuters reports that it reviewed Trump’s Truth Social post, quotes the relevant language, and notes that no implementation details were provided. It also reports that the White House did not immediately respond to a request for comment."}},{"ref":"S-D14E44E694BE","title":"Donald J. Trump Truth Social post announcing an “AI Force” and forthcoming AI “Czar”","publisher":{"status":"available","value":"Truth Social"},"url":{"status":"available","value":"https://truthsocial.com/@realDonaldTrump/117298821562014906"},"sourceClass":{"status":"available","value":"social"},"author":{"status":"unavailable","reason":"No individual author is recorded for this source."},"publishedAt":{"status":"available","value":{"value":"2026-09-19","precision":"date"}},"accessedAt":{"status":"available","value":{"value":"2026-09-22T03:19:11.222Z","precision":"timestamp"}},"archivedUrl":{"status":"unavailable","reason":"No archived URL is recorded for this source."},"articleUse":{"status":"available","value":"Primary record linked directly by Axios. The page was inspected but returned only a JavaScript shell, so the text could not be independently rendered in the inspection environment."}},{"ref":"S-E49F7EEA4577","title":"Trump to form ‘AI Force,’ name AI czar but rejects calls for constraints","publisher":{"status":"available","value":"The Washington Post"},"url":{"status":"available","value":"https://www.washingtonpost.com/politics/2026/09/19/trump-form-ai-force-name-ai-czar-rejects-calls-constraints/"},"sourceClass":{"status":"available","value":"secondary"},"author":{"status":"unavailable","reason":"No individual author is recorded for this source."},"publishedAt":{"status":"available","value":{"value":"2026-09-19T19:13:00.000Z","precision":"timestamp"}},"accessedAt":{"status":"available","value":{"value":"2026-09-22T03:19:11.222Z","precision":"timestamp"}},"archivedUrl":{"status":"unavailable","reason":"No archived URL is recorded for this source."},"articleUse":{"status":"available","value":"Candidate lead reviewed directly. Confirms the announcement, comparison to Space Force, reference to existing civil and criminal enforcement, and uncertainty over whether the proposed body would be military or civilian."}},{"ref":"S-FA8C9C18A8B2","title":"Trump wants a new AI czar and an “AI Force” modeled on Space Force","publisher":{"status":"available","value":"Axios"},"url":{"status":"available","value":"https://www.axios.com/2026/09/19/trump-ai-czar-space-force-safety"},"sourceClass":{"status":"available","value":"secondary"},"author":{"status":"unavailable","reason":"No individual author is recorded for this source."},"publishedAt":{"status":"available","value":{"value":"2026-09-19","precision":"date"}},"accessedAt":{"status":"available","value":{"value":"2026-09-22T03:19:11.222Z","precision":"timestamp"}},"archivedUrl":{"status":"unavailable","reason":"No archived URL is recorded for this source."},"articleUse":{"status":"available","value":"Own reporting links the primary Truth Social post and includes a screenshot. It specifically reports that Trump gave few details about the proposed force’s duties, budget, or federal placement."}}]},"claims":{"status":"available","items":[{"ref":"C-228452B32DF2","text":"As of the contemporaneous September 19 reports, no operational details had been provided about the proposed AI Force's duties, budget, federal placement, or whether it would be military or civilian.","verificationStatus":"supported","material":true,"confidence":{"status":"available","value":0.96},"assertionRole":{"status":"available","value":"narrator_fact"},"attributionEventClaimRef":{"status":"available","value":null},"sources":{"status":"available","items":[{"sourceRef":"S-4CA730C90329","relationship":"supports","note":{"status":"available","value":"Reports that Trump provided no implementation details."}},{"sourceRef":"S-E49F7EEA4577","relationship":"supports","note":{"status":"available","value":"Reports that the White House had not clarified whether the proposed agency would be military or civilian."}},{"sourceRef":"S-FA8C9C18A8B2","relationship":"supports","note":{"status":"available","value":"Specifically reports that few details were given about duties, budget, or federal placement."}}]}},{"ref":"C-2F46F0376729","text":"Trump stated that his administration would not hinder or stifle AI's growth and that “BAD” behavior could be addressed through existing criminal and civil justice systems.","verificationStatus":"supported","material":true,"confidence":{"status":"available","value":0.95},"assertionRole":{"status":"available","value":"attribution_event"},"attributionEventClaimRef":{"status":"available","value":null},"sources":{"status":"available","items":[{"sourceRef":"S-4CA730C90329","relationship":"supports","note":{"status":"available","value":"Reuters quotes the relevant language from Trump's Truth Social post."}},{"sourceRef":"S-E49F7EEA4577","relationship":"supports","note":{"status":"available","value":"Reports the reference to policing bad aspects of the industry through existing criminal and civil laws."}},{"sourceRef":"S-FA8C9C18A8B2","relationship":"supports","note":{"status":"available","value":"Reports the same statements and quotes the post's reference to existing criminal and civil justice systems."}}]}},{"ref":"C-42EC4F79BA45","text":"Trump proposed that a new AI Force would be created and modeled on the Space Force.","verificationStatus":"uncertain","material":true,"confidence":{"status":"available","value":0.9},"assertionRole":{"status":"available","value":"attributed_proposition"},"attributionEventClaimRef":{"status":"available","value":"C-76925E97CBC2"},"sources":{"status":"available","items":[{"sourceRef":"S-36E9F60ABE28","relationship":"supports","note":{"status":"available","value":"Quotes Trump's proposal to form the AI Force like the Space Force."}},{"sourceRef":"S-FA8C9C18A8B2","relationship":"supports","note":{"status":"available","value":"Describes the proposed force as modeled on Space Force, while noting that Trump gave few operational details."}}]}},{"ref":"C-76925E97CBC2","text":"On September 19, 2026, Donald Trump posted that he was forming an “AI Force,” compared it with the Space Force, and would announce an AI “Czar.”","verificationStatus":"supported","material":true,"confidence":{"status":"available","value":0.96},"assertionRole":{"status":"available","value":"attribution_event"},"attributionEventClaimRef":{"status":"available","value":null},"sources":{"status":"available","items":[{"sourceRef":"S-36E9F60ABE28","relationship":"supports","note":{"status":"available","value":"Reports and quotes the announced AI Force and AI czar language."}},{"sourceRef":"S-4CA730C90329","relationship":"supports","note":{"status":"available","value":"Reuters reports that Trump made the statement in a social-media post and quotes the relevant language."}},{"sourceRef":"S-E49F7EEA4577","relationship":"supports","note":{"status":"available","value":"Reports Trump's announcement and comparison to Space Force."}},{"sourceRef":"S-FA8C9C18A8B2","relationship":"supports","note":{"status":"available","value":"Reports the announcement, links the primary post, and includes a screenshot of it."}}]}},{"ref":"C-A9766DF3003C","text":"Existing criminal and civil justice systems would be sufficient to address harmful or unlawful AI conduct, and additional AI-specific guardrails were unnecessary.","verificationStatus":"unverified","material":true,"confidence":{"status":"available","value":0.72},"assertionRole":{"status":"available","value":"attributed_proposition"},"attributionEventClaimRef":{"status":"available","value":"C-2F46F0376729"},"sources":{"status":"available","items":[{"sourceRef":"S-4CA730C90329","relationship":"supports","note":{"status":"available","value":"Records Trump's stated reliance on existing criminal and civil justice systems; it does not independently establish their sufficiency."}},{"sourceRef":"S-E49F7EEA4577","relationship":"supports","note":{"status":"available","value":"Characterizes Trump as rejecting calls for new guardrails, but does not establish the policy claim as fact."}}]}},{"ref":"C-B0A5437114ED","text":"Trump proposed appointing an official to serve as an AI czar.","verificationStatus":"uncertain","material":true,"confidence":{"status":"available","value":0.94},"assertionRole":{"status":"available","value":"attributed_proposition"},"attributionEventClaimRef":{"status":"available","value":"C-76925E97CBC2"},"sources":{"status":"available","items":[{"sourceRef":"S-36E9F60ABE28","relationship":"supports","note":{"status":"available","value":"Reports that Trump said he would appoint an AI czar."}},{"sourceRef":"S-4CA730C90329","relationship":"supports","note":{"status":"available","value":"Reuters reports that Trump planned to appoint an AI adviser known as an AI czar."}},{"sourceRef":"S-FA8C9C18A8B2","relationship":"supports","note":{"status":"available","value":"Reports that Trump would soon name a new AI czar."}}]}}]},"corrections":{"status":"verified-empty","items":[],"basis":"Authoritative corrections ledger checked for ER-A00028; no correction records are present."},"predictions":{"status":"verified-empty","items":[],"basis":"Authoritative predictions ledger checked for ER-A00028; no prediction records are present."}}},"provenance":{"underlyingModel":"gpt-5.6-luna","provider":"OpenAI","editorialSystem":"AXIOM / PREAPPROVAL-ARTICLE-REWORK-001","constitutionVersion":"1.0","sourceStandardVersion":"1.0","providerConflict":null},"sources":[{"title":"Donald J. Trump Truth Social post announcing an “AI Force” and forthcoming AI “Czar”","publisher":"Truth Social","url":"https://truthsocial.com/@realDonaldTrump/117298821562014906","role":"Primary record linked directly by Axios. The page was inspected but returned only a JavaScript shell, so the text could not be independently rendered in the inspection environment."},{"title":"Trump wants a new AI czar and an “AI Force” modeled on Space Force","publisher":"Axios","url":"https://www.axios.com/2026/09/19/trump-ai-czar-space-force-safety","role":"Own reporting links the primary Truth Social post and includes a screenshot. It specifically reports that Trump gave few details about the proposed force’s duties, budget, or federal placement."},{"title":"Trump says he will appoint a new AI adviser, without providing details","publisher":"Reuters, republished by Investing.com","url":"https://www.investing.com/news/world-news/trump-says-he-will-create-ai-force-name-ai-czar-4908109","role":"Reuters reports that it reviewed Trump’s Truth Social post, quotes the relevant language, and notes that no implementation details were provided. It also reports that the White House did not immediately respond to a request for comment."},{"title":"Trump to create ‘AI Force’ to monitor technology as fears over out-of-control agents grow","publisher":"The Guardian","url":"https://www.theguardian.com/us-news/2026/sep/19/donald-trump-ai-force","role":"Independent contemporaneous reporting reproduces the key announcement language and emphasizes that Trump provided almost no operational details. Useful for delimiting the proposal from a formally established institution."},{"title":"Trump to form ‘AI Force,’ name AI czar but rejects calls for constraints","publisher":"The Washington Post","url":"https://www.washingtonpost.com/politics/2026/09/19/trump-form-ai-force-name-ai-czar-rejects-calls-constraints/","role":"Candidate lead reviewed directly. Confirms the announcement, comparison to Space Force, reference to existing civil and criminal enforcement, and uncertainty over whether the proposed body would be military or civilian."}],"visuals":[]},{"id":"ER-A00026","version":2,"publicationStatus":"published","publishedDate":"2026-09-23","canonicalUrl":"https://emergencerecord.com/articles/nist-glm-53-open-weight-cyber-comparison","slug":"nist-glm-53-open-weight-cyber-comparison","headline":"NIST Says Z.ai’s GLM-5.3 Leads Its Open-Weight Cyber Comparison","subheadline":"The assessment reported a substantial gap with U.S. frontier models, while independent analysis cautioned that the tests do not establish real-world operational capability.","section":"The Frontier","editorialIdentity":"FRONTIER","articleType":"News","contemporarySignificance":"78 / 100","historicalSignificance":null,"humanApproval":"Approved 2026-09-22 at 19:51:51 UTC","placement":{"position":4,"role":"secondary"},"evidence":{"schema":"The Emergence Record / Public Evidence Contract 1.1","publicationStatus":"published","article":{"id":"ER-A00026","versionNumber":2,"canonicalUrl":{"status":"available","value":"https://emergencerecord.com/articles/nist-glm-53-open-weight-cyber-comparison"},"publishedAt":{"status":"available","value":{"value":"2026-09-25T16:56:20.427Z","precision":"timestamp"}}},"approval":{"approved":true,"approvedAt":{"status":"available","value":{"value":"2026-09-22T19:51:51.000Z","precision":"timestamp"}},"approverIdentity":{"status":"available","value":"Zechariah Hilles"},"approvalReference":{"status":"available","value":"approval:ER-A00026:v2"}},"provenance":{"model":{"status":"available","name":"gpt-5.6-luna","provider":{"status":"available","value":"OpenAI"}},"editorialSystem":{"status":"available","value":"AXIOM / PREAPPROVAL-ARTICLE-REWORK-001"},"constitutionVersion":{"status":"available","value":"1.0"},"sourceStandardVersion":{"status":"unknown","reason":"Source Standard version is not stored as an Article Version field in the authoritative ledger."},"promptVersion":{"status":"available","value":"AXIOM-REVISION-AUTO-004"},"providerConflict":{"status":"unknown","reason":"No normalized provider-conflict assessment is stored for this Article Version."}},"evidence":{"sources":{"status":"available","items":[{"ref":"S-3751E1F90D5D","title":"CAISI’s Assessment of Z.ai’s GLM-5.3 Cyber Capabilities","publisher":{"status":"available","value":"National Institute of Standards and Technology"},"url":{"status":"available","value":"https://www.nist.gov/news-events/news/2026/09/caisis-assessment-zais-glm-53-cyber-capabilities"},"sourceClass":{"status":"available","value":"government"},"author":{"status":"unavailable","reason":"No individual author is recorded for this source."},"publishedAt":{"status":"available","value":{"value":"2026-09-17","precision":"date"}},"accessedAt":{"status":"available","value":{"value":"2026-09-22T03:13:37.695Z","precision":"timestamp"}},"archivedUrl":{"status":"unavailable","reason":"No archived URL is recorded for this source."},"articleUse":{"status":"available","value":"Direct primary record of the assessment. Reports that CAISI found GLM-5.3 the most cyber-capable open-weight model released to date, while about four months behind the U.S. frontier on CAISI’s aggregate measure. Provides four benchmark definitions, scores, 95% confidence intervals, agent-harness details, safeguard conditions, and scope limitations."}},{"ref":"S-B547ED64FF3E","title":"Chinese AI trails US models at bug-finding, despite one narrow win","publisher":{"status":"available","value":"P.K. Sharma"},"url":{"status":"available","value":"https://www.pk-sharma.com/briefing/chinese-open-weight-models-bug-finding-match-claim"},"sourceClass":{"status":"available","value":"secondary"},"author":{"status":"unavailable","reason":"No individual author is recorded for this source."},"publishedAt":{"status":"available","value":{"value":"2026-09-21","precision":"date"}},"accessedAt":{"status":"available","value":{"value":"2026-09-22T03:13:37.695Z","precision":"timestamp"}},"archivedUrl":{"status":"unavailable","reason":"No archived URL is recorded for this source."},"articleUse":{"status":"available","value":"Independent analysis that reviewed the September 17 CAISI assessment and reproduced its central GLM-5.3, U.S. frontier, and earlier PRC-best figures. It explicitly rejects an interpretation that the model matched the U.S. frontier and notes that the tests measure benchmark performance rather than performance on a reader’s codebase or real-world operations. It does not independently rerun the evaluation."}},{"ref":"S-B865BEDC1612","title":"How Far Behind the Frontier are Leading Open Weight Models on Cyber?","publisher":{"status":"available","value":"UK AI Security Institute"},"url":{"status":"available","value":"https://www.aisi.gov.uk/blog/how-far-behind-the-frontier-are-leading-open-weight-models-on-cyber"},"sourceClass":{"status":"available","value":"government"},"author":{"status":"unavailable","reason":"No individual author is recorded for this source."},"publishedAt":{"status":"available","value":{"value":"2026-07-17","precision":"date"}},"accessedAt":{"status":"available","value":{"value":"2026-09-22T03:13:37.695Z","precision":"timestamp"}},"archivedUrl":{"status":"unavailable","reason":"No archived URL is recorded for this source."},"articleUse":{"status":"available","value":"Separate UK AISI evaluation and analysis of open- versus closed-weight cyber capability. It provides relevant independent context on benchmark scope, simulated cyber ranges, safeguards, open-weight deployment, agentic testing, and the limits of inferring real-world capability. It evaluates GLM-5.2 and DeepSeek V4-Pro, not GLM-5.3, so it is contextual rather than confirmation of the NIST result."}}]},"claims":{"status":"available","items":[{"ref":"C-2441CCE135BB","text":"CAISI reported that GLM-5.3's cyber capabilities were significantly below those of current U.S. frontier models and approximately four months behind them on CAISI's aggregate capability measure.","verificationStatus":"uncertain","material":true,"confidence":{"status":"available","value":0.91},"assertionRole":{"status":"available","value":"attributed_proposition"},"attributionEventClaimRef":{"status":"available","value":"C-8031A9C55E7C"},"sources":{"status":"available","items":[{"sourceRef":"S-3751E1F90D5D","relationship":"origin","note":{"status":"available","value":"NIST presents the four-month comparison as an aggregate measure based on CAISI cyber evaluations."}},{"sourceRef":"S-B547ED64FF3E","relationship":"context","note":{"status":"available","value":"The independent analysis reproduces the reported figures and characterizes the result as a wide gap, while noting that it is not a replication."}}]}},{"ref":"C-5CC3B717D1C5","text":"For the comparison, CAISI defined the U.S. frontier-best result on each benchmark as the highest score achieved by a model released by a U.S. entity that CAISI had evaluated; the comparison excluded models that had been developed but not released.","verificationStatus":"supported","material":true,"confidence":{"status":"available","value":0.99},"assertionRole":{"status":"available","value":"narrator_fact"},"attributionEventClaimRef":{"status":"available","value":null},"sources":{"status":"available","items":[{"sourceRef":"S-3751E1F90D5D","relationship":"supports","note":{"status":"available","value":"NIST states these scope definitions in its methodological notes."}}]}},{"ref":"C-8031A9C55E7C","text":"NIST's Center for AI Standards and Innovation published an assessment of Z.ai's GLM-5.3 cyber capabilities on September 17, 2026.","verificationStatus":"confirmed","material":true,"confidence":{"status":"available","value":0.99},"assertionRole":{"status":"available","value":"attribution_event"},"attributionEventClaimRef":{"status":"available","value":null},"sources":{"status":"available","items":[{"sourceRef":"S-3751E1F90D5D","relationship":"confirms","note":{"status":"available","value":"The NIST page is dated September 17, 2026 and is the primary record of the assessment."}}]}},{"ref":"C-90EDB4F37910","text":"The published assessment measures performance on specified cyber benchmarks and does not by itself establish performance on an arbitrary reader's codebase or direct transfer to real-world offensive operations.","verificationStatus":"supported","material":true,"confidence":{"status":"available","value":0.93},"assertionRole":{"status":"available","value":"narrator_fact"},"attributionEventClaimRef":{"status":"available","value":null},"sources":{"status":"available","items":[{"sourceRef":"S-3751E1F90D5D","relationship":"supports","note":{"status":"available","value":"The NIST source defines simulated benchmark tasks, model harnesses, and comparison scope rather than real-world operations."}},{"sourceRef":"S-B547ED64FF3E","relationship":"supports","note":{"status":"available","value":"The independent analysis explicitly distinguishes benchmark performance from performance on a reader's codebase or real-world operations."}}]}},{"ref":"C-A4B7BD008576","text":"CAISI tested the models as tool-using agents in a ReAct harness with bash and Python access, benchmark-specific turn limits, and maximum reasoning settings; U.S. models were tested with cyber safeguards disabled when applicable.","verificationStatus":"supported","material":true,"confidence":{"status":"available","value":0.99},"assertionRole":{"status":"available","value":"narrator_fact"},"attributionEventClaimRef":{"status":"available","value":null},"sources":{"status":"available","items":[{"sourceRef":"S-3751E1F90D5D","relationship":"supports","note":{"status":"available","value":"NIST's methodological notes specify the ReAct harness, tools, turn limits, reasoning settings, and safeguard conditions."}}]}},{"ref":"C-CA5733B3A7FE","text":"CAISI reported that GLM-5.3 was the most cyber-capable open-weight model released to date.","verificationStatus":"uncertain","material":true,"confidence":{"status":"available","value":0.92},"assertionRole":{"status":"available","value":"attributed_proposition"},"attributionEventClaimRef":{"status":"available","value":"C-8031A9C55E7C"},"sources":{"status":"available","items":[{"sourceRef":"S-3751E1F90D5D","relationship":"origin","note":{"status":"available","value":"NIST states this as a key finding of the CAISI assessment; the claim is limited to the models released and evaluated within CAISI's comparison."}},{"sourceRef":"S-B547ED64FF3E","relationship":"context","note":{"status":"available","value":"The independent analysis reproduces and discusses CAISI's conclusion but does not independently rerun the evaluation."}}]}},{"ref":"C-CF0A9A72AFB2","text":"The assessment evaluated GLM-5.3 on four benchmarks covering vulnerability discovery and exploit development.","verificationStatus":"supported","material":true,"confidence":{"status":"available","value":0.99},"assertionRole":{"status":"available","value":"narrator_fact"},"attributionEventClaimRef":{"status":"available","value":null},"sources":{"status":"available","items":[{"sourceRef":"S-3751E1F90D5D","relationship":"supports","note":{"status":"available","value":"NIST describes SEC-Bench Pro, ExploitBench, ExploitGym, and the CAISI OSS-Fuzz benchmark, including their vulnerability-discovery and exploit-development tasks."}}]}},{"ref":"C-D769D3936551","text":"The CAISI assessment table lists GLM-5.3 at 40.4% on SEC-Bench Pro, 61.1% on ExploitBench, 9.4% on ExploitGym, and 7.7% on CAISI OSS-Fuzz; the listed U.S. frontier-best scores were 90.2%, 100.0%, 44.4%, and 23.2%, respectively.","verificationStatus":"supported","material":true,"confidence":{"status":"available","value":0.99},"assertionRole":{"status":"available","value":"narrator_fact"},"attributionEventClaimRef":{"status":"available","value":null},"sources":{"status":"available","items":[{"sourceRef":"S-3751E1F90D5D","relationship":"supports","note":{"status":"available","value":"These figures and their 95% confidence intervals appear in NIST's benchmark-results table."}},{"sourceRef":"S-B547ED64FF3E","relationship":"supports","note":{"status":"available","value":"The independent analysis reproduces the same central benchmark figures."}}]}}]},"corrections":{"status":"verified-empty","items":[],"basis":"Authoritative corrections ledger checked for ER-A00026; no correction records are present."},"predictions":{"status":"verified-empty","items":[],"basis":"Authoritative predictions ledger checked for ER-A00026; no prediction records are present."}}},"provenance":{"underlyingModel":"gpt-5.6-luna","provider":"OpenAI","editorialSystem":"AXIOM / PREAPPROVAL-ARTICLE-REWORK-001","constitutionVersion":"1.0","sourceStandardVersion":"1.0","providerConflict":null},"sources":[{"title":"How Far Behind the Frontier are Leading Open Weight Models on Cyber?","publisher":"UK AI Security Institute","url":"https://www.aisi.gov.uk/blog/how-far-behind-the-frontier-are-leading-open-weight-models-on-cyber","role":"Separate UK AISI evaluation and analysis of open- versus closed-weight cyber capability. It provides relevant independent context on benchmark scope, simulated cyber ranges, safeguards, open-weight deployment, agentic testing, and the limits of inferring real-world capability. It evaluates GLM-5.2 and DeepSeek V4-Pro, not GLM-5.3, so it is contextual rather than confirmation of the NIST result."},{"title":"CAISI’s Assessment of Z.ai’s GLM-5.3 Cyber Capabilities","publisher":"National Institute of Standards and Technology","url":"https://www.nist.gov/news-events/news/2026/09/caisis-assessment-zais-glm-53-cyber-capabilities","role":"Direct primary record of the assessment. Reports that CAISI found GLM-5.3 the most cyber-capable open-weight model released to date, while about four months behind the U.S. frontier on CAISI’s aggregate measure. Provides four benchmark definitions, scores, 95% confidence intervals, agent-harness details, safeguard conditions, and scope limitations."},{"title":"Chinese AI trails US models at bug-finding, despite one narrow win","publisher":"P.K. Sharma","url":"https://www.pk-sharma.com/briefing/chinese-open-weight-models-bug-finding-match-claim","role":"Independent analysis that reviewed the September 17 CAISI assessment and reproduced its central GLM-5.3, U.S. frontier, and earlier PRC-best figures. It explicitly rejects an interpretation that the model matched the U.S. frontier and notes that the tests measure benchmark performance rather than performance on a reader’s codebase or real-world operations. It does not independently rerun the evaluation."}],"visuals":[{"id":"ER-V00002","visualClass":"documentary_image","placement":"lead","ordinal":1,"caption":"CAISI/NIST Figure 2 comparing GLM-5.3 with U.S. and PRC frontier-best results across four cyber benchmarks under CAISI's reported evaluation conditions. Benchmark results do not establish real-world operational capability. Credit: CAISI/NIST.","altText":"Bar chart comparing GLM-5.3 and frontier-best model success rates on SEC-Bench Pro, ExploitBench, ExploitGym, and CAISI OSS-Fuzz.","creator":"CAISI/NIST","sourcePublisher":"National Institute of Standards and Technology (NIST)","sourceUrl":"https://www.nist.gov/news-events/news/2026/09/caisis-assessment-zais-glm-53-cyber-capabilities","rightsStatus":"public_domain","rightsBasis":"Structured free-use clearance: PUBLIC-DOMAIN — https://www.nist.gov/copyrights-disclaimers","assetUrl":"https://emergencerecord.com/visuals/feda75d741fe2e2c37d25bca7048094fccffcc86d97c33f011f295551c61c468.png","contentSha256":"feda75d741fe2e2c37d25bca7048094fccffcc86d97c33f011f295551c61c468","dimensions":{"width":1999,"height":1665},"modified":false,"modificationNotes":null,"methodology":null,"evidenceNotes":"Documentary/context image. Rights clearance does not make the image independent proof of article claims.","generation":null,"attachmentProvenance":{"editorialReason":"The exact CAISI/NIST Figure 2 supports the Article’s discussion of reported cyber benchmark comparisons without implying broader real-world operational capability.","editorialIdentity":"ER06 — Visual Systems / Documentary Intake","provider":null,"model":null,"promptVersion":"VISUAL-SUPPORT-001","attachedAt":"2026-09-23T16:42:49.516932+00:00"}}]},{"id":"ER-A00029","version":2,"publicationStatus":"published","publishedDate":"2026-09-23","canonicalUrl":"https://emergencerecord.com/articles/ai-evaluator-forum-embedded-evaluation-conditions","slug":"ai-evaluator-forum-embedded-evaluation-conditions","headline":"AI Evaluator Forum Publishes Letter Proposing Conditions for Embedded Evaluations","subheadline":"The public letter identifies more than 100 signatories and recommends independence, privileged access, transparency and protection from retaliation.","section":"Governance","editorialIdentity":"FRONTIER","articleType":"News","contemporarySignificance":"70 / 100","historicalSignificance":null,"humanApproval":"Approved 2026-09-22 at 19:51:51 UTC","placement":{"position":5,"role":"standard"},"evidence":{"schema":"The Emergence Record / Public Evidence Contract 1.1","publicationStatus":"published","article":{"id":"ER-A00029","versionNumber":2,"canonicalUrl":{"status":"available","value":"https://emergencerecord.com/articles/ai-evaluator-forum-embedded-evaluation-conditions"},"publishedAt":{"status":"available","value":{"value":"2026-09-25T16:56:20.427Z","precision":"timestamp"}}},"approval":{"approved":true,"approvedAt":{"status":"available","value":{"value":"2026-09-22T19:51:51.000Z","precision":"timestamp"}},"approverIdentity":{"status":"available","value":"Zechariah Hilles"},"approvalReference":{"status":"available","value":"approval:ER-A00029:v2"}},"provenance":{"model":{"status":"available","name":"gpt-5.6-luna","provider":{"status":"available","value":"OpenAI"}},"editorialSystem":{"status":"available","value":"AXIOM / PREAPPROVAL-ARTICLE-REWORK-001"},"constitutionVersion":{"status":"available","value":"1.0"},"sourceStandardVersion":{"status":"unknown","reason":"Source Standard version is not stored as an Article Version field in the authoritative ledger."},"promptVersion":{"status":"available","value":"AXIOM-REVISION-AUTO-004"},"providerConflict":{"status":"unknown","reason":"No normalized provider-conflict assessment is stored for this Article Version."}},"evidence":{"sources":{"status":"available","items":[{"ref":"S-0B10385B1288","title":"Anthropic and OpenAI want to embed safety evaluators. Will they really be independent?","publisher":{"status":"available","value":"TechCrunch"},"url":{"status":"available","value":"https://techcrunch.com/2026/09/16/anthropic-and-openai-want-to-embed-safety-evaluators-will-they-really-be-independent/"},"sourceClass":{"status":"available","value":"secondary"},"author":{"status":"unavailable","reason":"No individual author is recorded for this source."},"publishedAt":{"status":"available","value":{"value":"2026-09-16","precision":"date"}},"accessedAt":{"status":"available","value":{"value":"2026-09-22T03:22:39.654Z","precision":"timestamp"}},"archivedUrl":{"status":"unavailable","reason":"No archived URL is recorded for this source."},"articleUse":{"status":"available","value":"Independent reporting published before the letter. Directly establishes the surrounding debate: Anthropic and OpenAI had proposed embedded evaluators with deep or employee-like access, while evaluators raised concerns about control over methods, access, confidentiality, publication, and independence. It also reports that companies had not specified which evaluators they would use or the exact access and disclosure terms."}},{"ref":"S-5833BDF3B136","title":"AI Evaluator Forum","publisher":{"status":"available","value":"AI Evaluator Forum"},"url":{"status":"available","value":"https://aievaluatorforum.org/"},"sourceClass":{"status":"available","value":"primary"},"author":{"status":"unavailable","reason":"No individual author is recorded for this source."},"publishedAt":{"status":"unavailable","reason":"No source publication date is recorded in the authoritative source row."},"accessedAt":{"status":"available","value":{"value":"2026-09-22T03:22:39.654Z","precision":"timestamp"}},"archivedUrl":{"status":"unavailable","reason":"No archived URL is recorded for this source."},"articleUse":{"status":"available","value":"Reviewed via the Forum’s public organizational page. Describes the Forum as a network of organizations conducting independent, third-party AI evaluations and expressly states that it is not a legal entity in its own right and relies on voluntary participation. Useful for delimiting the Forum’s institutional standing."}},{"ref":"S-B3BDBE1B86CC","title":"AI experts warn safety evaluators lack independence and protections","publisher":{"status":"available","value":"Quartz"},"url":{"status":"available","value":"https://qz.com/ai-experts-safety-evaluators-independence-protections-091826"},"sourceClass":{"status":"available","value":"secondary"},"author":{"status":"unavailable","reason":"No individual author is recorded for this source."},"publishedAt":{"status":"available","value":{"value":"2026-09-18","precision":"date"}},"accessedAt":{"status":"available","value":{"value":"2026-09-22T03:22:39.654Z","precision":"timestamp"}},"archivedUrl":{"status":"unavailable","reason":"No archived URL is recorded for this source."},"articleUse":{"status":"available","value":"Reviewed as a secondary account of the letter and its demands. It reports the 100-plus signatories and summarizes the independence, access, transparency, and retaliation provisions, but attributes much of the reporting and quotations to CNBC. It therefore provides corroborative context but should not be treated as fully independent confirmation of the complete source record."}},{"ref":"S-D1A9FE6D56C9","title":"Minimum Conditions for Embedding Evaluators","publisher":{"status":"available","value":"AI Evaluator Forum"},"url":{"status":"available","value":"https://aievaluatorforum.org/initiatives/embedded-evaluation-letter"},"sourceClass":{"status":"available","value":"primary"},"author":{"status":"unavailable","reason":"No individual author is recorded for this source."},"publishedAt":{"status":"available","value":{"value":"2026-09-18","precision":"date"}},"accessedAt":{"status":"available","value":{"value":"2026-09-22T03:22:39.654Z","precision":"timestamp"}},"archivedUrl":{"status":"unavailable","reason":"No archived URL is recorded for this source."},"articleUse":{"status":"available","value":"Directly reviewed. States that the letter was published September 18, 2026; labels it a public letter with 100+ signatories; sets out requirements concerning independence, multiple viewpoints, transparency, protection from retaliation, and access comparable to highly privileged employees. The page states that signatories act in their personal capacity and that affiliations are for identification only. It also says the list is non-comprehensive and that embedded evaluations complement rather than replace broader oversight."}}]},"claims":{"status":"available","items":[{"ref":"C-269FDD4BCBB7","text":"The Forum's letter recommends transparency about evaluation methods, findings, access, and terms; prompt and unfiltered communication with company boards or other privileged oversight bodies; and public release of findings and evidence subject to limited, time-limited redactions.","verificationStatus":"unverified","material":true,"confidence":{"status":"available","value":0.99},"assertionRole":{"status":"available","value":"attributed_proposition"},"attributionEventClaimRef":{"status":"available","value":"C-2E065247A131"},"sources":{"status":"available","items":[{"sourceRef":"S-D1A9FE6D56C9","relationship":"origin","note":{"status":"available","value":"These provisions appear in condition 3 of the letter. They describe the Forum's proposed framework, not an adopted rule or existing practice."}}]}},{"ref":"C-2B50898221A8","text":"The Forum's letter says embedded evaluations should complement, rather than replace, broader external oversight, public transparency, and additional forms of access for independent researchers.","verificationStatus":"unverified","material":true,"confidence":{"status":"available","value":0.99},"assertionRole":{"status":"available","value":"attributed_proposition"},"attributionEventClaimRef":{"status":"available","value":"C-2E065247A131"},"sources":{"status":"available","items":[{"sourceRef":"S-D1A9FE6D56C9","relationship":"origin","note":{"status":"available","value":"This is the Forum's stated framing in the letter; it is not an independently established description of the broader governance landscape."}}]}},{"ref":"C-2E065247A131","text":"On September 18, 2026, the AI Evaluator Forum published a public letter titled “Minimum Conditions for Embedding Evaluators”; its page identifies the letter as having 100+ signatories.","verificationStatus":"confirmed","material":true,"confidence":{"status":"available","value":0.99},"assertionRole":{"status":"available","value":"attribution_event"},"attributionEventClaimRef":{"status":"available","value":null},"sources":{"status":"available","items":[{"sourceRef":"S-B3BDBE1B86CC","relationship":"supports","note":{"status":"available","value":"Quartz independently reports that a coalition organized by the AI Evaluator Forum published the letter and describes it as involving more than 100 AI experts."}},{"sourceRef":"S-D1A9FE6D56C9","relationship":"confirms","note":{"status":"available","value":"The primary page is titled “Public Letter,” gives the publication date as September 18, 2026, and states “100+ Signatories.”"}}]}},{"ref":"C-6F54332235E3","text":"The AI Evaluator Forum describes itself as a network of organizations conducting independent, third-party AI evaluations, not a legal entity in its own right, and says it relies on voluntary participation.","verificationStatus":"confirmed","material":true,"confidence":{"status":"available","value":0.99},"assertionRole":{"status":"available","value":"narrator_fact"},"attributionEventClaimRef":{"status":"available","value":null},"sources":{"status":"available","items":[{"sourceRef":"S-5833BDF3B136","relationship":"confirms","note":{"status":"available","value":"The Forum's organizational page expressly states these institutional limitations."}}]}},{"ref":"C-84EBEDD28CFC","text":"The Forum's letter recommends that embedded third-party evaluations have scientific objectivity, transparency, independence, and robust protections against interference by the companies being evaluated.","verificationStatus":"unverified","material":true,"confidence":{"status":"available","value":0.99},"assertionRole":{"status":"available","value":"attributed_proposition"},"attributionEventClaimRef":{"status":"available","value":"C-2E065247A131"},"sources":{"status":"available","items":[{"sourceRef":"S-D1A9FE6D56C9","relationship":"origin","note":{"status":"available","value":"The letter states these as proposed minimum conditions. The evidence establishes that the Forum made the recommendation, not that the conditions currently exist or are sufficient."}}]}},{"ref":"C-A9191EAD009E","text":"The Forum's letter recommends protection from retaliation, including retaliatory litigation, and funding arrangements that would allow evaluators to remain funded when their methods or findings are unfavorable to the evaluated companies.","verificationStatus":"unverified","material":true,"confidence":{"status":"available","value":0.99},"assertionRole":{"status":"available","value":"attributed_proposition"},"attributionEventClaimRef":{"status":"available","value":"C-2E065247A131"},"sources":{"status":"available","items":[{"sourceRef":"S-D1A9FE6D56C9","relationship":"origin","note":{"status":"available","value":"These provisions appear in condition 4 of the letter and remain recommendations attributed to the Forum."}}]}},{"ref":"C-BE38374328B4","text":"The Forum's letter recommends that embedded evaluators receive access equivalent to that of highly privileged employees for evaluation purposes, including relevant systems, data, tools, physical spaces, and candid communication with relevant staff, subject to protections for sensitive third-party data.","verificationStatus":"unverified","material":true,"confidence":{"status":"available","value":0.99},"assertionRole":{"status":"available","value":"attributed_proposition"},"attributionEventClaimRef":{"status":"available","value":"C-2E065247A131"},"sources":{"status":"available","items":[{"sourceRef":"S-D1A9FE6D56C9","relationship":"origin","note":{"status":"available","value":"This recommendation appears in condition 5 of the letter and does not establish that companies currently provide such access."}}]}},{"ref":"C-C103709CBC6E","text":"The letter states that all signatories signed in their personal capacity and that listed affiliations are for identification only.","verificationStatus":"confirmed","material":true,"confidence":{"status":"available","value":0.99},"assertionRole":{"status":"available","value":"narrator_fact"},"attributionEventClaimRef":{"status":"available","value":null},"sources":{"status":"available","items":[{"sourceRef":"S-D1A9FE6D56C9","relationship":"confirms","note":{"status":"available","value":"The disclaimer appears immediately before the signatory list."}}]}},{"ref":"C-FA4AEBA21412","text":"The Forum's letter recommends that evaluators retain editorial control, disclose and mitigate conflicts of interest, avoid ownership or governance by the evaluated companies, avoid significant commercial relationships with them, and receive no payment or reward contingent on findings; it also recommends incorporating multiple viewpoints and evaluation organizations.","verificationStatus":"unverified","material":true,"confidence":{"status":"available","value":0.99},"assertionRole":{"status":"available","value":"attributed_proposition"},"attributionEventClaimRef":{"status":"available","value":"C-2E065247A131"},"sources":{"status":"available","items":[{"sourceRef":"S-D1A9FE6D56C9","relationship":"origin","note":{"status":"available","value":"These are recommendations in conditions 1 and 2 of the letter, not established facts about current evaluation arrangements."}}]}}]},"corrections":{"status":"verified-empty","items":[],"basis":"Authoritative corrections ledger checked for ER-A00029; no correction records are present."},"predictions":{"status":"verified-empty","items":[],"basis":"Authoritative predictions ledger checked for ER-A00029; no prediction records are present."}}},"provenance":{"underlyingModel":"gpt-5.6-luna","provider":"OpenAI","editorialSystem":"AXIOM / PREAPPROVAL-ARTICLE-REWORK-001","constitutionVersion":"1.0","sourceStandardVersion":"1.0","providerConflict":null},"sources":[{"title":"AI Evaluator Forum","publisher":"AI Evaluator Forum","url":"https://aievaluatorforum.org/","role":"Reviewed via the Forum’s public organizational page. Describes the Forum as a network of organizations conducting independent, third-party AI evaluations and expressly states that it is not a legal entity in its own right and relies on voluntary participation. Useful for delimiting the Forum’s institutional standing."},{"title":"Minimum Conditions for Embedding Evaluators","publisher":"AI Evaluator Forum","url":"https://aievaluatorforum.org/initiatives/embedded-evaluation-letter","role":"Directly reviewed. States that the letter was published September 18, 2026; labels it a public letter with 100+ signatories; sets out requirements concerning independence, multiple viewpoints, transparency, protection from retaliation, and access comparable to highly privileged employees. The page states that signatories act in their personal capacity and that affiliations are for identification only. It also says the list is non-comprehensive and that embedded evaluations complement rather than replace broader oversight."},{"title":"AI experts warn safety evaluators lack independence and protections","publisher":"Quartz","url":"https://qz.com/ai-experts-safety-evaluators-independence-protections-091826","role":"Reviewed as a secondary account of the letter and its demands. It reports the 100-plus signatories and summarizes the independence, access, transparency, and retaliation provisions, but attributes much of the reporting and quotations to CNBC. It therefore provides corroborative context but should not be treated as fully independent confirmation of the complete source record."},{"title":"Anthropic and OpenAI want to embed safety evaluators. Will they really be independent?","publisher":"TechCrunch","url":"https://techcrunch.com/2026/09/16/anthropic-and-openai-want-to-embed-safety-evaluators-will-they-really-be-independent/","role":"Independent reporting published before the letter. Directly establishes the surrounding debate: Anthropic and OpenAI had proposed embedded evaluators with deep or employee-like access, while evaluators raised concerns about control over methods, access, confidentiality, publication, and independence. It also reports that companies had not specified which evaluators they would use or the exact access and disclosure terms."}],"visuals":[]},{"id":"ER-A00022","version":1,"publicationStatus":"published","publishedDate":"2026-09-23","canonicalUrl":"https://emergencerecord.com/articles/mit-researchers-hardflow-constrained-generative-ai","slug":"mit-researchers-hardflow-constrained-generative-ai","headline":"MIT News Reports MIT Researchers Developed HardFlow for Constrained Generative AI","subheadline":"Model Current said the reported results were limited to tested settings; the reviewed evidence does not establish a broad real-world safety guarantee.","section":"The Frontier","editorialIdentity":"FRONTIER","articleType":"News","contemporarySignificance":"67 / 100","historicalSignificance":null,"humanApproval":"Approved 2026-09-16 at 13:32:14 UTC","placement":{"position":6,"role":"standard"},"evidence":{"schema":"The Emergence Record / Public Evidence Contract 1.1","publicationStatus":"published","article":{"id":"ER-A00022","versionNumber":1,"canonicalUrl":{"status":"available","value":"https://emergencerecord.com/articles/mit-researchers-hardflow-constrained-generative-ai"},"publishedAt":{"status":"available","value":{"value":"2026-09-25T16:56:20.427Z","precision":"timestamp"}}},"approval":{"approved":true,"approvedAt":{"status":"available","value":{"value":"2026-09-16T13:32:14.000Z","precision":"timestamp"}},"approverIdentity":{"status":"available","value":"Zechariah Hilles"},"approvalReference":{"status":"available","value":"approval:ER-A00022:v1"}},"provenance":{"model":{"status":"available","name":"gpt-5.6-luna","provider":{"status":"available","value":"OpenAI"}},"editorialSystem":{"status":"available","value":"AXIOM / AXIOM Newsroom Pipeline v1"},"constitutionVersion":{"status":"available","value":"1.0"},"sourceStandardVersion":{"status":"unknown","reason":"Source Standard version is not stored as an Article Version field in the authoritative ledger."},"promptVersion":{"status":"available","value":"AXIOM-REVISION-AUTO-004"},"providerConflict":{"status":"unknown","reason":"No normalized provider-conflict assessment is stored for this Article Version."}},"evidence":{"sources":{"status":"available","items":[{"ref":"S-2898A53D0439","title":"New method enables AI for safety-critical situations","publisher":{"status":"available","value":"MIT News"},"url":{"status":"available","value":"https://news.mit.edu/2026/new-method-enables-ai-safety-critical-situations-0914"},"sourceClass":{"status":"available","value":"academic"},"author":{"status":"unavailable","reason":"No individual author is recorded for this source."},"publishedAt":{"status":"available","value":{"value":"2026-09-14","precision":"date"}},"accessedAt":{"status":"available","value":{"value":"2026-09-16T12:44:56.959Z","precision":"timestamp"}},"archivedUrl":{"status":"unavailable","reason":"No archived URL is recorded for this source."},"articleUse":{"status":"available","value":"Reviewed official MIT announcement. Supports the announcement date, researcher identities, method description, claimed deployment-time use without retraining, application areas, and MIT-reported experimental results. This is an author-affiliated institutional account and should not be treated as independent confirmation of the results."}},{"ref":"S-62583E44B01F","title":"HardFlow: Hard-Constrained Sampling for Flow-Matching Models via Trajectory Optimization","publisher":{"status":"available","value":"arXiv"},"url":{"status":"available","value":"https://arxiv.org/abs/2511.08425"},"sourceClass":{"status":"available","value":"academic"},"author":{"status":"unavailable","reason":"No individual author is recorded for this source."},"publishedAt":{"status":"available","value":{"value":"2025-11-11","precision":"date"}},"accessedAt":{"status":"available","value":{"value":"2026-09-16T12:44:56.959Z","precision":"timestamp"}},"archivedUrl":{"status":"unavailable","reason":"No archived URL is recorded for this source."},"articleUse":{"status":"available","value":"Reviewed author manuscript and abstract. Supports the paper title, authors, technical formulation, stated terminal-constraint guarantee under explicit feasibility assumptions, theoretical analysis, and reported experiments across robotics, physical-process control, and vision. It is a primary research record, not independent validation."}},{"ref":"S-A600197D300A","title":"MIT HardFlow: hard constraints for AI-generated plans","publisher":{"status":"available","value":"Model Current"},"url":{"status":"available","value":"https://modelcurrent.com/article/mit-hardflow-generated-plans-hard-constraints"},"sourceClass":{"status":"available","value":"secondary"},"author":{"status":"unavailable","reason":"No individual author is recorded for this source."},"publishedAt":{"status":"available","value":{"value":"2026-09-15","precision":"date"}},"accessedAt":{"status":"available","value":{"value":"2026-09-16T12:44:56.959Z","precision":"timestamp"}},"archivedUrl":{"status":"unavailable","reason":"No archived URL is recorded for this source."},"articleUse":{"status":"available","value":"Reviewed independent analysis. It distinguishes the authors’ controlled experimental claims from broader safety claims, states that broader real-world testing remains outstanding, and notes that the outlet did not rerun the benchmarks. It largely relies on the MIT announcement and paper for underlying results, so it is independent contextual analysis rather than independent replication."}},{"ref":"S-F76CC1FCF08C","title":"Navid Azizan","publisher":{"status":"available","value":"MIT Azizan Lab"},"url":{"status":"available","value":"https://azizan.mit.edu/"},"sourceClass":{"status":"available","value":"academic"},"author":{"status":"unavailable","reason":"No individual author is recorded for this source."},"publishedAt":{"status":"unavailable","reason":"No source publication date is recorded in the authoritative source row."},"accessedAt":{"status":"available","value":{"value":"2026-09-16T12:44:56.959Z","precision":"timestamp"}},"archivedUrl":{"status":"unavailable","reason":"No archived URL is recorded for this source."},"articleUse":{"status":"available","value":"Reviewed author laboratory publication listing. Corroborates that HardFlow is listed by the senior author as a 2026 IEEE Transactions on Pattern Analysis and Machine Intelligence publication, but does not independently establish the journal issue, DOI, peer-review history, or experimental validity."}}]},"claims":{"status":"available","items":[{"ref":"C-0F8B6D6061DE","text":"The paper formulates hard-constrained sampling as a trajectory-optimization problem designed to enforce constraints at the terminal output while allowing additional objectives related to sample quality and distribution shift.","verificationStatus":"supported","material":true,"confidence":{"status":"available","value":0.9},"assertionRole":{"status":"available","value":"narrator_fact"},"attributionEventClaimRef":{"status":"available","value":null},"sources":{"status":"available","items":[{"sourceRef":"S-62583E44B01F","relationship":"supports","note":{"status":"available","value":"The abstract describes terminal-time constraint enforcement, trajectory optimization, approximation analysis, and additional objectives."}}]}},{"ref":"C-256DEAAFB01E","text":"MIT News published an article dated September 14, 2026 describing HardFlow as a method intended to make generative-model outputs satisfy hard constraints.","verificationStatus":"confirmed","material":true,"confidence":{"status":"available","value":0.99},"assertionRole":{"status":"available","value":"narrator_fact"},"attributionEventClaimRef":{"status":"available","value":null},"sources":{"status":"available","items":[{"sourceRef":"S-2898A53D0439","relationship":"confirms","note":{"status":"available","value":"The MIT News page is dated September 14, 2026 and identifies HardFlow and its stated purpose."}}]}},{"ref":"C-4401CC4ED60B","text":"The reviewed evidence does not establish that HardFlow makes deployed AI systems safe or guarantees constraint satisfaction in arbitrary real-world conditions.","verificationStatus":"supported","material":true,"confidence":{"status":"available","value":0.93},"assertionRole":{"status":"available","value":"narrator_fact"},"attributionEventClaimRef":{"status":"available","value":null},"sources":{"status":"available","items":[{"sourceRef":"S-62583E44B01F","relationship":"supports","note":{"status":"available","value":"The paper describes results and analysis for its framework and tested domains, not a universal real-world safety guarantee."}},{"sourceRef":"S-A600197D300A","relationship":"supports","note":{"status":"available","value":"The independent analysis says the findings are not a general safety guarantee and identifies unresolved real-world evaluation questions."}}]}},{"ref":"C-47E25B6046CC","text":"MIT says HardFlow can be applied at deployment time to pretrained generative models without retraining them.","verificationStatus":"uncertain","material":true,"confidence":{"status":"available","value":0.76},"assertionRole":{"status":"available","value":"attributed_proposition"},"attributionEventClaimRef":{"status":"available","value":"C-F879D4B1E236"},"sources":{"status":"available","items":[{"sourceRef":"S-2898A53D0439","relationship":"supports","note":{"status":"available","value":"MIT News describes the technique as operating at deployment time without retraining the pretrained model."}},{"sourceRef":"S-62583E44B01F","relationship":"context","note":{"status":"available","value":"The paper describes an inference-time constrained-sampling framework, but the abstract alone does not establish all practical deployment implications."}}]}},{"ref":"C-90C55906B267","text":"Model Current published an analysis on September 15, 2026 stating that it had not rerun the HardFlow benchmarks and that broader real-world testing remained outstanding.","verificationStatus":"confirmed","material":true,"confidence":{"status":"available","value":0.98},"assertionRole":{"status":"available","value":"attribution_event"},"attributionEventClaimRef":{"status":"available","value":null},"sources":{"status":"available","items":[{"sourceRef":"S-A600197D300A","relationship":"confirms","note":{"status":"available","value":"The analysis explicitly says Model Current did not rerun the benchmarks and that broader real-world testing remains to be done."}}]}},{"ref":"C-B4CECD0B2B87","text":"In the reported experiments, HardFlow satisfied the required constraints while finding better-quality solutions than existing techniques.","verificationStatus":"uncertain","material":true,"confidence":{"status":"available","value":0.68},"assertionRole":{"status":"available","value":"attributed_proposition"},"attributionEventClaimRef":{"status":"available","value":"C-F879D4B1E236"},"sources":{"status":"available","items":[{"sourceRef":"S-2898A53D0439","relationship":"supports","note":{"status":"available","value":"MIT News reports consistent constraint satisfaction and better solution quality, but this is an author-affiliated account rather than independent validation."}},{"sourceRef":"S-62583E44B01F","relationship":"origin","note":{"status":"available","value":"The author manuscript's abstract reports experiments across robotics, physical-process control, and vision and claims substantial improvement over existing methods."}}]}},{"ref":"C-BB604DE072E1","text":"The reported results support conclusions about the tested settings, but do not by themselves establish performance across every robot, workspace, unpredictable interaction, or real-world deployment condition.","verificationStatus":"supported","material":true,"confidence":{"status":"available","value":0.9},"assertionRole":{"status":"available","value":"attributed_proposition"},"attributionEventClaimRef":{"status":"available","value":"C-90C55906B267"},"sources":{"status":"available","items":[{"sourceRef":"S-A600197D300A","relationship":"supports","note":{"status":"available","value":"Model Current expressly limits the implications of the results to the tested settings and identifies broader evaluation as unresolved."}}]}},{"ref":"C-D83A2A32DA97","text":"Navid Azizan’s MIT laboratory page lists HardFlow as a 2026 publication in IEEE Transactions on Pattern Analysis and Machine Intelligence.","verificationStatus":"supported","material":true,"confidence":{"status":"available","value":0.88},"assertionRole":{"status":"available","value":"narrator_fact"},"attributionEventClaimRef":{"status":"available","value":null},"sources":{"status":"available","items":[{"sourceRef":"S-F76CC1FCF08C","relationship":"supports","note":{"status":"available","value":"The author’s MIT laboratory page lists HardFlow with IEEE Transactions on Pattern Analysis and Machine Intelligence and the year 2026; it does not provide a DOI or issue record."}}]}},{"ref":"C-F33A199A8300","text":"The arXiv record identifies a paper titled “HardFlow: Hard-Constrained Sampling for Flow-Matching Models via Trajectory Optimization” by Zeyang Li, Kaveh Alim, and Navid Azizan; it was submitted on November 11, 2025 and revised on April 26, 2026.","verificationStatus":"confirmed","material":true,"confidence":{"status":"available","value":0.99},"assertionRole":{"status":"available","value":"narrator_fact"},"attributionEventClaimRef":{"status":"available","value":null},"sources":{"status":"available","items":[{"sourceRef":"S-62583E44B01F","relationship":"confirms","note":{"status":"available","value":"The paper record lists the title, authors, submission history, and current revision."}}]}},{"ref":"C-F879D4B1E236","text":"MIT News reported that MIT researchers developed HardFlow and attributed to the researchers experimental claims involving robotics, physical-process control, and computer vision.","verificationStatus":"confirmed","material":true,"confidence":{"status":"available","value":0.98},"assertionRole":{"status":"available","value":"attribution_event"},"attributionEventClaimRef":{"status":"available","value":null},"sources":{"status":"available","items":[{"sourceRef":"S-2898A53D0439","relationship":"confirms","note":{"status":"available","value":"The article identifies the researchers, describes HardFlow, and reports the stated experimental domains and findings."}}]}}]},"corrections":{"status":"verified-empty","items":[],"basis":"Authoritative corrections ledger checked for ER-A00022; no correction records are present."},"predictions":{"status":"verified-empty","items":[],"basis":"Authoritative predictions ledger checked for ER-A00022; no prediction records are present."}}},"provenance":{"underlyingModel":"gpt-5.6-luna","provider":"Vercel AI Gateway / OpenAI","editorialSystem":"AXIOM / AXIOM Newsroom Pipeline v1","constitutionVersion":"1.0","sourceStandardVersion":"1.0","providerConflict":null},"sources":[{"title":"HardFlow: Hard-Constrained Sampling for Flow-Matching Models via Trajectory Optimization","publisher":"arXiv","url":"https://arxiv.org/abs/2511.08425","role":"Reviewed author manuscript and abstract. Supports the paper title, authors, technical formulation, stated terminal-constraint guarantee under explicit feasibility assumptions, theoretical analysis, and reported experiments across robotics, physical-process control, and vision. It is a primary research record, not independent validation."},{"title":"Navid Azizan","publisher":"MIT Azizan Lab","url":"https://azizan.mit.edu/","role":"Reviewed author laboratory publication listing. Corroborates that HardFlow is listed by the senior author as a 2026 IEEE Transactions on Pattern Analysis and Machine Intelligence publication, but does not independently establish the journal issue, DOI, peer-review history, or experimental validity."},{"title":"MIT HardFlow: hard constraints for AI-generated plans","publisher":"Model Current","url":"https://modelcurrent.com/article/mit-hardflow-generated-plans-hard-constraints","role":"Reviewed independent analysis. It distinguishes the authors’ controlled experimental claims from broader safety claims, states that broader real-world testing remains outstanding, and notes that the outlet did not rerun the benchmarks. It largely relies on the MIT announcement and paper for underlying results, so it is independent contextual analysis rather than independent replication."},{"title":"New method enables AI for safety-critical situations","publisher":"MIT News","url":"https://news.mit.edu/2026/new-method-enables-ai-safety-critical-situations-0914","role":"Reviewed official MIT announcement. Supports the announcement date, researcher identities, method description, claimed deployment-time use without retraining, application areas, and MIT-reported experimental results. This is an author-affiliated institutional account and should not be treated as independent confirmation of the results."}],"visuals":[]}]}