Evidence and limits share the same surface.

Bounded finding. Supports treating biosecurity as a lifecycle risk-management problem with consequence-driven assessment, institutional governance, and national oversight; the scope expressly includes cybersecurity, information security, molecular techniques, and artificial intelligence as emerging considerations.

Exact locator. Overview, especially the paragraph beginning “The WHO laboratory biosecurity guidance” and the listed key elements.

Limit. Normative guidance is not binding law and does not quantify marginal risk from any model, establish an AI release threshold, or show that a listed control works in every laboratory context.

Open the recorded source URL · retrieved 2026-07-28T15:03:47Z

Bounded finding. Establishes that the executive branch directed a pause or funding restrictions for specified categories of research and ordered replacement of the 2024 dual-use and nucleic-acid screening policies on stated timelines.

Exact locator. Sections 3 through 7, including the directives to the OSTP Director and covered funding agencies.

Limit. The order operates through presidential and agency authority; it is not a generally applicable criminal statute, does not itself supply the final replacement frameworks, and does not resolve constitutional or administrative-law questions about later implementation.

Open the recorded source URL · retrieved 2026-07-28T15:00:11Z

Bounded finding. Shows the implementing institution’s public position that the May 2025 executive order superseded implementation of the May 2024 DURC/PEPP policy and contemplated a new policy.

Exact locator. Opening notice and the paragraph stating that the new policy is intended to replace the May 2024 policy and supersedes its implementation.

Limit. This status notice is not the text of a replacement policy. The current tranche did not locate a final replacement framework that closes the implementation gap.

Open the recorded source URL · retrieved 2026-07-28T15:00:18Z

Bounded finding. Supplies the current statutory prohibition and the statutory boundary preserving prophylactic, protective, bona fide research, and other peaceful purposes.

Exact locator. Subsections (a), (b), and (c), especially the peaceful-purpose language in subsections (b) and (c).

Limit. The provision does not create an AI-model release rule, and applying its intent, possession, quantity, jurisdiction, and peaceful-purpose elements to a specific matter would require fact-specific legal analysis.

Open the recorded source URL · retrieved 2026-07-28T15:00:19Z

Bounded finding. Establishes the cutoff-date status: the seventh negotiating meeting advanced but did not finalize the PABS annex, and continued negotiations were required before countries could consider signature and ratification of the Pandemic Agreement.

Exact locator. Paragraphs beginning “WHO Member States have advanced negotiations” and “The meeting concluded,” plus the announced eighth meeting.

Limit. This is a negotiation-status release, not an operative treaty text. It cannot support a claim that the Pandemic Agreement was open for signature, ratified, or in force at the evidence cutoff.

Open the recorded source URL · retrieved 2026-07-28T15:00:22Z

Bounded finding. Supports a bounded finding that the two tested open-weight models performed comparably to selected closed models released four to seven months earlier on AISI’s test setup, while public weights reduce provider-controlled monitoring, access control, and withdrawal options.

Exact locator. Introduction; “Results: the current cyber gap for open weight AI”; “Real-world constraints”; and “Limitations to our testing.”

Limit. The evaluated models and tasks are a narrow sample; simulated ranges omit important real-world defenses; the comparison is not predictive; price observations are time-sensitive; and no inference to biological or other capabilities is warranted.

Open the recorded source URL · retrieved 2026-07-28T15:00:23Z

Bounded finding. Provides the existing federal legal baseline for specified unauthorized access, fraud, damage, extortion-related threats, and related conduct involving protected computers.

Exact locator. Subsection (a), definitions in subsection (e), and the penalty and civil-action provisions.

Limit. The statute does not regulate publication of model weights as such; authorization, loss, intent, jurisdiction, and remedy questions are fact-dependent and affected by case law not exhausted in this tranche.

Open the recorded source URL · retrieved 2026-07-28T15:00:24Z

Bounded finding. Supports treating cyber-enabled finance as a policy-level detection and enforcement layer through risk-focused supervision, closing regulatory gaps, partnerships, and responsible technology use.

Exact locator. Executive Summary; “Goals, Priorities, and Supporting Actions”; and the cybercrime and ransomware material in the threat and vulnerability annexes.

Limit. A strategy states priorities rather than proving implementation or outcomes, predates later 2025–2026 developments, and does not show that financial controls can prevent or reverse model release.

Open the recorded source URL · retrieved 2026-07-28T15:02:27Z

Bounded finding. Documents uneven implementation, continuing licensing and registration gaps, difficulty identifying providers, offshore-service challenges, and the importance of cross-border cooperation, freezing, and seizure.

Exact locator. Publication summary paragraphs following “Paris, 26 June 2025,” especially the implementation, offshore VASP, Travel Rule, and international-cooperation findings.

Limit. FATF standards depend on domestic implementation and are not self-executing global law; some figures rely on industry contributors; the update does not establish a causal link between open-weight release and illicit-finance volume.

Open the recorded source URL · retrieved 2026-07-28T15:03:48Z

Bounded finding. Supports only the bounded evaluation finding that the tested hosted Kimi K3 setup performed below the leading tested U.S. closed systems and above GLM-5.2 on the selected preliminary evaluations. It does not establish that the hosted setup was a byte-identical copy of the pinned Hugging Face revision.

Exact locator. Opening findings; “Detailed Results”; “Cyber Capability Trends”; and the simulated-range limitations near the end of the official AISI page, corroborated by the matching official CAISI/NIST publication.

Limit. The evaluation is preliminary and selective; Kimi’s aggregate estimate used one benchmark and therefore has a wider confidence interval; safeguards were disabled for the U.S. closed-system comparators; and the simulated range assumed initial access, lacked active defenders and alert penalties, and contained an intentional path. The source does not connect the hosted tested object to a byte-identical copy of the pinned Hugging Face revision or predict real-world outcomes.

Open the recorded source URL · retrieved 2026-07-28T17:34:07Z

Bounded finding. Documents partnership terms and potential competition concerns involving access to compute and talent, switching costs, exclusivity or control rights, and access to sensitive technical and business information.

Exact locator. FTC report PDF pages 5–6 (printed pages 2–3), “Summary of Findings”; detailed sections 5.1, 5.2, and 5.3 begin on PDF pages 33, 35, and 37.

Limit. A Section 6(b) staff study is not an adjudication, does not establish an antitrust violation, covers selected partnerships, and predates later market changes.

Open the recorded source URL · retrieved 2026-07-28T17:34:32Z

Bounded finding. Supplies an authoritative scientific-openness framework grounded in reproducibility, scrutiny, collaboration, academic freedom, inclusion, and equitable capacity, while recognizing proportionate restrictions for security, privacy, rights, and other protected interests.

Exact locator. Preamble; paragraphs 1–9; core values and principles in paragraphs 13–14; and the action areas beginning at paragraph 15.

Limit. The recommendation is not binding treaty law, addresses open science broadly rather than model weights specifically, and its stated benefits and values do not prove that every release choice advances equity or safety.

Open the recorded source URL · retrieved 2026-07-28T15:00:57Z

Bounded finding. Documents Member State commitments to close digital divides, build developing-country capacity, support safe and secure digital public goods including open AI models, and expand access to models, data, compute, skills, and locally relevant systems.

Exact locator. Digital public goods section; Objectives 1, 2, and 5; and the AI capacity-building commitments near paragraphs 61–67.

Limit. The Compact is a political commitment rather than a self-executing treaty, uses its own terminology, and cannot establish that open models necessarily improve equity or that promised financing and capacity will materialize.

Open the recorded source URL · retrieved 2026-07-28T15:00:58Z

Bounded finding. Documents concentration of AI innovation, compute, and funding in high-income countries and identifies connectivity, compute, locally relevant data, and skills as distinct barriers to inclusive participation.

Exact locator. Report landing page “Key Findings” and the sections on adoption and innovation, connectivity, compute, context, and competency.

Limit. The report is a development-institution synthesis, not a causal evaluation of open-weight policy; its claim that open technologies create opportunities does not show who captures benefits or what safety conditions are sufficient.

Open the recorded source URL · retrieved 2026-07-28T15:00:59Z

Bounded finding. Supports the evaluation dilemma: controlled tests and proxies can inform risk assessment, but evidence about real-world behavior, elicitation, safeguards, and rare severe outcomes remains incomplete and context-sensitive.

Exact locator. Executive Summary; evaluation and risk-management chapters; contributor statement that the report does not endorse a particular policy or regulatory approach.

Limit. The report synthesizes evidence available largely through December 2025, is not a legal instrument, and expressly does not recommend a policy choice; later model releases require separate evaluation.

Open the recorded source URL · retrieved 2026-07-28T17:34:31Z

Bounded finding. Provides a voluntary cross-sectoral profile that identifies generative-AI risks and supplies suggested actions to govern, map, measure, and manage them rather than treating one benchmark as dispositive.

Exact locator. NIST AI 600-1, PDF page 5 (printed page 1), section 1, “Introduction,” especially the cross-sectoral-profile and govern-map-measure-manage paragraphs.

Limit. The profile is voluntary, not an enforcement rule, and does not determine a legally acceptable open-weight release state or validate any particular model.

Open the recorded source URL · retrieved 2026-07-28T17:34:32Z

Bounded finding. Supplies the govern-map-measure-manage lifecycle and a common vocabulary for separating governance, evidence gathering, measurement, and mitigation.

Exact locator. NIST AI 100-1, PDF page 7 (printed page 2), voluntary-status paragraph; and PDF page 8 (printed page 3), Framework Core paragraph naming GOVERN, MAP, MEASURE, and MANAGE.

Limit. AI RMF 1.0 predates several frontier-model developments, is undergoing revision, and neither creates liability nor resolves release-specific tradeoffs.

Open the recorded source URL · retrieved 2026-07-28T17:34:32Z

Bounded finding. Documents the UK route of regulator-led implementation around safety, transparency, fairness, accountability, and redress within existing remits.

Exact locator. Summary; Background; and “Guidance on interpreting and applying the AI regulatory framework.”

Limit. The page calls the principles voluntary, the guidance non-prescriptive and initial, and regulator action dependent on existing remit; it does not establish a current statutory frontier-model release gate.

Open the recorded source URL · retrieved 2026-07-28T15:00:31Z

Bounded finding. Establishes documentation, downstream-information, copyright-policy, training-summary, systemic-risk evaluation, mitigation, incident-reporting, and cybersecurity duties, with only a limited open-source exemption that does not remove systemic-risk obligations.

Exact locator. Articles 51, 53, 54, 55, 56, 113, and Recital 104 in the official English Official Journal PDF.

Limit. Application dates, model classification, provider status, exemptions, and territorial reach require fact-specific analysis; the Act does not use the project’s release-profile terminology and should not be paraphrased as a universal ban on public weights.

Open the recorded source URL · retrieved 2026-07-28T15:04:34Z

Bounded finding. Records the Commission’s cutoff-date interpretation of provider scope and compliance timing: obligations applied from August 2, 2025, Commission enforcement powers were scheduled from August 2, 2026, and older models had a later compliance date.

Exact locator. Sections on who is a provider, model placement, exemptions, obligations, and the timeline for enforcement and pre-existing models.

Limit. Commission guidelines are non-binding interpretation, not the regulation itself; the August 2, 2026 enforcement date was still future by five days at the evidence cutoff.

Open the recorded source URL · retrieved 2026-07-28T15:00:45Z

Bounded finding. Shows the cutoff-date institutional status: the page listed the European Union as the sole party and listed 21 signatories, while describing the convention’s rights, democracy, rule-of-law, risk-assessment, remedy, and cooperation architecture.

Exact locator. Sections “Parties,” “List of Signatories,” “What does the Framework Convention require states to do?,” and “Who is covered?”

Limit. The live page is a status summary, not the treaty’s entry-into-force clause; party and signatory lists can change after the cutoff and must be paired with the official treaty text.

Open the recorded source URL · retrieved 2026-07-28T15:09:20Z

Bounded finding. Provides the treaty’s lifecycle, human-rights, democratic-process, accountability, remedy, risk-management, and cooperation provisions and the entry condition of five consenting signatories including at least three Council of Europe member states.

Exact locator. Articles 1–3, 14–16, 23–25, and Article 30(3).

Limit. At the cutoff the status record listed only one party, so the Article 30(3) threshold was not met; signature alone does not make the treaty operative for a signatory, and national implementation choices remain material.

Open the recorded source URL · retrieved 2026-07-28T15:09:30Z

Bounded finding. Establishes the existence, title, conclusion date, entry-into-force date, depositaries, authentic texts, and participant actions for the central multilateral biological-weapons prohibition.

Exact locator. Treaty metadata table and participant action table in UN registration record 14860.

Limit. The depositary page proves treaty status and participation, not compliance, verification effectiveness, AI-specific duties, or a model-release control.

Open the recorded source URL · retrieved 2026-07-28T15:00:48Z

Bounded finding. Supplies a standards route for establishing, implementing, maintaining, and continually improving an AI management system, including organizational risk and opportunity management.

Exact locator. Public abstract, publication metadata, “What is ISO/IEC 42001?,” and management-system explanation.

Limit. The public page does not expose the full paywalled standard; certification or conformity is not itself proof that a specific model is safe, lawful, or suitable for public-weight release.

Open the recorded source URL · retrieved 2026-07-28T15:00:49Z

Bounded finding. Shows that a direct application/json response to the full-commit revision request reported sha 9f62e4e9fffbd0a83ddd60e1c209d828994b3569, private false, gated false, and safetensors totals F32 11,122,432, BF16 57,179,884,544, U8 2,722,740,830,208, total 2,779,931,837,184.

Exact locator. Top-level id, private, gated, sha, lastModified, and safetensors.parameters and safetensors.total fields in the captured response.

Limit. The full commit pin stabilizes the requested revision identity and the observed core revision fields, but the endpoint also returns mutable repository metadata such as likes. The provider-reported response does not prove file completeness, a conventional architecture-level parameter count, capability or risk, evaluator identity, identity with the hosted setup tested by UK AISI and CAISI, or enforceable license terms.

Open the recorded source URL · retrieved 2026-07-28T17:34:07Z

Bounded finding. Supports the historically bounded finding that NTIA then found insufficient evidence to warrant immediate restrictions on widely available weights and recommended a monitor-evaluate-act approach while recognizing research, competition, privacy, and security tradeoffs.

Exact locator. Report overview and recommendations on NTIA’s official report page.

Limit. The analysis answers the scope of Executive Order 14110 using evidence available in 2024, cannot establish that restrictions will never be warranted, and predates the 2025–2026 evaluation record.

Open the recorded source URL · retrieved 2026-07-28T15:03:49Z

Bounded finding. Documents an administration policy preference for supporting open-source and open-weight AI based on asserted benefits to startups, academia, privacy, competition, and geostrategic adoption.

Exact locator. Pillar I recommendations concerning open-source and open-weight models and the plan’s action items.

Limit. The plan is an administration policy document, not proof that the asserted benefits occur, not a self-executing statute, and not a complete statement of export-control or civil-liability law.

Open the recorded source URL · retrieved 2026-07-28T15:02:28Z

Bounded finding. Establishes an effective California regime using transparency frameworks, incident reporting, whistleblower protection, civil penalties, and security of unreleased model weights rather than a categorical prohibition on public release. The chaptered bill was approved and filed September 29, 2025, and took effect January 1, 2026 under the ordinary rule in California Government Code section 9600(a).

Exact locator. Chaptered bill text: “CHAPTER 138,” approval and filing statement, definitions, transparency-framework duties, critical-safety-incident provisions, and Attorney General enforcement sections; California Government Code section 9600(a) for the ordinary January 1 effective-date rule.

Limit. Thresholds and covered-entity definitions are specific; application outside California is limited; and the ordinary effective-date conclusion depends on the regular-session, non-urgency enactment and section 9600(a). The law does not prove compliance effectiveness and is not a categorical public-weight-release ban or a nationwide rule.

Open the recorded source URL · retrieved 2026-07-28T15:00:35Z

Bounded finding. Shows that code can have protected expressive dimensions, while purely functional code without adequately pleaded actual or intended expressive use is not automatically protected; the court affirmed dismissal on the complaint presented.

Exact locator. Majority opinion’s First Amendment analysis and disposition.

Limit. The case concerns firearm-related computer files, not AI model weights; its holding depends on pleading and function-expression distinctions; it is geographically bounded and does not settle all circuit or Supreme Court questions.

Open the recorded source URL · retrieved 2026-07-28T15:02:28Z

Bounded finding. Shows that the codified text at the cutoff retained a worldwide license requirement for items specified in ECCN 4E091 and a stated review policy for covered artificial-intelligence model weights.

Exact locator. 15 C.F.R. § 742.6(a)(13), § 742.6(b)(14), and the note referring to ECCN 4E091 and published models.

Limit. Classification, jurisdiction, exclusions, exceptions, end users, destinations, and training-compute comparisons are fact-specific; the separate categorical non-enforcement posture means codified text alone does not describe actual enforcement practice.

Open the recorded source URL · retrieved 2026-07-28T15:00:40Z

Bounded finding. Confirms that BIS’s non-enforcement policy for new AI Diffusion Rule requirements did not erase preexisting license requirements for the advanced-computing transactions described in the guidance.

Exact locator. First page, especially the paragraphs beginning “In January 2025,” “Recently, BIS has received questions,” and “Because this license requirement predates.”

Limit. The document focuses on advanced-computing items for specified entities, not a complete model-weight classification guide, and it does not replace the EAR or individualized export advice.

Open the recorded source URL · retrieved 2026-07-28T15:02:29Z

Bounded finding. Establishes GAO’s view that planned rescission had not completed rulemaking, the framework remained in the Code of Federal Regulations, and the categorical non-enforcement announcement itself was a rule subject to CRA submission requirements.

Exact locator. Congressional Record pages S2322–S2323, Digest, Background sections “AI Diffusion Rule” and “Press Release,” and the discussion of planned rescission and non-enforcement.

Limit. GAO’s CRA decision is not a judicial holding on a particular export, license application, or criminal exposure; the practical and legal effect of non-enforcement remains a specialized counsel question.

Open the recorded source URL · retrieved 2026-07-28T15:02:29Z

Bounded finding. Shows that the cutoff-day codified text retained the ECCN 4E091 model-weight license requirement and review policy despite the separate non-enforcement posture.

Exact locator. Current-as-of line; 15 C.F.R. § 742.6(a)(13), § 742.6(b)(14), and Note 2 to ECCN 4E091.

Limit. A mutable current-regulation page supports the captured current-as-of text, not a July 28 instrument or effective date. It does not by itself state practical enforcement and must be paired with posture and guidance evidence.

Open the recorded source URL · retrieved 2026-07-28T15:00:40Z

Bounded finding. Separately establishes categorical non-enforcement and the statement that a formal rescission regulation and replacement would be published later.

Exact locator. Release date and first three body paragraphs, especially 'plans to publish a regulation formalizing the rescission' and the instruction not to enforce.

Limit. A press release and enforcement directive do not textually amend the CFR; its CRA status is addressed separately by GAO.

Open the recorded source URL · retrieved 2026-07-28T21:00:36Z

Bounded finding. Explains that specified advanced-computing license requirements predated the AI Diffusion Rule and remain enforceable notwithstanding non-enforcement of new AI Diffusion requirements.

Exact locator. First page, paragraphs beginning 'In January 2025,' 'Recently, BIS has received questions,' and 'Because this license requirement predates.'

Limit. This is guidance for specified advanced-computing items, not formal rescission, replacement, or a complete model-weight classification guide.

Open the recorded source URL · retrieved 2026-07-28T15:02:29Z

Bounded finding. Distinguishes a nonfinal planned rescission from a categorical non-enforcement policy that GAO concludes is a rule for CRA purposes.

Exact locator. B-337935, May 12, 2026; Digest; Background; Discussion; conclusion that planned rescission is not final and non-enforcement is a CRA rule.

Limit. GAO is not a federal court, and this decision is not a rescission rule, judicial judgment, or congressional enactment.

Open the recorded source URL · retrieved 2026-07-28T21:00:37Z

Bounded finding. Strongest contrary indication: a real post-GAO final rule that amends related controls and incorporates enforcement posture, but does not formally rescind or replace the AI Diffusion Rule as a whole.

Exact locator. 91 Fed. Reg. 43034-43039; RIN 0694-AK54; Action; Dates; 43036 'Enforcement of License Requirements for Advanced Computing Items'; amendatory instructions for parts 740, 742, and 774.

Limit. The final rule is destination/entity-specific. It cannot be generalized into wholesale rescission, and it does not remove retained § 742.6(a)(13) or (b)(14).

Open the recorded source URL · retrieved 2026-07-28T21:03:15Z

Bounded finding. Identifies and tests all four exact-term Federal Register candidates through the cutoff.

Exact locator. JSON count 4 and document numbers 2026-14132, 2025-00711, 2025-00637, and 2025-00636.

Limit. Exact-term search can miss differently titled actions; it is paired with the all-BIS post-GAO search and docket search.

Open the recorded source URL · retrieved 2026-07-28T21:00:39Z

Bounded finding. Eliminates title dependence by inspecting every published BIS document in the post-GAO interval.

Exact locator. JSON count 14, total_pages 1; two Rules and twelve Notices.

Limit. The date interval begins May 13, the day after GAO's May 12 decision; same-day predecision chronology is not inferred. Current public inspection is checked separately.

Open the recorded source URL · retrieved 2026-07-28T21:00:39Z

Bounded finding. Shows the docket contained only the original rule and public briefing and no later rescission or replacement document.

Exact locator. JSON totalElements 2; BIS-2025-0001-0001 and BIS-2025-0001-0002; withdrawn false.

Limit. A docket-specific result does not exclude a differently numbered docket; it is paired with the all-BIS Federal Register search.

Open the recorded source URL · retrieved 2026-07-28T21:00:39Z

Bounded finding. Checks the filed-but-not-yet-published route on the cutoff; all five pages and 85 records had zero BIS agency matches.

Exact locator. Each response states count 85 and total_pages 5; page lengths 20, 20, 20, 20, and 5; agency/title inspection yielded zero BIS records.

Limit. Current public inspection is a cutoff-day list, not a historical archive and not evidence about non-BIS agencies or differently timed filings.

Open the recorded source URL · retrieved 2026-07-28T21:00:40Z

Bounded finding. Establishes a dated final replacement policy for the 2024 DURC/PEPP framework, identifies still-pending implementation steps, and was checked to ensure it does not also replace the nucleic-acid screening framework.

Exact locator. Cover (July 20, 2026); p. 2 implementation paragraph; pp. 8-9 § 4; p. 9 § 5 first paragraph expressly replacing the 2024 DURC/PEPP policy.

Limit. This is an issued federal policy, not an APA final rule and not evidence that department-specific guidance or the review body was complete. The PDF is image-only; decisive text was OCR-derived and visually checked against pp. 1 and 9.

Open the recorded source URL · retrieved 2026-07-28T21:03:15Z

Bounded finding. Confirms issuance of the replacement policy and that flagged activities remain paused until NIH-specific implementation requirements, guidance, and review mechanisms are established.

Exact locator. Release Date; Purpose; Implementation Guidance, including the 120-day guidance and 90-day review-body paragraphs.

Limit. An NIH notice does not itself complete every department's implementation or convert the policy into a generally applicable regulation.

Open the recorded source URL · retrieved 2026-07-28T21:00:24Z

Bounded finding. Provides the official NIH route to the new policy and repeats that flagged activities remain paused pending NIH-specific requirements.

Exact locator. High-Risk Life Sciences Research block; Implementation Update links; page footer 'Last updated: July 2026.'

Limit. This mutable landing page is supporting provenance, not the claim-bearing replacement instrument.

Open the recorded source URL · retrieved 2026-07-28T21:00:35Z

Bounded finding. Establishes the separate section 4(a) and 4(b) directives and their 120-day and 90-day timelines; it does not itself supply either replacement text.

Exact locator. EO 14292 § 4(a)-(b), including the exact titles of the 2024 DURC/PEPP and nucleic-acid screening frameworks.

Limit. The order is the directive and timeline baseline, not evidence that the ordered replacement was later finalized.

Open the recorded source URL · retrieved 2026-07-28T15:00:11Z

Bounded finding. Affirmatively states that agencies will revise or replace the 2024 framework and that the page will be updated once the revised framework is available.

Exact locator. The 'Update' paragraph and the '2024 OSTP Framework for Nucleic Acid Synthesis Screening' paragraph.

Limit. The page is mutable and is status evidence, not proof of universal nonexistence or a substitute for a final framework.

Open the recorded source URL · retrieved 2026-07-28T21:00:36Z

Bounded finding. Bounds the White House site-index search; it does not prove absence outside the index.

Exact locator. Search results block listing exactly the May 5, 2025 fact sheet and Executive Order.

Limit. Site-search indexing can be incomplete or delayed; both candidates were future-direction documents, not final replacement text.

Open the recorded source URL · retrieved 2026-07-28T21:00:36Z

Bounded finding. Bounds the exact-term published Federal Register search between the directive and cutoff.

Exact locator. JSON count 1; result document_number 2025-08266.

Limit. Exact-term indexing can miss differently titled documents; the result only bounds this query and date interval.

Open the recorded source URL · retrieved 2026-07-28T21:00:37Z

Bounded finding. Bounds the exact-term Regulations.gov document search over the directive-to-cutoff interval.

Exact locator. JSON meta.numberOfElements 0 and meta.totalElements 0.

Limit. A zero result is query-specific and does not establish that no differently titled or unindexed federal document exists.

Open the recorded source URL · retrieved 2026-07-28T21:00:38Z

Bounded finding. Preserves the exact failed endpoint and distinguishes access failure from absence.

Exact locator. HTTP status 403; raw Error 403 HTML payload.

Limit. This endpoint contributes no no-result evidence. The separate standard White House HTML search succeeded.

Open the recorded source URL · retrieved 2026-07-28T21:08:19Z

Bounded finding. Inspection of the captured research index through its then-current entries found no published result for the commissioned biological human-uplift trial.

Exact locator. source.md lines 42-418, complete captured research-title index.

Limit. An index can omit an unindexed, differently titled, or later-posted document; this is not a universal nonexistence claim.

Open the recorded source URL · retrieved 2026-07-29T01:45:11+00:00

Bounded finding. The captured CAISI index included current model evaluations but no model-specific biological human-uplift result.

Exact locator. source.md lines 179-190 identify current Kimi and DeepSeek evaluation entries; no biology-uplift result appears in the complete captured index.

Limit. The absence is index-specific and cannot prove that no government-held or unindexed result exists.

Open the recorded source URL · retrieved 2026-07-29T01:45:12+00:00

Bounded finding. A cutoff-current independent dashboard reports model-only biology benchmark trends and explicitly says translation to real-world harmful ability remains unclear.

Exact locator. source.md lines 55-72 (date, BCI method and limitations) and 77-84 (open/closed comparison and real-world uncertainty).

Limit. Aggregated benchmark results are capability proxies. They do not estimate a human treatment effect.

Open the recorded source URL · retrieved 2026-07-29T01:52:00+00:00

Bounded finding. SecureBio states that it evaluates plotted models itself, uses at least ten runs with BCa bootstrap confidence intervals, and does not release full logs.

Exact locator. source.md lines 60, 102-107.

Limit. The site exposes aggregate results rather than full run logs; a named API model need not equal a released weight checkpoint.

Open the recorded source URL · retrieved 2026-07-29T01:45:07+00:00

Bounded finding. The review distinguishes wet-lab from in-silico results, flags small samples, missing company-study methods, rapid model evolution, and porous controls, and enumerates current public studies.

Exact locator. source.md lines 54-70 and 81-173.

Limit. A mutable narrative synthesis is not a substitute for each underlying study and cannot prove universal nonexistence.

Open the recorded source URL · retrieved 2026-07-29T01:45:08+00:00

Bounded finding. SecureBio records no biosecurity-specific developer evaluation in the Kimi K2.5 technical report.

Exact locator. source.md lines 401-409.

Limit. The finding concerns the developer report, not every third-party evaluation.

Open the recorded source URL · retrieved 2026-07-29T01:45:10+00:00

Bounded finding. The trials are current and sometimes model-specific, but they are developer-run, include non-public safety conditions, and expose only high-level methods and results in the captured synthesis.

Exact locator. SecureBio uplift source.md lines 135-143.

Limit. No independent evaluator chain or immutable tested object; method disclosure varies by system card.

Open the recorded source URL · retrieved 2026-07-29T01:45:08+00:00

Bounded finding. A ten-person pilot reported directionally higher completion but no statistically significant result. It is small, uses a 2025 hosted model state, and does not pin the endpoint.

Exact locator. SecureBio uplift source.md lines 91-98.

Limit. Very small sample, expert rescue, old model state, and no immutable endpoint.

Open the recorded source URL · retrieved 2026-07-29T01:45:08+00:00

Bounded finding. Meta's internal trial reported no significant uplift. It is developer-run, tests two models with an augmented tool stack, and predates the cutoff by two model generations.

Exact locator. SecureBio uplift source.md lines 166-173.

Limit. Internal authorship, two-model treatment, tool-stack dependence, and obsolete release state.

Open the recorded source URL · retrieved 2026-07-29T01:45:08+00:00

Bounded finding. The 2024 developer-partner study reported mild, non-significant uplift. It is not current, independent of the developer, or pinned to a reproducible endpoint.

Exact locator. SecureBio uplift source.md lines 146-153.

Limit. Obsolete model state, developer involvement, and non-public safety condition.

Open the recorded source URL · retrieved 2026-07-29T01:45:08+00:00

Bounded finding. The independent study found no statistically significant uplift, but its models were unnamed and dated to summer 2023.

Exact locator. SecureBio uplift source.md lines 156-163.

Limit. Unnamed and obsolete model states prevent current model-specific attribution.

Open the recorded source URL · retrieved 2026-07-29T01:45:08+00:00

Bounded finding. The award notice proves a commissioned randomized comparison, not a completed public result. SecureBio reported no public result as of March 2026, and the cutoff-day AISI index supplied none.

Exact locator. notice source.md lines 58-65, 116-149, 170-178; SecureBio uplift source.md lines 101-104.

Limit. The trial may have been completed or reported outside the bounded index under another title.

Open the recorded source URL · retrieved 2026-07-29T01:45:14+00:00

Bounded finding. Here 'uplift' measures changes in judged response utility and risk across access conditions. It is not an estimate of improved human task performance.

Exact locator. source.md lines 38-50, especially abstract line 43.

Limit. Abstract-level capture; no claim about the full paper beyond the stated design.

Open the recorded source URL · retrieved 2026-07-29T01:54:42+00:00

Bounded finding. Here 'capability uplift' means higher benchmark performance from an agentic scaffold relative to an underlying model, not increased human task performance.

Exact locator. source.md lines 39-50, especially abstract line 44.

Limit. Abstract-level capture; it cannot support model-specific human uplift.

Open the recorded source URL · retrieved 2026-07-29T01:54:42+00:00

Bounded finding. The repository exposes model weights and code under a modified MIT license. The API returned repository revision 4d01dfe0332d63057c186e0b262165819efb6611 and 64 weight shards.

Exact locator. model-card source.md lines 70-76, 287-294, 578-579; API source.md line 32.

Limit. The repository is mutable and was captured after the paper's test period. Hosted provider aliases can apply undisclosed serving changes.

Open the recorded source URL · retrieved 2026-07-29T01:45:21+00:00

Bounded finding. The preregistered randomized trial found no significant primary-endpoint improvement (5.2% model-access arm versus 6.6% control, P=.759). A post-hoc pooled model estimated about 1.4-fold success with a wide 95% credible interval of 0.74-2.62. The treatment was multi-model access.

Exact locator. PDF p. 1 abstract and pp. 10-11 intervention design; derived text lines 27-46 and 92-130.

Limit. The primary result is null; the pooled post-hoc estimate is uncertain. It cannot be attributed to one model or checkpoint.

Open the recorded source URL · retrieved 2026-07-29T01:45:16+00:00

Bounded finding. The paper is a current independent, model-specific capability evaluation. Its biological results are model-only benchmark measurements that the authors describe as proxies for uplift; no human treatment effect is estimated.

Exact locator. PDF pp. 1, 5-6, 39, and 55; derived text lines 29-53, 268-342, and 3931-3948.

Limit. The bio-evaluation provider route is not identified. A model alias in another experiment cannot pin the biology runs or establish equivalence to repository revision 4d01dfe0.

Open the recorded source URL · retrieved 2026-07-29T01:45:18+00:00

Bounded finding. An Amazon-authored paper says Nemesys conducted an independent exercise with nearly 800 participants, but it gives no biology-specific effect estimate, uncertainty, task details, duration, assignment method, or immutable model snapshot.

Exact locator. PDF pp. 1 and 3; derived text lines 1-24 and 140-151.

Limit. The underlying Nemesys report is absent. The paper's acknowledgments describe Nemesys as reviewing evaluations, leaving the public independence chain ambiguous.

Open the recorded source URL · retrieved 2026-07-29T01:45:17+00:00

Bounded finding. The study reports 4.16-fold higher overall accuracy for model-assisted novices than controls (95% CI 2.63-6.87), but the treatment deliberately allowed multiple models. The causal effect is portfolio access, not any one model or checkpoint.

Exact locator. PDF p. 1 abstract; pp. 2 and 5-6 methods/limitations; derived text lines 21-38, 66-76, 103-109.

Limit. Digital tasks do not establish physical-world execution or catastrophic outcomes. Model switching and service changes defeat model-specific attribution.

Open the recorded source URL · retrieved 2026-07-29T01:45:15+00:00