Evidence and sources
Every claim on this site, and where it comes from
We argue that a record which cannot be checked by an outsider is not evidence. It would be strange to run a site that could not be.
How to read this
- Every external fact carries a source you can open and test yourself, with one exception: the analyst study on Category Definition is subscription research, described here and not linked.
- Claims that come from our own work do not ask you to take our word for them. Each one is marked method published where the method set out below can be repeated or refuted, and otherwise as what it is: an argument, a forecast or our account.
- Claims about what other systems do are held to the highest bar of all. The site carries as few as it can, and each one states what would disprove it.
- If you find a claim on this site that is not on this page, that is a defect. Tell us and we will either source it or remove it.
The method behind our own findings
- Period. Eight months of continuous searching, beginning in January 2026 and still running.
- Sources, all public. Company and product documentation, published patent applications and grants, standards and specification drafts, regulatory filings, and analyst material that has been publicly issued. Nothing behind a non-disclosure agreement is used to support a public claim.
- Tools. Current frontier language models for search and first-pass reading, at scale. Every placement is then read and adjudicated by a person, because a model is a proposer and not an arbiter, a distinction we of all people cannot ignore.
- Tests. The two on Category Definition, defined in the glossary: does the system stop the action, and can a third party check the decision without the cooperation of the system that produced it.
- What would falsify it. One system that answers yes to both. Not an argument, not a roadmap — one shipping system whose decision a third party can check without the producer’s help. Name it and we will publish the correction on this page, with the date. The same test applies to us. Our place in the upper right is a claim until a party you authorize has checked a ProofSmith decision without our help, which is what the demonstration is for.
The claims
| Claim | Where | Source |
|---|---|---|
| Forty percent of enterprises will demote or decommission autonomous AI agents by 2027, because of governance gaps identified only after production incidents occur | Why Now | A leading industry analyst firm, press release, 26 May 2026 |
| Enterprises treat AI agent governance as binary, either locked down or fully trusted | The Question, Part III | An analyst at the same firm, same release |
| More than 150,000 agents at an average global Fortune 500 enterprise by 2028, up from fewer than 15 in 2025 | Why Now | A leading industry analyst firm, press release, 28 April 2026 |
| Thirteen percent of organizations think they have the right AI agent governance in place | Why Now | A leading industry analyst firm, press release, 28 April 2026 |
| Every quotation and citation in the argument: the Declaration, the Constitution and the Bill of Rights, the Federalist, the Anti-Deficiency Act and three decisions of the Supreme Court | The Question, Parts II and III; Policy; Objections (paraphrased) | Declaration of Independence, 1776; Declaration of Independence, 1776, list of grievances; U.S. Constitution, 1787; Calder v. Bull, 3 U.S. 386, 388 (1798), Chase, J.; James Madison, Federalist No. 10, 1787; Gutierrez de Martinez v. Lamagno, 515 U.S. 417 (1995), opinion of the Court; James Madison, Federalist No. 51, 1788; Alexander Hamilton, Federalist No. 78, 1788; Johnson v. United States, 333 U.S. 10, 14 (1948), opinion of the Court; U.S. Const. amend. IV; U.S. Const. art. I §8; amend. X; Alexander Hamilton, Federalist No. 70, 1788; Anti-Deficiency Act, 31 U.S.C. §1341(a)(1)(A) — each verified at source, 12 September 2026 |
| What each of the seven instruments in the Obligations table asks for, who signs it, and what it rests on today | Why Now (Obligations); Objections; Sectors; Glossary | EU AI Act, Articles 43 and 47, Annex VI; Fed. R. Evid. 902(13), 902(14); NIST SP 800-37 Rev. 2, Risk Management Framework for Information Systems and Organizations, December 2018; SR 26-2 and attachment (Federal Reserve, FDIC and OCC), and OCC Bulletin 2026-13, 17 April 2026; Sarbanes-Oxley Act of 2002 §404, 15 U.S.C. §7262; Federal Power Act §215, 16 U.S.C. §824o; Regulation (EU) 2022/2554, Article 64 — verified 12 September 2026 — each verified at source, 12 September 2026 |
| EU AI Act dates: prohibitions and AI literacy from 2 February 2025, with later dates for some; general-purpose AI models from 2 August 2025; Annex III high-risk systems from 2 December 2027; Annex I products from 2 August 2028, as amended by the AI Omnibus in force since 27 July 2026 | Why Now | Regulation (EU) 2024/1689, Article 113 as amended; European Commission AI Act page — verified at source, 12 September 2026 |
| One hundred and seven agentic-AI deployments across nine industries, scored on six capabilities at five levels; forty-one at Level 2, sixty-six at Level 3, none at Level 4; the firm describes Level 4 as the level most likely to deliver transformative return, and the Level 3 deployments as planning, replanning and acting on production systems | Category Definition | A leading industry analyst firm, research study of 107 agentic-AI deployments across nine industries, 2026 (subscription research; the firm notes the examples were not extensively validated) |
| One hundred and fifty-one systems assessed; one hundred and forty placed on the two axes, eleven set off them; none of the one hundred and forty in the upper right | Home; Category Definition; Resources | Our register, version 5.3, 26 October 2026; method above; one counterexample settles it. The two-page summary and the Why ProofSmith memorandum carry the count as it stood on 24 September 2026: one hundred and eight assessed, one hundred placed, eight set offmethod published · falsifiable |
| The six arenas of the posture section: aviation, wrestling, auto racing, finance, law, warfighting | Home | Analogies, offered as argument, not as findingsargument, not a finding |
| Procurement moves before law does; the first serious failure will write the rules | Why Now, sections 04 and 05 | Our reading of how these regimes are adopted, stated as a forecast rather than a findingforecast, not a finding |
| We have not yet found a system that answers yes to both tests | Category Definition | Eight months of searching the public record against the two published tests. One hundred and forty systems on the axes and eleven set off them, in an internal register at version 5.3, 26 October 2026. Method set out above; sources are public and the search is repeatable. One counterexample settles it.method published · falsifiable |
| It is the test most systems fail | The Question, Part III | Same eight-month search of the public record, same two published tests, same falsification conditionmethod published · falsifiable |
| The six problems | The Problem; Sectors | Stated on the page as our working account, not as findings. Put to operators, revised against what they say, and some are always wrong — which is the point of asking rather than asserting.our account, not a finding |
| The characterization of each sector | Sectors | Drawn from the published obligations cited in the Obligations table on Why Now, every one of which is public and named, plus practitioner conversationsour account, not a finding |
| The defense authorization bill passed by the House on 22 July 2026 would direct the Department to update its autonomy policy to preserve human command responsibility for the use of force, identify the commanders responsible for authorizing it, and set requirements for auditability, traceability and accountability | Sectors | H.R. 8800, sec. 1524(b)(4) and (b)(5), as engrossed in the House, passed 22 July 2026 — verified at source, 12 September 2026 |
| Why deployments stop at Level 3: the person in the path is the accountability, and nothing yet takes the person’s place; Level 4 is where the value is | Category Definition | Our reading of the study, offered as argument, not as the firm’s findingargument, not a finding |
| Nearly every audit log, compliance report and governance dashboard in service today rests on records the acting system keeps about itself | The Question, Part I | Our observation, from the seven public regimes in the Obligations table on Why Now and the one hundred and fifty-one governing systems in the register. We have not surveyed logs and dashboards as such; the claim stands until one counterexample, which settles it.method published · falsifiable |
| No Chief Executive Officer and no Combatant Commander can prove an autonomous action stayed within authority; no system in service can | The Question, Part I; the two-page summary | Our finding under the two tests: eight months of the public record, the method above. One counterexample settles it.method published · falsifiable |
| Seven principles for governing AI; every model treated as an insider risk, with the controls kept outside the model | Category Definition | Satya Nadella, “Models as Insider Risks in the Super Intelligence Era,” article on X, 10 October 2026, paraphrased — verified at source, 12 October 2026 |
| None of the seven principles says who, outside the organization, can check its record | Category Definition | Our reading of the article as published. One sentence in it that names a party outside the organization who can check the record settles it.method published · falsifiable |
| Where we stand on each of the seven principles; that ProofSmith has built the architecture, and has the independence, that solves the problem; a working Arbiter, built and demonstrated | Category Definition | Our account. The demonstration is offered to any reader on request. Our place in the upper right remains a claim until a party you authorize has checked a ProofSmith decision without our help.our account, not a finding |
| A podcast segment of thirty-five minutes on how to make a model obey, and three examples from it | Category Definition | All-In Podcast, episode published 9 October 2026, at 7:40, 15:37 to 25:17, 28:56 and 1:23:33; paraphrased from the episode’s captions, no host quoted — verified at source, 12 October 2026 |
| The ProofSmith architecture is patent pending | Home; Category Definition | United States provisional patent application on file with the USPTO; the number and the date are not disclosed |
| Every quotation on AI is Running Wild | AI is Running Wild | The thirteen primary sources listed at the foot of that page, each linked — verified at source, 12 September 2026 |
| Of seventy-two published problem statements on AI agents, 2025 to September 2026, none states a third party’s ability to check a decision without the acting system’s cooperation as the problem to be solved. Four note its absence in their own case: a reviewer that depended on the actor’s cooperation, a self-assessment with no external review, a reviewing agent that shared the acting agent’s objection, and guidance that places independence in the validator rather than in the record | Seventy-two problem statements | Primary sources only, each re-verified at source on 12 September 2026; the list of seventy-two is published, with the two tests read against each statement; one counterexample settles itmethod published · falsifiable |
Last reviewed 13 September 2026.
Scope
What is not on this page
Nothing on this site describes how the architecture is built. That is deliberate. It is not modesty about the engineering: the material is unpublished, and publishing it is a decision with consequences we are not ready to take. What is here is the problem and the reasoning. How the record is kept, for how long, by whom it is produced, and how it is verified are matters for the engagement, not for this site. The biographical statements on the Team page are the founders’ own account and are not inventoried here. The rest is a conversation.