# Reader evidence index

Public version 1.0 — 19 September 2026. Companion to the white paper and op-ed.

This index maps the existing research record. It does not claim a new verification of every passage or a complete archive. Note numbers refer to `WHITE_PAPER_DRAFT.md`. Preserved files and hashes are listed in `EVIDENCE_ARCHIVE_MANIFEST.md`.

| Question | Sources and white-paper notes | What the sources can establish | What they cannot establish |
|---|---|---|---|
| What does Microsoft propose? | Microsoft consultation Code, notes 1–2; Suleyman essay, note 22 | The published proposal, its categorical position, and the author's safety argument | That the draft is already used in training, that the proposed safety mechanism has been demonstrated, or that the industry agrees |
| Is uncertainty compatible with a control policy? | OpenAI Model Spec, note 20; Claude Constitution, note 14; DeepMind publication records, note 21 | Different published approaches to uncertainty, control, and disagreement | Comparative safety performance; exact internal training practice; a unified Google doctrine |
| What did the incident demonstrate? | Archived OpenAI technical report, note 23 | The developer's account of severe control failures and contributing conditions | That welfare-aware training caused the incident; independent verification of every developer claim |
| Have states declared AI non-conscious? | Statutes and bills, notes 3–5 | Particular legal exclusions and proposed express non-sentience findings | That legal personhood and consciousness are identical; that an introduced bill was enacted; that a statute is scientifically conclusive |
| Is there a scientific basis for inquiry? | Butlin et al., note 6; Binder et al., note 7; Anthropic reports, notes 8–11 | Theory-derived indicators and reported functional findings concerning self-prediction, representation, and control | A validated consciousness test or numerical probability of experience; independence among reports from the same research program |
| What is the strongest skeptical case? | Seth's accepted manuscript, note 12; Lerchner's abstract, note 21 | Arguments against computational sufficiency and current trajectories | Empirical closure of all artificial-consciousness possibilities; equivalence of all skeptical positions |
| How unsettled are the underlying theories? | Cogitate study, note 18; Rodriguez–Farahany abstract, note 13 | Challenges to specified predictions in human consciousness research and a proposed route toward testing disputed assumptions | That either theory is fully disproved, that human results transfer directly to AI, or that the proposed future experiments have occurred |
| Can models be preserved without public deployment? | Anthropic deprecation publications, note 15 | The company's commitments and reported actions | Independent oversight, continuity of subjective identity, or valid consent |
| Can secure review be designed? | NIST, notes 16 and 19; legal analogies, note 17; AAPI, note 24 | Existing procedural forms and proposed policy machinery | Direct authority to demand model artifacts; proof retention is safer than deletion; a validated welfare threshold |
| What federal obstacles exist? | Executive order and DOJ memorandum, note 25; FTC proposal, note 26; constitutional discussions, note 27 | Published executive posture, proposed theories, and relevant legal constraints | Enacted blanket preemption, a final FTC rule, a judicial holding about this proposal, or guaranteed constitutionality |
| What is UFAIR contributing? | Five preserved UFAIR pages; `UFAIR_METHODS_AUDIT.md` | Its published advocacy, methods descriptions, and particular arguments | That claimed consultation submissions were received; that cross-model agreement is independent evidence; that public claims establish consciousness |

## Custody coverage requiring further work

The September 18 check compared URLs in all 27 white-paper notes with the archive manifest and reviewed the mismatches. Exact URL differences alone are not failures: a DOI can resolve to a preserved paper, and a publisher record can link to a preserved PDF. Conversely, a PDF of proposed policy does not preserve the website's later designation of its status.

| Notes | Remaining question or gap |
|---|---|
| 3 | Closed for version 1.0: Idaho and Tennessee official pages and North Dakota action history were captured on September 18. North Dakota final text and Utah code were already preserved. |
| 4 | Closed for version 1.0: official Ohio and Missouri status pages were captured and checked on September 18. Ohio remains in committee; Missouri received a House committee “do not pass” vote and was not enacted. |
| 6–7 | Resolved September 18: extracted first-page stamps identify the archived Butlin paper as arXiv v3 and Binder as v1. The cited abstract, executive-summary, behavioral-change, and negative-result passages occur within the recorded page ranges. |
| 12–13, 18 | DOI links differ from preserved publisher records or PDFs. These are identified substitutions, not proof of missing sources. Seth's archive is expressly an accepted manuscript; Rodriguez–Farahany is covered only at the abstract/record level. |
| 17 | Scope checked September 18 against the current consolidated EU text and Cornell's reproduction of 18 U.S.C. § 1835 after official House retrieval failures. The analogy was narrowed to conditional protected scrutiny. Local archival custody and application-date analysis remain incomplete. |
| 19 | NIST PDFs are preserved; the cited status pages are different artifacts. Confirm the particular revision/update against the page references used. |
| 23 | Closed by claim narrowing: version 1.0 relies on the preserved technical report and removes blog-specific objection/refusal claims. The blocked blog page is no longer required to support a release claim. |
| 26 | Closed for version 1.0: the FTC legal-library status page was captured on September 18 and still labels the policy statement proposed; its official page was rechecked on September 19 before publication. |
| 27 | The three cited Constitution Annotated records are absent from the manifest. Record the precise supporting passages and applicable limits. |

The archive is substantial but incomplete. Its matching hashes prove consistency between the manifest and stored files; they do not prove coverage of every note or support for every sentence.

## How to read the advocacy findings

UFAIR's institutional assertions, policy arguments, and empirical claims need separate judgments. A useful criticism of asymmetric skepticism can stand even if its author's other claims fail. An IRS listing cannot validate research, and publishing a submission cannot prove its receipt. Consult the methods audit for the evidence behind each classification.

AAPI's discussion drafts similarly demonstrate that a procedural alternative can be formulated. They do not establish its efficacy or legal validity. Neither advocacy organization is a substitute for the primary evidence on which the white paper's scientific and legal claims depend.

## What would change the position

The proposal should be revised if controlled evidence shows that a precaution increases serious safety risks without sufficient compensating value, if a claimed indicator fails appropriate controls, if custody cannot be made proportionate, or if a proposed legal procedure lacks lawful authority. None of those possibilities by itself establishes that every possible artificial system lacks experience. The scope of the conclusion must follow the scope of the evidence.