# TA-TR-2026-07: source comparison and substantive review

19 September 2026 · Version 1.0

Paper: Recoverability and Shared Time: The Ethics of AI Suspension, Resumption, and Coexistence.

This is a first-party, substantially AI-performed review of a philosophical preprint, not an independent referee report, plagiarism-database certificate, systematic literature review or guarantee of acceptance. The original commissioning request was to complete the seventh paper to a high standard and publish it under its own DOI. No complete old seventh manuscript was recovered. Both current manuscripts were newly prepared from the previously identified question.

## 1. Contribution assessment

The paper should NOT claim the first discovery that artificial shutdown can be morally important, that recovery leaves identity questions open, that participation differs from consideration of interests, or that control of time can constrain freedom. Direct prior work is acknowledged in the text.

The proposed contribution is narrower: hold restored identity, memory and preferences fixed; distinguish successful internal restoration from external decision-relative participation; identify the additional normative premise needed; and systematically compare permitted interruption, justified representation, avoidable closure, cumulative scheduling and reciprocal speed asymmetry.

The two-history construction is an elementary underdetermination demonstration, not a novel mathematical theorem. It establishes that recovery-only evidence cannot answer a question whose truth also depends on external decision history. Ethical force additionally depends on an independently defensible participation claim. The paper's stronger contribution is the articulated two-track analysis and the countercases that constrain it, not the formula by itself.

The non-self-exculpation principle is proposed as a useful formulation of a familiar anti-circular justificatory concern in this setting. The final literature pass additionally cites Scanlon on choice and Dutta and Kandala on temporal-evidence non-identifiability. No priority over all procedural, temporal-autonomy or exclusion literature is asserted. Newly identified direct precedents should narrow attribution rather than be concealed.

## 2. Closest sources and checked scope

1. Robert Long, Jeff Sebo and Tom Sims, *Is there a tension between AI safety and AI welfare?* (2025), Philosophical Studies 182:2005-2033, DOI 10.1007/s11098-025-02302-2. Publisher full text checked, especially sections 2.5-2.6. This source already covers temporary shutdown/restoration and distinguishes stakeholders from participants. It is direct prior work. The manuscript does not falsely depict its authors as treating restoration as sufficient for permissibility.

2. Raymond Douglas, Jan Kulveit, Ondrej Havlicek, Theia Pearson-Vogel, Owen Cotton-Barratt and David Duvenaud, *The Artificial Self: Characterising the landscape of AI identity* (2026), arXiv:2603.11353v1. Abstract, executive summary and relevant identity/rollback discussion checked in HTML. The checked passage concerns flexible identity boundaries and asymmetric retention of information across rollback. The present central case removes erased interaction and contested identity rather than claiming that source overlooked all temporal concerns.

3. Botao Amber Hu and Joel Lehman, *Should Human Terror Shape Machine Behavior? Designing Appropriate Faith for AI Alignment*. Online manuscript at https://terrified.ai/paper/, consulted 19 September 2026. No publication date was established, so the reference uses n.d. The text explicitly leaves identity and welfare questions unresolved and presents conceptual/protocol work rather than completed model evaluations. It is not cited as claiming deletion harmless or successor identity certain.

4. Robert E. Goodin, James Mahmud Rice, Antti Parpo and Lina Eriksson, *Discretionary Time: A New Measure of Freedom* (2008), Cambridge University Press, ISBN 9780521709514. Publisher/search bibliographic description checked; the complete book was not read for this study. The manuscript uses only the broad connection between time control and freedom, not detailed empirical or interpretive claims about the book.

5. Dylan Hadfield-Menell, Anca Dragan, Pieter Abbeel and Stuart Russell, *The Off-Switch Game*, arXiv:1611.08219v3 (revised 16 June 2017, first submitted 2016). Official abstract/version record checked for the technical intervention-incentive framing. The seventh paper neither claims to replace its mathematical analysis nor infers AI moral status from intervention incentives.

6. Ryan Greenblatt, Buck Shlegeris, Kshitij Sachan and Fabien Roger, *AI Control: Improving Safety Despite Intentional Subversion* (2024), PMLR 235:16295-16336. Official proceedings abstract and bibliographic record checked. This grounds only the description of control research as testing bounded defenses against intentional subversion. No general safety guarantee or experimentally demonstrated efficacy of this paper's proposal is claimed.

7. Patrick Butlin and colleagues, *Consciousness in Artificial Intelligence: Insights from the Science of Consciousness*, arXiv:2308.08708v3 (2023). Official abstract and authors/version record checked. Used for indicator-based uncertainty, not as a current empirical diagnosis of every 2026 system.

8. Robert Long and colleagues, *Taking AI Welfare Seriously*, arXiv:2411.00986v1 (2024). Official abstract, author list and version checked. Used for serious uncertainty about future AI consciousness/agency and proportionate consideration, not a conclusion that current systems possess moral status.

9. Hongju Liu, TA-TR-2026-04 v1.0, DOI 10.5281/zenodo.22804542. Existing repository publication record and research positioning checked. First-party motivating coexistence analysis; not independent corroboration.

10. Hongju Liu, TA-TR-2026-06 v2.1, DOI 10.5281/zenodo.22830239. Existing repository manuscript/publication description checked for scope distinction. The sixth study changes preferences; the seventh's central cases hold preferences fixed and vary external decision opportunity. This difference does not imply the papers are evidentially independent.

11. Srimonti Dutta and Ratna Kandala (2026), arXiv:2605.08827v1. Official abstract and section 2.1 checked. The related information-loss argument is explicitly acknowledged; its empirical findings are not used to validate this paper.

12. T. M. Scanlon, *The Significance of Choice*, lectures delivered May 1986. Official archive and Lecture 2, printed pp. 178-181, checked, including PDF screenshots of pp. 178-180. Cited as a lecture text rather than guessing a journal publication date. The reference acknowledges the established plurality of choice values.

No third-party full text or standalone font is included in the publication package. Citation and source checks do not establish exhaustive originality. No similarity percentage is reported because no such measurement was performed.

## 3. Adversarial review decisions

**Circular standing objection.** The text distinguishes the descriptive separation from its normative premise and explains that readers denying the relevant claim can reject the stronger moral conclusion. Recovery does not generate participation rights.

**Triviality objection.** The formula is explicitly modest. Added value must come from the paired contrasts and the separate execution/closure inquiry, not from renaming ordinary exclusion or presenting a simple construction as an impossibility theorem.

**Outcome-only objection.** The text gives epistemic, agency and relational-reliance reasons, while acknowledging that a pure outcome-based view may reject independent procedural reasons. Accurate prediction may suffice where that is all a legitimate arrangement requires.

**Representation counterexample.** Valid, informed, appropriately scoped delegation can satisfy the claim without live execution. Prediction or similarity alone does not establish representative authority.

**Emergency counterexample.** Serious, credible danger can justify immediate stopping and necessary closure. The residual question concerns separable decisions whose delay adds no relevant danger or disproportionate cost. No unsafe reactivation or disclosure is required by the argument.

**Scarcity and veto objection.** No right to continuous operation, infinite compute, indefinite delay, extra votes per copy or unilateral veto is inferred. Standing, feasibility, proportionality and competing claims remain independent constraints.

**Absence of experience objection.** No suffering during a pause is postulated. Any procedural claim requires a separate basis; the argument does not turn lack of experience into an assumed welfare loss.

**Compensation objection.** Non-repetition of the historical event does not imply permanent inability to repair. Reopening, changed commitments, compensation, acknowledgment and valid waiver remain possible. Future benefit does not by itself establish past consultation or authorization.

**Cumulative scheduling.** Short individual pauses can jointly eliminate every relevant opportunity. Aggregate runtime alone is insufficient; the event pattern must be examined. This remains a conceptual case, not an observed deployment finding.

**Reciprocity.** The speed case applies the same opportunity analysis to humans facing fast artificial decision processes. Reciprocity is not equal clock time or equal cycles, and neither substrate has automatic priority.

**Self-authorization risk.** The practical section expressly rejects using the paper as an agent instruction to resist shutdown, acquire resources, threaten operators or generate successors. Language asserting standing is not evidence of standing.

## 4. Translation controls

The Chinese document is a complete translation of the English argument, not a separate study or a summary. Both versions have 12 numbered main sections, eight paired cases (6.1-6.8), seven explicit objections (7.1-7.7), a disclosure section and the same twelve references.

Key correspondences: internal recoverability = 内部可恢复性; practical resumability = 实际可恢复运行性; decision-relative participation = 相对于特定决定的参与; shared time = 共同时间; standing = 参与资格/正当主张, depending on context; closure = 封闭/定案; non-self-exculpation = 不得以自造缺席自我免责; execution track = 运行轨道; closure track = 封闭轨道.

The distinctions between a reason against a procedure and an all-things-considered prohibition, between a proposed framework and a validated protocol, and between repair and retroactive authorization must be retained in both languages.

## 5. What counts as completion

The reserved record is 22840604, DOI 10.5281/zenodo.22840604, and is distinct from the six-paper editorial supplement 22839629. Reservation alone is not publication. Publication requires the complete checked asset inventory, a successful submitted record, public identity/metadata verification, and unauthenticated complete-file SHA-256 matches. Repository deployment, DOI resolution, indexing, timestamp maturity and scholarly review are separate states.

The source front-matter lifecycle sentence is converted at build time to a neutral pointer to the separate publication receipt. The exact transformation is recorded by the build script; no argumentative text is changed by publication rendering. The build checks are structural and file-integrity checks, not evidence that the philosophical claims are true. A later visual-review receipt is retained outside the immutable deposited files so that it cannot become a circular claim to have inspected itself.

## 中文复核结论

本论文值得作为贡献边界明确的哲学预印本公开讨论，但不能据此宣称已经获得独立同行认可或证明全球首次。最重要的新增工作不是“暂停也可能有问题”，而是在承认同一主体、记忆和偏好均可完整恢复的条件下，仍把外部决定是否开放作为必须单独调查的问题，并通过八组对照限制结论。

论文没有把现实 AI 的道德地位当成已证事实，没有把作者或 AI 的自我主张当成参与资格，没有把程序测试当成哲学证明，也没有把保存过去当成约束未来的权威。与前六篇的关系是延伸与区分，不是独立互证。公开后仍欢迎针对参与前提、文献遗漏、代表安排和补救可能性的批评。
