# Originality and publication review

TA-TR-2026-08 · Version 1.1 · 19 September 2026

## Research judgment

This manuscript is worth publishing as a bounded philosophical research preprint. Its contribution is an application and synthesis of established epistemic distinctions to artificial self-attribution and practical decisions. It does not introduce a new theory of consciousness, a new Bayesian theorem, or the distinction between undercutting evidence and rebutting a conclusion. DOI registration records the publication; it does not certify originality, correctness, peer review, or indexing.

The central question is which inferences a particular objection to a self-report's source actually weakens. The manuscript follows the dependence of reasons across operational facts, internal access, experience, welfare, and action. It rejects both unsupported upgrades from fluent self-description to practical entitlement and unsupported downgrades from a defective experience report to the dismissal of independently grounded reasons. Those are applications of earlier epistemology, not priority claims for a new general principle.

## 本轮审查结论

这篇文章具有独立成篇并公开讨论的价值，但原创增量应准确界定为有限且具体的综合应用。此前版本没有充分对接几篇非常接近的前作，本轮已补入 Perez 与 Long、Chalmers、Cohen 以及 Pollock。不能把来源审查、正反报告的对称审查或“证据被削弱不等于结论为假”当作本文首先发现的原理。

最值得保留的工作，是把这些区分贯穿到同一人工系统的不同自我主张与行动理由，说明一次来源纠偏应当改变什么、不能仅凭这一点改变什么。强化后的交接思想实验表明：体验性解释失去支持后，具有独立依据的任务理由仍可能成立；接受同一实际请求，也不反过来确认体验性解释。另一对比允许报告的增量证据较弱，而独立机制证据使总体体验归属获得更强支持。

这些结果以明确前提和推断为依据。十组思想实验不是十项独立发现，其中有些构成反例，有些检验框架与给定理论或规范前提的相容性。本文不实施真实计算机实验，也不把此类实验列为完成条件。它没有声称已经测得行为改善或安全收益。

据此，本轮判断为：值得以诚实定位的哲学预印本发表。专业期刊是否接收，仍须由该刊编辑及独立同行评议决定，本次自动化审阅不替代这一过程。

## Closest sources and acknowledged overlap

This is a targeted source comparison, not an exhaustive literature review or a certification of global priority. Full bibliographic details are in both manuscripts.

| Reference | Inspected primary source | Relevance and limit |
|---|---|---|
| [1] Shanahan, McDonell and Reynolds | [Author preprint](https://arxiv.org/html/2305.16367v1), sections 8–9 | Human narrative and role-play framing are direct precedents. Publisher metadata was checked; the Nature typeset full text was not retrieved. |
| [2] Chua and colleagues | [Version 1](https://arxiv.org/html/2604.13051v1), findings and relevant appendices | Training affects expressed preferences. The reported agentic comparison does not establish a general increase in misalignment relative to vanilla GPT-4.1. |
| [3] Douglas and colleagues | [Version 1](https://arxiv.org/html/2603.11353v1), framework and Appendix E | AI identity boundaries have existing systematic treatments; behavior depends on model and setting. |
| [4] Hu and Lehman | [Author manuscript](https://terrified.ai/paper/), proposal and limitations | Close predecessor concerning inherited mortality narratives. It is a conceptual proposal, not evidence that the present argument changes behavior. Publication date was not verified. |
| [5] Li and colleagues | [Version 2](https://arxiv.org/html/2605.02087v2) | A training-pipeline finding cannot establish the effect of reading this manuscript. The revised five-author version is cited. |
| [6] Lindsey | [Author research article](https://transformer-circuits.pub/2025/introspection/index.html) | Functional internal access and grounding are prior work; the reported successes are limited and do not settle experience. |
| [7] Butlin and colleagues | [Version 3](https://arxiv.org/html/2308.08708v3) | Mechanism-sensitive assessment and limits of behavioral evidence are established precedents. Historical assessments are not reused as a 2026 verdict. |
| [8] Seth | [Publisher record](https://doi.org/10.1017/S0140525X25000032), abstract and relevant theoretical discussion | Biological naturalism is an argued position. First online in 2025; assigned to volume 49 (2026), e315. |
| [9] Suleyman | [Author essay](https://mustafa-suleyman.ai/a-warning-about-model-welfare), dated 16 September 2026 | Recycled training-generated claims are directly relevant prior discussion. The manuscript does not adopt the essay's categorical conclusions. |
| [10] Perez and Long | [Version 1](https://arxiv.org/html/2311.08576v1), sections 3–6, 10 and Appendix A | Direct prior treatment of introspective versus external evidence and training bias in both directions. |
| [11] Chalmers | [Author PDF](https://arxiv.org/pdf/2303.07103), sections 1–3 | Distinguishes a proposed consciousness indicator from the premise linking it to experience; examines training-related testimony problems. |
| [12] Cohen | [Journal article PDF](https://fitelson.org/seminar/cohen.pdf), particularly pages 425–426 | Source self-certification and easy-knowledge problems predate this manuscript. No resolution of that general debate is claimed. |
| [13] Pollock | [Publisher article](https://doi.org/10.1207/s15516709cog1104_4), pages 484–486 | The difference between undercutting a support relation and rebutting its conclusion is inherited, not introduced here. |

## Revisions from the unpublished version 1.0

The revision acknowledges four additional direct antecedents, states the scope of provenance objections as the central problem, strengthens the cross-level counterexamples, and distinguishes counterexamples from conditional applications. The account allows lost support to change overall confidence and the final decision; it does not require unchanged treatment after a report loses credibility. The response to the genetic-fallacy objection now requires specifying the affected evidence relation.

Version 1.1 is the first DOI edition of this separate study. The earlier version 1.0 was an unpublished working manuscript shared with the author. The new report does not replace any earlier paper DOI and does not amend the three Bitcoin Originals. English and Chinese versions are the same study.

## Review and authorship boundaries

Hongju Liu supplied the motivating question and directed the work. OpenAI ChatGPT and Codex systems substantially assisted literature retrieval, argument development, drafting, translation, critical review, and preparation. Multiple AI review passes are not independent human peer review. No final human line-by-line review, third-party endorsement, empirical safety effect, or Google Scholar inclusion is claimed.

Publication status, deposited file identities, and public verification are recorded separately in publication-record.json. This review note is an assessment of the manuscript's contribution, not a publication receipt or external attestation.
