Uncertified Zeros and Correction Loops — A Framework for Drawing the Boundaries of AI Involvement by Structure, and User-Side Practice

Version note (2026-08-11; updated 2026-10-05): Both the Japanese original and this English translation are now at v0.9.11. Its claims are those of v0.9.10 (August 11, 2026), reflecting verification-series addenda E & W and the temperature-0 control, confirmed through six rounds of intra-family adversarial review plus three rounds of external (non-Claude-family) review; since then, a declaration of unreflected material was added at the end of §9 (August 31, 2026; not a revision), and v0.9.11 (October 5, 2026) aligned the revision counts following Revision Fifteen of the Companion Consideration. The English synchronization itself surfaced one residual inconsistency in the Japanese version line — a sentence unchanged since v0.8 that had become false because other things changed — which is corrected in v0.9.4 and recorded as the eleventh entry in the COI ledger. The Japanese version is the authoritative original, and where the two diverge the Japanese governs. The English text has not itself been reviewed: the synchronization was checked mechanically (90 checks at the synchronization to v0.9.4 and 93 at v0.9.10, including a JA↔︎EN clause-correspondence layer), but the accuracy of the translation is unexamined — review of the translation was dispensed with by the registrant’s adjudication of August 11, 2026, on the condition that the English paper state the authoritative status of the Japanese. Full workflow records, including the synchronization log: verification/v09-reflection-workflow/.

The sibling paper to Addendum II (incorporated into Version B v4) of the Sixth Work, Why Military AI Cannot Be Aligned. It starts from the same epistemological foundation as Addendum II — that “ε=0” regarding the violation probability ε is a universal proposition over the input space, for which a single counterexample suffices to disprove it, but for which proof (the direction deployers require) is systematically blocked — and translates that foundation into the domain of everyday life and policy in general, not limited to the military. It presents a framework that redraws “where AI must not be involved” — not by field name (medicine, the military) but by structure (reversibility, absorbability, verification cost, replication correlation) — together with the user-side five articles.

Files Included

The recommended reading order is: the paper itself → the Companion Consideration (the accessible version, including sections closer to everyday life such as romantic relationships and loss of relationships) → the verification table (the audit trail).

Disclosure of Verification Level (Important)

v0.9.10 of this paper has undergone the following verification (the only changes since are the declaration in §9 and the v0.9.11 alignment of revision counts): (1) adversarial audits by the Claude series (first through final, four instances in total, up to v0.8); (2) six rounds in total of adversarial review by four Claude-family instances (Opus 5) in the v0.9 reflection workflow (three rounds on the correspondence map plus three rounds of difference-limited review, with verbatim preservation and machine checks before submission); and (3) three rounds of out-of-lineage review (non-Claude, Gemini 3.6 Flash) — the first round independently recomputed the statistics and checked five external references, and caught one item that none of the four intra-family reviewers had caught (the catastrophe-conditional denominator); in the second round the out-of-lineage reviewer itself withdrew one of its first-round points; and the third (final) round approved every item.

Still outstanding (stated in §8 of the paper; the foremost items on the standing post-publication correction shelf): independent cross-checking by human reviewers, independent cross-checking against the primary records, and review by a statistics specialist have still not been carried out. The out-of-lineage review, too, was carried out at the level of this paper and the addendum reports, without reaching the primary records. The workflow’s COI ledger records, on both sides, the drafter’s lapses six through ten and the reviewers’ errors 1 through 12 — what this workflow could show is a record that errors were actually detected, not a certification of zero errors. Verification, counterexamples, and corrections from readers and experts, regardless of direction, will be logged in the public record.

On Distribution and Use

License: CC BY 4.0 (the whole repository). You are free to share and adapt this material, including commercially, provided you give appropriate credit. Reproduction and adaptation for educational and support purposes is therefore already permitted — what Appendix G originally described as a license “to be attached” is satisfied, and exceeded, by this existing license (Appendix G was brought into line with the actual state in v0.9.10).

All four layers of derivative material are now published, each synchronized with v0.9.10 of the paper (the correspondence maps and machine-check records are in the process records):

The level of verification differs by layer. The paper has passed six rounds of intra-family review and three rounds of out-of-lineage review, but the classroom materials have never been tested in an actual lesson, the Easy Japanese version has been reviewed neither by readers in its intended audience nor by specialists in Japanese-language education, and the English translations have not themselves been reviewed (the Japanese is authoritative; where the two diverge, the Japanese governs). Check before you use — that is this paper’s Technique 1 and Technique 2 applied to this distribution itself.

This is the English version. Where the English and Japanese editions differ in content, the Japanese edition (README.md) is authoritative.