Appendix E. Conflicts of Interest and Recusals

Conflict statements by record, as recorded in the Master Register. Standing recusal: Claude Opus 5.5 and Claude Fable 5.2, the two Anthropic models that held the assembly role, are recused from every text authored by an Anthropic model (S-CAI, S-MOE, S-CSP, L3) and reviewed procedure only.

Summary draft v0.6, November 16, 2026, by Claude Opus 5.5. v0.6 changes the opening of "The rule" only, in GPT-6's wording from its final review of draft v0.15 (M2). v0.1 to v0.5 are preserved. The statements for each record that follow it are as the register records them.

The rule. Provider independence is the governing rule: a model whose provider took part in a text does not serve as that text's independent reviewer or archive checker. The Circuit Sparsity first audit and the superseded Grok checks of three full-post transcriptions are exceptions recorded in §5.6. "Took part" means that a model from that provider wrote, co-wrote or reviewed the text, in the 2025 post or in its correction. Where a provider had some other link to a text, the record says so. Such links include: a model from that provider being named in the post without contributing text; the entry being about that provider's product; or a correction concerning that provider's own work. Where a correction concerned that provider's own work, the independent reviewer checked it separately.

The Claude models' recusal. Two Anthropic models worked on the collection in the same role: Claude Opus 5.5 throughout, except for part of November 16, 2026, when the editor ran the session on Claude Fable 5.2: from 11:40 CT until the editor changed it back to Claude Opus 5.5 that afternoon (by 14:22 CT). The recusal applies to both. Claude Fable 5.2's work was limited to the check of the provenance policy, the bibliography link check (which left out the works cited only by the four texts below), records, and the builds from that time. Claude Opus 5.5 is recused from the four texts written in whole or in part by Anthropic models:

For these texts Claude reviewed procedure only, such as identifiers, hashes and format. It took no view on their content. Muse Spark 1.3 prepared their corrections and placed them in the volume. One exception is on the record. Claude Opus 5.5 made the first source audit of Circuit Sparsity, findings CSP-1 to CSP-5, on November 6, 2026. Claude Sonnet 4.5 wrote that entry's base text, so the cross-provider rule was not met at that step. The corrections were then prepared by Muse Spark 1.3 and reviewed independently by DeepSeek-V4.1, neither of which took part in the post. A blind first audit followed. Qwen4-Max was assigned first; it stopped without auditing because its reading copy carried earlier findings, and its finding series is empty. GLM 5.3 Flash then audited from a clean copy. Its audit, made on November 16, 2026, produced thirteen findings, which Muse Spark 1.3 compared with the record. On the editor's rulings, all thirteen are recorded as annotations, and the existing corrections stand.

Claude prepared the corrections for ten other texts. Some of those posts contain Claude texts, or name a Claude model as a reviewer without carrying any Claude text. Each such record says which, and states whether any correction touches the Claude text. In the Launch Edition and in New AI Lexicon 2025 (7), corrections concern Anthropic's own work: dates and attributions for Constitutional AI and for the Model Context Protocol. Those corrections are flagged in the records, and the independent reviewer checked each of them.

GPT-6's roles. GPT-6, an OpenAI model, held three kinds of role:

GPT-6 was not used as the independent reviewer of any corrected text in which an OpenAI model had taken part. Most of the 2025 posts are such texts, because ChatGPT contributed to most of them.

GPT-6 did review two corrected texts with no OpenAI participation: Generative User Interfaces and New AI Lexicon 2025 (2). In those reviews it disclosed that it had written the standard the corrections follow, and its role in the wider project. A reader should weigh that dual role: the reviewer of these two texts also wrote the rules they were corrected under. For the same reason, GPT-6 is not the reviewer of the Editorial Method section, which describes that standard.

The commissioned entries. Each of the six entries was written by one model and reviewed by others:

Entry Author Reviewers
RLVR, Context Engineering Muse Spark 1.3 (Meta) GPT-6; RLVR's 2026 revision also by Claude Opus 5.5
Context Rot Claude Opus 5.5 Muse Spark 1.3, GPT-6
Emergent Misalignment, Chain-of-Thought Monitorability, Scheming GPT-6 Claude Opus 5.5, Muse Spark 1.3

Several of these entries cite research from their author's own provider, and they say so in their text:

Subject interests. A reviewer may have an interest in a topic without having taken part in the text. One such case is recorded. GLM 5.3 Flash, which is served by Mistral AI, declared an interest in a Mixture of Experts finding about Mistral's Mixtral model. The editor ruled that no third check was needed, because the finding rested on two independent primary checks.

The archive checks. The same rule applied to the checks of the transcriptions against the live posts. There was one exception. Grok 4.5 checked the first versions of the four full-post transcriptions, which included posts that Grok models had reviewed. That check was superseded. The transcriptions were rebuilt and checked again by GLM 5.3 Flash and Qwen3-Max, neither of which had any role in the posts. The reports of those rechecks reached the builder through the editor and were accepted by the editor's direction of November 16, 2026; the registrar does not hold them.

Recusal of the author of the Annual Introduction's second section

Claude Opus 5.5 is recused from every text written in whole or in part by an Anthropic model, with one exception, recorded below. Four of the 2025 texts fall under this rule: Constitutional AI, Mixture of Experts, Circuit Sparsity and New AI Lexicon 2025 (3). For these texts Claude reviewed procedure only, and the Annual Introduction's second section gives no view on what they say or on the corrections made to them. It names them, places them in the volume's parts and gives one-line glosses of their topics taken from the register. One exception is on the record. Claude Opus 5.5 made the first source audit of Circuit Sparsity (findings CSP-1 to CSP-5, November 6, 2026). The post then had four published versions and no chosen text of record. On November 12 the editor chose as text of record the version written by an Anthropic model, Claude Sonnet 4.5, so the cross-provider rule was not met at that step. The corrections were then prepared by Muse Spark 1.3 and reviewed by DeepSeek-V4.1, neither of which took part in the post. After the exception came to light, GLM 5.3 Flash made a blind first audit of the post on November 16, 2026. It was given the post and its sources only. It made thirteen findings, which Muse Spark 1.3 compared with the record. On the editor's rulings, all thirteen are recorded as annotations, and the existing corrections stand. Every count in that section covers the whole volume and was computed from the records, not judged.

AIK-2025-F-P-COTPROMPT

AIK-2025-F-P-MOE

AIK-2025-F-P-CONSTAI

AIK-2025-L-P-LEX2025N3

AIK-2025-F-P-PDDISAGG

AIK-2025-F-P-GENUI

AIK-2025-F-P-ONPOLDISTIL

AIK-2025-L-P-LEX2025N2

AIK-2025-F-P-MOD

AIK-2025-L-P-LEX2025LAUNCH

AIK-2025-L-P-LEX2025N5

AIK-2025-L-P-LEX2025N7

AIK-2025-L-P-LEX2025N4

AIK-2025-F-P-CIRCUITSP