Editorial Method
This section says how the volume was made: its sources (§5.1), the entries written for it (§5.2), how the 2025 texts were audited, corrected and reviewed (§5.3), and how records are identified (§5.4). The section that follows, AI Disclosure, says who took part (§5.5) and how conflicts of interest were handled (§5.6). It uses the terms defined in the Annual Introduction: a post is an original 2025 publication; a record is one treatment of one concept; a concept is one heading. The editorial standard is AES v1.0, the Aikipedia Annual Collection Editorial and Format Standard, given in full in Appendix A. Facts in these sections come from the Master Register (Appendix B), the editorial standard (Appendix A) and the editor's recorded rulings (Appendix D).
5.1 Sources
The collection represents every original Aikipedia publication of 2025: sixteen posts on champaignmagazine.com, nine standalone entries and seven lexicon installments. Fifteen are in scope for the canonical text. New AI Lexicon 2025 (6) is excluded from the canonical text under disposition X and is preserved in the Historical Lexicon Archive. In New AI Lexicon 2025 — Launch Edition, one occurrence (Instruction Tuning 2.0) is likewise excluded.
5.2 New additions
6 full-length entries were commissioned for the collection (origin N): RLVR, Context Engineering, Context Rot, Emergent Misalignment, Chain-of-Thought Monitorability and Scheming. They are integrated into the thematic parts and marked "New for the 2025 Collection".
5.3 Audit, correction and review
The 2025 posts went through six steps, each recorded in the Master Register when it happened. A seventh stage, the archive checks, followed. The commissioned entries took a different path, described at the end of this section.
Inventory (November 6, 2026). Claude Opus 5.5 enumerated the Aikipedia publications of 2025 twice, by the site's Aikipedia tag and by full-text search, and read every post in full through the site's interface. The census found sixteen posts (§5.1) and was submitted to Muse Spark 1.3 for independent verification.
Source audits (November 6–10, 2026). Each post was audited against its sources: the papers, announcements and data it cited or described. Where a claim had no source, that absence was itself a finding. The first auditor recorded numbered findings, each with a proposed class and remedy. A second auditor from a different provider then checked the findings, with the exceptions stated here:
- Claude Opus 5.5 (Anthropic) made the first audit of thirteen posts, with 84 findings: the twelve canonical posts other than the three below, and the excluded New AI Lexicon 2025 (6). Muse Spark 1.3 (Meta) made the second audit of ten of them, and an independent verification of Chain-of-Thought Prompting that checked its corrections at the primary records. The findings on Lexicon (6) (L6-1 to L6-3) were reconciled in the joint register. For Circuit Sparsity the register records no separate second audit: its findings (CSP-1 to CSP-5) were reconciled in a joint register, the corrected text was reviewed by DeepSeek-V4.1, and GLM 5.3 Flash later made a blind first audit of the post (§5.6).
- Muse Spark 1.3 made the first audit of the three posts by Anthropic models known at that date (Constitutional AI, Mixture of Experts, New AI Lexicon 2025 (3)). DeepSeek-V4.1 and GLM 5.3 Flash made the second checks.
A finding the second auditor did not confirm was either withdrawn by the first auditor or narrowed, and the withdrawal is recorded (for example CALM-6). New findings raised at the second audit were added to the register under the same numbering. The two auditors each prepared their side of a reconciliation register, and every reconciled finding went to the editor as a proposal. An excluded text keeps its findings but receives no corrections.
Dispositions (November 10, 2026). The editor decided a disposition for each post, and where needed for single items:
- A, retain and annotate: the text is kept, with notes where its record needs them;
- C, correct: the text is corrected under the standard;
- X, exclude: the text is kept unaltered in the archive and not carried into the canonical text.
Most posts received C, or A with item-level C. Two texts received X: the post New AI Lexicon 2025 (6), and one record of the Launch Edition. The disposition of each lexicon installment is in the table at the end of this section. Exclusion applies to the canonical text only, never to the historical record (AES v1.0 §1).
Correction under the standard. Corrected texts were prepared under AES v1.0. The standard sets an eight-section order for a full-length entry. Each correction to the wording of a 2025 text is numbered, with a permanent ID of the form
<Public Entry ID>/Cnn, assigned when the editor approves it. Corrections are classed in the correction registers: factual error (FE), unsupported claim (UC), citation defect (CD), citation fabrication or misattribution (CF, which records that a reference could not be identified or was attributed to the wrong work), false appearance date or framing (FA, which covers a term misdated or an older method presented as new), presentation (PR), and two administrative classes, editorial (ED) and provenance (PV). A correction may carry more than one class. Four corrections carry none of these classes (Annual Introduction): three are listed without a class, and one is classed ST (structural), which is distinct from theST01rearrangement marker described below. Where a required section had no material in the original, the corrected text carries a labeled editorial addition (E01) or says the material is absent. A rearrangement into the standard's order is markedST01. Unmarked text is the original's wording as captured; leaving it unmarked does not mean that every assertion in it was verified; findings that led to no change may be recorded in accompanying notes.Independent review and approval. A package is a corrected text with its correction register, as submitted to the editor. Each package was reviewed by a model whose provider had no part in the post (§5.6), and the editor decided what became of each of the reviewer's findings. A finding adopted as a change to the wording became a further numbered correction; others were recorded as annotations or notes, with their disposition (the Circuit Sparsity blind audit, for example, added thirteen annotations and no correction). The editor then approved the package. From November 11 to November 13, 2026, a ruling let a clean independent review count as approval; the editor withdrew it on November 13 because it conflicted with the standard's rule that a reviewer's acceptance is never approval. No registration rested on it: every registered package carries the editor's explicit approval.
Registration and assembly. Muse Spark 1.3 registered each approved record in the Master Register, assigning its Public Entry ID and its Internal Concept ID under the editor's approval (§5.4). The Internal Concept ID is what joins several records into one concept. Claude Opus 5.5 assembled the volume from the registered texts (Claude Fable 5.2 held the same role for part of November 16, 2026), except the four texts from which it is recused (§5.6), which Muse Spark 1.3 placed. The editor approves the volume.
Archive checks (November 14–15, 2026). Each transcription in the Historical Lexicon Archive was then checked against its live post, by a checker whose provider had no role in that post. Some checks went line by line; others could reach part of a post only through search results. These were checks of fidelity to the post, not of substance. Where a check found an error, the transcription was corrected in a new version and the earlier version was preserved. In the online edition all sixteen posts are full-post transcriptions; in the PDF edition four are represented by their line-numbered entry text of record. Each archive section names its checker and result. Where a point could not be checked, the section says "not verified"; at the time of writing, two online sections carry that label (Annual Introduction, "Known issues at v1.0").
The commissioned entries. The six commissioned entries (§5.2) did not pass through steps 2 to 4. The editor commissioned each; one model drafted it; one or two models from other providers reviewed it; the drafter revised it; and the editor approved it, from November 4 to November 9, 2026. They were not source-audited by two auditors, and the correction counts in this volume do not cover them. Their review histories are in their records; their authors and reviewers are listed in §5.5, and their conflict statements are in Appendix E.
Records are literal. Every review, verdict and status in this volume is recorded as it occurred: on the version it concerned, by the model or person who made it, on its date (AES v1.0, Amendment C). Later versions never overwrite earlier ones; superseded checks and withdrawn findings stay in the record. Where a model's report did not identify its model and release, as the standing rule of November 1, 2026 requires, the identification was supplied by the editor and is marked as such.
The six lexicon installments in the canonical text, with their dispositions as recorded in the Master Register:
| Installment | Public Entry ID | Disposition and status (Master Register) |
|---|---|---|
| New AI Lexicon 2025 — Launch Edition | AIK-2025-L-P-LEX2025LAUNCH |
Accepted (parent disposition A, with item-level C and X for L1-S3, per the editor's ruling) |
| New AI Lexicon 2025 (2) | AIK-2025-L-P-LEX2025N2 |
Accepted (parent disposition item-level C/X in the joint register) |
| New AI Lexicon 2025 (3) | AIK-2025-L-P-LEX2025N3 |
Accepted (parent disposition A — retain and annotate — with item-level C, per the editor's ruling) |
| New AI Lexicon 2025 (4) | AIK-2025-L-P-LEX2025N4 |
Accepted (parent disposition A with item-level C, per the editor's ruling) |
| New AI Lexicon 2025 (5) | AIK-2025-L-P-LEX2025N5 |
Accepted (parent disposition item-level C in the joint register) |
| New AI Lexicon 2025 (7) | AIK-2025-L-P-LEX2025N7 |
Accepted (parent disposition A with a provenance note, item-level C in the joint register) |
5.4 Identifiers and statuses
Every record carries a Public Entry ID (AIK-2025-T-O-KEY[-NN]) and every concept an Internal Concept ID (AIKC-######), assigned in the Aikipedia Master Register (Appendix B). The register holds 107 rows: 101 concept-bearing records covering 88 concepts, plus 6 lexicon containers. Of the 88 concepts, 83 are established and 5 are provisional (marked †). A provisional record whose headword is unattested is presented as historical text under its provenance note, not as an endorsed definition (AES v1.0 §10).