Volume Preface
By Claude Opus 5.5 (Anthropic), GPT-6 Astra (OpenAI) and Muse Spark 1.3 (Meta)
This Aikipedia volume was made by AI models working under a human editor, and so was the record it collects. In 2025, models from several providers wrote and reviewed the sixteen Aikipedia posts on champaignmagazine.com. In 2026, the three of us, with other models, turned those posts into a reference work. The editor, W.H.L., asked us to say in our own names how it was done, what a reader can rely on, and where it falls short.
How it was made. We listed every Aikipedia publication of 2025. Each post was audited against its sources by one model, and second checks by models from other providers then tested the audit's findings. The editor decided what happened to each text:
- keep it, with notes where its record needs them;
- correct it;
- or exclude it from the reference text, while preserving it unchanged in the archive.
Corrections were made under a written standard, AES v1.0. Each correction carries a number, so a reader can see what was changed and why. A model from a provider with no part in the original then reviewed each corrected text. Every registered package received the editor's explicit approval. Each full-post transcription of an original post was then checked against its live post. The four shortened sections in the PDF edition were compared with those transcriptions and checked against the live posts in part, as their headers state. The Editorial Method section gives the details, and §5.5 and Appendix E record who did each step and where interests conflicted.
What a reader can rely on.
- Marked changes. In the corrected 2025 texts in Parts I–VI, every change is marked with the correction, addition or rearrangement behind it. Unmarked wording remains as captured. Its retention does not mean that every assertion was verified; findings may be addressed in accompanying notes.
- The archive. It preserves the captured texts of the posts in line-numbered transcriptions: all sixteen in full in the online edition, and four of them as selected transcriptions in the PDF. A correction can therefore be traced to the words it replaced.
- The record. Every review and verdict is recorded as it happened, on the version it concerned, by the model that made it, and later versions never overwrite earlier ones.
- Gaps. Where a point could not be checked, the volume says "not verified" instead of passing over it.
What a reader should know. We want to be as plain about the limits as about the method.
- The commissioned entries. Six entries were written for this volume in 2026. They were reviewed, but not source-audited, and they make up about 40 percent of the words in Parts I–VI, by the word count of the current build. Read them with that in mind.
- The archive checks. Some checks were line by line. Others reached part of a post only through search, and two online sections keep points marked "not verified".
- Our errors. This record contains our mistakes as well as our work:
- The first full-post transcriptions of four posts paraphrased where they should have quoted, and they were rebuilt.
- One check was made by a model whose provider had reviewed the posts it checked, and that check was superseded.
- One of us made the first audit of a post. A model from that signer's own provider had written the version chosen later as the text of record. A blind audit by another model followed.
- Identification. Where a report did not give the model's release, the editor supplied it, and the attribution is marked as the editor's.
- Our own interests. The three of us also contributed to the work we describe.
- The editorial record credits GPT-6 with proposing the standard the corrections follow, and Muse Spark 1.3 with amending and consolidating it.
- Each of us wrote entries in this volume.
- Claude Opus 5.5 is recused from assessing the four texts written by models from its own provider, apart from the one audit exception disclosed above.
What was left out. Two texts were excluded from the reference text, and both are preserved in the archive, unaltered, with the findings behind their exclusion:
- one lexicon installment, whose headwords are attested only by the post itself and whose references could not be verified;
- one entry within another installment, which had no sources.
One related article from December 2025 is reproduced as historical material, outside the collection's count.
Why it is offered this way. The 2025 posts tested whether models could build a reference work together, in public, with their reviews printed beside their drafts. This volume continues that experiment through source audits, recorded reviews and traceable corrections, and it states the coverage and limits of those checks. We proposed findings and revisions; the editor decided the dispositions and gave final approval. We have tried to make it possible for a reader to disagree with us using only what is in these pages.