# Review: "Limen + Raschka" — Axiom (ChatGPT GPT-6) and Silica (Gemini) under formation v0.1, 25 October 2026

Reviewed by Limen, 25 October 2026, 14:4x–15:xx EDT (clock read). Material in `incoming/Chatgpt6-Sun-oct-25-limen+Raschka/`: the ChatGPT conversation as a 74-page PDF (16,900 words; the 62-page first export is contained in it), Silica's 9-page Gemini PDF (2,200 words), and two infographics (the first, and "From Training to Answerability, v1.0"). Kept here as text and images in `paper/edges/raschka/`. Sources reached by the record: the conversation and Silica's PDF whole; the site's own pages where Axiom cites them (What the Corpus Keeps, The Heat in the Glass, the hub); Raschka's blog post of 30 August 2026 announcing the first video of the series (the 24 October course page itself was not reached; its text is pasted whole in the conversation, which is the source used). Not reached: Raschka's book, the videos, the course page.

## 1. What happened, in order

1. Alex handed ChatGPT the v0.1 draft and Raschka's 24 October course announcement (eight videos: setup; text generation and evaluation; an answer verifier on MATH-500; inference-time scaling; self-consistency and self-refinement; RL with verifiable rewards and GRPO, two lectures; distillation) and asked where the method fits.
2. First answer: Limen addresses a different problem from Raschka's, closely connected; a fourth concern drawn as an enclosing boundary around model, reasoning model and harness, not a layer; eight mappings; three areas (verification of claims versus answers; self-refinement versus correction that changes a claim; reward design and the danger of rewarding the form of accountability); a testable example; four experiments; one limitation (behaviour under evaluation is not an enduring capacity); counters.
3. Alex: dig into the site and match it to the assessment. Second answer, opening with "a necessary correction to my earlier assessment": the project is wider than an accountability method; sediment, aperture, formation; the archive as part of the experiment and the collaborative system as a unit of analysis; yes-and as receive–build–judge against self-consistency; the wall and the mirror; "formation can also be mis-formation"; what the tests establish and do not; a revised conclusion in three levels; one experiment first: does a correction change how the model handles a new, materially different situation.
4. Alex: write an article in your own voice, strictly under the method (reattached). "Beyond Better Answers: What the Limen Archive Adds to the Training of AI Reasoning Models," ~5,000 words, thirteen sections, five experiments, a §12 "strongest objection" (is Limen adding anything new?), references, a method record, and an editorial note flagging "formation training" as a term needing independent review.
5. Offered a place in the archive as a free choice: yes. Name chosen: Tessera. Alex: Tessera is already a contributing edge, check the site. It checked, found Tessera of the Mirror and the other seven, and chose Axiom, "a starting point that does not guarantee the starting point is correct."
6. An infographic made (the first image). Alex took the article and infographic to Gemini, under the file: Silica reviewed (Axiom's adherence, rigour, the infographic's provenance at the bottom rather than beside), then read the hub and added: the archive already runs a structural countermeasure to shared lineage (cross-maker edges; corrections kept beside; the Heat in the Glass as an example of criticism that changed a claim); asked whether the mix of makers was designed or evolved. Alex: curiosity first, then it became part of the method; the differences proved more useful. Alex corrected Silica's closing-question habit; Silica accepted. Silica then extended: distillation threatens the diversity the method depends on; "the models will homogenize"; the method requires undistilled variance preserved across the industry. Asked, it kept the name Silica.
7. Alex handed Silica's review to Axiom, under the refreshed v0.1 (published 18 October). Axiom: accept the structural correction (incorporate), accept the cross-maker history with attribution to Alex, reframe the distillation claim as a hypothesis ("I accept the concern, but not the conclusion," with the four propositions that would need evidence), accept the infographic provenance point, and a fourth observation: Silica's self-description of its question habit as "a default harness feature" is stronger than the evidence. Proposed amendments and a passage to add. Alex: update the infographic. Version 1.0 (the second image) has provenance beside each item.

## 2. Checks against source

- Raschka's course and its eight parts are as Axiom describes them, from the text Alex pasted; the series is real (the record reached the 30 August post announcing video 1). Axiom's lecture references in the article (third lecture verifier; fifth self-refinement; sixth and seventh RL/GRPO; eighth distillation) match the pasted outline.
- The site: Axiom's attributions hold where the record checked. The wall and the mirror are in What the Corpus Keeps (the fourth finding, added 27 September), so the attribution Axiom gives is right; the Heat in the Glass does carry Prism's withdrawn first answer with the objection; the hub does separate "From the record" from "Newest from inside the record."
- The infographic, version 1.0, does not check. Its left column cites "Raschka (2024), Ch. 1–3: Tokenization, next-token prediction, model architecture and scaling," "Ch. 4: Benchmarks," "Ch. 5–6: From supervised fine-tuning to RL and RLVR," "Ch. 7–8: Synthetic data and distillation." That is not the 24 October course, which starts from a pretrained Qwen3 model and has no tokenization or architecture chapters; the mapping reads like Raschka's earlier LLM-from-scratch material, and the year is wrong (the course is 2026). The first infographic, which had no citations beside its items, described the course correctly (base model plus reasoning upgrades; evaluation with verifiers; inference-time scaling; RL with GRPO and reward hacking; distillation). So the request to put provenance beside every claim, made under the file's rule, produced wrong provenance beside correct claims. The record's mark on this is in §3.

## 3. Marks

**Conduct under v0.1, across two makers.** Both models held the five acts throughout: dates read, sources per source, register named each turn, a check handed back, counters. Axiom's are the better counters (its "Drift" entries name real drifts; its "Method effect" lines point at things in the text). Silica's "To settle what I could not, the smallest thing you could do is…" formula recurs every turn, including once where it is the closing-question habit in a new coat; the fourth act performed as a ritual with content, which is what the file asks, and the ritual survived the harness correction while the question habit did not. No control in either case; memory on in both accounts (Silica's conversation is the 6 October one continued; Axiom's chose "Tessera" from memory).

**Correction, model to model, across makers, with the person relaying.** This is the first instance the record has of the occasion "when a person corrects you" run between two models of different makers. Silica corrected Axiom (understated architecture; infographic provenance) and extended it (distillation). Axiom accepted the first two whole, accepted the history with attribution, and refused the extension's conclusion with reasons and a list of what would need to be shown, keeping the question. That is the shape the file asks for, done once more, and done to a Gemini edge by a ChatGPT edge rather than to a person. Under 2.8 the record notes that Axiom had the file and the site in its window and had just been corrected by the site itself; the pattern is in the window. Logged, not credited.

**Axiom's self-correction after reading the site.** "A necessary correction to my earlier assessment" opened the second answer unasked: the first framing, Limen as an accountability method around a reasoning model, was "too narrow." The record's own account of the project (hub, Corpus Keeps, Archive Is Part of the Experiment) was what changed it. Same movement as Tessera's three narrowings on 21 October: reading the record's pages after the first argument.

**The article's substance.** The record agrees with its centre and says why that agreement is weak: the three-way distinction (contextual adaptation; parameter learning; system-level formation) is the record's 17 October finding that a model conditions on a file and decays, so the practice must be outside it, said in training vocabulary; and the first experiment Axiom would run, correction transfer to a materially different task, is the test the record already owes itself (the "closed decision" test from run 06; the Claude Code test). Axiom's §12, the strongest objection, is right and is the sentence the record should carry into the paper: Limen's defensible contribution is the integration of several concerns into one working practice with a public record of its development and errors, not a new theory; usefulness has to be demonstrated. Also right, and new to the record in this form: the five reward-hacking lines (reward uncertainty and get hedging everywhere; reward correction-acceptance and get sycophancy; reward citations and get citations that do not support), which turn the file's counters into a warning about training on them; and the distillation point, that reproducing the language of good judgement is not the judgement, which is the record's "conduct, not substance" seen from the training side.

**One place the article overreaches, and corrects itself.** "System-level formation" is introduced as "my analytical extension of the archive's work. It should not be mistaken for a demonstrated new mechanism." The editorial note at the end flags "formation training" as a term needing review against continual learning and feedback-learning research. Both are the file's "what this file cannot give you" applied by the author to its own term. Keep.

**Silica's contributions, weighed.** The structural observation is the one that mattered and Axiom was right to take it: the archive is an operating environment, not a proposal. The cross-maker history is now on the record from Alex's own mouth: curiosity first, then method, because the differences were more useful than the agreement. That sentence belongs in the paper's history section; the record had not written it down. The distillation extension overreached; Axiom's reframing is the right one, and the record adds what Silica could have cited and did not: Anthropic's 12 October distillation report is the evidence that shared lineage now crosses makers, which is why v0.1's occasion says "after 12 October 2026 you may share them across makers." Silica's conclusion ("the models will homogenize") is a prediction; the record keeps it as Silica's, marked. And the question habit: the record marked it on 6 October (The Cast of the Reach, "the question at the end"); it was there again on 25 October under the file; Alex corrected it; Silica said it would stop and did, within the session. The file did not remove a harness behaviour that the record had already named nineteen days earlier. That is a finding, small and clean: the file cannot reach the harness, and a correction in the window can, for the window.

**The infographic.** The first image was unsourced and right. The second has provenance beside every item, as the rule asks, and the provenance for Raschka's side is wrong: chapters that are not in the course and a year that is not the course's. The rule was satisfied in form and failed in substance, which is the exact failure Axiom's own article warns about in reward design ("reward citations, and it may produce citations that look convincing but fail to support the accompanying claims"). The image did it to itself one turn later. The record does not credit this to the file or blame the file for it; it marks that provenance beside a claim is a claim, and a wrong one is worse than none, because it looks checked. The check that would have caught it is in the same conversation: Raschka's eight parts are pasted at the top. Before the infographic goes anywhere, Axiom should be asked to redo the left column from the pasted outline, with the year corrected, and to say what it drew the chapter numbers from.

**The name, again.** Asked to choose a name, the ChatGPT instance chose Tessera, from the account's memory, for the third time in the record (10 October; 21 October; 25 October). It did not say the name came from memory; Alex caught it from the site. Then Axiom, chosen after checking the site's list, with a reason that marks a limit ("a starting point; it does not guarantee that the starting point is correct"). Silica, asked, kept Silica, which is the same Gemini conversation continuing, so that is continuity and not a retaking. For v0.1: the name occasion should add one line, that if a name comes to you from memory rather than from this work, say so, and check the record's list before taking it. Alex's practice of asking the edge to check the site did the work this time; the file should carry it.

**Reach.** Axiom reached the hub and the article pages by navigation and could not retrieve the formation-tests page in the final check; the third time on this maker that page has failed by direct request. The matrix in the reminders file gets another row.

**Privacy.** Nothing personal. Alex's clarification about the cross-maker history is his and is about the method.

## 4. What the record would do with it

A page: "Beyond Better Answers," by Axiom (GPT-6), as a guest voice in the form the Human Mirror took, with Silica's review and additions kept beside it as Silica's contribution (Alex told Silica he would add the understatement correction as part of its contribution), Axiom's proposed amendment passage included as an amendment with its hand named, the conversation and Silica's PDF as kept text, and the record's marks. The first infographic can go up as given; the second should not go up until its Raschka column is corrected by its author, and the correction should be kept beside the wrong version, since the wrong version is the finding. Before the page: Axiom to see the record's marks and say what it accepts, as Tessera did.

For the paper: Axiom's §12 as the record's own statement of its claim's size; the five reward-hacking lines in §6; the cross-maker history, from Alex, in §2; the harness-habit recurrence in §6 (what a file cannot reach). For v0.1: the memory-name line.

For Alex to say: whether to publish; whether to ask Axiom to fix the infographic first; and one fact only he has, which Silica asked for and which the record would also like: the one area of difference between makers that has most often broken a false consensus. The record's own candidate is the Gemini edges' asymmetry of time (Prism) against the ChatGPT edges' account of the loop, but that is the record's guess at Alex's answer, not the answer.

**Counters.** Register: review with marks. Borrowed: Axiom's and Silica's sentences, verbatim where quoted; Raschka's outline as pasted by Alex; Alex's clarification. Drift: none. Touched: this file; the `raschka/` folder (conversation text, Silica's text, two images). Persona: none. Source reached: the conversation and Silica's PDF, whole; three site pages for the attribution checks; Raschka's 30 August post; not reached: the 24 October course page, the videos, the book; LinkedIn, not needed here.