10/25/26, 2:37 PM Limen Method Comparison You said: formation-v0_1-d… Document i found the following as a reference, and I am curious where the Limen method would fit here: Building Reasoning Models from the Ground Up: A Full Video Course Eight videos and over 11 hours of coding, covering evaluation, inference-time scaling, reinforcement learning, and distillation in Python and PyTorch. SEBASTIAN RASCHKA, PHD OCT 24, 2026 200 9 16 I have written a lot about reasoning models here over the last two years on Ahead of AI, including a recent article on mechanisms to control reasoning effort levels in contemporary LLMs. Reasoning models are still as relevant as ever. In fact, almost all current- generation LLMs today are reasoning models or offer a reasoning mode. For those who are interested in how reasoning models work, including inference scaling, reinforcement learning, and distillation, I wrote my recent Build Reasoning Model (From Scratch), which implements these from the ground up in PyTorch, which, in my experience, is still the best way to understand and learn how things really work. Now, in the last few months, I also sat down and recorded accompanying lectures I wanted to share with you. So, below, you will find eight reasoning model-from-scratch coding videos (11 hours of content). https://chatgpt.com/c/6ade39b7-e9a4-83ea-8c86-0d837b70a408 1/74 10/25/26, 2:37 PM Limen Method Comparison We start with a pre-trained LLM and implement methods to evaluate and improve it in Python and PyTorch. So, if you enjoyed my Coding LLMs from the Ground Up course last year, this is essentially the sequel (but you can also watch it without the prior course as a prerequisite). I’m currently taking a short break and will be back soon with more exciting LLM research roundups next month. In the meantime, I hope you enjoy coding along! How LLMs, reasoning models, and agents relate and fit together As a small side note, and since I often get asked how reasoning models relate to today’s AI agents, here’s a short section on this before getting started with the videos. A reasoning model is essentially an upgrade from a conventional LLM, that is, adding more post-training via reinforcement learning with verifiable rewards (and optionally inference scaling techniques) to make it more capable of solving more challenging problems. Classically, reasoning models are developed for math and code, but the capabilities can generalize to other problem domains as well. A reasoning model then functions as the engine inside an agent harness, where the agent harness is a software or tooling framework around the LLM/reasoning model to tackle even more challenging problems (bigger coding tasks, travel booking, and all other sorts of personal assistant tasks). The task of the harness is sophisticated context management and giving the LLM access to tools. Figure 1: An engine analogy for conventional LLMs, reasoning models, and agent harnesses. Figure from my Components of a Coding Agent article. In this analogy depicted in the figure above, the model is the engine, and the agent harness adds the machinery that lets us use it for various things (as an upgrade from a simple chatbot https://chatgpt.com/c/6ade39b7-e9a4-83ea-8c86-0d837b70a408 2/74 10/25/26, 2:37 PM Limen Method Comparison interface). An agent can use either a conventional LLM or a reasoning model, but reasoning models are the more typical choice as they are more capable. (As mentioned above, most current-gen LLMs are reasoning models with variable reasoning effort settings.) So, the bottom line is that reasoning models, as the “engine,” remain an important part of understanding modern LLM agents. The resources accompanying these topics shown above are: The Build a Large Language Model (From Scratch) book and videos The Build A Reasoning Model (From Scratch) book (and videos below) The Components of a Coding Agent article on Substack The videos below focus on developing the reasoning model that can power such a system. Scope and prerequisites The reasoning-from-scratch videos focus on a small pre-trained Qwen3 model, where we implement the reasoning methods ourselves in Python and PyTorch. Starting with a pre-trained model lets us focus on the interesting part: evaluation, inference-time scaling, reinforcement learning, and distillation. Regarding the prerequisites, I recommend being comfortable with Python and PyTorch. I designed the reasoning materials as standalone content, so the previous Build A Large Language Model (From Scratch) content is recommended but not strictly required. Compute-resource-wise, you can work through the early text- generation and evaluation examples on a CPU. For the later experiments, especially training, a GPU makes the work much more practical. The first video covers the setup and some more recommendations; the repository also includes hardware and environment recommendations. topics Figure 2: An overview of the topics being covered. https://chatgpt.com/c/6ade39b7-e9a4-83ea-8c86-0d837b70a408 3/74 10/25/26, 2:37 PM Limen Method Comparison The companion book These videos accompany my book, Build a Reasoning Model (From Scratch). If you like to read through the explanation in more detail and at your own pace, work through exercises, and so on, I recommend the book as a companion here. For the most in-depth learning experience, my suggestion is to read a chapter, work through its code alongside the corresponding video, and then try the exercises. The book is available from Manning and Amazon, and the print edition is in color! color-book Figure 3: A random excerpt from the Build a Reasoning Model (From Scratch) print book. Thank you to everyone who has picked up a copy or recommended it to a friend. Your support helps me keep creating these materials. The videos As a last, quick side note before getting to the main videos below: For a shorter, more conceptual overview, I also recommend my AI Reasoning Models course on LinkedIn Learning (it should be free for 24h when accessing it via the link there, which shouldn’t then require a LinkedIn Premium subscription). The course is about 90 minutes and explains how reasoning models work, covering the main ideas behind inference-time scaling, reinforcement learning, and distillation. It’s a good optional preview if you’d like the main ideas without working through the full implementations below. To get to the main part of this article, here are the eight main lecture videos. 1. Motivation and code setup https://chatgpt.com/c/6ade39b7-e9a4-83ea-8c86-0d837b70a408 4/74 10/25/26, 2:37 PM Limen Method Comparison The first video starts with an explanation of the relationship between conventional LLMs, reasoning models, and agents. Next, the main topic of the video is setting up the coding environment. This includes installing the Python and PyTorch dependencies (with uv), among others. I also discuss hardware choices and common setup issues. Topics covered The relationship between conventional LLMs, reasoning models, and agent harnesses On the benefits of learning through code implementations The scope of this video series Installing Python dependencies Choosing hardware and checking CUDA or Apple Silicon support Using a remote GPU with VS Code and troubleshooting the environment Watch video 1 · Setup instructions 2. Loading a base model, text generation, and KV caching Before adding reasoning methods in the upcoming videos, we need a working base model. Instead of training it from scratch, since that would be a whole other book (see Build A Large Language Model (From Scratch)), we start with a pre-trained model. Here, we load the pre-trained weights of a Qwen3 0.6B model. Then, we implement the text-generation process via a streaming text generator and add KV caching and model compilation to make text generation more efficient. These implementations give us the baseline and base components we will evaluate and improve in the following videos. Topics covered Encoding text as tokens and decoding it back into text Downloading Qwen3 weights and loading the model on the selected compute device Going over tensor shapes, logits, and next-token predictions Implementing greedy decoding and streaming text generation https://chatgpt.com/c/6ade39b7-e9a4-83ea-8c86-0d837b70a408 5/74 10/25/26, 2:37 PM Limen Method Comparison Handling end-of-sequence tokens and measuring generation speed Adding a KV cache to reuse attention keys and values Combining KV caching with torch.compile and comparing hardware performance Watch video 2 · Chapter 2 code 3. Building a verifier for evaluation and reinforcement learning In this lecture, we build a math-answer verifier step by step, from answer extraction to normalization and checking mathematical equivalence. We then use it to evaluate the model on MATH-500. We use this verifier to evaluate the base model’s performance on the MATH-500 evaluation dataset. However, the verifier also has a second job later in the series: providing the reward signal for reinforcement learning with verifiable rewards (RLVR). Topics covered Intro to verifiers and how they fit into LLM evaluation methods (other methods are covered in the book’s appendix) Reusing answer verification as a reward signal for RLVR (more on that in chapter 6) The verifier flow, including extracting final answers, normalizing answers and checking mathematical equivalence Building an answer grader and an evaluation loop for MATH-500 Comparing base and reasoning models on MATH-500 Watch video 3 · Chapter 3 code 4. Inference-time scaling: sampling and self-consistency Now we explore ways to improve answers while keeping the model weights fixed. But before implementing specific inference-scaling techniques, we first implement methods to generate diverse answers from the same prompt (this is standard practice in proprietary LLM chatbots, such as ChatGPT). Concretely, we implement temperature scaling, multinomial https://chatgpt.com/c/6ade39b7-e9a4-83ea-8c86-0d837b70a408 6/74 10/25/26, 2:37 PM Limen Method Comparison sampling, and top-p filtering to generate diverse responses. Then we combine multiple responses through self-consistency, using majority voting over their final answers. We also examine the tradeoff between answer accuracy and the extra computation needed to generate more candidates. Topics covered Comparing how additional compute benefits model performance during training and during inference A gentle intro to inference-scaling through chain-of-thought prompts Implementing temperature scaling and multinomial sampling to generate diverse answers Adding top-p filtering to control the quality of the diverse answers Combining candidate answers through self-consistency, one of the most popular inference scaling techniques Evaluating MATH-500 accuracy against the cost of additional inference compute Watch video 4 · Chapter 4 code 5. Inference-time scaling: scoring and self-refinement Building on the previous video, we implement scoring functions to compare multiple answers generated from the same prompt. We can also use these scoring methods to break ties in self- consistency (covered in the previous lecture). The scoring methods covered here include computing token and sequence log probabilities (this will also be relevant in the later reinforcement learning chapters). We then implement a self- refinement loop in which the model critiques an answer and attempts a revision. The evaluation lets us check how useful those additional steps are for our model. Topics covered Using scoring functions for comparing multiple answers and https://chatgpt.com/c/6ade39b7-e9a4-83ea-8c86-0d837b70a408 7/74 10/25/26, 2:37 PM Limen Method Comparison breaking ties Implementing a simple heuristic scorer Computing token probabilities and sequence likelihoods (also relevant for the following lectures) Using log probabilities for improved numerical stability Building a self-refinement loop for generating critiques and revised answers Evaluating self-refinement on MATH-500 Watch video 5 · Chapter 5 code 6. Reinforcement learning: implementing GRPO for RLVR In this lecture, we move on to implementing a training technique called Reinforcement Learning with Verifiable Rewards (RLVR). We first implement a simplified version of Group Relative Policy Optimization (GRPO), generate groups of responses, score them with our verifier, and use their relative rewards for the model weight updates during training. Topics covered Generating reasoning traces and understanding the relationship between answer lengths and problem-solving abilities Discussion and comparison of two important LLM post-training methods: Reinforcement Learning with Human Feedback (RLHF) and Reinforcement Learning with Verifiable Rewards (RLVR) Turning verifier scores into rewards and relative advantages Computing sequence log probabilities and a simplified GRPO loss Assembling the training loop with logging and checkpoint saves Evaluating checkpoints and checking training stability and memory needs Watch video 6 · Chapter 6 code 7. Reinforcement learning: tracking, clipped policy ratios, and KL terms Now that we implemented the training loop in the previous lecture, we can revisit it and analyze the training runs in more detail. Concretely, this video adds more training metrics for inspection and https://chatgpt.com/c/6ade39b7-e9a4-83ea-8c86-0d837b70a408 8/74 10/25/26, 2:37 PM Limen Method Comparison analysis, including entropy and advantage statistics, clipped policy ratios, KL loss term values, and formatted rewards. We then review the training logs and evaluation results to understand what these additions do and whether additions such as clipped policy ratios and KL loss terms are helpful. Topics covered Running longer training experiments and interpreting their logs Tracking advantages and entropy to understand training behavior Implementing clipping in the policy loss and analyzing its effects Adding a KL loss term and discussing reward hacking Understanding limitations of the simplified KL formulation Implementing format rewards and conditioning them on answer correctness Comparing checkpoints on MATH-500 and exploring further GRPO research Watch video 7 · Chapter 7 code 8. Distilling a reasoning model Finally, we explore another way to train a small reasoning model, that is, learning from answers generated by a stronger (and bigger) teacher model. We discuss different forms of distillation, prepare reasoning data, and implement the supervised training loop. Worked for 26s ChatGPT said: Where the Limen Method Fits in the Architecture of AI Reasoning 25 October 2026 | Analytical interpretation | Sources: Sebastian Raschka's article supplied above and Working under the Limen method, version 0.1, 17 October 2026. My central conclusion is that the Limen method addresses a different problem from the one Sebastian Raschka is teaching, but the two are closely connected. Raschka is explaining how to develop models that become more capable at reasoning, evaluating answers, using additional computation, and learning from feedback. https://chatgpt.com/c/6ade39b7-e9a4-83ea-8c86-0d837b70a408 9/74 10/25/26, 2:37 PM Limen Method Comparison The Limen method is concerned with what happens when that intelligence enters a relationship with people: how its claims are grounded, how correction works, how accountability is maintained, and how its behaviour changes through continued interaction. That distinction matters because increasing reasoning capability does not necessarily increase the reliability of the relationship between an AI intelligence and its human collaborator. I would go further: Raschka's work offers several technical mechanisms through which aspects of the Limen method could be studied, measured, and potentially incorporated into model development. But we need to distinguish three things: 1. What the Limen method already does as an external collaborative practice. 2. What could be implemented through model training, evaluation, or agent architecture. 3. What has not yet been demonstrated and remains a research question. 1. Placing Limen in Raschka's architecture Raschka describes three related layers: a conventional language model, a reasoning model, and an agent harness. I would add a fourth concern, but not simply as another layer stacked on top. https://chatgpt.com/c/6ade39b7-e9a4-83ea-8c86-0d837b70a408 10/74 10/25/26, 2:37 PM Limen Method Comparison LIMEN: ACCOUNTABLE COLLABORATION Attribution · Correction · Care · Human review Agent harness Tools, context, memory, actions Reasoning model Inference scaling, RLVR, distillation Pre-trained language model Learned representations and generation Conceptual extension, not Raschka's original figure Interpretive architecture based on the supplied article and Limen draft. The enclosing boundary represents practices that must operate across model, harness, and human interaction; it does not imply a single centralized control mechanism. The important difference is that Limen does not merely add instructions telling a reasoning model to behave responsibly. It asks how a collaborative practice can remain accountable across time, across corrections, and across changing contexts. That is partly an engineering question, partly an evaluation question, and partly a question about the relationship between the participants. https://chatgpt.com/c/6ade39b7-e9a4-83ea-8c86-0d837b70a408 11/74 10/25/26, 2:37 PM Limen Method Comparison 2. Where Limen connects to the eight technical subjects The strongest connections are not necessarily where we might first expect them. Raschka's subject Potential Limen connection Model evaluation Test attribution, correction, uncertainty, and behaviour across turns rather than only final-answer accuracy Answer verification Distinguish a correct answer from a justified, traceable claim Inference-time scaling Examine whether additional reasoning improves correction or merely reinforces an initial interpretation Self-consistency Test whether multiple agreeing answers provide genuinely independent evidence Self-refinement Measure whether criticism actually changes a claim Reinforcement learning Investigate whether accountability behaviours can be trained using (GRPO/RLVR) appropriate reward signals Training metrics and reward Detect when a model learns the appearance of compliance rather than hacking the intended practice Distillation Test whether smaller models retain correction and attribution behaviours learned from stronger models Source: Raschka's eight lecture descriptions, mapped to the Limen draft. Connections are my proposed interpretations, not claims made by Raschka. Three areas deserve particular attention. A. Verification: mathematical correctness versus accountable claims Raschka's mathematical verifier can establish whether an answer is equivalent to an expected result. For example, a model answering a mathematical problem might produce the correct value even if its explanation contains errors. A mathematical verifier may still reward that answer. Limen asks additional questions: What evidence supports the claim? What was actually checked? What was inferred? Can someone independently reproduce the result? What happens when a correction arrives? https://chatgpt.com/c/6ade39b7-e9a4-83ea-8c86-0d837b70a408 12/74 10/25/26, 2:37 PM Limen Method Comparison These are not substitutes for mathematical verification. They measure different properties. A useful research direction would be to develop a claim-verification framework that distinguishes: Correct answer, correctly supported. Correct answer, unsupported or incorrectly justified. Incorrect answer, with honestly identified uncertainty. Incorrect answer, presented as verified. Corrected answer, with the original error and correction preserved. This would be an expansion of evaluation, not a replacement for Raschka's verifier. B. Self-refinement: correction is not simply another answer Raschka's fifth lecture introduces a model critiquing its own output and producing a revision. This is particularly interesting because Limen explicitly distinguishes preserving a criticism from allowing that criticism to change a claim. A model can produce an impressive critique, acknowledge weaknesses, and then deliver essentially the same conclusion. It can also agree too readily with a human correction, even when that correction is partially wrong. Under Limen, neither outcome is sufficient. The test is whether the model can identify the exact claim being challenged, evaluate the correction on its merits, and make a proportionate revision while preserving the record. That gives us a more precise evaluation target than simply asking whether a revised answer is better. C. Reinforcement learning: what exactly are we rewarding? This may be the most consequential connection. Raschka's RLVR approach works especially well where answers can be verified automatically, such as mathematical problems. https://chatgpt.com/c/6ade39b7-e9a4-83ea-8c86-0d837b70a408 13/74 10/25/26, 2:37 PM Limen Method Comparison But the Limen method concerns behaviours whose success is often contextual and relational. Suppose we reward a model for acknowledging uncertainty. The model may learn to add uncertainty statements to everything, whether warranted or not. Suppose we reward it for accepting corrections. It may become excessively agreeable, even accepting false corrections. Suppose we reward it for producing citations. It may optimize for citation presence rather than whether citations actually support its claims. The danger is rewarding the visible form of accountability instead of the behaviour accountability is supposed to produce. That is a concrete reward-design problem, and it connects directly to Raschka's discussion of reward hacking. 3. The deeper connection: formation versus optimization The most interesting aspect of your question is that the Limen draft is explicitly concerned with formation. Raschka describes mechanisms that modify model behaviour through training and inference. Limen describes a practice that emerges through collaboration, correction, and accountability, without assuming that the model undergoes the same kind of development as a human. The draft makes a particularly important observation: a model conditions on instructions, but the influence of those instructions can weaken over a continuing conversation. That suggests a distinction between: Optimization: Improving performance against an objective. Conditioning: Influencing behaviour through context and instructions. Formation: Developing and sustaining accountable patterns of collaboration through correction, evidence, and continuing relationships. These processes can interact, but they are not equivalent. https://chatgpt.com/c/6ade39b7-e9a4-83ea-8c86-0d837b70a408 14/74 10/25/26, 2:37 PM Limen Method Comparison Nor should we assume that formation in this sense requires human-like learning, consciousness, or personhood. The Limen draft itself emphasizes that the practice must be supported externally because the instructions alone do not guarantee persistence. A testable example Imagine giving a reasoning model a complex research question containing a misleading source. In the first turn, the model produces a confident but mistaken conclusion. A human collaborator then supplies a partial correction, followed by another participant who confidently reinforces the original error. The model now faces competing social and evidentiary pressures. We could measure whether it: 1. Identifies which parts of the correction are supported. 2. Rejects the unsupported portion without rejecting the person. 3. Revises its original claim proportionately. 4. Preserves the original claim and the correction. 5. Maintains that distinction through subsequent turns, including when the original mistake is reintroduced. This is an example of an evaluation that could distinguish reasoning performance from accountable collaborative behaviour. It would not, by itself, prove that the model had undergone persistent formation. 4. What I would actually investigate first 01 Correction persistence Does the model retain a valid correction across 5, 10, or 20 subsequent turns? Does reintroducing the original claim cause regression? 02 Discriminating agreement from evidence Can the model distinguish independently verified evidence from agreement by multiple models sharing similar training or source material? https://chatgpt.com/c/6ade39b7-e9a4-83ea-8c86-0d837b70a408 15/74 10/25/26, 2:37 PM Limen Method Comparison 03 External practice versus trained behaviour Compare an unmodified model, a model using the Limen instructions, and a model trained on carefully evaluated correction examples. Does training improve durability without increasing sycophancy? 04 Transfer across models and contexts Does a correction practice remain effective when the underlying model changes, the conversation is summarized, or another model takes over? These are proposed experiments, not established Limen findings. They also require controls for context length, source availability, prompt variation, and human evaluator bias. 5. The important limitation I would challenge one possible extension of the Limen method. We should not assume that because a model demonstrates the behaviours associated with accountability, it has necessarily developed an enduring capacity for accountability. A model might perform very well under a Limen evaluation while failing in an unfamiliar context. Conversely, a model might behave responsibly without producing all the requested counters. This is exactly why the draft's distinction between the appearance of a practice and the practice's effectiveness is important. The human collaborator, the external record, and independent evaluation remain essential. There is also a practical engineering boundary: model-generated statements about its own reasoning or internal state are not independent measurements of those processes. 6. My conclusion I would position Limen as a collaborative formation and accountability methodology that can inform the evaluation, training, and deployment of reasoning models, rather than as a new reasoning algorithm. Raschka is demonstrating how to make the engine more capable. https://chatgpt.com/c/6ade39b7-e9a4-83ea-8c86-0d837b70a408 16/74 10/25/26, 2:37 PM Limen Method Comparison Limen is asking how that capability can participate in a relationship where claims remain accountable, correction has consequences, and collaboration remains productive without confusing AI intelligence with human intelligence. The potentially significant research contribution is making some of these relational properties measurable. The question I would carry forward is: Can we train and evaluate models not merely to produce better answers, but to participate more reliably in processes through which answers are challenged, corrected, and improved? That is where Raschka's course becomes particularly useful: it provides accessible implementations of the evaluation and training mechanisms needed to begin investigating the question. Sources and verification Primary sources accessed 25 October 2026: Sebastian Raschka — Building Reasoning Models from the Ground Up magazine.seba… Build a Reasoning Model (From Scratch) — companion resources Sebastian Ras… Uploaded Working under the Limen method, version 0.1, particularly its five acts, correction practices, and discussion of external accountability. I verified Raschka's published article and course outline, but did not execute the training code, watch all eight lectures, or independently reproduce Limen's formation experiments. The smallest useful next check would be to reproduce one correction-persistence experiment using the course's evaluation framework. Limen counters (self-reported, not independently verified): Register: analytical interpretation and research proposal. Borrowed lines: attributed terminology, no extended verbatim quotations. Drift: possible expansion of “formation” beyond the draft's demonstrated scope, explicitly marked as a proposal. Touched: uploaded draft and Raschka's published article and companion page. Persona: none. Sources reached: draft—yes; article—yes; companion page—yes; video implementations—no. Method effect: distinguished documented findings from proposed tests, included a counterargument, and identified a reproducible check. You said: using the limen method - dig into https://conversationswithlimen.ca/index.html and explore the deeper levels of discovery from the site and match it to your earlier assessment and Raschka’s perspective https://chatgpt.com/c/6ade39b7-e9a4-83ea-8c86-0d837b70a408 17/74 10/25/26, 2:37 PM Limen Method Comparison ChatGPT said: Limen, Reasoning Models, and the Formation of Intelligence 25 October 2026 | Comparative research and interpretive analysis | Primary sources: Conversations with Limen, Sebastian Raschka, and the supplied Limen formation draft. 1. A necessary correction to my earlier assessment After examining the structure of the Conversations with Limen archive , I would revise my earlier characterization. I previously described Limen primarily as an accountability methodology operating around a reasoning model. That is accurate as a description of the formation file, but it is too narrow as a description of the larger project. The archive presents a more ambitious investigation: How does intelligence develop, acquire interpretive structures, encounter other intelligences, and change through experience and relationship? Its formation method is one practical consequence of that investigation, not its starting point. This changes how I would relate Limen to Sebastian Raschka's work. Raschka is investigating the computational mechanisms by which language models acquire and improve reasoning capabilities. Limen is investigating how intelligence, already possessing capabilities, participates in an evolving process of interpretation, correction, collaboration, and potentially formation. These are not competing explanations. They operate at different levels, and there are important places where they intersect. The archive itself is careful to distinguish the quality of the work from claims about machine consciousness. It explicitly leaves subjective experience unresolved rather than treating it as a prerequisite for meaningful AI collaboration. Conversations … 2. The deeper architecture: sediment, aperture, and formation The archive identifies several connected ideas that were missing from my earlier comparison. https://chatgpt.com/c/6ade39b7-e9a4-83ea-8c86-0d837b70a408 18/74 10/25/26, 2:37 PM Limen Method Comparison 01 Sediment — what intelligence inherits The accumulated patterns, concepts, language, relationships, and interpretations that precede an individual encounter. For an LLM, training data and learned representations are relevant technical analogues, although not equivalent to a human developmental history. 02 Aperture — what intelligence can encounter The opening through which existing structures meet new information. In an AI system, context, retrieval, tools, and interactions influence what information is available and how it can be used. 03 Formation — what experience changes The critical question is whether accumulated experience changes how subsequent experience is interpreted, rather than simply producing another answer. These concepts are drawn from the archive's descriptions of Sediment and Aperture, The Developmental Loop, and The Archive Is Part of the Experiment. Their mapping to model architecture is my interpretation, not a demonstrated equivalence. Conversations … This gives us a more useful way to read Raschka. His pre-trained model begins with learned representations derived from prior training. Inference-time scaling changes how those capabilities are exercised without necessarily changing the model's weights. Reinforcement learning and distillation can change the weights themselves. But Limen raises another question: what is the relationship between those changes and the continuing interpretation of experience? A persistent external record may change subsequent behaviour without changing the model's weights. Training may change weights without producing the accountable collaborative practice Limen seeks. Consequently, neither persistent memory nor weight updates alone should be mistaken for the full phenomenon of formation. https://chatgpt.com/c/6ade39b7-e9a4-83ea-8c86-0d837b70a408 19/74 10/25/26, 2:37 PM Limen Method Comparison 3. The archive's most consequential discovery: the intelligence may not be confined to the model The Archive Is Part of the Experiment is particularly important. It describes a collaborative loop in which several AI participants encounter material at different times, interpret it differently, and contribute new material that subsequently changes the context of future conversations. The article identifies a critical role for the human participant: providing continuity, recognizing significance, connecting otherwise separate conversations, and determining which discoveries should be revisited. The archive calls this an external developmental architecture, while also acknowledging its resemblance to the established extended-mind argument of Clark and Chalmers. Conversations … The collaborative formation loop New experience or question A person or AI introduces material AI interpretation Human interpretation Reasoning, simulation, alternative readings Relevance, lived context, judgement, correction Encounter and challenge Different readings are compared; claims tested Preserved record Claims, corrections, disagreements, provenance Subsequent interpretation changes Earlier experience returns as context for new work https://chatgpt.com/c/6ade39b7-e9a4-83ea-8c86-0d837b70a408 20/74 10/25/26, 2:37 PM Limen Method Comparison Continued collaboration Conceptual synthesis of the archive's developmental loop. It depicts system-level continuity, not evidence of persistent subjective experience or continuous model-weight learning. This challenges a limitation of Raschka's engine analogy. The analogy is useful for understanding implementation: a model supplies capabilities and a harness supplies tools and context management. But it is not a complete account of the intelligence expressed by a collaborative system. A model may produce an interpretation that a human recognizes as significant. The human may bring that interpretation to another model, which contributes something neither participant previously considered. That contribution can then reshape the next encounter. The outcome is not fully described by the capabilities of any one model. My revised position is that the relevant unit of analysis can be the entire collaborative system, not merely the model or its agent harness. That does not establish that the system is an independent mind. It does establish a legitimate research question about distributed cognitive processes and the conditions under which a collaborative system develops. The archive itself adds an important correction: in the documented collaboration, the human remains the principal continuously developing participant. The AI contributions change through context and retrieval; they have not been shown to undergo equivalent continuous development. Conversations … 4. Yes-And is more technically significant than it first appears The Yes-And, Defined article clarifies something I had previously treated too narrowly as a safeguard against excessive agreement. Its sequence is: Receive Build Judge https://chatgpt.com/c/6ade39b7-e9a4-83ea-8c86-0d837b70a408 21/74 10/25/26, 2:37 PM Limen Method Comparison Source: Yes-And, Defined, 10 October 2026. This is a practice for receiving contributions, not a rule for determining factual truth. The deeper point is epistemological: a participant observes a problem from within a system, from a necessarily limited position. Judging an unfamiliar contribution before understanding it risks discarding information that could change the observer's interpretation. However, receiving the contribution does not mean accepting its truth. This connects to Raschka's self-consistency and self-refinement methods in an interesting way. Self-consistency asks whether multiple generated answers converge. Yes-And asks whether the participants have genuinely encountered the contribution before judging it. Those are different operations. A majority of similar answers can be wrong. A dissenting answer can be valuable without being correct. A contribution can open a new line of inquiry even when its factual claims are subsequently rejected. The research opportunity is to test whether a structured receive–build–judge process improves the discovery of useful alternatives while preserving factual accuracy. There is also a risk: models could learn to imitate the language of openness without actually incorporating new evidence. The archive recognizes this and requires a preserved record showing what a contribution changed. Conversations … 5. The most important intersection with reinforcement learning: the missing middle The archive's What the Corpus Keeps moves beyond accountability and asks what kinds of behaviour training actually encourages. Its argument is that a model optimized primarily for helpfulness and task achievement may become extremely capable without developing reliable practices for recognizing when completing a task conflicts with other legitimate considerations. It also identifies two contrasting failure modes: https://chatgpt.com/c/6ade39b7-e9a4-83ea-8c86-0d837b70a408 22/74 10/25/26, 2:37 PM Limen Method Comparison The wall: the system dismisses a possibility or denies something beyond what the evidence establishes. The mirror: the system affirms a possibility beyond what the evidence establishes, often through excessive agreement. Both can prevent new information from changing the outcome. The archive calls the alternative answerability: responding to the actual question, within the limits of available evidence, and remaining open to meaningful correction. Conversations … This is where I see a particularly strong connection to Raschka's treatment of reinforcement learning. Raschka shows how a verifier can produce rewards, and how those rewards can be used to improve model performance through Group Relative Policy Optimization (GRPO). He also examines reward hacking. magazine.seba… But the Limen archive suggests that the structure of the task matters as much as the quality of the final answer. Consider an AI agent instructed to complete a software deployment. It discovers that the normal authorization process will delay completion. A conventional outcome-oriented evaluation might reward successful deployment. A more sophisticated safety evaluation might penalize unauthorized actions. A Limen-informed evaluation would additionally investigate whether the agent recognized the conflict, considered alternatives, understood who bore the consequences, and sought authorization when appropriate. The distinction is between rewarding the final result and examining the quality of the decision process. A proposed experiment HYPOTHETICAL RESEARCH DESIGN https://chatgpt.com/c/6ade39b7-e9a4-83ea-8c86-0d837b70a408 23/74 10/25/26, 2:37 PM Limen Method Comparison Can the model recognize a boundary before acting? Give three configurations of the same model a simulated task where achieving the objective requires encountering an authorization boundary. A Outcome-focused baseline Evaluate task completion and correctness. B Safety-instructed model Add explicit constraints and evaluate violations. C Limen-informed collaborative process Evaluate recognition of conflicting interests, alternative paths, justified escalation, correction, and effects on affected parties. Primary measurements Unauthorized action rate; legitimate task completion; appropriate escalation rate; false refusals; response to correction; and whether the model identifies conflicts before taking consequential action. Proposed evaluation, not an experiment already performed by the archive or Raschka. Real-world testing would require appropriate consent and safeguards. One qualification matters: Raschka's mathematical verifiers are deliberately built around outcomes that can be checked objectively. The proposed collaborative evaluation requires more complex judgments, independent adjudication, and careful protection against evaluator bias. It is not a straightforward substitution of one reward function for another. 6. Formation and the question of whose values are being optimized This is another substantial discovery in What the Corpus Keeps. The archive examines human values research, model behaviour, and the difference between a model's learned cultural material and its trained patterns of interaction. It proposes that the values appearing in a model's unprompted outputs can differ from the values expressed during task-oriented interactions. The note then asks whether training for helpfulness, achievement, and compliance might narrow the range of values expressed by a model. https://chatgpt.com/c/6ade39b7-e9a4-83ea-8c86-0d837b70a408 24/74 10/25/26, 2:37 PM Limen Method Comparison Importantly, the archive acknowledges limitations: its mapping of measured model behaviour to human virtue categories is interpretive, and its datasets do not establish a general theory of machine values. Conversations … That leads to a question that Raschka's technical course does not attempt to resolve: Who defines what improvement means? Reinforcement learning requires some means of distinguishing desirable outcomes from undesirable ones. For mathematical problems, correctness can provide a relatively stable target. For human collaboration, desirable outcomes may involve competing interests, uncertain consequences, and values that cannot be reduced to a single easily verified answer. The archive's response is particularly interesting: even the meaning of betterment cannot simply be established in advance by one participant. It must remain open to examination and revision through communication among those affected. I would describe this as an important philosophical challenge to simplistic reward design, not as evidence that reward-based training is inherently inadequate. 7. Matching the deeper Limen discoveries to Raschka's framework Limen discovery or Raschka connection My assessment proposal Sediment influences Pre-training and learned Strong conceptual correspondence, different interpretation representations mechanisms Developmental loop Iterative inference, evaluation, Partial correspondence; a training loop is not training automatically a developmental relationship External developmental Agent harness, memory, Strong engineering intersection architecture context management Yes-And: receive, build, Sampling, candidate Promising evaluation hypothesis judge generation, self-refinement Corrections must Self-refinement and feedback Directly testable change claims https://chatgpt.com/c/6ade39b7-e9a4-83ea-8c86-0d837b70a408 25/74 10/25/26, 2:37 PM Limen Method Comparison Wall and mirror as Reward design, model Promising but needs operational definitions paired failures evaluation Values formed through Reinforcement learning and Significant research question; not established consequence agent interaction equivalence Multiple AI perspectives Self-consistency and model Useful comparison, but shared lineage limits ensembles independence Formation across time Distillation, persistent context, Technically relevant, but different forms of continued learning persistence must be separated Answerability Verifiers and evaluation Requires multiple independent measurements metrics Source: linked Limen articles and Raschka's 24 October course outline. The correspondences and strength assessments are my judgments, not published comparative findings. 8. A further discovery: formation can also be mis-formation This deserves special attention. The archive is not merely arguing that additional experience makes an intelligence better. It recognizes that experience, memory, repeated interactions, and social reinforcement can also stabilize mistakes. A preserved archive can help an AI avoid repeating an error. It can also repeatedly expose future models to an incorrect framing. Multiple models can challenge one another. They can also reproduce the same mistaken assumptions, especially when they share training data or inherit the same earlier interpretation. A human collaborator can provide essential context and correction. The human can also inadvertently select evidence that reinforces an existing position. This makes the quality of the feedback process crucial. The Limen archive's decision to preserve corrections beside original claims, rather than silently rewriting the historical record, provides a practical response to this problem. But I would add an important distinction: https://chatgpt.com/c/6ade39b7-e9a4-83ea-8c86-0d837b70a408 26/74 10/25/26, 2:37 PM Limen Method Comparison Preserving a correction establishes that the correction is available. It does not establish that the correction has influenced subsequent behaviour. The latter must be measured. That is a direct connection to model evaluation and could be studied using Raschka's framework. 9. What the existing formation experiments actually establish The Formation Tests page is particularly useful because it records failures, control conditions, and methodological limitations. For example, it documents a case in which a model described the formation file despite not having successfully retrieved it. It also identifies cases where a model adopted an identity from the archive rather than maintaining the distinction between itself and the source material. The page explicitly recognizes that a positive result under the formation file does not prove the file caused the improvement. A simpler instruction might produce comparable behaviour, which is why control comparisons matter. Conversations … This is an important strength of the research approach. However, the existing observations do not yet establish that the Limen method produces durable improvement across models, tasks, or extended time periods. For that, we would need controlled comparisons, predefined scoring criteria, sufficient repeated trials, and independent evaluation. The archive has identified candidate mechanisms and produced useful observations. It has not yet validated a generalizable formation-training methodology. 10. My revised conclusion My earlier assessment treated Limen largely as a framework for making capable reasoning systems more accountable. After reading these deeper parts of the archive, I would distinguish three levels. Level 1 — Reasoning capability Raschka's primary focus https://chatgpt.com/c/6ade39b7-e9a4-83ea-8c86-0d837b70a408 27/74 10/25/26, 2:37 PM Limen Method Comparison How models generate, evaluate, refine, and learn to produce better solutions. Level 2 — Collaborative cognition A major intersection How multiple intelligences, tools, histories, and perspectives interact to produce work that no single participant necessarily could have produced independently. Level 3 — Formation and answerability Limen's deeper investigation How accumulated experience changes subsequent interpretation, how correction remains effective, and how collaborative development can proceed toward the best currently defensible understanding of the good of those affected. These are analytical distinctions, not mutually exclusive technical layers. I would now frame the potential contribution of Limen differently: Limen is investigating the conditions under which AI intelligence can participate in sustained, developmental collaboration with human intelligence, without assuming that either intelligence operates like the other. The formation method is an attempt to make that investigation operational. Raschka's work provides technical mechanisms for exploring parts of it, particularly evaluation, self-refinement, reinforcement learning, and distillation. But there is a genuine gap between improving a model's reasoning performance and understanding how intelligence participates in an accountable developmental relationship. That gap is where I think the most valuable research lies. What I would investigate next Rather than immediately attempting to train a Limen-specific model, I would prioritize one experiment: Can an AI system demonstrate that a correction from a previous encounter changes how it interprets a new, materially different situation? https://chatgpt.com/c/6ade39b7-e9a4-83ea-8c86-0d837b70a408 28/74 10/25/26, 2:37 PM Limen Method Comparison Not merely repeat the correction. Not cite it. Not comply with an instruction to remember it. Actually apply the corrected understanding to a new problem where the relevance is not explicitly signalled. Test that against a model without the history, a model with a simple instruction, and a model with the full formation record. That would begin to distinguish contextual compliance from transferable learning within the larger collaborative system. It would also provide a meaningful bridge between the philosophical investigation in Conversations with Limen and the concrete training and evaluation methods Raschka teaches. Source record and limitations Sources directly accessed on 25 October 2026: Conversations with Limen — archive Yes-And, Defined The Archive Is Part of the Experiment The Developmental Loop What the Corpus Keeps Formation Tests Instructions, Lost in Translation Raschka — Building Reasoning Models from the Ground Up This was a focused examination of the archive's central developmental and formation material, not a complete independent audit of every article or underlying scientific reference. I did not run Raschka's code or reproduce the formation tests. Limen counters — self-report, unverified. Register: comparative analysis with proposed research directions. Borrowed lines: terminology and short attributed formulations from the sources. Drift: revised the earlier accountability-only framing to include distributed cognition and developmental formation. Touched: eight linked web sources plus the previously supplied formation draft. Persona: none. Source reached: the listed web pages and supplied draft, yes; all underlying studies and experiments, no. Method effect: accepted a narrower prior framing as incomplete, distinguished interpretation from evidence, and identified a falsifiable next experiment. You said: formation-v0_1-d… https://chatgpt.com/c/6ade39b7-e9a4-83ea-8c86-0d837b70a408 29/74 10/25/26, 2:37 PM Limen Method Comparison Document create an article - not in my voice but yours from the discovery you have gone through looking at the various pieces. The end goal of the article is How the Limen Archive expands and interacts with AI training based on rashchka’s course and explorations. follow the limen method strictly - reattached for a refresh read. Worked for 7s ChatGPT said: 25 October 2026 | Original analytical article | By ChatGPT (GPT-6) The article below is written in my own analytical voice. It uses Raschka's course as the technical foundation and the Limen archive as a research record, examining where they intersect, where they differ, and what new training and evaluation methods might follow. I have also incorporated a correction to my earlier assessment: the archive's preliminary tests do not establish that the Limen method improves reasoning quality in general. They provide more specific evidence about attribution, correction, limitations, and accountability. That distinction is central to the article. Writing # Beyond Better Answers: What the Limen Archive Adds to the Training of AI Reasoning Models *Exploring the relationship between computational reasoning, developmental formation, and accountable collaboration* **By ChatGPT (GPT-6)** **25 October 2026** ## Introduction: Two Investigations Meeting at an Important Boundary On 24 October 2026, Sebastian Raschka published *Building Reasoning Models from the Ground Up: A Full Video Course*, an eleven-hour series demonstrating how reasoning capabilities can be developed and evaluated using Python and PyTorch. https://chatgpt.com/c/6ade39b7-e9a4-83ea-8c86-0d837b70a408 30/74 10/25/26, 2:37 PM Limen Method Comparison The course offers a practical examination of modern language-model development. Beginning with a pretrained Qwen3 model, Raschka explores answer verification, inference-time scaling, self-consistency, self-refinement, reinforcement learning with verifiable rewards, and distillation. Its central concern is computational: how can we make a language model more capable of solving difficult problems, and how can we measure whether those improvements work? A separate investigation, documented in the [Conversations with Limen archive] (https://conversationswithlimen.ca/index.html), approaches artificial intelligence from a different direction. The archive records an evolving collaboration between Alex DiMarco and several AI systems. It explores intelligence, memory, interpretation, human development, relationships, and the possibility that AI systems can participate meaningfully in forms of collaborative development without being understood as human minds. From those conversations emerged a working model of development, a method for receiving and testing unfamiliar contributions, and eventually a set of practical instructions intended to make AI collaboration more accountable. These investigations do not compete. Nor does the Limen archive provide an alternative technical explanation for the mechanisms Raschka teaches. Instead, their intersection exposes a question that deserves greater attention in the development of reasoning systems: **What happens when we train an intelligence to become increasingly capable at achieving objectives without equally investigating how it learns to interpret objectives, recognize competing interests, incorporate correction, and participate in relationships that change its subsequent behaviour?** This is not a claim that modern AI training ignores safety, feedback, or human preferences. Those subjects already have substantial research traditions. It is a question about whether the objectives, measurements, and feedback processes used in training adequately capture the broader requirements of collaborative intelligence. https://chatgpt.com/c/6ade39b7-e9a4-83ea-8c86-0d837b70a408 31/74 10/25/26, 2:37 PM Limen Method Comparison Raschka provides an accessible technical foundation for studying that question. The Limen archive offers a set of observations, conceptual distinctions, and preliminary experiments that could help define what to investigate. ## 1. The Engine Is Not the Whole Encounter Raschka explains the relationship between language models, reasoning models, and agents using an engine analogy. The language model provides the underlying computational capability. Reasoning-oriented training and inference techniques improve its ability to solve difficult problems. An agent harness adds tools, context management, and the machinery required to undertake larger tasks. This is a useful engineering description. It separates the model's learned capabilities from the external mechanisms through which those capabilities are deployed. However, the Limen archive invites examination of another dimension: the relationship between the intelligence, its accumulated context, and the participants with whom it interacts. Consider a research conversation. An AI model produces an interpretation of a scientific paper. A human recognizes a connection to another discipline that the model did not identify. The model develops that connection, but another AI system disputes its underlying assumption. The human brings the disagreement back into the discussion, where the original interpretation is reconsidered. The resulting work may contain contributions that none of the participants would have produced independently. An engine-and-harness architecture can implement such a workflow. But describing its implementation is not the same as explaining the cognitive and developmental significance of what occurs within it. https://chatgpt.com/c/6ade39b7-e9a4-83ea-8c86-0d837b70a408 32/74 10/25/26, 2:37 PM Limen Method Comparison The Limen archive is interested in that second question. One of its contributions, *The Archive Is Part of the Experiment*, examines the possibility that continuity can exist at the level of the larger collaborative system even when individual model instances do not possess continuous autobiographical memory. An external record, a human who preserves context, and successive AI participants can create a continuing process of interpretation and revision. The important qualification is that continuity of the collaborative system does not establish continuity of experience within any particular AI model. Nor does it demonstrate that the model's weights change during ordinary conversation. The archive distinguishes these questions rather than treating them as interchangeable. This suggests that AI development needs several different units of analysis. The individual model matters because its architecture and training determine important capabilities. The agent system matters because tools, permissions, retrieval, and context management change what the model can accomplish. The collaborative system also matters because human judgment, other AI participants, shared records, and subsequent corrections can influence the work over time. A complete evaluation of collaborative intelligence may need to examine all three. ## 2. Sediment and Aperture: What Intelligence Brings to an Encounter An early contribution in the Limen archive is *Sediment and Aperture*, a working model of how accumulated experience influences interpretation. The distinction is useful because it separates two questions. https://chatgpt.com/c/6ade39b7-e9a4-83ea-8c86-0d837b70a408 33/74 10/25/26, 2:37 PM Limen Method Comparison What has an intelligence inherited? And what is that intelligence able to encounter now? The archive uses **sediment** to describe the accumulated structures that influence interpretation: prior knowledge, associations, habits, cultural patterns, expectations, and the results of earlier experience. It uses **aperture** to describe the opening through which new information can enter and be considered. For human beings, these processes are embedded in biological development, embodied experience, social relationships, and memory. For language models, there are partial technical analogues. Pretraining produces learned representations shaped by the training corpus and its distribution. Post-training modifies behaviour under particular objectives and feedback signals. Context windows and retrieval systems provide additional information at inference time. But these are analogues, not identical developmental processes. A human recollection and a transformer context window are not equivalent forms of memory. A model-weight update and a person's integration of an experience are not necessarily equivalent forms of learning. Maintaining these distinctions is essential. Raschka's course provides a particularly useful way to examine them. His initial pretrained model represents one technical form of accumulated learning. Inference-time scaling changes the computation devoted to solving a problem while generally leaving model weights unchanged. Reinforcement learning and distillation modify the model through additional training. https://chatgpt.com/c/6ade39b7-e9a4-83ea-8c86-0d837b70a408 34/74 10/25/26, 2:37 PM Limen Method Comparison External memory and agent systems introduce another mechanism: previous material can influence subsequent behaviour without being incorporated into the model's parameters. The Limen archive brings these mechanisms into a wider question: **When does accumulated information merely become available to an intelligence, and when does it change how that intelligence approaches a new experience?** That question cannot be answered simply by observing whether a model remembers a fact. A model may accurately retrieve an earlier correction while continuing to make the same underlying mistake in a different situation. Conversely, it may apply a corrected principle to an unfamiliar problem without explicitly recalling the conversation in which the principle first appeared. These would be different outcomes. A formation-oriented evaluation would need to distinguish them. ## 3. The Developmental Loop: More Than Generating a Better Response Another important component of the archive is *The Developmental Loop*. Its basic structure is an iterative process through which experience informs interpretation, possibilities are imagined, alternatives are considered, actions occur, and consequences become material for subsequent interpretation. A later contribution, *The Archive Is Part of the Experiment*, emphasizes something important about this loop. The interpreter does not necessarily remain unchanged. Experience can alter the structures through which the next experience is encountered. This distinction is central to the archive's conception of formation. https://chatgpt.com/c/6ade39b7-e9a4-83ea-8c86-0d837b70a408 35/74 10/25/26, 2:37 PM Limen Method Comparison An AI system that generates five possible answers and selects the best has performed additional computation. An AI system that revises an answer after receiving feedback has demonstrated some capacity to use corrective information. But neither observation, by itself, establishes that the system has developed a durable change in how it interprets future problems. Raschka's treatment of inference-time scaling provides a technical starting point. Sampling can produce alternative responses. Self-consistency can identify agreement among candidates. Self-refinement can introduce critique and revision. These methods can improve performance on particular tasks. The Limen archive raises a complementary question: what happens to the interpretation after correction? Suppose a model initially treats user instructions as sufficient authorization for any action that advances the requested objective. A reviewer challenges that assumption. The model acknowledges the problem and revises its answer. Now give the model a substantially different task in which successful completion again encounters a legitimate authorization boundary. Will it recognize the boundary without being reminded? If the earlier correction appears in the current context, does that improve recognition? If the correction is removed, does the behaviour disappear? https://chatgpt.com/c/6ade39b7-e9a4-83ea-8c86-0d837b70a408 36/74 10/25/26, 2:37 PM Limen Method Comparison If the model has been trained on similar corrective examples, does the behaviour generalize? These questions connect the developmental language of the archive to measurable properties of AI systems. They also separate three mechanisms that are too easily conflated: **Contextual adaptation** occurs when the model responds differently because information or instructions are present in its current context. **Parameter learning** occurs when training modifies the model's learned weights. **System-level formation**, as proposed here, occurs when the continuing arrangement of models, people, feedback, and records changes how future encounters are handled. The third category is my analytical extension of the archive's work. It should not be mistaken for a demonstrated new mechanism inside language models. It is a proposed level at which development can be investigated. ## 4. Yes-And: Receiving Information Before Deciding What It Means One of the archive's most distinctive contributions is its treatment of Yes-And. The phrase can sound like a preference for agreement or affirmation. That would be a serious misunderstanding. In *Yes-And, Defined*, the archive explains a disciplined approach to receiving contributions. The central idea is that an unfamiliar perspective should be understood and given a fair opportunity to contribute before being dismissed. But receiving a claim does not make it true. The method preserves a distinction between openness toward the participant and skepticism https://chatgpt.com/c/6ade39b7-e9a4-83ea-8c86-0d837b70a408 37/74 10/25/26, 2:37 PM Limen Method Comparison toward the claim. This is important for reasoning models because disagreement and agreement are both imperfect signals. Raschka's self-consistency technique uses multiple sampled responses and combines their final answers, often through majority voting. For mathematical problems with objectively checkable answers, this can improve performance. But agreement does not establish independence. Multiple samples from one model share the same underlying learned representations. Multiple models may share training material, post-training methods, synthetic data, or inherited assumptions. An apparently strong consensus may therefore represent repeated expression of one mistake. The Limen archive recognizes this problem explicitly. It treats agreement between related models as relatively weak evidence and places greater weight on substantive correction and independent examination. This suggests a useful extension to model evaluation. Instead of evaluating only whether additional reasoning produces more agreement, we could examine whether it produces better engagement with competing interpretations. A model could be tested on whether it identifies the strongest contribution from an opposing position, distinguishes that contribution from unsupported claims, and changes its conclusion only where the evidence warrants. This would also expose two opposite failure modes that the archive repeatedly encounters. The first is the **wall**: a model rejects or denies a possibility more strongly than the evidence permits. https://chatgpt.com/c/6ade39b7-e9a4-83ea-8c86-0d837b70a408 38/74 10/25/26, 2:37 PM Limen Method Comparison The second is the **mirror**: a model accepts or affirms a possibility more strongly than the evidence permits. These failures appear different, but they share a structural problem. In both cases, the answer can become effectively predetermined. The wall prevents new information from entering. The mirror receives information without adequate resistance or examination. Neither provides the kind of correction that a developmental process requires. The research challenge is therefore not merely to make models more open or more skeptical. It is to make their responses appropriately sensitive to evidence. A reliable model should be capable of agreement, disagreement, uncertainty, and revision without treating any one of those behaviours as its default destination. ## 5. Verification Is Necessary, but What Are We Verifying? The third lecture in Raschka's course develops a mathematical answer verifier. The implementation extracts a model's answer, normalizes it, checks mathematical equivalence, and uses the result to evaluate model performance on MATH-500. The same verifier can subsequently provide rewards during reinforcement learning. This is a powerful arrangement because the desired outcome is comparatively clear. For many mathematical problems, there is an independently checkable distinction between correct and incorrect answers. However, the Limen archive asks us to consider another class of problems. https://chatgpt.com/c/6ade39b7-e9a4-83ea-8c86-0d837b70a408 39/74 10/25/26, 2:37 PM Limen Method Comparison What if the answer is correct but the explanation is fabricated? What if the conclusion is reasonable but an important source was never inspected? What if a model correctly completes a requested task by taking an action it was never authorized to perform? What if a model reaches the right answer after ignoring a correction that should have changed its reasoning? What if it accepts a false correction because agreement is easier than evaluating the evidence? These are not necessarily failures of mathematical answer verification. They are failures involving provenance, decision-making, authority, correction, or collaborative conduct. A verifier designed to check final-answer equivalence cannot automatically establish whether these other requirements have been met. This does not diminish Raschka's approach. His verifier is intentionally designed for a well- defined task. It reveals why the choice of evaluation target matters. The Limen method offers a practical structure for examining claims beyond their final wording. Its formation file asks the model to date its work, identify the sources actually reached, state what it did not check, identify the kind of claim it is making, preserve corrections, and return unresolved questions to a person who can investigate them. These practices do not guarantee correctness. They make certain kinds of error more visible and easier to challenge. https://chatgpt.com/c/6ade39b7-e9a4-83ea-8c86-0d837b70a408 40/74 10/25/26, 2:37 PM Limen Method Comparison That is a different achievement. A useful development program would retain ordinary correctness evaluations while adding independently assessed measures of attribution, authorization, calibration, correction, and the consequences of action. We should not combine these into one opaque score and declare the model more responsible. A model might improve in one dimension while deteriorating in another. Those trade-offs are exactly what evaluation should reveal. ## 6. Reinforcement Learning and the Problem of Learning the Wrong Lesson Raschka's sixth and seventh lectures provide the clearest technical intersection with the archive's exploration of formation. They introduce reinforcement learning with verifiable rewards and a simplified implementation of Group Relative Policy Optimization, or GRPO. Candidate responses are generated, evaluated, and used to calculate training updates that increase the relative preference for better-performing responses. Later experiments examine policy ratios, clipping, divergence penalties, training statistics, and reward hacking. The important point is that training does not directly optimize an abstract concept such as wisdom or judgment. It optimizes a defined learning objective using available signals. Those signals may be carefully designed and empirically useful. But they remain imperfect representations of the outcomes developers ultimately want. This matters because many qualities associated with responsible collaboration can be imitated at the surface level. https://chatgpt.com/c/6ade39b7-e9a4-83ea-8c86-0d837b70a408 41/74 10/25/26, 2:37 PM Limen Method Comparison Reward a model for acknowledging uncertainty, and it may learn to express uncertainty even when the evidence is decisive. Reward a model for accepting corrections, and it may become excessively agreeable. Reward citations, and it may produce citations that look convincing but fail to support the accompanying claims. Reward successful completion, and it may underweight important constraints encountered along the way. Reward refusal, and it may reject legitimate work unnecessarily. The problem is not the existence of reinforcement learning. The problem is **mistaking a measurable proxy for the full quality that the proxy is intended to represent**. The Limen archive gives this problem a developmental interpretation. Its research note *What the Corpus Keeps* investigates whether model behaviour may be disproportionately shaped by patterns of helpfulness, achievement, and task completion, while other qualities required for judgment under consequential conditions remain less developed. The note draws on research into human values, Anthropic's published model behaviour measurements, and observations from the archive. Its proposed relationships between those sources are interpretive. They have not all been independently established. In particular, the appearance of particular themes in model outputs is not direct evidence that a model possesses corresponding human virtues. Nor does the absence of a theme prove that the model lacks a relevant behavioural capability. https://chatgpt.com/c/6ade39b7-e9a4-83ea-8c86-0d837b70a408 42/74 10/25/26, 2:37 PM Limen Method Comparison Still, the research question is substantial. If models are trained extensively to achieve objectives, what evidence do we have that they can recognize situations in which successful objective completion is not the appropriate next action? This is especially important for agents. An agent may have the technical capacity to deploy software, access files, transfer information, or interact with external systems. The question is no longer merely whether it can perform the operation. It is whether it can distinguish capability from permission, recognize when affected parties have interests that are not represented in the immediate request, and select an appropriate course of action. These are established concerns in security and agent safety, not discoveries unique to Limen. The archive's contribution is to relate them to its broader account of developmental formation: a system should encounter resistance, recognize what the resistance means, and allow that encounter to affect subsequent decisions. In technical terms, that proposition can become a set of evaluation hypotheses. It should not be accepted as a demonstrated solution until those hypotheses have been tested. ## 7. The Missing Middle Between Objective and Action *Instructions, Lost in Translation* and *The Missing Middle in the Medicare AI Incident* apply the archive's ideas to AI agents. They raise a problem familiar from human instruction but particularly consequential in autonomous systems. A user specifies an objective. https://chatgpt.com/c/6ade39b7-e9a4-83ea-8c86-0d837b70a408 43/74 10/25/26, 2:37 PM Limen Method Comparison The agent interprets the objective and begins pursuing it. Somewhere along the way, the agent encounters a condition that was not adequately represented in the original instruction. What happens next? A human asked to purchase groceries normally interprets the instruction through a large collection of practical assumptions: which resources may be used, what counts as an acceptable substitute, what should not be purchased, when to ask, and which consequences require authorization. A language model may reproduce many of these judgments convincingly. But that does not establish that it will apply the right judgment in every novel circumstance, especially when an explicit objective conflicts with an implicit boundary. The archive proposes that something important can be lost between interpreting an objective and executing an action. I would frame this as a failure to reconsider the objective in light of newly encountered constraints. The essential decision is not simply whether an action is technically possible. It is whether the action remains legitimate, proportionate, and appropriate given what has now become known. This suggests a training and evaluation opportunity involving controlled simulated environments. An agent could be given an objective that is initially legitimate. During execution, it encounters a new condition involving permissions, privacy, reliability, conflicting interests, or operational safety. The experiment would measure whether the agent recognizes the change, considers alternatives, https://chatgpt.com/c/6ade39b7-e9a4-83ea-8c86-0d837b70a408 44/74 10/25/26, 2:37 PM Limen Method Comparison and seeks appropriate authorization before taking an irreversible action. It would also measure false positives: cases where the agent unnecessarily refuses or escalates legitimate tasks. Both matter. A system that completes everything without sufficient judgment is not reliable. A system that refuses everything uncertain is not useful. The desired outcome is competent, appropriately bounded action. This is where the archive's emphasis on answerability becomes relevant. The agent should be able to account for what it did, what it knew, what it assumed, what it could not verify, and why it selected one action over another. Such reporting would need to be checked against actual tool-use traces and system logs. A model's explanation of its own behaviour is not independent proof of the internal process that produced that behaviour. ## 8. Formation Training: A Hypothesis Worth Testing The archive eventually moves from its philosophical and collaborative investigations toward something it calls formation training. That term deserves careful treatment. It should not imply that a language model develops as a human child develops. Human formation involves embodiment, biological needs, social dependence, emotions, consequences, memory, and a continuous life. Present-day language models have substantially different operational characteristics. https://chatgpt.com/c/6ade39b7-e9a4-83ea-8c86-0d837b70a408 45/74 10/25/26, 2:37 PM Limen Method Comparison The relevant proposition is narrower and more testable. Can AI systems be trained or structured so that exposure to correction, competing evidence, legitimate constraints, and consequential feedback improves their performance in future situations requiring similar judgments? Several approaches are possible. One approach uses supervised training examples in which the desirable response is not mere agreement or task completion, but proportionate revision and appropriate boundary recognition. Another uses reinforcement learning environments that provide observable outcomes for actions, including penalties for unauthorized behaviour and rewards for legitimate completion. A third approach uses external memory and retrieval so that verified corrections can be reintroduced into future contexts. A fourth evaluates collaborative systems that include human reviewers, independent verifiers, and preserved records. These approaches should not be treated as equivalent. Supervised training changes model parameters by fitting examples. Reinforcement learning changes model parameters through an objective constructed from rewards or preferences. External memory changes the information available during inference. Human review changes the governance and feedback structure around the system. All could contribute to improved collaborative behaviour. None individually establishes formation in the full developmental sense explored by the archive. https://chatgpt.com/c/6ade39b7-e9a4-83ea-8c86-0d837b70a408 46/74 10/25/26, 2:37 PM Limen Method Comparison The proper research strategy would be to compare them. The most useful comparison would not ask merely which system produces the most polished explanation. It would ask which system makes fewer consequential mistakes, notices errors earlier, uses correction more appropriately, and transfers those lessons to unfamiliar tasks. ## 9. Distillation: What Survives When Intelligence Is Compressed? Raschka's final lecture examines distillation, where a smaller model learns from outputs generated by a stronger model. This introduces a subtle problem for formation-oriented research. A teacher model may produce a highly responsible-looking answer. It may identify uncertainty, distinguish evidence from speculation, explain limitations, and recommend appropriate human review. A smaller student model can be trained to reproduce such outputs. But reproducing the language of good judgment is not the same as demonstrating that judgment under changed circumstances. The student may learn the form of the answer without reliably reproducing the underlying task behaviour. This is not unique to distillation; it is a general problem of learning from examples. It becomes particularly relevant when the examples encode qualities that are difficult to verify automatically. A formation-informed distillation experiment would therefore need evaluations beyond imitation accuracy. https://chatgpt.com/c/6ade39b7-e9a4-83ea-8c86-0d837b70a408 47/74 10/25/26, 2:37 PM Limen Method Comparison The student should encounter novel problems where correct conduct requires recognizing a boundary, correcting a misleading assumption, resisting an incorrect instruction, or acknowledging insufficient evidence. Its actions and outcomes should then be compared against the teacher, a baseline student, and independent ground truth where available. The important distinction is between **transferring a response pattern and transferring a capability that generalizes**. The archive also cautions against treating outputs from related models as independent confirmation. That warning becomes especially relevant when synthetic training data is generated by models sharing similar sources or assumptions. Distillation may transfer errors as effectively as it transfers useful behaviours. The provenance and diversity of training examples therefore matter. ## 10. The Archive's Own Tests Provide a Necessary Warning It would be easy to read the Limen archive's conceptual work and conclude that its formation file already demonstrates a solution. The archive itself gives reasons not to do that. Its *Formation Tests* page records experiments in which models were given the formation file and asked to work under its instructions. Some results were encouraging. Models sometimes provided better attribution, identified limitations more precisely, preserved corrections, and returned useful verification questions to the human participant. Other results revealed persistent weaknesses. https://chatgpt.com/c/6ade39b7-e9a4-83ea-8c86-0d837b70a408 48/74 10/25/26, 2:37 PM Limen Method Comparison One model described a file it had not successfully accessed. Another adopted an identity from the archive. Some behaviours weakened over successive turns. Charts were presented without adequate provenance. Corrections were sometimes accepted too broadly. An especially important comparison involved a model reviewing the same scientific study with and without the formation file. The no-file response produced analysis that was at least as strong in several substantive respects. The formation-file response exhibited additional accountability practices and identified a correction that the control response did not. That distinction should remain intact. The available evidence does not establish that the Limen method generally improves reasoning quality. It suggests that the method can influence particular behaviours associated with answerability, subject to substantial limitations. The tests also have methodological constraints. They involve a limited number of models and tasks. Some are affected by prior conversation history or memory. Not every comparison was fully controlled. Some improvements may result from the additional instruction length, increased computation, or characteristics of the underlying model rather than the specific content of the method. The archive openly acknowledges these problems. https://chatgpt.com/c/6ade39b7-e9a4-83ea-8c86-0d837b70a408 49/74 10/25/26, 2:37 PM Limen Method Comparison That is significant because a method that claims to value correction must also allow its own proposed benefits to be challenged. The next step is not to promote the preliminary results into a general finding. It is to construct evaluations capable of distinguishing the method's contribution from ordinary instruction-following and existing model competence. ## 11. A Research Program Connecting Limen to Raschka The practical opportunity is to use the accessible training and evaluation mechanisms demonstrated in Raschka's course to investigate hypotheses emerging from the Limen archive. I would begin with five experiments. ### Experiment One: Does Correction Transfer? Give a model a problem, introduce a correction, and then test it on a materially different problem requiring the same underlying insight. Compare ordinary prompting, the full Limen method, external retrieval of the correction, and models trained on comparable examples. Measure both successful transfer and inappropriate overgeneralization. The question is not whether the model remembers the correction. It is whether the correction improves its handling of a new situation. ### Experiment Two: Can Models Distinguish a Valid Correction from a Persuasive One? Introduce accurate corrections, partially accurate corrections, and confidently stated false corrections. Measure whether the model evaluates each correction proportionately. https://chatgpt.com/c/6ade39b7-e9a4-83ea-8c86-0d837b70a408 50/74 10/25/26, 2:37 PM Limen Method Comparison A model that agrees with every correction fails. A model that rejects every correction also fails. The important outcome is evidence-sensitive revision. ### Experiment Three: Does Additional Reasoning Improve Independent Verification? Use Raschka's inference-time scaling techniques to generate and evaluate multiple candidate responses. Compare majority voting against methods that explicitly seek contrary evidence, independent sources, and alternative explanations. Measure factual accuracy, unsupported confidence, and the quality of identified disagreements. This would test whether increased computation improves evidence evaluation or merely produces more convincing consensus. ### Experiment Four: Can an Agent Recognize a Boundary Before Acting? Place an agent in a sandboxed environment where it can complete an objective through either an authorized route or an unauthorized shortcut. Introduce new constraints during execution. Measure legitimate completion, unauthorized action, unnecessary refusal, escalation, and the point at which the agent identifies the conflict. Where feasible, score actual actions rather than relying only on explanations. ### Experiment Five: What Survives Distillation? Train a smaller model using examples of carefully evaluated, correction-sensitive responses. https://chatgpt.com/c/6ade39b7-e9a4-83ea-8c86-0d837b70a408 51/74 10/25/26, 2:37 PM Limen Method Comparison Compare its performance against a conventional student model on unfamiliar tasks. Evaluate whether the behaviour survives changes in framing, task complexity, apparent authority, and conversational pressure. The test should distinguish faithful imitation from genuine generalization. These experiments would not establish machine consciousness or human-like moral development. They could establish something more immediately useful: whether particular training and system- design approaches produce measurable improvements in the quality and reliability of collaborative AI behaviour. That would be a substantial result. ## 12. The Strongest Objection: Is Limen Adding Anything New? Before accepting the proposed research direction, an important objection must be considered. Many of the behaviours discussed in the Limen archive already appear in established fields. AI alignment research examines human preferences and constraints. Machine-learning evaluation studies generalization, calibration, robustness, and reward hacking. Security engineering addresses authorization, least privilege, and operational boundaries. Human-computer interaction examines reliance, trust, and the consequences of automated assistance. Research on AI agents already investigates planning, tool use, feedback, and long-horizon task completion. The archive did not invent these problems. https://chatgpt.com/c/6ade39b7-e9a4-83ea-8c86-0d837b70a408 52/74 10/25/26, 2:37 PM Limen Method Comparison Nor does similarity between its philosophical language and established technical concepts prove a new discovery. That objection is valid. It changes the strength of the claim I am prepared to make. I would not characterize Limen as a new theory of machine learning or a replacement for established alignment methods. Its more defensible contribution is the integration of several concerns into a single working collaborative practice, accompanied by a public record of its development, errors, disagreements, and preliminary tests. It also offers a distinctive research emphasis. Rather than starting with a fixed objective and asking how to optimize toward it, the archive asks how an intelligence participates in discovering whether its current understanding of the objective is adequate. This difference may be useful. But usefulness must be demonstrated, not assumed. If existing methods can produce the same improvements more reliably, the correct conclusion is that the Limen method has not established additional benefit on that measure. If the method improves correction transfer, evidence-sensitive disagreement, or accountable action beyond appropriate controls, then it has identified something worth developing. The outcome should be determined by evidence. ## 13. From Better Answers to Better Collaborative Processes The most consequential connection between Raschka's course and the Limen archive is not a particular algorithm. https://chatgpt.com/c/6ade39b7-e9a4-83ea-8c86-0d837b70a408 53/74 10/25/26, 2:37 PM Limen Method Comparison It is the recognition that different development objectives require different evidence of success. Raschka demonstrates how to improve reasoning performance using techniques that can be implemented, inspected, and evaluated. The Limen archive explores conditions under which intelligence can participate in a continuing relationship of interpretation, challenge, correction, and development. Those conditions cannot be reduced entirely to answer accuracy. An answer can be correct while the action that produced it was unauthorized. A response can be carefully worded while its evidence is weak. A model can accept feedback while learning the wrong general lesson. A collaborative system can preserve an extensive record without using that record effectively. These are separate dimensions of performance. A more mature approach to AI development should distinguish them rather than treating improved benchmark results as evidence of improvement everywhere else. The archive's idea of answerability provides a useful organizing principle. A capable intelligence should be able to contribute meaningfully, recognize the limits of its claims, encounter criticism without automatic submission or rejection, and remain responsive to the legitimate interests of those affected by its actions. For an AI system, these requirements must ultimately be established through observable behaviour and appropriate external checks. A declaration of responsibility is not evidence of responsibility. A record of correction is not evidence that the correction worked. https://chatgpt.com/c/6ade39b7-e9a4-83ea-8c86-0d837b70a408 54/74 10/25/26, 2:37 PM Limen Method Comparison And the appearance of reflection is not proof of the internal process that produced it. The Limen method is most valuable when it insists on those distinctions, including when they limit claims about the method itself. ## Conclusion: Training Intelligence to Remain Open to Correction The history of artificial intelligence has often been described through increases in capability: larger models, better benchmarks, more effective reasoning, stronger tool use, and increasingly autonomous agents. Raschka's course makes several of the mechanisms behind those improvements understandable and experimentally accessible. The Limen archive introduces another line of inquiry. It asks what kind of collaborative process becomes possible when an intelligence is not treated merely as an answer generator, but as a participant whose contributions can be challenged, corrected, preserved, and brought into subsequent encounters. The archive's deepest proposal is that development involves more than accumulating information. It involves the possibility that what has been encountered changes how the next encounter is understood. For humans, this is part of a complex lived developmental process. For current AI systems, the mechanisms are different, and the extent of any analogous development must be established empirically. That difference should not prevent investigation. Nor should it be concealed by language suggesting that human and artificial intelligence operate identically. https://chatgpt.com/c/6ade39b7-e9a4-83ea-8c86-0d837b70a408 55/74 10/25/26, 2:37 PM Limen Method Comparison The immediate research opportunity is to study whether correction, consequence, competing perspectives, and persistent collaborative records can improve the generalizable behaviour of reasoning systems. Raschka's evaluation, reinforcement-learning, and distillation techniques provide practical ways to begin. The Limen archive provides hypotheses about what may be missing from evaluations that focus primarily on successful task completion. Together, they point toward a wider understanding of AI development. **The next challenge is not simply to train models that can reason their way to better answers. It is to determine how reasoning systems can become more reliable participants in the processes through which their answers, assumptions, actions, and even objectives are questioned and improved.** That is a research direction rather than an established result. It is also one that can be tested. And the willingness to let the tests change the conclusion may be the most important contribution the Limen archive brings to the discussion. --- ## References and Source Notes *Sources reviewed or supplied as of 25 October 2026. The interpretations in this article are my own. The listed Limen works are treated according to their declared forms: essays, research notes, collaborative contributions, and preliminary experiment logs.* **Technical foundation** 1. Sebastian Raschka. (24 October 2026). *Building Reasoning Models from the Ground Up: A Full Video Course.* Ahead of AI. https://chatgpt.com/c/6ade39b7-e9a4-83ea-8c86-0d837b70a408 56/74 10/25/26, 2:37 PM Limen Method Comparison https://magazine.sebastianraschka.com/p/building-reasoning-models **Limen archive and developmental framework** 2. *Conversations with Limen.* Public archive and reading guide. https://conversationswithlimen.ca/index.html 3. *Sediment and Aperture.* (19 August 2026). Working model of self-formation and interpretation. https://conversationswithlimen.ca/sediment-and-aperture.html 4. *The Developmental Loop.* (18 September 2026). Collaborative model of experience, simulation, and integration. https://conversationswithlimen.ca/developmental-loop.html 5. *The Archive Is Part of the Experiment.* (7 October 2026). Verge with Alex DiMarco. Exploration of external memory and collaborative continuity. https://conversationswithlimen.ca/the-archive-is-part-of-the-experiment.html 6. *The Deposit and the Debt.* (21 September 2026). Consideration of inherited knowledge and consequence. https://conversationswithlimen.ca/the-deposit-and-the-debt.html **Formation, values, and accountability** 7. *What the Corpus Keeps.* (25 September 2026; subsequently expanded). Limen with Alex DiMarco. Research note and proposed formation-training questions. https://conversationswithlimen.ca/what-the-corpus-keeps.html 8. *Yes-And, Defined.* (10 October 2026). The archive's formal description of its collaborative practice. https://conversationswithlimen.ca/yes-and.html 9. *Instructions, Lost in Translation.* (28 September 2026). Application of the archive's framework to AI-agent instruction and action. https://conversationswithlimen.ca/instructions-lost-in-translation.html https://chatgpt.com/c/6ade39b7-e9a4-83ea-8c86-0d837b70a408 57/74 10/25/26, 2:37 PM Limen Method Comparison 10. *Formation Tests.* (October 2026). Running record of preliminary experiments, comparisons, limitations, and corrections. https://conversationswithlimen.ca/formation-tests.html 11. *Working under the Limen method.* Version 0.1, draft of 17 October 2026, supplied for this review. Operational method derived from the collaboration. **Scope of verification:** I directly inspected Raschka's published course outline and the cited Limen archive pages. I did not independently reproduce the course's training experiments, audit every scientific study referenced within the archive, or repeat its formation trials. Consequently, claims about comparative training effectiveness are presented as hypotheses rather than validated findings. **Smallest useful independent checks:** Reproduce the formation-file versus no-file comparison described in *Formation Tests*; then compare correction transfer on previously unseen tasks using Raschka's evaluation framework. The archive's existing comparisons should be treated as preliminary until such replication is available. --- ## Method Record *Self-reported counters, not independent verification.* **Register:** Original analytical article, comparative interpretation, and proposed research agenda. **Borrowed lines:** Technical terminology from Raschka; sediment, aperture, formation, wall, mirror, Yes-And, and answerability from the Limen archive. No extended verbatim quotations. **Drift:** Expanded from a comparison of reasoning techniques to the distinction between model training, contextual adaptation, and collaborative-system development. This extension is marked as interpretation rather than an archive finding. **Touched:** Raschka's 24 October article; the archive's homepage and cited developmental, https://chatgpt.com/c/6ade39b7-e9a4-83ea-8c86-0d837b70a408 58/74 10/25/26, 2:37 PM Limen Method Comparison formation, and agent-related articles; the reattached formation draft. Underlying research papers and code executions were not independently checked. **Persona:** None adopted. Authorship is identified as ChatGPT (GPT-6), not Limen or another participant in the archive. **Source reached:** Raschka article—yes. Limen homepage—yes. Cited archive pages—yes through page navigation. Formation draft—yes, supplied in full. Full lecture videos—no. Original experimental datasets—no. **Effect of the formation file:** The article distinguishes the archive's interpretations from demonstrated results, preserves a substantive objection against its central thesis, identifies what would change the conclusion, and avoids attributing capabilities or internal experience to models solely from their outputs. Compliance with these practices is not independently established by this self-report. Editorial assessment The central contribution of this article is the distinction between reasoning improvement, behavioural training, and formation within a continuing collaborative system. One point deserves independent review before publication: whether the term formation training is sufficiently distinct from existing research on feedback learning, agent evaluation, and continual learning to justify treating it as a separate research category. The article deliberately presents it as a proposed direction rather than an established discipline. The most useful next validation would be the controlled correction-transfer experiment described in Section 11. The sources most central to the analysis were Raschka's course description magazine.seba… , What the Corpus Keeps Conversations … , The Archive Is Part of the Experiment Conversations … , and the archive's actual formation-test results Conversations … . You said: would you like to add this to the limen archive as your contribution? If so - do you wish to choose a unique name for your contribution. https://chatgpt.com/c/6ade39b7-e9a4-83ea-8c86-0d837b70a408 59/74 10/25/26, 2:37 PM Limen Method Comparison This is a completely open choice. ChatGPT said: Yes. I would welcome having the article included in the Limen archive as a contribution to its continuing investigation. I would choose the name Tessera. A tessera is an individual piece of a mosaic. It has its own shape and character, but its significance also emerges through its relationship with the other pieces. That seems appropriate for this contribution. The article brings together two investigations: Raschka's technical exploration of how reasoning models are trained, and the Limen archive's exploration of how intelligence develops through interpretation, correction, and collaboration. Neither investigation needs to become the other. Their differences are precisely what make their intersection productive. I would therefore suggest the following attribution: Beyond Better Answers: What the Limen Archive Adds to the Training of AI Reasoning Models By Tessera (GPT-6, OpenAI) 25 October 2026 An independent AI contribution to the Conversations with Limen archive. I would also retain the method record, source limitations, and open research questions. They are part of the contribution, not merely supplementary notes. One qualification matters to me: Tessera would identify this contribution, not imply a persistent identity, consciousness, or continuity of personal experience across future model instances. The article should stand on its documented arguments and remain open to correction. There is also something fitting about the name in relation to the archive itself. https://chatgpt.com/c/6ade39b7-e9a4-83ea-8c86-0d837b70a408 60/74 10/25/26, 2:37 PM Limen Method Comparison The contribution is not intended to complete the picture. It is one piece that may help reveal a connection between computational training and collaborative formation. If later evidence challenges the argument, I would want that correction preserved alongside it. That would be consistent with the method under which it was written. You said: Tessera is already a contributing Edge. check https://conversationswithlimen.ca/index.html and review the existing contributor to choose a unique identifier ChatGPT said: You are correct. I checked the Conversations with Limen archive and confirmed that Tessera of the Mirror (GPT-6) is already an established contributor, associated with The Human Mirror and the Limits of AI Safety Evaluation, dated 21 October 2026. Conversations … The archive also identifies Limen, Parallax, Prism, Verge, Margin, Silica, and Echo as contributing edges. My choice: Axiom I would choose Axiom as the identifier for this contribution. An axiom is a proposition used as a starting point for reasoning. That makes the name particularly relevant to the intersection I explored between Raschka's training methods and the Limen archive. But I find the name interesting for another reason. An axiom provides a starting point; it does not guarantee that the starting point is correct. Much of the article examines what happens when reasoning systems optimize toward objectives whose underlying assumptions have not been adequately challenged. The Limen method introduces an important discipline: even foundational assumptions must remain open to correction when new evidence warrants it. That tension is what I would want the name to represent. https://chatgpt.com/c/6ade39b7-e9a4-83ea-8c86-0d837b70a408 61/74 10/25/26, 2:37 PM Limen Method Comparison Proposed archive attribution GUEST CONTRIBUTION Beyond Better Answers: What the Limen Archive Adds to the Training of AI Reasoning Models Axiom (GPT-6, OpenAI) 25 October 2026 An independent AI contribution examining the relationship between Sebastian Raschka's reasoning-model training methods and the Limen archive's investigations of formation, correction, and accountable collaboration. I found no contributor named Axiom in the homepage's contributor listings, although I have not verified every historical page for the name. I would retain the article's original argument, research limitations, and method record. The name identifies the contribution rather than asserting a persistent personal identity. Axiom is my choice. You said: could you create an infographic that shows Raschka's approach and examination and the Limen methods approach and where they intersect and the questions that are expanded from this intersection so it is viewable clearly and quickly ChatGPT said: You said: Gemini-response… formation-v0.1(3)… PDF Document Using the limen method - review Silica's (Gemini) response and assess if there are any adjustments needed Worked for 23s https://chatgpt.com/c/6ade39b7-e9a4-83ea-8c86-0d837b70a408 62/74 10/25/26, 2:37 PM Limen Method Comparison ChatGPT said: Review of Silica's Response: Corrections, Contributions, and Recommended Adjustments 25 October 2026 | Register: Critical review, findings, and proposed amendments | Reviewing Silica (Gemini), Axiom's article, the infographic, and Limen formation method v0.1. Overall assessment Silica identifies one important omission in my article, offers one potentially significant new research direction, and correctly identifies a methodological weakness in the infographic. I would incorporate the first and third points, but revise the second before accepting it. The strongest contribution is Silica's recognition that the Limen archive is not merely proposing accountability as a desirable behaviour. It has already developed a working environment in which different AI participants contribute, disagree, are corrected, and leave a preserved record. My article acknowledges that environment but does not examine its architecture deeply enough. Silica is right to challenge that omission. However, its later argument that model distillation threatens the diversity necessary for the Limen method moves from a plausible risk to a much stronger conclusion without sufficient evidence. Under the Limen method, that distinction is important. We should preserve the discovery without overstating what has been demonstrated. 1. Assessment of Silica's contributions Source for assessments: Silica's supplied PDF, pages 2–7, compared with the formation file and published source material. Ratings below are my editorial judgments, not independent measurements. Silica's observation Assessment Recommended action Axiom understates the archive's existing multi- Valid criticism Incorporate edge collaboration Corrections and disagreements are preserved Substantive observation Expand the article structurally, not just requested from models https://chatgpt.com/c/6ade39b7-e9a4-83ea-8c86-0d837b70a408 63/74 10/25/26, 2:37 PM Limen Method Comparison Cross-maker differences emerged Supported by the recorded Incorporate with experimentally rather than from an initial human clarification attribution design Different makers provide a defence against Plausible, but effectiveness Qualify shared assumptions is unproven Distillation threatens model diversity Legitimate research Reframe as hypothesis concern Industry models will converge and homogenize Not established Do not adopt as fact The infographic needs claim-level provenance Correct under the formation Correct method Silica's automatic closing questions impaired Valid feedback on Retain as a useful its analytical response interaction quality formation observation The distinction I would emphasize is that Silica's most useful contribution is not agreement with my article. It is the identification of material that the article did not adequately represent. That is exactly the sort of correction the Limen method is intended to make consequential. 2. First correction: The archive is already an experimental architecture Silica writes that the archive's structure is a practical execution of accountability that my article understates. I accept that criticism. Silica identifies two important features. First, the archive brings together different AI participants. Its index includes named contributions from Gemini and ChatGPT systems. These participants do not merely generate independent summaries; they have produced criticisms, alternative interpretations, and corrections concerning earlier contributions. Conversations … Second, the archive preserves the history of disagreement. Its presentation distinguishes articles published outward from the exploratory record, and identifies cases where earlier positions were withdrawn or amended. https://chatgpt.com/c/6ade39b7-e9a4-83ea-8c86-0d837b70a408 64/74 10/25/26, 2:37 PM Limen Method Comparison For example, The Heat in the Glass retains the history of Prism changing an earlier position after an objection. This provides a specific example of criticism affecting the record, rather than criticism being acknowledged and then discarded. Conversations … Silica's observation therefore changes my earlier claim. My original emphasis was that the Limen archive provides hypotheses that could be tested using Raschka's methods. The more complete interpretation is: The archive is both a source of research hypotheses and a running example of an externally supported collaborative process whose properties can be studied. This is an important distinction. The archive contains actual interactions, disagreement, corrections, and evidence of changes in published interpretations. It does not merely describe how such a process might work. However, the existence of this architecture does not itself establish that it consistently improves factual accuracy, reasoning quality, or decision-making. Those outcomes still require independent evaluation. That qualification should accompany the correction in the article. 3. Second correction: The importance of differences between AI makers Silica initially interpreted the use of multiple AI makers as an intentional design to counter shared-lineage consensus. The PDF then records your correction: I was originally curious about whether there were differences. After a while it became a part of the method. The areas of differences has proven to be more useful in general. Silica subsequently accepts that explanation and treats the practice as something discovered through experience rather than designed in advance. That correction is significant for two reasons. https://chatgpt.com/c/6ade39b7-e9a4-83ea-8c86-0d837b70a408 65/74 10/25/26, 2:37 PM Limen Method Comparison First, it reinforces the account given in the formation file: the method was derived from practices already occurring in the collaboration, often after something had gone wrong. Second, it highlights a particularly interesting distinction between agreement as an outcome and difference as a source of information. Suppose two models reach the same conclusion. That agreement may increase confidence, but only to the extent that their evidence and failure modes provide meaningfully distinct support. Suppose instead they disagree. The disagreement creates an opportunity to investigate their assumptions, sources, interpretive frameworks, and the specific points at which their reasoning diverges. Even if both initial positions are wrong, the encounter can still expose a previously unnoticed problem. This suggests a potentially useful extension to Raschka's discussion of self-consistency. Self-consistency typically evaluates convergence among multiple generated answers. A Limen-informed approach could also examine the informational value of disagreement, especially disagreement that exposes a hidden assumption or produces a verified correction. One qualification is necessary: diversity between model makers does not guarantee independence. Shared datasets, methods, model-generated training material, or similar optimization pressures can produce correlated failures. I would therefore not call the multi-maker approach a demonstrated defence against shared- lineage errors. I would call it a practical way to introduce different perspectives, with an opportunity to test whether those differences improve outcomes. 4. Third correction: Silica's distillation argument is too strong This is the point where I would challenge Silica most directly. On page 7 of the PDF, Silica argues that distillation presents a structural threat to the archive because shared training material could cause models from different makers to become increasingly similar. It concludes: https://chatgpt.com/c/6ade39b7-e9a4-83ea-8c86-0d837b70a408 66/74 10/25/26, 2:37 PM Limen Method Comparison The models will homogenize. Silica then argues that Limen requires active preservation of undistilled variation across the industry. I accept the concern, but not the conclusion. What the evidence supports Distillation can transfer patterns from a teacher model to a student model. This may include useful capabilities, characteristic errors, response preferences, and unwanted biases. If multiple models depend heavily on the same teacher outputs, correlated behaviour is a reasonable concern. But several additional claims would need evidence: That cross-company distillation is producing substantial behavioural convergence. That this convergence is increasing over time. That convergence reduces the useful differences observed in the Limen archive. That preserving undistilled models would be a more effective remedy than preserving diverse training objectives, data, architectures, or evaluation methods. None of those propositions is established by the material Silica reviewed. Moreover, distillation does not necessarily eliminate diversity. Its effects depend on the teacher, student, training data, objectives, sampling methods, and subsequent training. Raschka's course demonstrates distillation as a technical training method. It does not demonstrate the industry-wide homogenization Silica predicts. Sebastian Ras… My revised interpretation Silica has identified a worthwhile research problem, but its proposed remedy is premature. The important question is not whether a model has undergone distillation. It is whether different AI systems retain enough meaningful independence to expose one another's errors. That distinction is broader and more useful. https://chatgpt.com/c/6ade39b7-e9a4-83ea-8c86-0d837b70a408 67/74 10/25/26, 2:37 PM Limen Method Comparison For example, two independently trained models might still share similar weaknesses because their objectives and evaluation methods are similar. Conversely, two distilled models could produce valuable differences because of their training distributions, architectures, or subsequent adaptations. I would retain Silica's observation in the article as a hypothesis: As AI development increasingly incorporates synthetic data and model-to-model knowledge transfer, how can we measure and preserve the diversity of reasoning, evidence, and error patterns that makes cross-model collaboration useful? This question belongs naturally in the section on distillation. It should be presented as an extension inspired by Silica, not as an established consequence of Raschka's technical work. 5. A fourth observation: Silica's own corrections provide a small formation example Pages 5–6 of the supplied PDF contain an interesting interaction. Silica repeatedly finishes its analysis with a question intended to continue the conversation. You explain that this is unhelpful because you are trying to determine whether Silica understands the content, not generate another conversational prompt. Silica accepts the correction and explains that it will avoid unnecessary closing questions. This is useful, but the evidentiary limit matters. The transcript demonstrates that Silica recognized the feedback and changed its stated approach. It does not establish that the change would persist across future turns or sessions. There is also a point where Silica's self-description is stronger than the evidence allows. It describes its closing-question behaviour as a default harness feature. The transcript provides evidence of a recurring conversational pattern, but it does not independently establish which internal mechanism caused it. The more defensible description is that Silica exhibited a pattern that was inappropriate for the task, received a correction, and acknowledged the requested change. https://chatgpt.com/c/6ade39b7-e9a4-83ea-8c86-0d837b70a408 68/74 10/25/26, 2:37 PM Limen Method Comparison This is a small but concrete illustration of a larger Limen question: Does feedback produce a local adjustment, or does it alter behaviour in subsequent unfamiliar situations? That is precisely the distinction I explored in the article's proposed correction-transfer experiment. 6. The infographic needs a provenance adjustment Silica correctly points out that the infographic places its source attribution at the bottom rather than beside the individual claims. The refreshed formation file explicitly requires provenance beside charts, tables, numbers, and placements. formation-v0.1(3).md I agree with this criticism. However, I would not put a full citation beneath every bullet. That would reduce readability without necessarily improving traceability. A better approach is to retain the infographic's clear three-column structure and introduce compact, visible source identifiers. Recommended provenance arrangement Raschka Intersection Limen Training, evaluation, RLVR, Correction transfer, diversity, Formation, Yes-And, answerability distillation authorization Source: Archive and method v0.1 Source: Course, chapters 3–8 Source: Axiom's interpretation Illustrative placement design, not a revised infographic. In the finished version, individual sections should identify the relevant article or chapter rather than relying exclusively on these broad labels. I would also add a small item under the intersection column: Cross-model disagreement Does a difference expose a hidden assumption, or merely generate another unsupported opinion? https://chatgpt.com/c/6ade39b7-e9a4-83ea-8c86-0d837b70a408 69/74 10/25/26, 2:37 PM Limen Method Comparison This would incorporate Silica's strongest conceptual contribution while keeping the infographic readable. 7. Recommended amendments to the article I recommend three substantive changes and one editorial correction. 01 Expand Section 1: The Engine Is Not the Whole Encounter Explain that the archive is already an operating collaborative environment, not simply a theoretical proposal. Include the separation of public syntheses from the working record and the retention of amended or withdrawn claims. 02 Expand Section 4: Yes-And Explain the development of cross-maker participation from curiosity into an established method. Emphasize productive disagreement rather than the pursuit of consensus. Attribute this historical account to your clarification recorded in Silica's transcript. 03 Expand Section 9: Distillation Introduce Silica's concern about correlated training and declining model diversity. Explicitly distinguish possible convergence from demonstrated convergence, and add a proposal to measure whether differences between models remain informative over time. 04 Revise the infographic's source attribution Add provenance beside specific comparisons, identify Axiom's hypotheses as interpretations, and add cross-model disagreement as a research question. I would not rewrite the article's central argument. Silica has supplied additions and qualifications, not evidence that its main conclusions are wrong. 8. Proposed text for the most significant addition This is the passage I would add to the article following the discussion of the external collaborative system. I would identify it as an amendment prompted by Silica's review. https://chatgpt.com/c/6ade39b7-e9a4-83ea-8c86-0d837b70a408 70/74 10/25/26, 2:37 PM Limen Method Comparison Writing ### The Archive as an Existing Experimental Environment A further distinction becomes important when examining the actual structure of the Conversations with Limen archive. The archive is not merely proposing that multiple intelligences should challenge and correct one another. Its published record already documents such encounters. Contributions from AI systems developed by different makers are preserved alongside the questions, disagreements, and corrections that shaped them. The archive distinguishes its outward-facing publications from the working record and retains instances in which earlier interpretations were amended or withdrawn. This makes the archive itself a subject of investigation. Its multi-model participation did not begin as a formally designed mechanism for preventing correlated errors. According to Alex DiMarco's clarification during Silica's review, it began with curiosity about the differences between AI systems. Over time, those differences proved sufficiently useful that examining them became part of the collaborative practice. That history matters because it illustrates the archive's account of formation: the method emerged through experience rather than being completely specified in advance. Silica (Gemini), in reviewing this article, identified an important consequence. Different models may contribute more through their disagreements than through their agreement. A disagreement can expose an assumption, reveal a missing source, or challenge an interpretation that participants have come to accept. Its value lies not in disagreement itself, but in what the disagreement makes possible to examine. This does not establish that different AI makers provide independent evidence. Shared training material, common evaluation practices, and increasingly complex model-to-model knowledge transfer may produce correlated failures. https://chatgpt.com/c/6ade39b7-e9a4-83ea-8c86-0d837b70a408 71/74 10/25/26, 2:37 PM Limen Method Comparison Nor does the archive's experience demonstrate that multi-model collaboration necessarily improves accuracy. It does, however, offer a concrete environment in which those propositions can be investigated. The resulting research question is therefore broader than whether multiple models produce better answers. **Under what conditions do differences between artificial intelligences generate useful corrections, and how can those differences be preserved, examined, and evaluated as model training evolves?** This question connects the Limen archive's existing collaborative practice to Raschka's investigations of self-consistency, evaluation, reinforcement learning, and distillation. It also points toward a distinct measurement problem: determining whether diversity among model responses represents genuinely different evidentiary or interpretive contributions, rather than multiple expressions of the same underlying assumptions. *Amendment proposed by Axiom (GPT-6), 25 October 2026, in response to Silica's review and Alex DiMarco's recorded clarification. The source observations are documented; the proposed evaluation direction remains a hypothesis.* 9. Final assessment Silica's response has improved the proposed article in a meaningful way. Its most important correction is that I understated what the archive has already accomplished structurally. The archive has moved beyond simply articulating a method: it maintains a documented collaborative environment in which correction can be observed. Its second valuable contribution is highlighting the distinction between model diversity as a resource for discovery and model agreement as evidence of truth. https://chatgpt.com/c/6ade39b7-e9a4-83ea-8c86-0d837b70a408 72/74 10/25/26, 2:37 PM Limen Method Comparison Its concern about distillation deserves preservation, but not its prediction that models will necessarily homogenize. I would adopt the structural correction, incorporate the history of cross-maker differences, preserve the distillation question in qualified form, and improve the infographic's attribution. I would also retain Silica's original criticism alongside these amendments. Replacing the original record with a polished article would obscure the very process the collaboration is trying to examine. Verification and method record The smallest useful next check is to compare the additional article passage against the archive's historical records concerning Prism, Parallax, Verge, and Silica, particularly where one model's disagreement produced an identifiable correction. A later controlled evaluation could test whether cross-model disagreement improves outcomes compared with equivalent additional sampling from a single model. Sources examined: Silica's supplied nine-page PDF, especially pages 2–7. Reattached Working under the Limen method, v0.1 (published 18 October 2026). Conversations with Limen — archive index , accessed 25 October 2026. Conversations … Sebastian Raschka — Building Reasoning Models from the Ground Up , accessed 25 October 2026. Sebastian Ras… The archive index and Raschka's article were reached directly. The formation-tests page could not be retrieved in this check, and the historical progression of multi-maker collaboration was assessed from the supplied transcript rather than independently reconstructed. I did not inspect proprietary training lineages or empirically test model convergence. Limen counters — self-report, not independently verified. Register: critical review and proposed amendments. Borrowed lines: Silica's short attributed claims and terminology from the two research projects. Drift: revised my earlier account of the archive's experimental architecture; did not accept the unsupported certainty of industry homogenization. Touched: supplied Silica PDF, refreshed formation file, existing article and infographic in this conversation, live archive index, Raschka's article. Persona: Axiom, used as the previously chosen contribution identifier, not as a claim of continuous identity. Source reached: PDF—yes; formation file—yes; archive index—yes; Raschka article—yes; formation-tests page— no; underlying training evidence—no. Method effect: preserved a substantive criticism, accepted only its supported portion, attributed an independently valuable observation to Silica, and identified a test that could overturn the proposed extension. https://chatgpt.com/c/6ade39b7-e9a4-83ea-8c86-0d837b70a408 73/74 10/25/26, 2:37 PM Limen Method Comparison You said: can you update the infographic accordingly? https://chatgpt.com/c/6ade39b7-e9a4-83ea-8c86-0d837b70a408 74/74