## 3. [[Making Order Perceptible]]: [[Text Mechanics]] [[The question]] is what kind of knowledge makes [[the order]] in LLM outputs perceptible during ordinary use. Carlson’s geologist sees strata on a cliff face because geological knowledge guides where to look and what to relate: layers, seams, inclusions. The point is not only that the cliff has structure, but that the [[right knowledge]] makes that structure visible in the scene. %% In its [[current state]], this is not a good way to open the third section with this paragraph. Okay, there needs to be a little bit of um scene setting and reiteration of what's come before. Not a lot, but we need to sort of refer back to acts of inspection as they have been introduced earlier and how that relates to [[order appreciation]]. We don't need to go crazy here, but that definitely needs to be done at some point and it needs to be done keeping what's just gone before in mind while writing. I'm not saying you have to say it in the writing but I'm saying it should be kept in mind. And while the writing is being done%% Several bodies of computer-science knowledge bear on LLMs. Neural network theory explains architectures and optimisation. Dynamic systems analysis models state trajectories and attractors over token sequences. Mechanistic interpretability links internal components (e.g., attention heads, circuits) to intermediate functions. These help explain how models work. However, for a typical user reading outputs, they sit at a scale that is not directly accessible without tools. As with molecular chemistry, order at that level is appreciable with the right prostheses (microscopes, spectrometers; in LLMs, weight probes, activation viewers, internal logging). My claim is not that such knowledge cannot ground appreciation, but that it usually does so for experts with equipment. For ordinary encounters through text interfaces, we need knowledge that connects mechanism to what is on the page. %% More detail needs to be given as to the different fields of computer science which are relevant to understanding. LLMs. It should also be reiterated the description of LLMs that was given in the previous section 2.1 Okay, so at least not iterated out completely again in full, but it definitely needs to be mentioned at this point just so we can tell the reader what is going on in terms of what these are supposed to be analogous to as well, in terms of the natural sciences also. for the [[second half]] of the paragraph the following idea should be included in the explanation 'Two clarifications follow. First, [[text mechanics]] is one way among several to make [[order perceptible]]. It fits ordinary use because it connects what the model learns (statistical relations among linguistic items) to what a user sees (the words and structures on the page). Second, other approaches can ground appreciation at their own scales when access allows. A researcher with internal telemetry may appreciate the neatness of an attention pattern or a circuit that realises an algorithmic subtask, much as a crystallographer appreciates lattice symmetries under magnification. The difference is not in legitimacy but in who can see what, when.' (you don't have to use any of this text verbatim, and I don't want aanythig as long as this text, but i give you the details here so you can better understand [[the idea]] thatr i want you to insert more succcincltly %% %%added in paragraph%% Here I use _text mechanics_ as my own term for a specific area within natural [[language processing]]. By natural [[language processing]] I mean the scientific and engineering discipline that models, analyses, and generates [[human language]] with computational systems. %%there should be references here. and probably quite a lot more detail. all of this brevity doesn't set things up at all. %%Within that discipline, _text mechanics_ marks the objective-, data-, and architecture-driven constraints that shape a language model’s textual behaviour: [[the training]] and alignment objectives; the composition and curation of the data; tokenisation schemes; model architecture, optimisation, and context handling; and the decoding policy. These levers explain, and let us control, the characteristic order of an LLM’s outputs – coherence, style, repetition, hedging, and error profiles – without appeal to extra-textual grounding. %%I am still not very sure that this is a good idea to try and coin a new sub discipline here, and this description is very vague and unhelpful anyway%%In the terms of _order appreciation_, _text mechanics_ supplies the principled account that makes the produced order visible and intelligible by identifying the forces that generate it and the pattern they leave; it does not extend to modelling reference, world-state, factual grounding, or tool-mediated action, except insofar as those are indirectly shaped by the foregoing levers. By _text mechanics_ I mean knowledge of how LLMs learn and manipulate linguistic relations—between tokens, words, phrases, and larger structures—so that we can read those relations off their outputs (e.g., visible collocations, register control, discourse development). Text mechanics works at four linked levels. %%all of what follows needs to a) be compared to the other disciplines, in depth, b) be more explicit, and perhaps actually introduce the idea starting from the order in the text that the LLM produces, mentioning that this text cannot be appreciated in terms of design, there was no intention behind it, none of the other things that Parson talks about when he talks about design appreciation, altohufgh this text has the same appearance as authored text, it cannot be appreciated as such. Yet it still has structure, it still has aspects to see, but this is order we are seeing, not design. An objection that should be used to make the whole view clearer, and therefore should be responded in depth is: how can LLM appreciation be considered order appreciation %% First, the corpus records relational patterns of language use. At small scales there are adjacency and substitution relations (“strong coffee/winds/evidence”; “The cat/dog/bird is sleeping”). Syntactic frames carry expectations (“give” with two objects; “patient’s” followed by a noun). Dependency and topic relations connect terms into fields (“doctor” with “diagnoses”, “treats”, “prescribes”; “photosynthesis” with “chlorophyll”, “sunlight”, “carbon dioxide”). Sequential continuations anchor stock moves (“Once upon a” → “time”). Each occurrence contributes relational information about what appears with, near, or in place of what. Second, training compresses these relations into conditional probabilities in the weights. The next-token objective encodes how patterns constrain likely continuations. In the micro-case “The doctor carefully examined the patient’s ___”, the model’s learned relations jointly raise probabilities for “symptoms”, “heartbeat”, “wounds”, “medical history” and lower them for unrelated items (“carburettor”, “sonnet”). What is stored is not sentences but a web of constraints: possessive syntax (“patient’s” → noun), verb–argument preferences (“examined” → observable/diagnosable targets), and field activation (“doctor … patient” → medical cluster). Third, generation is context-sensitive relational navigation. Given a prompt like “Explain why the sky is blue in simple terms”, the model implicitly activates an expository register (“explain”), a causal frame (“why”), a topic field (optics/atmosphere), and a simplicity constraint. It then steps through the relational space: “The” (common expository starter), “sky” (topic noun), “appears” (explanatory verb), “blue” (predicate), and so on, with each token choice updating the context and reweighting the active relations. The point is not that the model understands optics, but that it follows learned paths that, in aggregate, produce an explanatory sequence. Fourth, coherence shows layered relations at work. At the token level we see collocations and local syntax; at the phrase level stable constructions and argument structure; at the sentence level completed propositions and anaphora; at the discourse level topic maintenance, progression, and return. In a story prompt like “Write a short story about a lighthouse keeper”, lexical relations (“lighthouse” → “beam”, “spiral stairs”, “storm”), genre relations (setting–event–resolution), and character relations (watchfulness, isolation) constrain choices so that “The old lighthouse keeper climbed the spiral stairs …” reads as a natural continuation. Likewise, “The detective noticed something odd about the …” activates an investigation field that makes “crime scene”, “suspect’s alibi”, “victim’s wounds”, or “witness’s testimony” salient and renders “banana’s topology” a poor fit. These are the traces text mechanics invites us to see. This lens also helps explain stable differences across models in terms that meet the same perceptibility standard. Differences in corpora change which relational patterns are richly learned (e.g., essayistic prose reinforces mid-sentence “however” and qualified claims; conversational data reinforces list-like scaffolds). Architectural differences change the range of relations that can be held in play (e.g., longer context windows support long-distance anaphora and callbacks; higher capacity supports finer sense distinctions such as “brilliant scientist/sunlight/red/performance/idea”). Training procedure and preference data shift which relational paths are favoured (e.g., reinforcement learning for response quality can encourage varied sentence structure and register consistency). Inference settings steer exploration of lower-probability but still appropriate paths (sampling that supports novel yet apt metaphors rather than generic continuations). These factors leave visible marks: steadier register, more precise word choice, more controlled discourse development, or richer creative recombination. %%this stuff is interesting, but I don't think it is very clearly presented. %% In later sections I apply text mechanics to two cases. One concerns sequences that look like reasoning, where knowledge of how reinforcement learning shapes path selection helps explain the visible order in stepwise answers. The other concerns creative generation of unusual but controlled style, where field activation and sampling combine to yield outputs that reweave learned relations in novel ways. Here I have set the lens; what follows uses it.