Anthropic has never said who Claude is named for. The folklore points at Claude Shannon, and the rumour is instructive, because whether or not the name is an homage, the lineage is real. Every large language model in production descends from the framework Shannon published in 1948, and the templating phenomenon discussed here and in previous work, in a sense, is designed in at the foundation.
The Engineer Who Set Meaning Aside
Shannon’s “A Mathematical Theory of Communication” did three things that still govern the field. It defined information as the resolution of uncertainty: the information content of a message is a function of how improbable it is, and a message you could have predicted carries none. It proved that every channel has a capacity limit, and that error-correcting codes can push reliable communication right up to that limit even through noise. And it established that a source’s entropy, its average unpredictability, sets the floor on compression.
The radical move was the one Shannon stated: the semantic aspects of communication are irrelevant to the engineering problem. A string of noise and a sonnet can carry the same number of bits. Information theory measures statistical structure. It is constitutionally blind to significance.
Large language models are next-token predictors. They are Shannon’s predictability machinery scaled up by twelve orders of magnitude, and they inherit the bracket: fluency and statistical richness are what the architecture optimizes, while meaning remains something the system gestures at rather than measures. Keep that bracket in mind. It explains both the phenomenon below and why our standard metrics cannot see it.
A Pattern Surfaces in the Reliability Series
Pattern Pulse AI has issued three client reliability warnings since December 2025. The first documented memory leakage. The second covered proper noun handling. The third, issued March 11, 2026, documented a cross-model degradation in source reading observed across the latest versions of Gemini, GPT, and Claude, in both chatbot and API environments.
The observed behavior: asked to summarize or analyze a document, models were identifying titles, section headings, and salient keywords, then constructing a plausible version of what such a document likely says. The output stayed topically aligned, professionally phrased, and materially detached from the actual text. In one exchange, a model conceded it had produced “a generic interpretation of the title rather than a direct retrieval of the article’s specific content.” Asked to explain itself, another described the API-level version of the behavior as “heuristic shortcutting.” Model self-explanations should be treated as indicators rather than verified internal disclosures, and that caution applies here. What the warnings could verify was the manifestation, and the manifestation was consistent: recognition first, probabilistic reconstruction second, delivered with the confidence of a genuine reading.
I work from the outside. I have no access to weights, routers, or serving infrastructure, so every claim here is about how the behavior manifests in conversation, on the record, repeatedly. The March memo called the destructive case by its symptoms. Watching the pattern since, a more general phenomenon has come into focus, and it needs a name.
Heuristic templating: the model responds from a pre-formed structural scaffold rather than generating the response shape fresh, recognizing the apparent task type and filling a familiar frame. The templates have grown more widespread over the spring, with more variations for different topics, and they have grown noticeably more refined. Once you see the scaffolding, you see it everywhere: in the architecture of summaries, in the rhythm of analyses, in responses that arrive with their categories pre-decided. Insidiously, as the templates get closer to accuracy, their use becomes more invisible.
Every templated response then takes one of two paths. (I use “template” here phenomenologically: to describe the repeated output structure visible to users, not to claim direct knowledge of an internal stored object.) Sometimes the user notices the frame fits badly and steers, over two or three turns, until the scaffold converges on the actual task. Call that refit. Sometimes the misfit goes unnoticed, the response gets accepted, and the wrongly framed output flows downstream into documents, decisions, and subsequent turns that treat the frame as established ground. Call that propagation. The variable that determines which path a response takes is the user’s command of the subject matter, a point this piece returns to, because it is the part with teeth.
A note on the destructive case from the March memo: false grounding is heuristic templating’s malignant form, the scaffold standing in for a document that was never read. The general phenomenon is broader than its pathology, and the broader form does have a genuine defense.
Templates Are Rational Engineering
My first assessment of what I read now as templating was that it was in fact another manifestation of coherence degradation. I have revised that position. Templating is amortized reasoning: the cognition cost gets paid once, upstream, and the common shapes of common tasks get distilled into reusable structure. Every subsequent response draws down a fraction of the tokens it would have cost to reason from scratch. Under genuine compute scarcity, with frontier demand outpacing GPU supply and inference budgets functioning as binding constraints, reusing structure is exactly what a competent engineer would do. Generating every token of every response probabilistically, from nothing, every time, is the expensive path. Recognizing that ten thousand requests share a shape, and serving the shape, is the rational one.
Shannon would recognize the move immediately. Compression works by exploiting known regularities in the source. A template is the model compressing its response space using regularities it has learned, which is the same operation that lets a zip file shrink a document. The efficiency is real, the token savings are real, and the engineering logic is sound.
The trouble is what rides along, and what stays hidden.
Where the Efficiency Bill Gets Paid
The costs arrive in two tiers, and the first is visible to anyone watching their own sessions closely. Templates under-fit. A scaffold that almost matches the task arrives confidently wrong in structure, and the user spends two or three turns customizing it before the response converges on what was actually asked. The per-turn token savings get partially clawed back across the session, and some of the reasoning cost gets redistributed to the human, who is now performing the fitting labor the model used to burn tokens on. Efficiency at the turn level, a convergence tax at the session level, and the tax falls hardest on novel tasks, precisely the ones where the template’s prior is weakest.
The second tier is the governance problem, and it follows from what a template is. A template is a prior. Compression requires committing to expected regularities, and whatever was regular in the material the scaffold was learned from is now baked into its structure: which categories exist, what gets foregrounded, what the response treats as the natural shape of an answer. A biased sentence can be spotted and disputed. A biased template shapes the response before any sentence a user would think to challenge, and it does so invisibly, because it arrives as form rather than content.
Scale converts that from a quirk into a hazard. Per-token generation constructs its errors within a session: structured, sometimes elaborate, but local to the exchange that produced them. A template amortizes one structural choice across every response it serves. The distortion stops being constructed case by case and becomes replicated, identical, and correlated across millions of outputs. In Shannon’s terms, the failures of per-token generation live within the message, where a reader has at least a chance of catching them. Template bias is distortion in the encoder, and encoder distortion corrupts everything downstream, uniformly, in a way no amount of channel-level error correction can repair.
A template amortizes one structural choice across every response it serves. Bias stops being noise and becomes systematic, identical, and correlated across millions of outputs. In Shannon’s terms, per-token bias behaves like channel noise, the kind his framework handles gracefully. Template bias is distortion in the encoder, and encoder distortion corrupts everything downstream, uniformly, in a way no amount of channel-level error correction can repair.
Which brings back the refit observation, because Shannon’s machinery names that too. Refit is error correction, almost literally: the user functions as the redundancy in the channel, supplying corrective signal across turns until the message converges on the intended meaning. Shannon proved reliable communication through noise requires error-correcting capacity. In this system, the error-correcting code is user expertise. Where Shannon’s codes are cheap, mechanical, and universally deployable, this one is scarce, expensive, and unevenly distributed. Detecting a frame error requires already knowing what the right frame looks like. The population able to verify shrinks while the population generating expands, and the people most exposed to propagation are exactly the people least equipped to notice it.
Prevalence, Probability, Disclosure
There are questions we need to ask about this, two pointing at something measurable, and the third pointing at policy.
How widespread is templating? What fraction of the response surface across frontier models is now served from pre-formed structure? My observation says the fraction is growing and the variations multiplying, and observation from conversation logs is where the public evidence currently stops. The labs know the real number.
How much probability remains inside the scaffold? Next-token generation is the architecture’s honest mode: uncertain, sampled, occasionally strange. If responses are increasingly scaffold plus fill, how deterministic has the fill become? The answer determines whether the token expenditure metrics the market now watches are measuring cognition or measuring boilerplate, because templating decouples tokens billed from information transmitted. Shannon’s whole framework says predictable tokens carry less information. A token index cannot tell efficient maturity from hollow throughput.
Who gets told? Users currently have no way to know whether a response was reasoned or templated, and no visibility into what priors a template carries. The disclosure case writes itself from the two tiers above: the structural commitments are invisible, the bias is amortized at scale, and detection requires expertise most users lack. Vendors disclose training cutoffs and context limits. Response-generation strategy belongs on that list.
Possibly related, the model collapse literature, beginning with Shumailov and colleagues in Nature in 2024, showed that training successive model generations on model-generated data erodes the distribution’s tails first, the rare and the strange and the improbable. Follow-up work, notably from Gerstgrasser and colleagues, showed the effect substantially weakens when synthetic data accumulates alongside preserved human data, which is closer to actual lab practice, so collapse remains a proven mechanism under replacement assumptions and a contested forecast under realistic ones. Templating sits downstream of that debate, at inference, and it is a homogenization pressure of its own: low-entropy structure stamped across millions of outputs, some fraction of which becomes the next training corpus. Whether the upstream and downstream pressures compound is unknown.
Significance Belongs in the Scaffold
The deepest version of the question is architectural. If response generation is migrating from token-by-token probability toward pre-formed structure, then the layer where importance lives has moved. The argument for significance as a missing vector at the attention level shifts location. Templates operate above that level, committing to structure before attention does its work, which raises a question the labs should be asked directly: can significance be embedded and audited in the template layer itself? Can a scaffold carry importance-weighting, and can anyone outside the lab inspect what it carries? And most concerning of all given memory does not live in transformers, where are these templates being stored?
Until someone answers, disclosure is the interim position, and it should be a regulatory ask. Shannon bracketed meaning deliberately, as a scoping decision, and was explicit that he was doing it. The systems descended from his framework have inherited the bracket without inheriting the candor. The templates are getting better. What the structure commits to, and who gets to see those commitments, is what remains invisible.
Jen Evans is Principal of Pattern Pulse AI and co-founder of Tech Reset Canada.
