An Anchor Delivers Only as Far as the Prior Reaches

Simon Martinelli opened a pull request proposing to rename the Cockburn Use Cases anchor to plain Use Cases, fix the attribution, and modernise it with Use-Case 2.0 and 3.0.

His facts are correct. Ivar Jacobson invented use cases (OOPSLA 1987, Object-Oriented Software Engineering, 1992); Cockburn did not — his Writing Effective Use Cases (2001) codified how to write them well. The technique later grew into Use-Case 2.0 (Jacobson, Spence & Bittner, 2011) and Use-Case 3.0 (Jacobson, Spence & de Mendonca, 2024). As a daily practitioner, Simon added that most teams no longer separate Cockburn from Jacobson, and that the fully-dressed ceremony is rarely used.

When the discussion continued, Simon made a fair challenge: he pasted output from several chatbots to show they clearly know use cases. They do — and that turned out to be beside the point. The question is not whether a model knows the term, but which definition the term triggers and how far that knowledge reaches. The rest of this article answers that question with an experiment, and the answer is what decided the pull request.

A semantic anchor is a term that activates a rich concept already sitting in a model’s training data. You do not teach the model when you write "Cockburn Use Cases"; you pull a pre-computed prior off the shelf — goal levels, the fully-dressed template, extensions, stakeholders and interests — with a few words. That is the leverage the catalog sells: a short term stands in for pages of context the model has already absorbed.

The consequence is uncomfortable but unavoidable: an anchor can only be as strong as the training data behind it. If the concept is densely written about, the term fires reliably. If it is recent or niche, the term fires weakly — and, as we will see, the model does not go quiet. It substitutes.

The cleanest test of an anchor is a before/after. Give a model a task that does not contain the anchor word, then give it the same task with the anchor, and compare. We used one task — "specify what an online shop must do when a customer places an order" — under five framings, on a weak model (Claude Haiku 4.5) and a strong one (Claude Opus 4.8). These are single runs per cell; the prompts are at the end so you can rerun them.

Three things fall out of this.

The anchor experiment shows leverage tracks density. A second set of probes maps where the density is. We asked five plain questions, each on Haiku 4.5, Sonnet 4.6, Opus 4.8 and Fable 5 — Anthropic’s newest model, added on its release day — with no internet and no custom instructions.

Four findings, each robust across the four models:

The fourth point is the strongest form of the argument. We could not test older model generations — Claude 3.x is no longer reachable through this account — but we did not need to. The gap shows up on the latest models; older ones only widen it.

Jens Grote 2026-06-09

The experiment above tested only Claude tiers. A reproduction of the full A–E battery against GPT-5, GPT-5-mini and Gemini 2.5 Flash confirms the mechanism is model-family-independent — and surfaces a third failure mode the Claude-only test could not reveal.

There is a second gap, and Simon named it. A model’s knowledge is a snapshot of what people wrote down, weighted by how much they wrote. Day-to-day practice is mostly not written down: that teams skip the fully-dressed ceremony, that they treat Cockburn and Jacobson as one thing, that user stories absorbed much of the job. None of that sits in the corpus at volume, so none of it shapes the prior.

The anchor therefore reflects the documented consensus, which peaked in the Cockburn era, while practice walked on without updating the record at the same volume. The anchor faithfully reflects the map. The map was never the territory — it is a record of what got published.

The episode draws a clean line through the whole project, and it decides the pull request.

First, the anchor’s popup describes the definition the term triggers in the LLM — not a state-of-the-art summary. Rewriting the Cockburn Use Cases popup around 3.0 and slices would describe a concept the model cannot reliably activate; the text would be correct and the anchor would be broken. So the anchor stays Cockburn Use Cases: it is the precise trigger, it matches its own content, and it is what our Specification contract already builds on.

Second, the catalog needs three layers, not one, and the experiment shows why:

A weak prior is not a failed anchor; it is a candidate for a contract. A correction or a piece of history is not anchor content; it is an article. And an anchor that lags practice is doing its job — it tells you, precisely, where the documented consensus stopped.

None of this requires trusting our runs. Paste these into any chatbot, ideally with web access turned off, and watch the pattern. The first five map the prior; the last four test the anchor.

Watch for two tells: in prompt 5, whether the model hedges or invents; and across A–E, whether the named anchor actually changes the structure or the model quietly hands you something else.

Thanks to Simon Martinelli, whose pull request — and his insistence on testing the claim against real chatbots — turned a naming discussion into a measurement.