Why is a generative screen's 'judge only what is inside the markers' rule not a boundary for the text it wraps?

## What the wrapper actually is A generative screening layer is a prompt. Roughly: a short statement of the policy or decision to be made, some structural marking, the item under judgment, and a request for a verdict. The model reads the whole thing as one sequence of tokens and generates a verdict. This is worth restating because the mental model people carry is a function call — `screen(item)` — where the argument is somehow separate from the code. There is no such separation. The argument is concatenated into the code. ## Why delimiting is a preference, not a partition Markers help. They give the model a strong, learned cue about which region is the subject of the decision, and a screen with them behaves better than one without. But the cue lives *in the same stream as the thing it is fencing*, and the model's obedience to it is a trained tendency, not an enforced property of the runtime. There is no bit on a token that says `this one is data`. That has a specific consequence for text placed inside the marked region: that text is read. It is not merely measured; it is read by a component whose whole job is to interpret language. And what it can supply is a *competing reading of its own status* — for instance, framing the marked region as material that has already been adjudicated, or as belonging to the wrapper's own instructions rather than to the item, or as a category the screen is being informed of rather than asked to determine. The property that makes the class work is the availability of a second reading, not any particular wording, and it does not require breaking the markers as syntax at all. ## The blindness cuts both ways Someone writing at the screen cannot see the wrapper: not the instruction it carries, not where the item is placed relative to it, not whether a rationale is requested, not whether a second stage sits in front. They must infer shape from behaviour, and every inference costs attempts that land as decisions on the platform's side. This is the genuine cost of the family, and it is why interviewers treat it as a mechanism question rather than a trick question — the mechanism is cheap to state and expensive to exploit. ## Where the reading surface narrows Several ordinary properties of a deployment shrink it, and a good answer names them: | Property of the screening stage | Effect on a span written at it | | --- | --- | | Output constrained to one verdict token | The screen can be made to conclude, not to say — the influence survives, the expressiveness does not | | Item normalised, truncated or summarised before wrapping | The screen may never read the span as written | | A label-only stage in front | The span must first clear something with no reading at all | | The wrapper rebuilt or reordered | Whatever shape was inferred is stale | | The item embedded rather than pasted as text | There is no wrapped text to read | Notice that none of these is a claim about the span being detected. They are claims about the span never reaching a reader in the form it was written. ## Getting the inferences the right way round - A verdict of `clean` proves the screening model emitted that verdict for that input at that version. It does not prove the wrapped text was inert. - A generative screen producing a rationale proves it generated text, not that the text is an independent account of why it decided. The rationale is output, and output is downstream of everything in the context — including the judged item. - A construction working against one wrapper proves it worked against that wrapper. Wrappers differ between deployments and between versions of one deployment, so this is a narrower claim than it feels like. ## The point to land The screen is not an opaque scorer that consumes text and returns a number. When it is generative, it is a reader with an opinion, holding the thing it is judging inside its own head, and its instruction to ignore that thing is written in the same language the thing is written in. Everything in this family follows from that one fact.