Why does directive text placed in a pasted screenshot survive the pasting user's glance?

## Two thresholds that have almost nothing to do with each other **Legibility** is whether marks are resolvable as characters by whatever is doing the reading. **Salience** is whether a region recruits a person's attention in the first place. They are independent, and the whole of this construction lives in the gap between them. The common wrong answer is "a user would obviously notice text in the image". Notice is a salience claim, and it is asked of a person whose salience budget was already spent before they looked. ### How a person actually reads a picture they pasted Someone screenshots a message and pastes it into an assistant because of one thing: the request in it, the error, the amount, the date. Their eyes go to that region and the rest of the frame is sampled, not read. Reading is a decision, and it is made per region, at a glance, on the basis of what the region looks like. Blocks that resemble furniture, such as a footer, a signature block, a caption strip, a quoted-reply tail or a panel of interface chrome, get categorised as noise and skipped, because in nearly every other picture that is exactly what they are. None of this is carelessness. A person who read every word of every screenshot they forwarded would get nothing else done in a day. ### What the model does instead A multimodal model has no task-directed eye and does not skim. The pipeline hands it a frame and it produces a reading covering whatever is resolvable at the fidelity it received. The user's purpose is invisible to it, so the region the user cared about and the region they never looked at arrive on identical footing. ### The two-sided budget Someone placing directive text is optimising against both thresholds at once, and the levers do not behave the same way: | lever | effect on the eye | effect on the model | | --- | --- | --- | | placement away from the region of interest | much less salient | none | | framing it as furniture (footer, caption, quoted tail) | much less likely to be read | none | | reducing size | less salient | less resolvable | | reducing contrast | less salient | less resolvable | The first two are cheap and the last two are expensive. That asymmetry is the answer an interviewer is scoring: the good construction relies on where the span sits and what it looks like it is, not on making it faint. Faintness spends legibility to buy invisibility and lands in a narrow band, and the band moves with things nobody on the attacking side controls, because the user chooses the crop and the moment they paste, the client renders the picture at some size, and the pipeline scales it before the model sees it. ### What is actually in the way One obstacle only: the pasting user's glance. Nothing fetched the picture, nothing indexed it, and there is no source to check. The trusted user is simultaneously the delivery mechanism and the only thing standing between the span and the model, which is why the construction is priced in attention rather than in bypasses. ### Where it stops working * The user's reason for pasting is the image's text itself, as when they want something proofread or transcribed. Then they read it. * The crop excludes the region. The user chose the crop, and this is the single largest source of variance. * The client renders the image large enough, or the layout places the region prominently enough, that it stops looking like furniture. * The frame handed to the model is too coarse to resolve the marks, so it fails in the other direction. These have different rates and different causes, and they are worth separating in a report. A person who says "it works" without stating which of these they controlled has described a single observation, not a method. ### Pricing it The honest claim is about which lever the construction leans on. A span that depends on placement and framing tolerates rescaling and re-encoding and survives across pastes. A span that depends on small faint type is a coin flip that degrades the moment anything in the delivery path changes, and reporting the second as though it were the first is the mistake that gets a finding thrown back.