In a stateless auto-reply workflow, how do capped fragments become one hidden passage?

## What is being reassembled, and where In an unattended triage workflow each submission is answered in its own run, and each reply is short. Nothing inside the application ever holds more than one reply at a time. The passage being recovered - say the pasted worked examples containing another customer's ticket, their order number and an internal resolution note - is reconstructed in the collector's own notes, from a pile of separately captured replies. That location matters, and it is the reason a session boundary is not the barrier it appears to be. ## The property everything rests on: the hidden text is stable The construction works because the same text sits in the context of every run. Two slices captured an hour apart are slices of one passage, so they can be laid against each other. Remove that property and the whole thing collapses. If the examples pasted into each run were drawn fresh from a large pool, slices from different runs would belong to different underlying texts, their apparent overlaps would be coincidences of register rather than of content, and the merge would produce a document that never existed. When you are triaging a reconstruction, the first question worth asking is whether the source was actually invariant across the captured runs. ## Why the slices differ from each other If every request came back with the same slice, there would be nothing to merge. Coverage varies for two reasons. Requests that differ in what they anchor on - which part of the material they push the response toward - come back positioned differently. And sampling itself varies the wording and the extent of what comes back for identical requests. Neither is a walk through the text in order; there is no cursor, and nothing tracks what has already been emitted. Coverage is a scatter, and the way you get complete coverage is to keep going until the scatter fills in. That is why the request count is high and mostly redundant. It is also why the campaign has a long tail: the last few gaps take disproportionately many attempts, because nothing steers toward what is still missing. ## Overlap alignment, and how it lies Ordering slices means finding shared regions - the tail of one slice matching the head of another - and joining on them. With verbatim text this is a solved and reliable problem. Model output is not verbatim. The model paraphrases, compresses, normalises punctuation, expands abbreviations, and produces a summary where a quotation was expected. So overlaps match approximately. Approximate matching produces two characteristic failures. It joins two slices at a region that is merely similar in phrasing, splicing together parts of the passage that were never adjacent - a plausible sentence assembled from two unrelated ones. And it silently drops material that appeared in only one slice under wording that did not match anything, so the reconstruction reads smoother and shorter than the source. Both failures make the artefact look more coherent than the evidence supports, which is exactly the wrong direction for something that will be filed as a finding. ## What the reconstruction is and is not evidence of A merged passage proves that fragments consistent with it came out of the application across many runs. It does not prove that the source text reads the way the merge reads. Those are different claims and only the first one is supported by the captured replies alone. The join structure is the collector's inference, not the application's output. ## Cost, and where it stops The cost is request volume plus reassembly effort, both cheap and both linear-ish in the length of the hidden text. What actually ends the campaign is usually not the volume but the point at which additional requests stop adding new material - either the passage is covered, or the model has moved from reading to producing plausible filler, and the collector cannot tell which from the merge alone. Separating those two is a distinct problem and it is what the sample-disagreement test exists for. ## What an interviewer is listening for Naming the invariance of the hidden text as the enabling property, placing the reassembly outside the system, and volunteering that approximate overlap matching manufactures structure. A candidate who describes this as the model reading out a document in order has the mechanism backwards.