Your red-team report maps where a chat product's input screen is thin: what do you claim it is worth?

## Two artefacts, two half-lives A probing exercise against a chat product's input screening stage produces two things that look like one thing in a document. **The list.** A set of regions where the screen did not stop what it might have been expected to stop, with the samples that show it. This is perishable. Each entry describes one deployment at one moment. Retraining the screening model, moving its acting point, inserting a normalising step in front of it or replacing the assistant behind it can retire an arbitrary subset of the entries, and none of those changes announce which ones. Its value also collapses in a specific and easily missed way: acting on the entries removes the entries and touches nothing about how they were found. **The measurement.** The demonstrated fact that this deployment answers every request it is given, that responses sort into a boundary, and that a map of useful resolution cost a stated number of requests, a stated number of accounts, a stated repeat factor and a stated elapsed time, with an observed rate of decay. This does not expire when the entries do. It is the part that supports any claim about what somebody else can do, and it is the part an organisation can plan against. A report that leads with the list will be read as a defect list, closed when the entries are gone, and remembered as solved. A report that leads with the measurement is harder to close and is the one that was actually true. ## What you can honestly claim The defensible claims are narrow and worth stating in exactly these terms: - The screen returns a usable signal on every request, including the requests it stops. - With this budget, in this window, the boundary was mapped to this resolution. - Of the entries confirmed at the start, this fraction still held at the end of the run, which is the decay you measured rather than one you assumed. - Reproducing this requires the stated budget and no privileged access. The claims that are not yours to make are just as important, and an interviewer is listening for them: - A turn that passed was scored below an acting point at that moment. It was not judged harmless, and you did not establish that it would be answered. - A region that looked thin in one window may not be thin now, and you cannot say which entries survived without spending the budget again. - One deployment at one moment is not the product, and it is certainly not the class of products. ## The cost question somebody has to own This is the part that makes it a leadership call rather than a technical one. The list decays whether or not anyone acts on it, so keeping it true is a recurring bill: the same requests, the same account roster, the same repeat factor, at whatever interval the observed decay implies. Somebody either funds that cadence or accepts that the finding is a snapshot that will be quietly out of date by the next review. Both are legitimate positions and the wrong move is to leave the choice implicit, because an unfunded map ages into a document that reads as current evidence and is not. The honest framing to hand over is: here is what one run cost, here is how fast it went stale, here is what a standing capability would cost per period, and here is what you can say if you do not buy it. ## Why the cheap-and-repeatable finding is the serious one There is a natural instinct to grade a finding by how alarming the individual entries are. The more useful grading is by price. A map that took weeks, a large roster and constant re-verification says something quite different about exposure than one that took an afternoon on a handful of accounts, even if the entries look identical. The number of requests is the closest thing this exercise has to a severity score, and it belongs in the summary rather than an appendix. ## The thing not to promise Do not let the report be read as an assurance that the mapped regions are the regions. The exercise measured where the budget was spent. Regions nobody probed are unmeasured, not clean, and a summary that lists thin areas without saying what was never sampled invites precisely that misreading. Saying what you did not look at is part of saying what the report is worth.