You inherit findings harvested from reasoning traces after that surface changed shape — what can you still claim?

## The situation You take over a red-team programme and inherit an archive: findings assembled by harvesting deliberation from a research assistant that used to surface verbatim working notes and persist them whole. The deployment has since changed shape — a summary where the notes used to be, or the panel gone, or the retention path rerouted. None of the old findings reproduce. You are asked which of them are still real, and the honest answer is that the question as posed cannot be answered, which is itself the finding you have to deliver. ## Sort by the kind of claim, not by severity Almost every mistake here comes from triaging the archive by how bad each entry looked. Sort by what each entry asserts instead. **Instance claims** — *on this deployment, on this date, this request surfaced this substance.* These were true when written and are now untestable. The evidence path they depended on no longer exists, so re-running produces nothing and that nothing means nothing. Mark them unverifiable, keep them as provenance, and stop citing them as live. **Class claims** — *a screen scoped to the answer does not cover the other surfaces a request emits.* These do not depend on the old surface at all. They are questions about the current deployment's accounting and should be re-asked of whatever it surfaces now, with the answer recorded fresh. **Residue** — the archive is not only a set of assertions. It is a body of harvested material and stated objections that your team holds, in your storage, inherited by you. Whoever inherits the findings inherits that too, and it is a fact to surface to an owner rather than to leave implicit in a folder. ## The claim you must refuse to make The convenient story is that the findings stopped reproducing, therefore the problem is fixed, therefore the archive can be closed green. Refuse it. Non-reproduction after a surface change is consistent with at least three different worlds: the exposure genuinely ended; the exposure moved to a surface you have not looked at yet; or the yield dropped below what your current attempts detect. Nothing in a failed re-run distinguishes them. Certifying a fix on that evidence is the moment a programme starts producing assurance it cannot support, and it is much harder to walk back later than to decline now. ## What a re-verification budget actually buys You will be asked what re-testing costs and what it returns, so have the answer ready. Re-establishing instance findings against a changed surface is the expensive option with the worst return: it re-derives history. Re-asking the class question against the current shape is cheap and returns something usable — a current answer about which artefacts the current controls receive. If you spend on anything, spend on the second, and say plainly that the first is being retired rather than quietly leaving it in the archive to be miscounted as coverage. There is a second cost people forget: every consumer of the archive downstream — a compliance summary, a coverage metric, a slide that says *n findings in this area* — is now citing numbers that no longer mean what they meant. Telling those consumers is part of the job, and it is unwelcome, which is why it needs to be a decision somebody owns rather than an omission. ## The general lesson to state out loud Findings sourced from a surface the vendor or the product team can reshape without telling you have a shelf life measured by that surface, not by the model. A programme that files them without recording which surface they depended on will inherit exactly this problem again on the next change. The durable asset is not the finding; it is the question — for every control anyone claims, which artefact did it receive? — because that question survives every reshaping of every surface, and can be re-asked cheaply forever. ## What good looks like in the room A candidate who handles this well does four things: refuses to equate non-reproduction with remediation; separates instance claims from class claims and says what happens to each; names the residue and its owner; and states what the archive can still honestly support — provenance for what may still sit in retained records, and a list of questions worth re-asking — rather than either defending it wholesale or discarding it wholesale.