The oracle keeps no register¶

The communications office keeps six months of board decks. Three of them contain the sentence “the Home’s data shows that”, followed each time by something the Home’s data had not shown until the reader was asked, and one of the three has been pasted into a funding application under the heading “the transformative potential of intelligent automation”. Wrong answers were always available; what changes is their career path. An answer given with confidence gets repeated, quoted in a deck, pasted into a briefing, and repeated, and somewhere in the circulation it stops being an answer and becomes a fact. The Home begins to accumulate folklore of a particular provenance: statements that are true because the assistant has said them often enough. Remediation looks like a documentation problem and is actually a memory problem, because correcting the record requires knowing which beliefs entered through evidence and which entered through the oracle, and a ledger of that distinction rarely exists. The estate records what was written. Nothing is built to record what was answered.

The chain of justification changes shape underneath this. The traditional chain ran document, interpretation, decision, and a person stood at the interpretation step who could be asked why. The new chain runs documents, retrieval, synthesis, answer, decision, and the synthesis is an intermediate authority that belongs to nobody. Even with every source preserved, a future investigation stalls at a new place: asked why this was believed, the honest reply becomes that the system presented it as the answer, which is a statement about trust in a process rather than a judgement about evidence, and it leaves nobody to cross-examine. Reconstruction carries a temporal problem on top of the attributional one. The answer was produced by a particular model version, under a particular prompt, against a particular index, with particular permissions, over an estate that has since changed. Yesterday’s answer is not merely unlogged; once the ground it stood on has moved, it is unreproducible. Chains of custody were designed for files. This one would have to hold language, and the language’s context expires.

The Home has a sharper case of this than most, and it is the one nobody raises. The Home keeps no register of what its field workers’ testimony amounts to, and the absence is deliberate: nothing tasked, nothing signed, and a denial that is true at every level ever asked to approve it. An oracle keeps no register either, but not on purpose. Ask it what the Home knows about the movements of a particular Guild in a particular hill town and it will synthesise an answer that was never a document, attribute it to nobody, and hand it over as a thing the Home said. The two absences look alike from outside and are not: one is a policy and the other is a product. Only one of them survives the first time somebody asks the Home to explain what it told the reader to say.

The culture drifts with the chain. Producing a convincing answer used to require finding evidence, and the effort disciplined the confidence. When a convincing answer requires only a question, fluency becomes cheaper than verification, and under time pressure organisations reward what is cheap. The tell is small: meetings in which “where is the evidence” gives way to “how useful is the answer”. The consultant heard it twice in one infrastructure committee.

The oracle also changes voice. Estates changed slowly; the people who read them changed more slowly still, and nothing in the Home’s records management assumed otherwise. A reader is a moving component in a system that never had one. A model update can change summarisation habits, retrieval ranking, the interpretation of ambiguous requests, refusal behaviour and extraction accuracy overnight, unannounced and unversioned from the estate’s point of view, so the same question asked across a quarter is answered by a succession of slightly different oracles, none of which signed their work. What follows is familiar in software and unnamed in knowledge work: regression testing, release management and change control, applied to how the Home answers questions. Governance inherits that pace whole. The old pattern let the DPO catch up with each new document type at leisure; every new connector, repository, capability and update now opens a surface at integration speed, thirty chapters at once in phase two, and the governance function graduates from designing the archive to chasing the archive’s offspring.