[{"data":1,"prerenderedAt":4},["ShallowReactive",2],{"article-body-same-inputs-same-number":3},"\nHere is a test any enterprise can run this afternoon. Take a question\nwhose answer feeds real decisions, what was our exposure to this\nsupplier last quarter, how many open high-severity findings do we\ncarry, and ask your newest AI tool twice. If the two answers differ,\nyou have learnt something important about where that tool may be\nallowed to sit.\n\nGenerative models are, by design, samplers. Fluency is the product and\nvariation is the mechanism, and for language tasks that is exactly\nright. But a number that can change on refresh has a specific set of\nthings it cannot do: it cannot be reconciled by two people who received\nit on different days, it cannot be reproduced for an auditor, and it\ncannot be traced when the decision it fed goes wrong. AI errors are\nalready reaching boards and external audiences often enough to be\nbenchmarked (the verification gap),\nand errors are survivable when they are reproducible. An error nobody\ncan regenerate cannot even be diagnosed.\n\n## What determinism buys, concretely\n\nSame inputs, same number, every time. Hold an intelligence system to\nthat sentence and four governance properties follow directly:\n\n- **Reconcilability.** Two viewers, or the same viewer twice, can\n  agree on what the system said, which is the precondition for\n  disagreeing productively about what to do.\n- **Auditability.** Any historical figure can be regenerated from its\n  inputs and checked. The audit trail is not a log of claims but a\n  recomputable fact.\n- **Attributability.** When the number moves, it moved because an\n  input moved, and the driver can be named. Explanation stops being a\n  story and becomes a calculation.\n- **Accountability.** A decision made on the number can be reviewed\n  against exactly the number that was shown, not a best guess at what\n  the model probably said that day.\n\n> An answer that cannot be regenerated cannot be governed. It can\n> only be believed or disbelieved, which is not a control\n> environment.\n\n## The placement line\n\nNone of this is an argument against language models. It is an argument\nabout placement, and the line is clean once stated.\n\n**Language belongs at the edges.** Turning a plain-language question\ninto a precise query. Summarising a resolved analysis for a human\nreader. Drafting the narrative around the figures. These are language\ntasks, ambiguity is native to them, and generative models do them\nbetter than anything before. The words of a summary may vary between\nruns; that is native to language too. What must not vary is the\nsubstance beneath them: the facts, figures and citations are\ndeterministic, the narration adds nothing to them, and the version\nthat was shown is kept.\n\n**The middle must be deterministic.** The retrieval that selects the\ndata, the computation that produces the figure, the aggregation that\nrespects the viewer's permissions, the model that issues a calibrated\nprobability: this pipeline runs the same way every time, and the\ngenerative layer is never allowed to *be* the number, only to talk\nabout it.\n\nThe failure mode of the current copilot wave is collapsing that line:\nletting the sampler both find the data and state the figure, then\nwondering why two executives are holding different totals. The fix is\nnot better prompting. It is architecture that routes each kind of work\nto the kind of machine that can be held accountable for it.\n\nA useful procurement question falls out of this: ask a vendor which\nparts of their answer pipeline are deterministic, and how they prove\nit. A system built on the right line answers with an architecture\ndiagram. A system built on vibes answers with a roadmap.\n\nProphesee is built on that line. Deterministic computation and\ncalibrated models produce the figures, and generative language serves\nat the edges. Same inputs, same number, every time. To see where the\nline runs on your own stack, [start here](/contact).\n",1786799035483]