[{"data":1,"prerenderedAt":4},["ShallowReactive",2],{"article-body-the-model-passport":3},"\nA number lands in front of a decision maker: this customer's churn risk\nis 82%, this shipment is high risk, this control is likely to fail. If\nthe organisation is about to act on it, a short list of questions ought\nto be answerable in about a minute. Which model produced this? Which\nversion? Trained on what data, frozen when? How has it performed since,\nand where did it miss? Who owns it, and who approved it for exactly\nthis use?\n\nIn most enterprises, answering that list is a forensic project. Models\nare retrained in place, training data is a moving query rather than a\nfrozen artefact, ownership is a team alias, and approval is a meeting\nthat left no structured trace. The number arrived from nowhere in\nparticular, and decisions are being made on it.\n\n## What the passport contains\n\nThe remedy is a document the model cannot travel without. Every model\nthat reaches production in Prophesee carries a passport with six\nentries, and the record is append-only: corrections add pages, nothing\nis ever rewritten.\n\n- **Immutable version.** The exact artefact making predictions, not\n  the family it belongs to. Retraining creates a new version with a\n  new passport page, never a silent replacement.\n- **A named owner.** A person, not a team alias. Accountability that\n  cannot be pointed at an individual diffuses into nobody.\n- **The frozen training slice.** The precise data the model learned\n  from, preserved as it was at training time, so any later question\n  about bias, leakage or drift can be answered against the actual\n  evidence rather than a reconstruction.\n- **The backtest, misses included.** Performance on data the model\n  never saw, with the failures on the record next to the successes,\n  and the placebo runs proving the result stands apart from luck.\n- **The calibration curve.** Whether the model's stated probabilities\n  matched observed frequencies, republished every cycle so drift is\n  visible when it happens.\n- **The approval chain.** Who reviewed what evidence and signed off on\n  which uses, so scope creep, a model approved for triage quietly\n  becoming a model used for pricing, is detectable by construction.\n\n> Append-only is the load-bearing property. A provenance record that\n> can be edited is a narrative, not a record.\n\n## Why the timing has changed\n\nThe passport was always good engineering: it is what lets an\norganisation debug a bad quarter of predictions, retire a drifted\nmodel without archaeology, and distinguish its working AI from its\ndecorative AI. What has changed is the outside world. The EU AI Act's\nobligations are phasing in through 2026, with penalties reaching €35\nmillion or 7% of global turnover, and its documentation requirements\nfor higher-risk uses (technical documentation, data governance,\nrecord-keeping, human oversight) are passport questions in\nregulatory language.\n\nAn organisation whose models already carry passports meets that kind\nof obligation by exporting what it has. An organisation without them\nfaces the worst version of compliance: reconstructing provenance\nafter the fact, for models nobody fully remembers, under deadline.\nThe same divide is opening with auditors and, increasingly, with\nenterprise customers, who have begun asking vendors passport\nquestions in procurement.\n\nA subtler benefit shows up inside the walls. When every model carries\nits record, arguments about whose number to trust change character.\nThe question stops being whose model has the more senior sponsor and\nbecomes which passport shows the better recent calibration, which is\na question with an answer.\n\nThe Model Passport is standard issue in Prophesee's Foresight engine.\nNo model reaches production without one, and the record ships with\nevery prediction. To inspect a passport against your own models' paperwork,\n[start here](/contact).\n",1786799035573]