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