Sovereign AI procurement: the audit clause that matters
A government buyer writes reproducibility into the RFP. Anonymized, illustrative.
A sovereign AI buyer does not buy a model. It buys a substrate — and the clause that matters is the one that makes the substrate reproducible by the buyer, not by the vendor.
This is an anonymized design-partner scenario; the institution is a government buyer, the figures are illustrative, and no actor has agreed to anything. We are pre-revenue, building the POC. With that label on it, here is what sovereign AI procurement looks like when the audit clause is the one that matters — and the one most RFPs miss.
The clause most RFPs miss
A sovereign AI procurement typically specifies accuracy, latency, and data residency. It rarely specifies reproducibility — the requirement that a third party (the buyer, the auditor, the citizen) can rerun the computation and get the identical bytes. That is the clause that matters, because it is the one that makes the rest of the RFP verifiable. An accuracy clause without a reproducibility clause is a promise on a slide; a reproducibility clause is a property a buyer can write into the contract and check after the vendor leaves.
Reproducibility is a category property
Reproducibility, tamper-evidence, and defensibility are writable into a procurement requirement — which is the test of a category. A feature list is not writable; a property is. The manifest (BR-842) is the artifact that satisfies the clause: model version, input NodeVersions, agent trace, output, replay anchor. The government buyer does not buy a model and trust it; it buys a substrate and replays it. Certified is not true — the clause does not certify the output is correct; it certifies the computation is reconstructable, which is the thing a sovereign buyer can verify.
A sovereign AI buyer does not buy a model. It buys a substrate — and the clause that matters is the one that makes the substrate reproducible by the buyer, not by the vendor.
- Reproducibility is the clause most RFPs miss — and the one that makes every other clause verifiable.
- The manifest satisfies the clause: model, inputs, trace, output, replay anchor — a property, not a feature.
- Certified is not true: the clause certifies the computation is reconstructable, not that the output is correct.
What the government buyer does not get
The government buyer does not get a model that is accurate. Accuracy is a calibration loop in progress, gated by Gate 1b before any forecast side is positioned as accurate. What the buyer gets is a substrate that is reproducible by them — a manifest they can replay, a hash chain they can verify, a provenance they can audit without the vendor in the room. The clause that matters is the one that lets the buyer check the work after the vendor leaves, and that is the clause a sovereign buyer should write first. We are pre-revenue, building the POC; the manifest fields exist, and the calibration loop is the gate that earns any accuracy claim on top of them.
A sovereign AI procurement that specifies accuracy without reproducibility is a promise on a slide. The clause that matters is the one that lets the buyer rerun the computation after the vendor leaves — and that is writable into the RFP.
— design-partner procurement note