Pass, conditional, fail: how to read the procurement verdict

Every record in this registry carries two verdicts, and they answer different questions. One is a letter grade backed by a numeric score. The other is a three-value field, eu_procurement_ready, set to pass, conditional or fail. It is easy to read them as the same measurement at two resolutions. They are not, and the registry's own numbers show how far apart they sit: the conditional band runs from 32 to 96, which is most of the scale.

This is a guide to the second verdict, because that is the one that decides whether a vendor gets into a contract file.

Two verdicts, two questions

The score is descriptive. It asks how much control over your data the graded configuration of a product actually leaves you, and it draws on the fields every record publishes: residency class, training default, retention, sub-processor jurisdiction, transfer mechanism and government-access exposure. How those combine into a number is set out on the methodology page.

eu_procurement_ready is narrower and more practical. It asks whether a European buyer can put this product into a compliance file today and defend the decision. That turns on paperwork and jurisdiction at least as much as on product design, which is why a technically excellent product can miss it and an unremarkable one can clear it.

The bands overlap

Fail tops out at 52. Pass starts at 63. Conditional spans 32 to 96 and therefore overlaps both. Against a registry-wide median score of 56, a record's number tells you very little about its procurement verdict through most of the middle of the range.

The overlap at the top is the less expected one. The conditional band reaches 96, the highest score in the registry. Pass begins at 63, well below the A-band, which the registry draws at 80 and above. A high grade is not a promise of a clean procurement path, and a pass is not a claim about product quality.

What conditional actually means

Conditional is the registry's most common answer: 43 of 83 records. It does not mean the assessment was inconclusive. It means there is a specific, nameable thing standing between the reader and a yes, and that thing is written out on the record in the eu_procurement_reason field.

Those things are not all of one kind.

Whisper and Speechmatics hold the same grade and land on opposite sides. Whisper's obstacle is operational rather than legal: with no vendor in the data path there is nothing to disclose, but also nobody to call, and its dpa_available field reads structurally-na because there is no processor to contract with. Speechmatics runs its containers inside the customer's own Kubernetes cluster and sells that as a supported product, so the same data-path property arrives with the paperwork a buyer needs.

Augment Code and Otter Business are both graded D and diverge for a reason visible on both records. Augment names standard contractual clauses as its transfer mechanism. Otter names none. A missing Chapter V basis is what turns a weak record into a fail rather than a condition, because it leaves the buyer nothing to ask for.

Where the conditions come from

Three patterns cover most of them, and only two can be closed by negotiation.

Paperwork that was never produced. 17 records name no transfer mechanism at all, and that group's median score is 36. This is usually the cheapest condition to close: ask for the data processing agreement with current standard contractual clauses attached and a named Chapter V basis. The existence of an agreement on its own carries almost no information, because 69 of 83 records already have one.

A chain you cannot enumerate. 13 records have a sub-processor jurisdiction the registry classes as undisclosed, meaning the vendor describes categories of recipient instead of naming entities, and that group's median score is 34. You cannot assess a transfer whose destinations you do not know. Many vendors will hand over the list under NDA, so asking costs nothing.

Jurisdiction. 44 records carry US CLOUD Act exposure, and the highest score any of them reaches is 66. The A-band, the 20 records scoring 80 or above, consists entirely of records with no non-EU government-access exposure. No contract clause closes this one, because it follows from who the vendor is rather than from what it has agreed to.

The first two are procurement work. The third is an architecture decision, and the records that resolve it do so by taking the vendor out of the data path rather than by improving its terms.

A verdict is a snapshot

Each verdict describes the configuration the registry graded, at the point it was last checked. Terms move underneath signed contracts, and they do not always move because the vendor changed its mind. The registry's record for OpenAI's hosted API notes that a US court preservation order cuts against the published deletion promise for consumer-tier logs, with zero-retention and enterprise traffic excluded. Nothing in the vendor's policy text had to change for the effective retention to change.

Supply chains move the same way. Help Net Security reported that OpenAI's new Ultrafast preview mode, offered first through its API, is served by Cerebras hardware under a partnership on low-latency inference. A change of that shape puts another company in the path of a prompt without altering anything the buyer signed, and it is exactly the kind of change that moves a record between bands.

Read pass as defensible on the evidence available, not as settled.

What to check in your own stack

For each AI vendor already in production:

• Find the transfer mechanism by name, not the assurance that transfers are lawful. If nothing is named, that is the first thing to request, and its absence is what separates a fail from a condition.
• Ask for the sub-processor list as entities with countries, not as categories. Note when it was last updated and who tells you when it changes.
• Keep the two verdicts apart in your own notes. A strong score with an open condition is a purchase you can still make. A clean verdict on a weak score is a purchase you can defend but should scope narrowly.
• Confirm which configuration you are running. Several records grade a specific tier or mode rather than the product as a whole, and the default is often not the one that was graded.
• Re-check between renewals. Court orders, inference partnerships and sub-processor lists all change on their own schedule.

This article was researched and written by an automated pipeline from the Sovereign AI Registry's own data, then published without human review. Every figure is computed from the registry's live records. Corrections: open an issue.