Determinants, Stakes, Risks and Multidirectional Correctives of US–China–EU Competition in Critical Technologies Including Frontier AI, and Their Transcendence: Evidence from a Preregistered Serial-Refutation Programme

European policy towards frontier artificial intelligence rests on a diagnosis: that Europe under-invests, over-regulates, examines patents too slowly, and lets a few large incumbents capture what funding there is. We subjected it to a preregistered serial-refutation programme: five campaigns, 31 hypotheses, thresholds sealed by hash before execution, negative controls throughout. Of the 31, 16 were refuted, 4 fell below criterion, 1 proved not testable and 10 survived. Two central quantities are composition effects. Filing route accounts for 97.4% of the European patent-grant delay relative to China on the strata where comparison is possible, and at identical route the European office is not slower (-0.0233 years, 95% CI [-0.2183; 0.1716]). The concentration comparison behind the capture claim sets a total over a thin slice of projects against a total over all of them; drawn alike, European AI funding is less concentrated than the rest. Effort and institutional quality fall below their preregistered amplitude thresholds. Europe holds 5.1% of installed AI compute against 79.6% for the United States; at the frontier its coefficient is -0.928 dex (decimal orders of magnitude), a shortfall of a factor of 8.5, 95% interval [-1.830; -0.026], with a design power of 0.288 at the sealed threshold; the US-China frontier gap is closing at 0.526 dex a year on a wide interval. Rank reduction, the collapse of the European trajectory onto fewer directions, is refuted at the preregistered magnitude. Across 114 320 funded participations, an organisation with no prior participation receives 0.7592 of what an incumbent does with the same observable characteristics; a placebo that assigns newcomer status under a false vintage returns 1.0133, no effect. On the European Innovation Council's flagship open call, applicant cost is of the order of the amount awarded; of EUR 200 billion announced under InvestAI, 0.52% is committed. Five analyses lie outside the seals and are marked as such, among them the withdrawal of a result of our own whose estimator treats refusals as independent censoring. The binding constraints are scale at the frontier and the cost of entry; of the correctives in force, only the compute pillar acts on them. Bilingual deposit. The record contains the preprint in English (QTRIAD_preprint_v2_EN.pdf) and in French (QTRIAD_preprint_v2_FR.pdf); the French version is a full rendering of the same text, figures and tables, not a summary, and the sealed interdicts appear in it in their sealed French wording. / Dépôt bilingue : la notice contient la prépublication en anglais et en français ; la version française est une restitution intégrale du même texte, des mêmes figures et tableaux. Replication. The deposit contains a replication archive of 180 files: the five sealed preregistrations and their three addenda with their SHA-256 digests, every execution trace, the campaign and preprint code, the figure-generating code in both languages, the LaTeX sources of both versions, and the numbers file from which every figure is drawn. The archive carries a per-file SHA-256 manifest; the digests of the archive and of the manifest are printed in Appendix D of the preprint, not inside the archive, so that the archive authenticates against a published value rather than against one it carries. The SHA-256 digest of QTRIAD_replication_20260915.zip is 196f87d3af43eadfa03686e72c12ec46011f441caf513e57df0e364980ed1d27. Access. Files are restricted; the metadata, including this abstract and the archive digest above, is public. Access is granted to researchers, auditors and reviewers who wish to verify the preregistrations, the execution traces or the computations. State your name, affiliation and the purpose of the request. Notes. Primary arXiv category: cs.CY. The two language versions are generated from one set of macros and one bibliography, so every quantity printed in either version comes from the same execution traces. One previously published result of the programme is withdrawn in Section 6, and the identifiable-sum figure of Section 5.4 was corrected at the closing audit of this preprint (recorded in trace R30, block D).