La IA generativa y la frontera de la firma en el comercio de servicios de EE. UU.: prueba ciega de cota de P1 + A-USD con importaciones del BEA por afiliación

Abstract

We pre-register a single, blind bound test of the joint hypothesis P1 + A-USD. P1, a proposition of a verifiability model of offshoring, says that after generative AI (ChatGPT, 30 Nov 2022), in AI-exposed service categories, the share of tasks a rich buyer country buys from unaffiliated foreign parties falls relative to the share bought within corporate groups. BEA data are in dollars, and the model's own cost table implies a price channel of the opposite sign: for tasks that stay in their channel, AI lowers the unit value of captive (per-hour governance) supply proportionally more than that of third-party (per-deliverable governance) supply, which raises the unaffiliated/affiliated dollar ratio if demand is inelastic. A-USD is the auxiliary assumption that spending per task that stays in its channel changes in the same proportion in both channels (e.g., unit elasticity). Under P1 + A-USD the dollar ratio falls; the pure task-content rival predicts no change. P1 holds in the model's two polar regimes; in the intermediate regime the ordering is not guaranteed. Data: U.S. imports of services by type and affiliation, BEA International Services Table 2.3 (world), vintage of 7 July 2026, fixed by SHA256; pre-period 2006-2022, base 2022, post 2023-2025. The confirmatory statistic (R1) is θ_E = mean(y_2023..2025) − y_2022 with y = ln(unaffiliated) − ln(affiliated) for the exposed basket {computer services 9.2, information 9.3, professional and management consulting 10.2}, chosen by a 2-of-3 rule over three external exposure classifications published in 2023 and 2025 that do not use BEA trade data; the cutoffs, the hand-made bridges and the application to BEA types were set by AI agents after viewing the classifications and pre-period BEA levels (never post-period values); a strict 3-of-3 basket is reported. Nulls are fitted on each series' own 2007-2022 annual changes (random walk with and without drift, analytic and bootstrap; the maximum p-value is used). R1: θ_E ≤ −0.1707 (−15.7%). A fall counts as support only if it also appears in a fixed-weight (2022) index of within-type ratios and is larger in sign than in low-exposure services. A robust verdict also requires significance (10%) against the low-exposure basket, robustness without software licences and cloud, without computer services, and without intra-group management charges (10.2.1.3), significance (10%) with the U.S.-parent denominator, and post-period measurement diagnostics, because BEA replaced cell suppression with undocumented coarsening exactly for 2023-2025. A bounded null requires the one-sided 95% lower bound (the end of a two-sided 90% interval) to exceed −23% for the exposed basket and, net of the frozen census-to-sample jump, against the low-exposure basket; 23% is a reference size taken from an earlier test, not a model prediction. Seven levels with pre-written readings, including what each means for the thesis, are fixed in code. Power, from pre-period data only, assuming the measurement checks pass (so the figures for robust support and bounded null are upper bounds): the 80% minimum detectable fall is 22%-35% for R1, 26%-39% for any support and 30%-44% for robust support across seven noise scenarios; the plausible effect is 2%-10%. With no effect, the most likely outcome is inconclusive (37%-80%) or bounded null (16%-62%); false support ≤ 3.2% (3.6%-8.1% if post noise were 1.5 times the pre noise), false contrary ≤ 0.8%, false bounded null under a true 23% fall ≤ 2.0%. If BEA rounded cells to 100 million or more, a bounded null becomes nearly unreachable (≤ 12%) and the test is in practice "support or inconclusive". A common artefact of 18%-26% in all services passes as (mostly non-robust) support 12%-32% of the time. The test is a complementary bound, not a test designed to confirm P1 at plausible sizes. The single real run happens only after this registration, behind a gate file written by the author, with a lock file; its integrity rests on the author comparing the manifest hash with this registration. Manifest SHA256 (MANIFIESTO_BEA.txt): 528663eb6ba9e88c7d816755ad4147ead00040ac7dded895b4db774907930461. Single blind run after registration, behind an author-written gate file. Integrity of the run rests on comparing this hash with the local manifest before running.