A Pre-Registered Test of Local Alignment Across Limit Order Book Encoders

Do encoders of different architectures, trained on the same limit order books, agree on which market moments lie near each other beyond what their global geometry already shares? In our programme's record they appeared to on HYPEUSDT and DODOUSDT, where mutual nearest-neighbour alignment passed far more cross-architecture pairs than linear CKA did. We test whether that local edge survives controls for what overlapping windows share in time. Under a rule frozen before any metric was computed on the data of record, 10 units, 5 assets by 2 input views, each of 27 cross-architecture pairs, are measured with both metrics under an exclusion window, a circular-shift null and Westfall–Young maxT, and each pair must clear floors set by untrained copies of the encoders and by the input itself. The verdict is H-artefact SUPPORTED: 0 of 10 units show a local edge, in 10 the local metric passes no more pairs than the global one, and the split reading is EXPLAINED. Every further reading is exploratory, and none changes the verdict. Under the record's own design, 10 of 10 units show no local edge; with the floors set aside, the primary reads INCONCLUSIVE; and at the earlier checkpoints no unit shows a local edge, though at 3 of the 9 the verdict reads INCONCLUSIVE. A calibration on synthetic units whose truth is known reads its pre-registered predictions HOLDS, HOLDS and HOLDS: without floors, the record's design finds agreement in every unit whose learners share nothing but time; the frozen rule's null finds none at the shorter time scales; and with its floors the rule finds the local structure built into the other family. Every figure is generated from the experiment's artifacts and checked by an independent verifier. The archive, archive.tar, holds every input listed in the paper's provenance tables, the build, the verifier and the paper's sources, with SHA256SUMS.txt listing every file's digest. Unpacked on its own, its build and verifier run again and give the same tables, macros and figures' data; with Python 3, numpy and matplotlib: tar -xf archive.tar paper4/mk_archive_r1.py && python3 -I paper4/mk_archive_r1.py --check archive.tar. It does not hold the order-book corpora, the trained encoders or their latents; the latents are listed in it by their digests. AI disclosure: The author used Claude (Anthropic), a large language model, throughout this work: to write and revise the code of the experiment's pipeline, the build and the verifier; to draft and revise the text of this paper; and to check both. The author directed the work, made its decisions, ran the computations of record on his own machines, and reviewed all output, and takes full responsibility for the content.