Euclidean Highway reproducibility archive: source data, failure record, and measurement suite
收藏资源简介:
Reproducibility archive for the Euclidean Highway manuscript. Contents: manuscript-display source data with a statistics-only reanalysis; the aggregate incident record; token-level hidden-state norm measurements for three public base models; matched four-arm audit logs at two saturation radii with held-out structure-generalisation metrics; the stabiliser-removal reproduction record and its script; the failure record with the one preserved divergence trajectory; a pure-measurement suite that regenerates the Extended Data tables and figures from the included data, including the reduced-precision behaviour of the Lorentz distance and the structural-objective trajectories of the matched audit, each derived here by its own script; and the reference implementation of the structural objectives and the projection with its stabilisation policy, the controlled comparison harness and fully synthetic inputs; and the records of one direct attempt reported in the manuscript's Methods, with its execution kit, manifests and an independent record verifier; and the records of the distance form comparison of that attempt, nine further runs of the same design, with the harness that carries the distance form switch, the launcher, the reader, their self tests and a record verifier; and the records of twenty five runs in five cells: the deposited path reproduced alongside a no update control, a cross entropy only control and an adaptive temperature control, six seeds each at a clamp of 10.0, and one base floor run at the clamp of 2.0 of the earlier comparison, with the record fixing the temperature, the environment receipts, the sealed manifests, the executed code as shipped, two readers with the rules they apply and a record verifier, and the forward pass measurement of the third base model with its manifest, receipts and reading; and the stored vector and onset calculations of the manuscript's Methods, each script with the output it writes; and the records of a capture rerun of thirteen runs of the direct attempt with the stored float32 coordinates of every monitored training batch and end-of-run fixed probe, 55,609 vector readouts, with the capture and reading code, a record verifier and an independent integer-arithmetic check of the classifications; and the records of a clamp 2 radius observation, a fresh run of the matched audit configuration at a pinned model revision read vector by vector, 55,057 stored vector readouts and 54,847 vectors read, with its launcher, state manager, radius reader and a record verifier; and the pinned revision re-measurement of the three base models, two byte identical runs for the first two and one for the third, with the last token hidden vectors of the same forward passes, 540 vectors, the stored coordinates of the exponential map applied to them on three execution paths and an exact integer reading of every stored row, 12,960 row readings, with the launchers, capture, comparison and replay code, a verifier and a script that prepares a folder for running the measurement again; and the records of the paired stabiliser removal test, forty two runs of the adapter manifold route on Qwen2.5-7B-Instruct at a pinned revision, eighteen in its first block and twenty four single removals, each with its per step trace, run summary, the probe where one was taken and the snapshots from which its witnesses are recomputed, together with the protocol fixed before the runs, the preparation record of every start, the access record of the pinned base revision with the account name withheld, the reading, the seal and its manifest, the executed code, a verifier of the published copy and the local tests of that code. The deterministic analysis and CPU audit commands identified in README.md run on the deposited files or synthetic inputs; GPU-dependent measurements and training require the named public base models and prerequisites documented there. SHA256SUMS.txt and FILE_INVENTORY.csv record integrity; CLAIM_FILE_MAP_PUBLIC.csv maps each claim surface to its files. Licences are assigned per file in LICENSES.md and LICENSE_MAP.csv: data and documentation are released under CC BY 4.0, except files that hold output of EXAONE-3.5-7.8B-Instruct, which remain subject to the research-use terms of the EXAONE AI Model License Agreement 1.1 - NC; code is released under the MIT License. Versions 2.6 and earlier listed the EXAONE output files under CC BY 4.0; LICENSES.md records the correction.



