| --- |
| title: Reproduction - Expressivity-Efficiency Hybrid Sequence Models |
| emoji: 📐 |
| colorFrom: blue |
| colorTo: green |
| sdk: static |
| pinned: false |
| short_description: Six claims audited, arXiv 2603.08859 |
| tags: |
| - trackio |
| - open-reproductions |
| - icml2026-repro |
| - paper-82EJxJzG6r |
| --- |
| |
| # Local ICML reproduction audit |
|
|
| This package audits the six frozen registered claims for arXiv:2603.08859v1 / OpenReview `82EJxJzG6r`. It is deliberately isolated and local: no Hugging Face Space, campaign ledger, or remote validation target is created or changed. |
|
|
| Run the complete deterministic package with: |
|
|
| ```sh |
| python3 run_all.py |
| ``` |
|
|
| That runs every evidence route twice, compares the eight result artifacts byte-for-byte, builds all pages, runs `official_validator.py` (the authoritative local-package gate, not an ICML/OpenReview/author validator), writes `SEMANTIC_V4.json` with 12 local semantic checks and `VALIDATION.json`, and finally writes `MANIFEST.sha256`. |
|
|
| The proposed (not created) target slug is `repro-hybrid-seq-82ejxjzg6r`. |
|
|
| Evidence categories are deliberately distinct: |
|
|
| - Claims 1–2: literal theorem audit and direct numerical witnesses. |
| - Claims 3–4: independent finite implementations of written constructions, plus direct execution of author notebook cells recorded separately. |
| - Claims 5–6: exact authored-source table audit only; learned training is not claimed as rerun because the supplied training path hard-codes CUDA and this host has no CUDA device. |
|
|
| Important limitations are in each claim page. In particular, Claim 1 is falsified at the literal printed scope, Claim 5’s `.999` is not silently rewritten to perfect accuracy, and Claim 6’s Figure 5 and Figure 6 tasks are not conflated. |
|
|
