Data and code for: No Free Universality — A Pre-Registered Null for Asset-Invariant Microstructure Representations in an Off-the-Shelf LLM
收藏资源简介：
<p>Replication materials for the paper <em>No Free Universality: A Pre-Registered Null for Asset-Invariant Microstructure Representations in an Off-the-Shelf LLM</em> (Chan, 2026).</p><p>A pre-registered test of the Platonic Representation Hypothesis on cryptocurrency limit-order-book microstructure. Five spot pairs (BTC, ETH, SOL, HYPE, DODO) on one venue are sampled contemporaneously at 5 Hz (21,950 matched windows); each window is rendered into an identical, asset-blind nine-feature line and read by a frozen Mistral-7B-Instruct-v0.3 encoder. Under criteria locked before any representation was read, cross-asset representational alignment shows no excess over the input-feature floor (0/10 pairs) and the depth signature is inverted.</p><p><strong>Contents.</strong></p><ul><li><code>PHASE1A_PREREG.md</code> — measurement pre-registration (verdict criteria, locked 2026-07-02).</li><li><code>PHASE1_D3_PREREG_v0_2.md</code> — pre-registered discriminating follow-up (committed 2026-07-03).</li><li><code>phase1a_analysis.py</code> (v2), <code>fulln_spot.py</code> (v2.1) — analysis code.</li><li><code>phase1a_curves.npz</code>, <code>phase1a_results.json</code>, <code>fulln.json</code> — per-layer curves and result artifacts.</li><li><code>phase1a.log</code>, <code>phase1a_v2.log</code> — run logs (the first includes the voided v1 run retained as the numerical-protocol case study).</li><li><code>corpus/</code> — the 32-key feature-snapshot corpus (JSONL, derived from public exchange market-data streams) with its SHA-256 manifest.</li><li><code>reps_mistral7b_v03_peakL12.h5</code> — peak-layer (L12) representation slice, both pooled positions, all five assets (~1.1 GB, gzip fp16).</li><li><code>extract_reps.py</code>, <code>make_peak_slice.py</code> — representation extraction and slice scripts; the full 56 GiB representation file is regenerable from the public model at the pinned revision.</li><li><code>SHA256SUMS.txt</code> — checksums for every released file.</li></ul><p>No personally identifiable information is present; all market data derive from public order-book and trade streams. Corpus run_id 20260628T173943Z; encoder Mistral-7B-Instruct-v0.3 at revision c170c708c41dac9275d15a8fff4eca08d52bab71.</p>
