Dataset Viewer

The dataset viewer is not available because its heuristics could not detect any supported data files. You can try uploading some data files, or configuring the data files location manually.

Asset-1 reduced feature matrices — "The Gauge Is the Obstacle"

Tier-2 artifacts of the Asset-1 adapter-bank study (TASUMER MAF): Bielec & Carlson, "The Gauge Is the Obstacle: Task Identity, Cross-Family Transfer, and Merge-Conflict Prediction from Adapter Weights Alone" (2026).

The bank: 480 LoRA adapters (2 families × 6 tasks × 40 seeds; Qwen2.5-1.5B-Instruct and Llama-3.2-1B-Instruct), trained under a pre-registered protocol with a completeness interlock and independent Director verification. Code, per-item verification bundle (Tier 1), and the paper source live in the public repository: https://github.com/tasumermaf/rhombic (paper under paper/).

Files

file size SHA-256
features_qwen2.5-1.5b.npz 90.6 MB 80365ac3b00c958a9cdb04ca7d783197efca29b5876fd5b59039399c4004c9ef
features_llama3.2-1b.npz 56.7 MB 49daac1c0d416ee0cb4a11bc8fef0bf96a59b32cbf05a9ff9d307a0d2b635dd5

Each .npz holds the reduced per-representation feature matrices (float32) exported by the frozen D1 feature functions (scripts/asset1_d1_identifiability.py at rhombic commit 638f4a8), so a from-scratch classifier re-run bit-matches the paper's H1/H2 results: raw-flattened, GL(r)-gauge-canonical ('full' variant, proj_dim=16, proj_seed=0), and vocab-signature representations, keyed by run per the bank manifest.

Provenance

  • Bank manifest closed 480/480 COMPLETE, 2026-07-20T16:21:19Z.
  • Verification-bundle archive SHA-256: c1891d50fc7c48c27d6fb606d667f1f1120105ed55cb7414c8e4b3d874acfadd.
  • Every headline number in the paper was independently re-derived from the per-item bundle by the study's Director; the paper's Reproducibility section documents the scope of each re-derivation.
  • The paper's post-acceptance adversarial audit (117 verified findings, all dispositioned) is in the repository at paper/audit/round-2/ADVERSARY_LADDER_LEDGER_ASSET1.md.

The adapter bank itself (480 trained runs, ~weights) is not published for size; the manifest, per-run configs, and the trainer that reproduces it are in the repository.

Downloads last month
2