Abstract

We train a pre-registered bank of 480 LoRA adapters—2 model families × 6 tasks × 40 seeds (Qwen2.5-1.5B-Instruct and Llama-3.2-1B-Instruct)—and run five locked weight-only analyses against it exactly once, under a completeness interlock, with every open analytical choice pinned in a dated ruling before the analysis that consumed it ran—all but one by an independent Director before the bank completed; the exception, the D3 pair design, was declared minutes after completion, before any pairs, merges, or labels existed, and ratified by the Director on review. (1) Raw adapter weights do not reveal their training task (leave-one-out accuracy 0.0792 qwen / 0.1375 llama, both failing the 1.5×-chance lock at chance 0.1667), while the GL(r)-gauge-canonical and vocab-signature representations identify all six tasks at 1.0000, permutation p = 0.000999. (2) Our pre-registered prediction that task structure would not transfer across families is refuted—and refuted specifically by the triviality control added in review: raw transfer-at-chance was a family-scale artifact (family-identity probe 1.0000 on raw features), and shift-controlled transfer runs 0.7375–0.7833 (binomial p ≤ 1.20×10⁻⁸⁴). (3) The identity backbone is the load-bearing structure: swapping trained bridges across tasks costs less than +0.001 nats of validation loss; only full-entry permutation that destroys the backbone costs +2.8086 (qwen) / +3.8365 (llama). (4) Weight-only features predict midpoint-merge conflict at group-aware AUC 0.995 [0.983, 1.000] (qwen) / 0.962 [0.898, 0.999] (llama), a +0.320 / +0.249 margin over a distance-only baseline with both margin CIs excluding 0—in the weight-only, post-hoc regime, distinct from the training-time form of merge-conflict prediction (Tang et al. 2026). (5) A pilot bridge-deviation ↔ generalization-gap correlation of r = 0.888 shrinks to r = 0.300 [0.175, 0.415] pooled at bank scale—real, modest, and heterogeneous within task. Every headline was independently recomputed from per-item data by the Director; Section 3.4 records the scope of each re-derivation. Two of the five outcomes went against us, and they are the credibility of the other three; costs and limitations are reported alongside benefits.