Frozen Forecast: Sign-Kernel Alignment Ranks Held-Out Adapters' Output-Matched Learning (SCAFFOLD, E5)
Abstract
A forecast frozen on 2026-10-17 10:47 +0400 (SHA-256 0400cbfb03436aa54c54a3b59bb11fd9c252475cbdb34e11c91272f0cc8b1012), before any B-learning run with these adapters, for the preprint Head Start or Process? Decomposing How an Experience-Acquired Abstraction Shapes Later Learning in a Small Language Model (doi:10.5281/zenodo.23432800). This record is its public registration. In Stage 2 of that study (Qwen3-1.7B), the t = 0 sign-gradient kernel alignment F_sign of the output-matched frame ranked six experience-trained LoRA adapters' output-matched learning at Spearman 0.94 on both B tasks. That was a post hoc observation, one of four statistics tried. This record tests it on the four adapters held out since the first preregistration (P1_20260925-28). Predicted ranking (rank 1 = most output-matched learning): P1_20260925: F_sign 34.0 (B_ind, predicted rank 1); 46.6 (B_ext, predicted rank 1) P1_20260926: F_sign 25.5 (B_ind, predicted rank 4); 13.6 (B_ext, predicted rank 2) P1_20260927: F_sign 26.6 (B_ind, predicted rank 3); 9.3 (B_ext, predicted rank 3) P1_20260928: F_sign 28.2 (B_ind, predicted rank 2); 4.1 (B_ext, predicted rank 4) Test. The A_matched arm on ten new B seeds (301-310) with the Stage-2 recipe and budgets. The statistic is the mean over the two tasks of Spearman(F_sign, mean G(A_matched)) across the four adapters, against an exact permutation null. It is confirmed at ≥ 0.80 (p = 0.031), disconfirmed at ≤ 0, and inconclusive otherwise. The full text is in FSIGN_FORECAST.md. Use of AI assistance. Substantial AI assistance was used throughout this project. The question of whether an experience-acquired abstraction changes the process by which a later abstraction is formed, rather than only supplying information, is the author's. An AI system (Claude, Anthropic), working under the author's direction, audited the earlier studies, reviewed the literature, designed the study, wrote the preregistrations and code, ran the experiments and analyses, and drafted the paper. A separate instance of the same system reviewed the draft; that review led to the post hoc analyses E3 and E4. The preregistrations record that the author delegated the design decisions to the AI system. The author directed the work and is responsible for its content.