Termination Behavior Across Fixed-Prefix Interface Conditions in a Single 4B QLoRA-SFT Lineage: Termination Rates, Early Stop Mass, and Endpoint Saturation Under Prompt-Loss-Weight Manipulation

Abstract

Corrected technical preprint, version 1.1. English is the primary version; a faithful Japanese companion version and both supplements (EN/JA) are included. This version supersedes v1.0 (DOI 10.5281/zenodo.21967093), whose files were restricted after a premature release. The correction reasons are listed in the v1.0 record and in PAPERB_CORRECTION_NOTICE_V1_1.md inside this package. No files of v1.0 were silently replaced or deleted. We report three linked measurements in one 4B QLoRA-SFT lineage, separating termination rate, position-wise stop mass, and the saturation of a binary endpoint under prompt-loss-weight (PLW) manipulation. C1. Termination rates separated completely across two fixed-prefix conditions: RAW terminated in 0/24 sequences, THINK_EMPTY in 24/24. These are descriptive pooled counts, not 24 independent topics: the same eight topics were repeated at each of three PLW levels, and the contrast was 0/8 versus 8/8 within every level. C2. The simple account that the prefix causes earlier stopping by increasing stop-token mass at the beginning of generation was not supported within the measured window. Across the first 32 generated positions, the paired difference in joint stop mass (THINK_EMPTY − RAW) was negative for all eight topics (median −6.702249 nat; 0/8 positive pairs). C3. The preregistered binary endpoint was saturated at both boundaries across PLW 0.0 / 0.1 / 1.0 (RAW 0/8, THINK_EMPTY 8/8) and could not discriminate PLW-level differences. Claim ceiling. This paper does not claim a causal mechanism, equivalence among PLW conditions, absence of PLW effects, or generalization beyond the tested lineage, seed, and topics. The 32-position stop-mass window did not reach the observed stopping region (median 195.5, minimum 139 tokens). The frozen raw artifact contains the label PLW_NOT_A_CONFOUND; that label exceeds the evidence and is explicitly rejected in the paper and in a machine-readable correction sidecar. Reproduction scope. All reported aggregate values can be recomputed from the frozen raw records and code included in the package. The three QLoRA adapters are not included, so independent rerunning of generation is not supported and is not claimed. Corrections from v1.0 (full list in PAPERB_CORRECTION_NOTICE_V1_1.md): tested scope stated in title and abstract; pooled counts identified as repeated observations; author name in the Bai et al. citation corrected to He Bai; causal / equivalence-implying / no-effect wording removed; stopping-region and sequence-length values corrected; publication status of two 2024 references updated; verified checksums and status text.