Pre-registration (Study 1b, the controls): Is It the Specification? A Seven-Arm Controls Study of Specification Preambles for LLM Code Generation

Pre-registration, version 1.0, for Study 1b, the controls. The parent study (arXiv:2609.23270, dataset DOI 10.5281/zenodo.22598204) found that a 267-word specification preamble cut defect findings in LLM-generated code. It was single-turn and had no control arms. This study asks whether that effect comes from the specification’s content, or from extra text, the reasoning it invites, a generic secure-coding instruction, or the delivery format. It also asks whether the effect survives when functional correctness is scored alongside security. Seven arms: bare, frame inline, a neutral length-matched preamble, reason-first, a generic secure prompt, staged delivery, and the frame plus a security block. Six hypotheses. Five have pre-stated refuters and are tested per model. One is a non-inferiority test (H4, 10 percentage point margin on correct-and-clean, tested pooled across models with a task-block bootstrap, single verdict). H1, H2, H3 and H6 carry a base-rate reachability guard. The verdict rule is registered: SUPPORTED, REFUTED or PARTIAL per model, and SUPPORTED or NOT SUPPORTED for H4. Analysis: exact sign test per model, Holm correction within each hypothesis, 2,000-resample bootstrap, seed 20260905. Block bootstrap for pooled figures, descriptive except for H4. A joint correct-and-clean metric. Two frame-blind scanners (Bandit 1.9.4 unless superseded, Semgrep p/python offline, pinned in v1.1). Arm order randomised per task and model from the registered seed. A registered nondeterminism repeat, descriptive only. A minimum-detectable-effect table anchored on published counts. The record holds the pre-registration (PDF and DOCX), the seven verbatim arm texts (armA.txt to armG.txt), the harness note, the lexical audit script for arm C, and SHA-256 sums. All twelve files, including the empty armA.txt and SHA256SUMS.txt, are in Study1b_FREEZE_FILES_20261102.zip. The PDF is also attached separately for preview. The design is frozen on the date of this deposit and precedes any generation. The checkers, unit oracles, scoring code, scanner pins and protocols follow as a dated v1.1 of this record before any model is called. The abstract was submitted to the SANER 2027 Registered Reports Track (Stage 2 with the Empirical Software Engineering journal) on 30 October 2026 under the Stage-1 title “Is It the Specification? A Pre-Registered Seven-Arm Controls Study of Specification Preambles for LLM Code Generation”. No data had been collected at the time of deposit. Version 1.0 (dated 2 November 2026, deposited 4 November 2026). Author: Sandeep Dhuri, Independent Researcher, New Jersey, USA, ORCID 0009-0008-1829-2431.