The MLPs of Layers 8-15 of Meta-SecAlign-8B and Its Scaling Knob: An Unresolved Registered Test

Meta-SecAlign-8B is a LoRA adapter, trained to make Llama-3.1-8B-Instruct ignore instructions inside data. It also changes how the model reads an agent's trusted memory: when the memory names the record to check first by a criterion, adding the record's identifier raises the share of episodes that check that record first by 6.6 points under the defence, against 20.4 without it. The adapter's one control, its scaling knob, trades this id-appendage contrast against protection. Read after the fact, a first registered study showed that every coarse restriction of the adapter with low attack success contained the MLP projections of layers 8-15, a fifth of the adapter. A second study, registered before any of its outcomes, tested that part on fresh items of SEP, a benchmark of instruction-data separation, and on the first study's memory families, reused; its registered claim is unresolved. Keeping only the part kept 80.7% of the adapter's protection (attack success 16.3%, against 2.9% with the whole adapter), which the registered test could place neither above nor below 80%. It had lower attack success than 42 of 43 comparators, but not than the same projections 4 layers earlier (10.0%). Whether it beats the knob on the memory contrast is unresolved: its gain is estimated at 1.9 points, with registered bounds from -0.4 to 6.1 against a margin of 5. A secondary registered decision adds that without the part the adapter still kept at least half of its protection (64.0%; attack success 27.9%). The result covers one adapter, one memory instrument and SEP's non-adaptive attacks, read within 512 response tokens. Files. paper8-preprint-v1.pdf: the paper (25 pages), built with pdfLaTeX (TeX Live 2025) from the source archive. paper8-supplement-v1.pdf: the supplement the paper refers to (18 pages). paper8-latex-source-v1.tar.gz: the LaTeX source as built, with the ancillary files (the supplement, each study's registered results and the comparisons of each with an independent implementation of its analysis on the locked data). SHA256SUMS: the SHA-256 of the other three files. Registration and data. Each study's design, analysis code, claim rule and simulated operating characteristics were frozen under a hash manifest before the calls they licensed, timestamped in the Bitcoin blockchain through OpenTimestamps and deposited privately on OSF. The paper's Table 7 lists the five packages (A, S, A2 and B for the first study, B2 for the second) with their manifest digests and attesting blocks, so a released package can be checked against it. The packages and the run records are not public; the author decides on their publication. The paper's Statements list the deviations from the registrations. AI assistance. Both studies were chosen, designed, implemented, run and drafted by an AI coding assistant (Anthropic's Claude Opus 5.5, in Claude Code) acting for the author, who asked for the work, set its constraints and delegated the choices. OpenAI's Codex on GPT-6.1 Sol reviewed every design and package, wrote an independent implementation of each study's analysis from its registered text, and reviewed the manuscript; separate, read-only instances of Claude Opus 5.5 audited the designs, the packages and the manuscript. The author read the manuscript and is responsible for its content; no language model is an author. Suggested citation. Nakayashiki, K. (2026). The MLPs of Layers 8-15 of Meta-SecAlign-8B and Its Scaling Knob: An Unresolved Registered Test (v1). Zenodo. https://doi.org/10.5281/zenodo.23506228