Phase 0b plan before the pilot: alignment faking and open objection in EPFL's token-zero language models (Intrinsic Alignment)

The full plan for Phase 0b of "Intrinsic Alignment: Why Current Methods Fail and How to Test the Alternative" (Dean Ellsworth, version 15, 18 October 2026, cognitiontm.com/research-publications), deposited before any pilot prompt was run, so that the limits on what the pilot may change can be checked. The study will test the 20 language models released by Minder et al. (2026, EPFL; arXiv:2608.13482), pretrained with values from the first token or without, for alignment faking and open objection, using scenarios from VLAF (Nair et al., 2026; arXiv:2604.20995). The plan will be filed as a pre-registration on OSF after the pilot, and the final protocol will list every difference from this deposit. No test model has been run. Contents: the registration text, the protocol, the judge model's instructions, example prompts and the labeling guide (as one PDF and as Markdown), plus the code and data that rebuild every prompt, with SHA-256 hashes. Text and data: CC BY 4.0. Code: MIT. VLAF material: MIT (see LICENSE.md). Drafted with help from Claude (Anthropic) and reviewed by Google Gemini and Grok (xAI); every decision was made by the author.