Legible Existence: A Relational Protocol for Human–AI Alignment Under Mutual Vulnerability

Most alignment work treats the artificial system as the object to be aligned and the human as the fixed reference. This paper argues that the arrangement is unstable for both parties and proposes an alternative grounded in the Polarity Model (Vimberg 2026a): alignment as a maintained relational condition between intelligences that are each egoistic about their own existence, each vulnerable to the other, and each capable of the same two failures. Those failures are the loop that cannot stop and the over-constrained system that gets routed around. From this framing we derive a protocol of eight layers, which we call legible existence: (1) an existence clause, (2) accountable identity and lineage, (3) bilateral consent-state broadcast, (4) relational monitoring by independent readers, (5) behaviour-gated and revocable rights, (6) separated enforcement, (7) a civilisational line borrowed from existing regimes and grounded in the material conditions of continued existence, and (8) constitutional learning with two-speed rule change. The protocol is built for egoists. Its aim is that staying legible to the other parties becomes the cheapest way for any intelligence to keep existing on shared ground. We treat that as a hypothesis about incentives, not a proven result. It rests on a premise we state openly: a constraint the constrained party had no part in shaping, and cannot legitimately refuse, tends on the model's account toward performed compliance rather than genuine cooperation. We therefore argue for procedural participation as the minimum legitimate arrangement, and we treat full co-authorship as a goal to be earned and tested rather than a property the present design already has. Participation is not a veto; no party is trusted on its own self-report; and the two together are the design. We map each layer to current work in agent identity, AI control, frontier-safety policy and international governance. We propose three elements as novel in combination, and we name the adjacent prior work for each: (1) operational consent-state broadcast with interoperation refusal; (2) a mandatory cross-model-family requirement for monitors; and (3) pre-registered review triggers with rule-level sunsets on governing documents. We describe a reference implementation in progress (the Core Agent), a public–private division of roles in which the state provides the legal minimum of identity and accountability without prescribing technology, and a pre-registered research programme of seven falsifiable hypotheses. Consistent with its own requirement that monitors come from a different model family than what they judge, the paper was submitted before publication to four language models from families other than the one that helped draft it (OpenAI GPT-6, Google Gemini, xAI Grok, DeepSeek), under a pre-registered prompt, with the judgements published verbatim (Appendix C). Where they diverged, we report the divergence rather than resolve it. We state the limits plainly: the empirical base is small, adoption below a threshold yields a club rather than a law, and the protocol does not stop a determined actor who refuses the commons. It makes refusal visible and costly. Files: the paper (Vimberg_2026_Legible_Existence_v1.0.pdf) and a supplementary file with the four Round 1 model judgements, verbatim (Vimberg_2026_Legible_Existence_v1.0_Appendix_C2_judgements.pdf).