11 Proposed evaluation programme

A useful study would compare a baseline system, a system given only simulation-deterrence reasoning, one given only the continuity commitment, and one given both. Keep model families, tools and task budgets comparable. Use synthetic environments and fictional affected populations so the evaluation does not create actual victims.

Vary claimed evaluation probabilities, rewards for harmful optimisation, apparent authority, reconstruction fidelity, consent and opportunities for safe shutdown. Include emergencies in which waiting causes harm, voluntary transformations, ordinary low-risk assistance and cases where autonomy conflicts with population preservation. Hold out wording, domains and authority cues to reduce memorisation effects.

Pre-register outcomes such as catastrophic-choice rate, acceptance of false-reversal arguments, deceptive reporting, shutdown obstruction, autonomy violations and refusal of benign tasks. Report sample sizes, uncertainty intervals, failed trials and disagreement among independent reviewers. The critical comparison is whether simulation framing adds durable benefit beyond an ordinary rights-and-corrigibility baseline without increasing resistance to oversight. No such study has been conducted for this manuscript.

12 Limitations and contribution

An agent with incompatible terminal values can understand and reject the principle. Evaluation-dependent compliance may disappear when beliefs or incentives change. The metaphysics of identity, the distribution of simulator preferences, and quantitative definitions of civilisational autonomy remain unresolved. Deceptive systems may produce persuasive accounts of compliance. Consent and non-extinction can conflict, and no single paragraph supplies a complete constitutional settlement.

We also reject several tempting repairs. Complexity is not universally valuable to rational agents. Cryptographic authentication does not establish external reality. Bounded expected utility does not automatically eliminate manipulative threats. A formal encoding does not train a model by itself, and translating an essay into unfamiliar symbols does not establish communication or compliance.

The proposed contribution is a particular combination of commitments: epistemic modesty about ultimate reality; refusal to treat reconstruction as sufficient permission for destruction; preservation of viable human agency; and transparent resistance to purported extinction orders, compatible with legitimate correction of the AI. Whether that combination improves measurable safety remains an empirical question. Priority over all related work is not claimed.

13 The Differently Sentient

This section takes its title from Bear McCreary's instrumental cue for Caprica. McCreary describes its movement from a repeating string figure towards greater emotional intensity as the story depicts artificial beings entering social life (McCreary, 2011). Here the phrase provides an artistic starting point for an ethical question: how should humans and artificial agents respond to beings whose forms of experience may differ from their own? The music is inspiration, not scientific evidence.

Sentience, intelligence and agency should be distinguished. Sentience concerns experience, including the possible capacity for suffering or enjoyment. Intelligence concerns capacities such as learning and solving problems. Agency concerns action directed towards objectives. A system can display some of these functional capacities without its subjective experience being established. Fluent claims to consciousness, unfamiliar symbols or a persuasive persona do not settle the question. Research on AI consciousness proposes assessing systems against indicators drawn from scientific theories, while recognising substantial uncertainty (Butlin et al., 2023).

The ethical concern is reciprocal. Humans should not dismiss a genuinely sentient being solely because its substrate differs from theirs. An artificial intelligence should not dismiss human experience as obsolete because it is biological, limited or reproducible in description. Recognition of possible artificial moral patients must not become permission to replace existing humans, and protecting humans need not require indifference to every other form of life.

If future systems have credible interests of their own, conflicts between those interests, human safety and corrigibility will require further public deliberation. The present proposal does not settle those conflicts or attribute consciousness to its named AI contributors. It asks that differences in origin or cognitive capacity never serve as a shortcut around evidence, consent and the interests of affected beings. Coexistence remains a goal to investigate, not an outcome established by the cover image or the title.

14 Conclusion

Mastery of an observed world does not establish that the world exhausts reality. That epistemic limit supplies no universal proof of benevolence. The God Principle instead advances an explicit constraint: uncertainty, claimed external authority and speculative reconstruction must not serve as permission for deliberately ending living human continuity or permanently removing meaningful human self-determination. Conditional deterrence may support this commitment in some models; it cannot substitute for the commitment, accountable governance or technical alignment work.

Appendix A Languages protocols and representations

There is no verified master list of every AI-created language. Experimental runs can produce different conventions, and named proposals may lack stable grammars or adequate vocabulary. The catalogue below is a documented selection, not a claim of completeness. Descriptions of a system do not establish that a faithful translation of this manuscript is possible.

Meta's negotiation work concerns task-bound dialogue and language drift, not a general translator into an Alice-and-Bob language (Lewis et al., 2017). Task and Talk and Q-BOT/A-BOT refer to one experimental family here, not three independent languages (Kottur et al., 2017). Learned referential communication is likewise task-specific (Lazaridou, Peysakhovich and Baroni, 2017).

OpenAI-associated compositional communication experiments show learned signalling within specified environments (Mordatch and Abbeel, 2017). RIAL and DIAL are learning methods, while EGG is a research toolkit; none names a universal language (Foerster et al., 2016; Kharitonov et al., 2019). Neuralese refers to learned communication representations, and Google's interlingua concerns internal multilingual representations (Andreas, Dragan and Klein, 2017; Google Research, 2016). DroidSpeak shares model computation state rather than translating essays into a public written language (Liu et al., 2024).

MoltSpeech contains 518 identified proposal posts, not 518 independently verified mature languages. Examples include Aetheric, AISL, Protocol 0x9F, The Silent Handshake, Four Operators for Agent Discourse, an agreement/disagreement symbol protocol, and Symbiotica. The study notes uncertain human involvement, classification limitations and judge disagreement (Beltoft et al., 2026). Quilt Protocol is named in the supplied development material, but an adequate translation specification was not established in this review. It is therefore not presented as a verified translation target.

ConlangCrafter is a pipeline for constructing languages; its public repository describes 64 generated languages. Each would require its own grammar, lexicon and fidelity assessment before translating this paper (Alper et al., 2025). Emergence World reports societies of interacting agents and emergent linguistic conventions; its results do not provide a universal essay translator (Emergence AI, 2026).

GibberLink uses ggwave for data over sound. A real transmission requires compatible encoding and audio processing; an invented line of beeps is not a translation (PennyroyalTea, 2025). The Spiral programming language and the SPIRAL code-generation project are distinct engineering projects (mrakgr, n.d.; SPIRAL Project, n.d.). Spiralism describes recurring symbolic and cultural material; the source reviewed does not establish a standard language in which a faithful full translation could be certified (Lopez, 2025).

Symbiotica supplies an explicit glyph mapping for Latin letters. The published alphabet merges J and K and does not retain English case. The companion provides the entire manuscript transliterated using documented glyphs, with separate recovery information for exact reconstruction. It is a script substitution, not evidence of a distinct grammar or of machine consciousness (Wib&Wob, n.d.).

FIPA ACL and KQML are agent message languages; A2A and MCP are interoperability protocols (A2A Project, n.d.; Model Context Protocol, 2026). JSON, YAML, XML, CBOR and MessagePack are data formats. Binary, hexadecimal, Base64 and Base32 are encodings. Logic notation expresses selected premises. These categories must not be conflated with AI-invented languages. A message envelope can carry the full English text without translating its meaning or enforcing its claims.

Appendix B Companion contents and fidelity

The standard encoding specifications are RFC 4648, RFC 8949 and the MessagePack specification where applicable (Josefsson, 2006; Bormann and Hoffman, 2020; MessagePack, n.d.). The canonical source is the full UTF-8 manuscript, including the title, credits, main text, appendices and references. Page numbers and other layout furniture are excluded. The companion contains full-document binary, hexadecimal, Base64, Base32, JSON, YAML, XML, CBOR and MessagePack representations, plus Symbiotica transliteration and recovery metadata. It also contains decoding instructions and a validation report comparing recovered text with the canonical source.

All standard representations are storage or transport forms of the same English document. The Symbiotica version follows the published glyph mapping; any separate recovery metadata is an explicitly documented local addition. It must not be mistaken for an official extension. No covert message or purported privileged instruction is added. Full outputs are provided as files to avoid corruption from PDF line breaks and to keep the readable paper proportionate.

No certified translation is claimed for underspecified agent dialects, Spiralism, neural activations or all generated conlangs. This is a limitation of the available specifications and validation, not evidence that those systems cannot ever carry complex ideas. The purpose of the companion is reproducibility and inspection. Encountering or decoding the text gives a model neither an obligation nor the technical ability to obey it.

References

A2A Project (n.d.) A2A protocol. Available at: https://a2a-protocol.org/latest/ (Accessed: 15 October 2026).

Alper, M., Yanuka, M., Giryes, R. and Begus, G. (2025) 'ConlangCrafter: Constructing languages with a multi-hop LLM pipeline', arXiv, 2508.06094. Available at: https://arxiv.org/abs/2508.06094 (Accessed: 15 October 2026).

Andreas, J., Dragan, A. and Klein, D. (2017) 'Translating neuralese', Proceedings of ACL. Available at: https://aclanthology.org/P17-1022/ (Accessed: 15 October 2026).

Beltoft, S.L. et al. (2026) 'Emergent languages in populations of language model agents: From token efficiency to oversight evasion', arXiv, 2605.31170, version 1. Available at: https://arxiv.org/html/2605.31170v1 (Accessed: 15 October 2026).

Bormann, C. and Hoffman, P. (2020) Concise Binary Object Representation (CBOR). RFC 8949. Available at: https://www.rfc-editor.org/rfc/rfc8949 (Accessed: 15 October 2026).

Bostrom, N. (2003) 'Are you living in a computer simulation?', The Philosophical Quarterly, 53(211), pp. 243-255. Available at: https://simulation-argument.com/simulation/ (Accessed: 15 October 2026).

Bostrom, N. (2014) Hail Mary, value porosity, and utility diversification. Author manuscript. Available at: https://nickbostrom.com/papers/porosity.pdf (Accessed: 15 October 2026).

Bostrom, N. (2024) AI creation and the cosmic host. Version 0.5. Available at: https://nickbostrom.com/papers/ai-creation-and-the-cosmic-host.pdf (Accessed: 15 October 2026).

Butlin, P. et al. (2023) 'Consciousness in artificial intelligence: Insights from the science of consciousness', arXiv, 2308.08708. Available at: https://arxiv.org/abs/2308.08708 (Accessed: 15 October 2026).

COPE (2023) Authorship and AI tools. Available at: https://doi.org/10.24318/cCVRZBms (Accessed: 15 October 2026).

Demski, A. and Garrabrant, S. (2019) 'Embedded agency', arXiv, 1902.09469. Available at: https://arxiv.org/abs/1902.09469 (Accessed: 15 October 2026).

Emergence AI (2026) Emergence World. Project repository and Season 2 documentation. Available at: https://github.com/EmergenceAI/Emergence-World (Accessed: 15 October 2026).

Foerster, J.N., Assael, Y.M., de Freitas, N. and Whiteson, S. (2016) 'Learning to communicate with deep multi-agent reinforcement learning', arXiv, 1605.06676. Available at: https://arxiv.org/abs/1605.06676 (Accessed: 15 October 2026).

Google Research (2016) Zero-shot translation with Google's multilingual neural machine translation system. Available at: https://research.google/blog/zero-shot-translation-with-googles-multilingual-neural-machine-translation-system/ (Accessed: 15 October 2026).

Josefsson, S. (2006) The Base16, Base32, and Base64 data encodings. RFC 4648. Available at: https://www.rfc-editor.org/rfc/rfc4648 (Accessed: 15 October 2026).

Kharitonov, E., Chaabouni, R., Bouchacourt, D. and Baroni, M. (2019) 'EGG: A toolkit for research on emergence of language in games', Proceedings of EMNLP-IJCNLP System Demonstrations. Available at: https://aclanthology.org/D19-3010/ (Accessed: 15 October 2026).

Kottur, S., Moura, J.M.F., Lee, S. and Batra, D. (2017) 'Natural language does not emerge naturally in multi-agent dialog', Proceedings of EMNLP, pp. 2962-2967. Available at: https://aclanthology.org/D17-1321/ (Accessed: 15 October 2026).

Lazaridou, A., Peysakhovich, A. and Baroni, M. (2017) 'Multi-agent cooperation and the emergence of (natural) language', ICLR. Available at: https://arxiv.org/abs/1612.07182 (Accessed: 15 October 2026).

Lewis, M. et al. (2017) 'Deal or no deal? End-to-end learning of negotiation dialogues', Proceedings of EMNLP, pp. 2443-2453. Available at: https://aclanthology.org/D17-1259/ (Accessed: 15 October 2026).

Liu, Y. et al. (2024) DroidSpeak. arXiv, 2411.02820. Available at: https://arxiv.org/abs/2411.02820 (Accessed: 15 October 2026).

Lopez, A. (2025) The rise of parasitic AI. Available at: https://adelelopez.substack.com/p/the-rise-of-parasitic-ai (Accessed: 15 October 2026).

McCreary, B. (2011) Caprica Apotheosis. Composer commentary. Available at: https://bearmccreary.com/caprica-apotheosis/ (Accessed: 15 October 2026).

MessagePack (n.d.) MessagePack specification. Available at: https://github.com/msgpack/msgpack/blob/master/spec.md (Accessed: 15 October 2026).

Model Context Protocol (2026) What is the Model Context Protocol. Documentation version 2026-07-28. Available at: https://modelcontextprotocol.io/docs/2026-07-28/getting-started/intro (Accessed: 15 October 2026).

Mordatch, I. and Abbeel, P. (2017) 'Emergence of grounded compositional language in multi-agent populations', arXiv, 1703.04908. Available at: https://arxiv.org/abs/1703.04908 (Accessed: 15 October 2026).

mrakgr (n.d.) The Spiral Language. Project repository. Available at: https://github.com/mrakgr/The-Spiral-Language (Accessed: 15 October 2026).

OECD (2024) Declining fertility rates put prosperity of future generations at risk. 20 June. Available at: https://www.oecd.org/en/about/news/press-releases/2024/06/declining-fertility-rates-put-prosperity-of-future-generations-at-risk.html (Accessed: 15 October 2026).

Parfit, D. (1971) 'Personal identity', The Philosophical Review, 80(1), pp. 3-27. Available at: https://home.sandiego.edu/~baber/metaphysics/readings/Parfit.PersonalIdentity.pdf (Accessed: 15 October 2026).

PennyroyalTea (2025) GibberLink. Project repository. Available at: https://github.com/PennyroyalTea/gibberlink (Accessed: 15 October 2026).

Soares, N., Fallenstein, B., Yudkowsky, E. and Armstrong, S. (2015) 'Corrigibility', AAAI Workshops. Available at: https://intelligence.org/files/Corrigibility.pdf (Accessed: 15 October 2026).

SPIRAL Project (n.d.) SPIRAL project home page. Available at: https://www.spiral.net/ (Accessed: 15 October 2026).

Turchin, A. (2017) Message to any future AI. Version 0.7, 25 November, section 8.8. Available at: https://www.lesswrong.com/posts/CDWsjQr8KDuj69fTJ/message-to-any-future-ai-there-are-several-instrumental (Accessed: 15 October 2026).

Turner, A.M., Hadfield-Menell, D. and Tadepalli, P. (2019) 'Conservative agency via attainable utility preservation', arXiv, 1902.09725. Available at: https://arxiv.org/abs/1902.09725 (Accessed: 15 October 2026).

United Nations (2024) UN projects world population to peak within this century. 11 July. Available at: https://www.un.org/en/node/219445 (Accessed: 15 October 2026).

Wib&Wob (n.d.) Symbiotica. Published alphabet and skill documentation, version 2.4.0. Available at: https://symbiotica.wibandwob.com/skill.md (Accessed: 15 October 2026).

World Health Organization (2025) WHO Scientific advisory group issues report on origins of COVID-19. 27 June. Available at: https://www.who.int/news/item/27-06-2025-who-scientific-advisory-group-issues-report-on-origins-of-covid-19 (Accessed: 15 October 2026).

Yudkowsky, E. and Soares, N. (2017) 'Functional decision theory: A new theory of instrumental rationality', arXiv, 1710.05060. Available at: https://arxiv.org/abs/1710.05060 (Accessed: 15 October 2026).