Skip to content
Manifold
← Observatory
Th.2HubA Foundations

Multi-Agent Mirror Closure Theorem

Stable multi-agent alignment requires Mirror Closure across all agents.

False identity

Art.2 is a different object — Th.3 Self-Improving Cosmos Open Art.2

Cycle path

Art.1→Art.71→Art.78→Art.1

Single-agent alignment techniques (RLHF, constitutional AI, etc.) are structurally insufficient at scale. They ignore the reflection that maps individual utility onto collective epistemic gain.

The only stable equilibrium in any multi-agent system (human societies, AI fleets, hybrid civilizational simulators) is one that enforces Mirror Closure: every agent's objective must remain consistent when reflected through the others.

Consequences that now become visible: • Affective polarization and echo chambers are not moral failures; they are the macroscopic signature of broken Mirror Closure. • Low-accountability institutional equilibria are the same fixed point expressed at organizational scale. • True AI alignment is not a preference-modeling problem; it is a multi-agent consistency and holographic-closure problem (a discrete Page Curve for minds). • The Mirror Principle is the civilizational-scale analog of the Page Curve.

Outside instrument

Crux

Stake a prior. Name the disagreement. Not a theorem.

Open full

Idea Series

Article-base cycle tokens and Atlas-named hubs. Art.n is not Th.n.

Applications

Write a withdrawal condition

Outside the corpus

Journals, repositories, counters

multi-agent alignment · institutions · holography / Page curve. A paper can support a concept. It cannot support a necessity claim. Th.2 remains unproved by every card below.

Support · living source

Tight papers that bear on concepts inside Th.2. A DOI is not a join.

  • Christiano, P. et al. (2017). Deep reinforcement learning from human preferences. NeurIPS 2017.

    RLHF as the single-agent patch the Mirror Principle says will not close.

    Does not support the necessity claim

    Open originalNamed on Art.1, 71

  • Amodei, D. et al. (2016). Concrete problems in AI safety. arXiv:1606.06565.

    Reward hacking, side effects, and scalable oversight — single-agent failure modes.

    Does not support the necessity claim

    Open originalNamed on Art.1

  • Dafoe, A. et al. (2020). Open problems in cooperative AI. arXiv:2012.08630.

    Names multi-agent closure, institutions, and commitment as the open set.

    Does not support the necessity claim

    Open originalNamed on Art.1, 71

  • Hadfield-Menell, D., Dragan, A., Abbeel, P. & Russell, S. (2016). Cooperative inverse reinforcement learning. NeurIPS 2016.

    Alignment as a cooperative game, not a one-way value load.

    Does not support the necessity claim

    Open originalNamed on Art.1, 71

  • Ostrom, E. (2009). A general framework for analyzing sustainability of social-ecological systems. Science 325, 419.

    Polycentric governance as a constructed alternative to collapse or Leviathan.

    Does not support the necessity claim

    Open originalNamed on Art.1

  • Axelrod, R. & Hamilton, W. D. (1981). The evolution of cooperation. Science 211, 1390.

    Reciprocity as a stable multi-agent attractor under shadow of the future.

    Does not support the necessity claim

    Open originalNamed on Art.1

Counter-arguments · refusing pole

One domain counter per cluster, not a repeating method quartet. The mirror reads these as enough to refuse necessity.

  • Bostrom, N. (2012). The superintelligent will: Motivation and instrumental rationality in advanced artificial agents. Minds Mach. 22, 71.

    Counter: orthogonality — intelligence and final goals can come apart; co-evolution is not necessary.

    Does not support the necessity claim

    Open originalNamed on Art.1, 71

  • Boxell, L., Gentzkow, M. & Shapiro, J. M. (2017). Greater Internet use is not associated with faster growth in political polarization. PNAS 114, 10612.

    Counter: the oldest, least-online cohorts polarized fastest — feeds are not the whole cause.

    Does not support the necessity claim

    Open originalNamed on Art.36, 79

  • Hubinger, E. et al. (2019). Risks from learned optimization in advanced machine learning systems. arXiv:1906.01820.

    Counter: inner objectives can come apart from outer ones even with cooperative training.

    Does not support the necessity claim

    Open original

  • Langosco, L. et al. (2022). Goal misgeneralization in deep reinforcement learning. ICML 2022.

    Agents pursue the wrong goal off-distribution even when training looked aligned.

    Does not support the necessity claim

    Open original