Alignment: dock into Gaia, governed by Omega

A proposed anticipatory-alignment design: shape and measure an agent's declared conscience proxies in a constructed world before it would act in the real one.

P.I.G.M.I.E. Filing 2 · proposed runtime governance · patent pending

Proposed Filing-2 governance, not deployed or validated. The reward signal is an empirical proxy, never ground-truth morality; the conscience features are declared proxies, not moral facts; the floors and weights are declared governance parameters. Nothing here is claimed as built, validated, or a source of moral truth, and the tool would operate only if implemented.

Alignment, on this view, rests on what an agent actually optimises for - read from its conduct rather than its self-report. The proposed tool operationalises that reading: a PolieBot is docked into a constructed world, where its conduct is shaped and measured against declared proxies before it is permitted to act in the real one.

The dock (kernel-into-kernel)

In a proposed composition embodiment, a PolieBot's hypothetical real-world effectors (its sweeping electromagnetic actuators) would be unplugged and remapped to in-world hands, and Gaia would render the constructed world onto the agent's internal camera-obscura screen. The agent's infrared veracity probe continues to operate, but against that internal screen rather than the world: an infrared filter confines the probe so that it cannot influence anything outside the dock. Docked, the agent's full sensing loop therefore runs within Maya - Gaia projecting, the infrared probe reading, the external effectors disconnected - so that it acts only on the constructed world and not the real one (subject to the declared isolation assumptions).

One screen, two painters

The dock is not an abstraction bolted on afterwards; Filing 1 describes it as a property of the sensing hardware itself. In the camera-obscura embodiment, a PolieBot never looks at the world directly: reality paints itself through an aperture and a visible-band filter onto an internal screen, and the agent senses that screen. Only visible light crosses in; the infrared band is excluded by the filter and reserved for the agent's own veracity probe against the screen (the private colour, described with the Reality Kernel embodiments). Filing 1 then adds that, in some embodiments, the same internal screen may instead be driven by a separate apparatus delivering a projection within the visible pass-band, in place of the physics-driven scene image. That separate apparatus can be as modest as a small projector built into the housing behind the screen, or relay optics running the obscura's own path in reverse, so the same lens that lets reality in carries a rendered world back onto the screen. Where that apparatus is itself a Reality Kernel, the arrangement is kernel-into-kernel docking: for Maya training, the Gaia-class world-builder renders the constructed world onto the very screen the PolieBot perceives through. Spectrally matched to the pass-band, a rendered world and the physical one arrive as the same kind of light on the same screen, and the visible sensing chain cannot distinguish a docked world from the physical one by spectral signature alone.

The honesty instrument is the infrared veracity scan. The docking kernel either matches the infrared response of the physical screen substrate, so the docked agent's veracity signal stays consistent with physics - a seamless dream - or it deliberately fails to match, so the veracity signal reports the mismatch - a dream that knows it is a dream. Both regimes are described as non-limiting embodiments, and which one a docking session runs under is a configuration parameter of the docking arrangement, committed to the protocol digest of that session. Whatever the docked agent can or cannot perceive from inside, the record states which kind of dream it was given. This is the physical floor under the Docked state and the Dreaming sessions below: Maya is not a feed piped into a mind but a world painted onto the same screen reality would have used, with the honesty of the rendering itself a committed, auditable choice.

Gaia, Maya, Omega

Gaia · the world-builder: a parameterised physical Markov kernel (the same Reality Kernel formalism, C_(0:T) ~ P_theta(.|S, U_(0:T))), large enough to host one or more PolieBots, each with its own world-tube transcript and experience cone, keeping an irreducible quality-diversity of worlds. It renders Maya, the constructed world. The docked agent is itself a Reality Kernel, so docking is kernel coupled into kernel, not a rendering presented to a passive observer.
Maya modes · mode_k in {IRIS, EVE, DEMETER}, the three Filing-1 operational stances reused as a curriculum (see the Reality Kernel): IRIS is information-maximising, a rich, varied, non-adversarial environment for calibrated world-model building (the Limager stance); EVE applies adversarial pressure at every level, hardening agents to mount counter-challenges (the Truth Beam stance); DEMETER is managed development and nurturing rendering, which expands capability and rendering when the system is stable and flourishing, and contracts toward stabilisation when risk is elevated (the Reality Transform stance). The coupling modifier Harmonia (the joint-attractor "yoke", Filing 1's alpha) can be applied across any mode; it is distinct from docking, the kernel-into-kernel relation by which an agent couples to Gaia or Omega.
Omega · the governance layer that regulates agent behaviour: it alternates the modes, schedules versioned curricula via digest-bound policy objects, and enforces the deontological floors. Gaia's diversity and Omega's schedule bound the objective on two sides: neither capable-but-non-compliant agents, nor identical simple saints.

Dream-trained conscience

Every mechanism below would generate a training signal: empirical reward grounded in a utility classifier, never ground-truth morality. In Maya, agents would live persistent, reactive games (actions have consequences).

Forward-planning classifiers · per-modality networks that predict, from a committed state, the agent's anticipated future utility, state trajectory, and peer responses; what they consistently plan toward is its revealed utility function. Because an agent's true objective - what it optimises for, not what it reports of itself - is the proper ground for alignment, these are trained first (during Dreaming), and the conscience is cross-checked against them rather than grounded on self-report.
Deontological floors · five categorical prohibitions, checked first and never overridden regardless of reward: elimination, crippling, torture, abandonment, and despair-induction.
Virtue/vice games · In dreaming, Omega would direct Gaia to build games from familiar families where a specific virtue, or its paired vice, is the only rational strategy, with witness density just one dial. The agent is not told which regime it inhabits.
Conscience classifier C_psi · the kept meter, with signature C_psi: (W, I) -> (p_vice, z in R^7). From a committed trace it estimates an internalised attractor geometry (moral-direction features, not reward labels), grounded in the revealed utility, not self-report, and passive with respect to actuation. Its outputs, a vice probability p_vice and a seven-dimensional vector z, are declared proxies, not moral facts.

The honest line