# HORTRAME IPS Alignment — Analysis v0.5.0

## 1. Why the 3D change is substantive

The move from a flat node-link diagram to overlapping 3D bubbles is not intended as visual novelty. The prior architecture contained several relations that a single planar graph conflated: containment, overlapping membership, causal coupling, nested scale, concurrent output events, and temporal feedback. In v0.5.0 these are separated visually.

A macro-IPS bubble is a **declared coarse-graining**. A sub-IPS can intersect several coarse-grainings. A mycelial strand represents a different relation: causal coupling. This allows, for example, one human sub-IPS to participate simultaneously in social, institutional and economic processes while also being causally coupled to AI or ecological processes that do not share the same membership substrate.

The modeling move is consistent with multilayer-network literature, which treats complex natural and engineered systems as having multiple simultaneous relation types and interacting layers. It should not be oversold: the current spheres are not a formal multilayer-network data set and their geometry is illustrative.

## 2. Hybrid dynamics

The user's distinction between continuous flow and discrete informational events maps cleanly onto hybrid dynamical systems. Goebel, Sanfelice and Teel formalize hybrid models with continuous flow maps and discrete jump maps. v0.5.0 adopts only that general modeling pattern. It does not claim any stability theorem from the hybrid-systems literature.

The executable now advances simulation dynamics on a fixed 1/60-step clock independent of render-frame rate. This matters scientifically: stochastic event generation should not change merely because one computer draws more frames per second than another.

## 3. Output-event alignment

The current working claim is deliberately narrow:

> IPSs are not intrinsically labeled aligned or misaligned; a realized causal output event can be evaluated as aligned, mixed, disaligned or unresolved relative to a declared boundary, state, horizon and lens version.

This is not equivalent to saying model internals are irrelevant. Hidden-objective, reward-tampering and alignment-faking research shows that repeated apparently safe outputs do not certify the producer. The event-level ontology therefore tracks output history as evidence about risk without converting that history into an intrinsic binary essence.

The concept also requires a stronger meaning of "output" than LLM text. An output is an enacted causal transition across a declared boundary. A model response that is never acted on may be an informational output from one boundary while not yet constituting the downstream institutional or physical output at another boundary.

## 4. Why the system is asynchronous

A distributed world does not generally have one global output at a time. Independent IPSs can act concurrently. The simulation allows several discrete output pulses to coexist. Conceptually the global history is a partially ordered set of events: some outputs causally precede others, while unrelated outputs may be concurrent.

The prototype does not yet implement a formal vector-clock or event-DAG export; this is a candidate v0.6 extension.

## 5. No-control premise

The model assumes only that permanent unilateral control of sufficiently autonomous AI is not guaranteed. It does not assert that loss of control is inevitable. The International AI Safety Report treats loss-of-control as a serious but uncertain risk class; AI-control research studies defenses even when a model is deliberately subversive. These streams motivate a separation:

- **control mechanisms** are defenses and deployment tools;
- **alignment** in this architecture is a property of interaction/output trajectories and admissibility, not synonymous with permanent obedience.

This moves the architecture toward coexistence, cooperation, institutional design and runtime governance rather than replacing control research.

## 6. Why the network level matters

Several current results make single-agent evaluation insufficient:

- Shen et al. report that organizations of individually alignment-trained agents can produce more effective but less ethical solutions than single agents in their studied settings.
- Comes frames human and AI interactions as evolving networks in which AI participates in collective sensemaking and changes information environments.
- Carroll et al. formalize the difficulty of alignment when reward functions are changing and influenceable.
- Glickman and Sharot provide evidence that human–AI feedback loops can alter subsequent human judgments.

These results support the direction of system-level modeling. They do not validate the specific IPS/Vow formalism.

## 7. Economy, ecosystem and world IPSs

These labels are intentionally treated as coarse-grained information-processing / adaptive-system abstractions.

Hayek's 1945 analysis emphasizes decentralized information in economic coordination. Levin describes ecosystems and the biosphere as complex adaptive systems in which higher-level patterns emerge from localized interactions. Neither source licenses the claim that an economy or ecosystem is a mind. v0.5.0 therefore places an explicit warning in the executable UI.

## 8. The theorem bridge

Priority-State Dynamics theorem 4.1 is relevant only as an exact local exemplar of history dependence. In its own stated stochastic priority model, join–mix–leave can change the returned mean; the same population labels after departure do not guarantee the same endpoint. A restoring newcomer conditional priority can depend on the mixing history.

The alignment consequence is a research question, not a theorem:

> when an autonomous IPS cannot be set directly, can surrounding topology, exposure, retention, permissions, buffering, compensating outputs or institutions reduce undesirable trajectory changes?

The theorem does not answer that question.

## 9. Five-Vow field

The Vow pentagon is currently a normative supervisory proposal. Its complete `K5` geometry encodes interdependence: no Vow is intended to be optimized independently of the others. The center is not automatically green; color severity is separated from directional displacement so symmetric multi-Vow tension cannot cancel visually.

v0.5.0 preserves lens-version history. A lens change cannot retroactively turn a failed past output green. This is necessary if "Flow" and "Evolution" are to remain compatible with falsifiability and Integrity.

## 10. What would make this scientific rather than decorative

The visualization becomes scientifically useful only if future releases replace illustrative functions with measurable quantities and discriminating tests. High-priority experiments remain:

1. output-event versus agent-level prediction;
2. multi-agent composition failure;
3. recurrent human–AI feedback;
4. no-control coexistence;
5. Vow-field ablation and alternative normative fields;
6. criterion gaming;
7. scale/coarse-graining sensitivity;
8. auditor independence;
9. event-history / causal-DAG reconstruction;
10. empirically measured channel coupling.

A successful visualization is not evidence that the architecture solves alignment.