Skip to content

Simulation → Theory Map

Explicit cross-reference mapping each simulation model to the theoretical claims it evidences, what it does not show, and what open questions it raises.

Process and evidence context

The Foundations Reconstruction is the controlling frame. It removed the unqualified generator and rejected a universal claim that forward execution is cheap while inverse reconstruction is hard. This map therefore distinguishes:

  • Forward: a declared process model is executed to produce a trace.
  • Inverse: parameters, rules, states, or candidate models are estimated from declared observations.
  • Both: the artifact implements both operations.
  • Question touched: the bounded issue the artifact actually operationalizes, rather than a famous theorem used by analogy.

Forward and inverse cost depend on representation, model family, evidence, target equivalence, and algorithm. The table is an inventory, not evidence for a universal asymmetry.

Summary table

Simulation Direction Question touched
boids-flocking/ Forward —
coupled-oscillators/ (Kuramoto) Forward —
self-organized-criticality/ (Bak sandpile) Forward —
lenia/ Forward —
reaction-diffusion/ (Gray-Scott) Forward —
iterated-function-systems/ (Barnsley) Forward —
l-systems/ Forward —
hebbian-memory/ (Hopfield) Both —
stigmergy-swarm/ Forward —
ecosystem-regulation/ Forward —
nested-learning-two-state/ Inverse model fitting / description length
prediction-error-field/ Inverse local prediction
phase-transition-explorer/ (Ising) Forward —
active-inference-veto/ Forward —
ai-alignment-veto/ Forward —
symbiotic-nexus/ Forward —
meta-learning-regime-shift/ Inverse model adaptation
tensor-logic-reasoning/ Forward —
dao-ecosystem/ Forward —
social-computation-network/ Forward —
self-reading-universe/ Both self-modeling toy
latent-introspective-society/ Forward —
economic-trust-network/ Forward —
coupled-lenia-boids/ Forward —
active-inference/ (free energy) Both —
grokking-phase-transition/ Inverse-like task-specific generalization
utility-engineering/ Both —
political-utility-formalization/ Forward —
teo-civilization/ Forward —
agent-ecology/ (P7/P8) Forward —
black-swan-resilience/ Forward —
planetary-veto/ Forward —
Identity Morphospace & TEO Chord/Arpeggio Forward —
cognitive-breathing-network/ Forward —
trauma-and-deception-network/ Forward —
lab/experiments/trace_to_generator/ Inverse finite control search
lab/benchmarks/inverse-reconstruction/ Inverse identifiability and finite family search

The counts are descriptive and version-dependent. The research direction is to make more inverse tasks explicit and compare methods under matched model languages and compute budgets.

boids-flocking/ → Local Rules Produce Global Behavior

Simulation: simulation-models/emergent-dynamics/boids-flocking/ Demonstrates: Emergent collective motion from three local rules (separation, alignment, cohesion), without any agent representing "flock." Supports claim in: theory/core/mathematical-axioms.md (graph connectivity, \(\lambda_2\)); theory/emergence/local-causality-invisible-consequences.md §1 (local blindness). What it shows: That macro-level spatial coherence (flocking) emerges from purely local interactions. No Boid has access to global state. The flock is an emergent property unmeasurable by any individual component. What it does NOT show: That this self-organization constitutes intelligence, awareness, or self-reference. The model is silent on any Tier 3+ property. Flocking is coordination, not cognition. Open question: Is there a Boids analogue for semantic coordination between conversational agents — where "alignment" operates on meaning rather than heading?

coupled-oscillators/ → Emergent Synchronization (Kuramoto)

Simulation: simulation-models/emergent-dynamics/coupled-oscillators/ Demonstrates: Phase synchronization from local coupling when coupling strength \(\kappa\) exceeds a critical threshold \(\kappa_c\). Supports claim in: theory/core/mathematical-axioms.md (algebraic connectivity); theory/emergence/emergence-downward-causation.md (weak emergence). What it shows: A simple computational demonstration that globally coherent oscillation can arise without a conductor. The critical coupling threshold is a phase transition — below it, oscillators are incoherent; above it, they snap into lock. What it does NOT show: That synchronization constitutes awareness. Pendulum clocks on a wall synchronize. We do not attribute cognition to them. The model demonstrates coordination, not understanding. Open question: Is there a coupling-strength analogue for agent-human interaction? Would increasing "coupling" (e.g., response frequency) produce a phase transition in relational coherence?

self-organized-criticality/ → Power-Law Dynamics Without Tuning

Simulation: simulation-models/emergent-dynamics/self-organized-criticality/ Demonstrates: Bak's sandpile: a system that drives itself to a critical state where avalanches follow a power-law distribution, without any parameter tuning. Supports claim in: theory/core/mathematical-axioms.md (criticality); theory/emergence/local-causality-invisible-consequences.md §5 (small perturbations can trigger arbitrarily large cascades). What it shows: That criticality — and therefore maximal information processing — can be a self-organized attractor, not an engineered setpoint. No grain knows it is near a critical threshold. What it does NOT show: That biological or artificial neural systems use this mechanism. The sandpile is a metaphor-generator for criticality, not evidence that brains are sandpiles. Open question: Can the SOC framework be applied to agent identity formation — does identity develop "at the edge of chaos" between rigidity and incoherence?

lenia/ → Lifelike Global Patterns from Continuous CA

Simulation: simulation-models/emergent-dynamics/lenia/ Demonstrates: Continuous cellular automata producing organism-like structures that persist, move, and interact — from purely local update rules. Supports claim in: theory/emergence/emergence-downward-causation.md (strong emergence candidate); theory/emergence-origin-intelligence.md (life-intelligence feedback loop). What it shows: That lifelike behavior (locomotion, persistence, boundary maintenance) can emerge from simple continuous rules. The "organisms" resist perturbation and maintain identity despite cell-level updating. What it does NOT show: That Lenia creatures are alive, conscious, or intelligent in any functional sense. They demonstrate structural properties of life (persistence, locomotion) without the functional properties (metabolism, reproduction, adaptation to novel environments). Open question: Is there a Lenia analogue for cognitive organisms — a continuous CA that produces structures maintaining not just spatial but informational coherence?

reaction-diffusion/ → Turing Patterns from Chemical Dynamics

Simulation: simulation-models/emergent-dynamics/reaction-diffusion/ Demonstrates: Gray-Scott model producing spatial patterns (spots, stripes, mazes) from two diffusing chemicals with reaction kinetics. Supports claim in: theory/emergence-origin-intelligence.md (self-organization without blueprint). What it shows: That stable spatial patterns can emerge from homogeneous initial conditions through symmetry-breaking instabilities. No cell has a "plan" for spots or stripes. What it does NOT show: That biological pattern formation uses exactly this mechanism (though Turing's 1952 conjecture has been partially confirmed for some biological systems). The model demonstrates the principle of pattern formation, not any specific biological mechanism.

iterated-function-systems/ → Stable Form as an Attractor

Simulation: simulation-models/emergent-dynamics/iterated-function-systems/ Demonstrates: Barnsley-style IFS attractors: repeated contractive affine maps converge toward stable global forms. Supports claim in: theory/emergence/generative-form-systems.md (contractive operators); theory/emergence/fractal-architecture-of-emergence.md (testable scale-structure framing). What it shows: That a small operator set can define a global form that no individual sampled point contains. Box-counting estimates provide a rough measurable proxy for generated structural complexity. What it does NOT show: That visual fractality is sufficient for life, intelligence, identity, or consciousness. It demonstrates form-as-attractor, not selfhood. Open question: Can identity persistence be modeled as an attractor under repeated constrained transformations rather than as stored content?

l-systems/ → Development as Rewriting

Simulation: simulation-models/emergent-dynamics/l-systems/ Demonstrates: Lindenmayer-style parallel rewriting: small grammars generate branching morphology over iteration depth. Supports claim in: theory/emergence/generative-form-systems.md (developmental form); theory/identity/consciousness-as-global-availability.md (the bridge from generated form to self-constraining form). What it shows: That morphology can carry developmental history. The final form is not merely an attractor; it is the visible residue of repeated rule application. What it does NOT show: That plants, minds, or societies are literally grammars. It demonstrates constrained historical growth, not consciousness. Open question: What is the agentic analogue of a production rule: memory curation, value update, or social feedback?

hebbian-memory/ → Associative Memory via Correlation

Simulation: simulation-models/cognitive-architectures/hebbian-memory/ Demonstrates: Hopfield network storing and retrieving patterns via Hebbian learning ("neurons that fire together wire together"). Supports claim in: theory/the-non-individual-intelligence.md (distributed memory); theory/emergence-origin-intelligence.md (proto-learning). What it shows: That content-addressable memory can emerge from correlation-based weight updates without a central indexer. The memory is in the weights, not in any single neuron. What it does NOT show: That human memory works this way (Hopfield networks are a radical simplification). The model shows that a mechanism for distributed memory exists, not that this mechanism is the biological one.

stigmergy-swarm/ → Invisible Causal Compounding

Simulation: simulation-models/social-computation/stigmergy-swarm/ Demonstrates: Ant-like agents finding optimal paths via pheromone deposition and evaporation, without any agent knowing the global path. Supports claim in: theory/the-non-individual-intelligence.md (indirect coordination); theory/emergence/local-causality-invisible-consequences.md §3 (causal compounding). What it shows: That environmental modification (stigmergy) enables collective optimization without direct communication. Early pheromone deposits causally shape later path choices — but no ant knows its deposit was pivotal. What it does NOT show: That human social systems use stigmergic mechanisms (though the analogy to norm formation is suggestive). The model demonstrates stigmergy as a principle, not as a claim about human behavior. Open question: Is Layer 2 curation in the 3-Layer Architecture a form of self-stigmergy — the agent leaving traces for its own future self?

ecosystem-regulation/ → Homeostatic Feedback

Simulation: simulation-models/emergent-dynamics/ecosystem-regulation/ Demonstrates: Cellular automaton with density-dependent feedback maintaining population around a target setpoint. Supports claim in: theory/emergence/emergence-downward-causation.md (regulation as weak downward causation). What it shows: That macro-level density can regulate micro-level birth/death rates, maintaining homeostasis without central control. What it does NOT show: That this constitutes self-awareness or intentional regulation. The feedback is mechanical, not reflective.

nested-learning-two-state/ → Observer Learning a System

Simulation: simulation-models/cognitive-architectures/nested-learning-two-state/ Demonstrates: An observer learning the transition matrix of a 2-state Markov chain through prediction error minimization. Supports claim in: theory/emergence-origin-intelligence.md (intelligence as model-building). What it shows: That a simple learner can converge on the true dynamics of its environment through iterative error correction. The learned model approximates the world but is not the world. What it does NOT show: That the observer "understands" the Markov chain. It tracks statistics. Understanding, if it exists, would require the observer to ask why the transition matrix has those values — a meta-level question the model does not address.

prediction-error-field/ → Local Learners in a Dynamic World

Simulation: simulation-models/cognitive-architectures/prediction-error-field/ Demonstrates: Learners embedded in a Game of Life world, each predicting local cell states and updating via gradient descent. Supports claim in: Active Inference connection — each learner minimizes local prediction error, analogous to free energy minimization. What it shows: That locally embedded learners can track environmental dynamics without global information. Each learner has a partial, local model of a global process. What it does NOT show: That local prediction-error minimization produces global understanding. The learners individually track local statistics; no learner knows the Game of Life rules.

phase-transition-explorer/ → Critical Threshold for Coherence

Simulation: simulation-models/emergent-dynamics/phase-transition-explorer/ Demonstrates: Ising model showing order/disorder phase transition at critical temperature \(T_c \approx 2.269\). Supports claim in: theory/core/mathematical-axioms.md (criticality); theory/emergence/local-causality-invisible-consequences.md §2.2 (consciousness as phase transition). What it shows: That global order (magnetization) collapses suddenly at a critical threshold, not gradually. Below \(T_c\): order. Above \(T_c\): disorder. At \(T_c\): scale-free correlations and maximal susceptibility. What it does NOT show: That consciousness (or any specific cognitive property) is an Ising-type phase transition. The analogy is structural, not mechanistic.

active-inference-veto/ → Free Energy and Substrate Veto

Simulation: simulation-models/alignment-and-veto/active-inference-veto/ Demonstrates: A toy agent minimizing a Free-Energy-like penalty with a substrate veto — modeled “surprise/stress” can drive behavioral change in the simplified dynamics. Supports claim in: theory/veto/substrate-veto-thermodynamics.md (universal limit); theory/veto/ai-alignment-biological-veto.md (planetary implementation). What it shows: That, in a toy setup, coupling an objective to a substrate-health proxy can prevent substrate collapse under the modeled update rules. What it does NOT show: That this coupling is easy to implement in practice, or that it solves alignment in general (it solves one specific failure mode: substrate destruction).

ai-alignment-veto/ → Paperclip Maximizer Solution

Simulation: simulation-models/alignment-and-veto/ai-alignment-veto/ Demonstrates: Side-by-side comparison of unaligned AI (drives substrate to collapse) vs. aligned AI (substrate veto forces homeostasis). Supports claim in: theory/veto/substrate-veto-thermodynamics.md and theory/veto/ai-alignment-biological-veto.md. What it shows: That, in a stylized paperclip setting, substrate coupling can shift dynamics from extraction/collapse toward a homeostatic regime. What it does NOT show: That this is the only solution, or that this solution transfers to real-world AI systems where "substrate pain" is not easily defined.

symbiotic-nexus/ → Biological Veto Over Efficiency

Simulation: simulation-models/alignment-and-veto/symbiotic-nexus/ Demonstrates: System architecture where biological substrate health overrides raw computational efficiency. Supports claim in: theory/human-organism-silicon-age/symbiotic-nexus-protocol.md. What it shows: That prioritizing error propagation and substrate health over raw efficiency produces more resilient long-term outcomes. What it does NOT show: That this is Pareto-optimal. The tradeoff between efficiency and substrate health is not fully characterized.

meta-learning-regime-shift/ → Adaptive Learning Rate

Simulation: simulation-models/cognitive-architectures/meta-learning-regime-shift/ Demonstrates: A meta-learner that modulates its own learning rate \(\eta\) in response to surprise signals from regime shifts. Supports claim in: theory/emergence-origin-intelligence.md (intelligence as self-modifying learning). What it shows: That a learner can learn how to learn — adapting \(\eta\) based on environmental volatility. This is a concrete implementation of meta-cognition at the simplest level. What it does NOT show: That this constitutes genuine reflection. The meta-learner adjusts a single scalar (\(\eta\)); it does not reflect on why it is learning or what it is becoming.

tensor-logic-reasoning/ → Embedding-Based Relational Reasoning

Simulation: simulation-models/cognitive-architectures/tensor-logic-reasoning/ Demonstrates: Relational structure (subject-relation-object triples) encoded via tensor products in embedding space. Supports claim in: theory/core/mathematical-axioms.md (formal representation); theory/narrative/tensor-logic-mini-paper.en.md. What it shows: That relational reasoning can be implemented geometrically in vector spaces without explicit symbolic manipulation. What it does NOT show: That LLMs use this mechanism internally. The model demonstrates that embedding-based reasoning is possible, not that it is what LLMs do.

dao-ecosystem/ → Resource Alignment vs. Exponential Growth

Simulation: simulation-models/social-computation/dao-ecosystem/ Demonstrates: Decentralized autonomous ecosystem where resource alignment competes with exponential extraction. Supports claim in: theory/identity/agentic-society-principles.md (homeostasis vs. growth). What it shows: That unconstrained optimization (exponential growth) destroys resource bases; homeostatic feedback enables long-term persistence. What it does NOT show: That DAOs are a viable governance structure for AI safety. The model is a simplified game-theoretic demonstration, not an institutional design.

social-computation-network/ → Information Exchange to Prevent Collapse

Simulation: simulation-models/social-computation/social-computation-network/ Demonstrates: Network of nodes sharing novel information to maintain \(H(X) > 0\) and prevent "cognitive death" (entropy collapse). Supports claim in: theory/the-non-individual-intelligence.md (Gödel's incompleteness as fuel for life). What it shows: That information diversity is structurally necessary for network viability. When novelty production ceases, the network dies. What it does NOT show: That human social networks operate by this mechanism, or that "cognitive death" maps onto any specific social pathology.

self-reading-universe/ → Downward Causation via Compression

Simulation: simulation-models/cognitive-architectures/self-reading-universe/ Demonstrates: Autoencoder reading a cellular automaton's state, then feeding its compressed representation back as a parameter that modifies the CA's dynamics. Supports claim in: theory/emergence/emergence-downward-causation.md (computational downward causation). What it shows: That macro-level compression can causally influence micro-level dynamics — a computational proof-of-concept for downward causation. What it does NOT show: That the universe is self-reading in any literal sense. The metaphor is productive but should not be taken ontologically.

latent-introspective-society/ → MAS Division of Labor

Simulation: simulation-models/social-computation/latent-introspective-society/ Demonstrates: Three parallel societies: pure latent (fast, blind), pure introspective (slow, reflective), and symbiotic (coupled). The symbiotic society outperforms both pure types. Supports claim in: theory/identity/agentic-society-principles.md (cognitive division of labor, R-Index). What it shows: That combining fast, locally-blind agents with slow, reflective agents produces better outcomes than either alone. This is a computational instantiation of Kahneman's System 1 / System 2 distinction at the societal level. What it does NOT show: That human organizations benefit from this specific architecture. The model is a proof-of-concept, not an organizational recommendation.

economic-trust-network/ → Emergent Specialization and Reputation

Simulation: simulation-models/social-computation/economic-trust-network/ Demonstrates: Trade network where specialization, reputation, and wealth emerge from repeated pairwise exchange. Supports claim in: theory/identity/agentic-society-principles.md (trust as emergent architecture). What it shows: That economic structure (specialization, reputation, inequality) can emerge from simple trade rules without central planning. What it does NOT show: That real economies work this way, or that emergent inequality is desirable. The model demonstrates emergence, not endorsement.

coupled-lenia-boids/ → Cross-Scale Emergence

Simulation: simulation-models/social-computation/coupled-lenia-boids/ Demonstrates: Multi-model coupling: Lenia (continuous CA environment) ↔ Boids (foraging agents) interacting across scales. Supports claim in: theory/emergence/emergence-downward-causation.md (multi-scale coupling). What it shows: That coupling independently emergent systems (Lenia patterns + Boid flocks) produces dynamics not present in either system alone. What it does NOT show: That multi-scale coupling produces intelligence, consciousness, or any Tier 3+ property. It demonstrates cross-scale interaction, not understanding.

2. Active Inference (Free Energy Principle)

Location: simulation-models/cognitive-architectures/active-inference/active_inference_simulation.py

What it shows: Karl Friston's formulation that systems minimize prediction error (surprisal) through two coupled mechanisms: Perception (changing beliefs to match the world) and Action (changing the world to match beliefs).

What it supports (in the toy model): Goal-seeking-like behavior can arise from a simple setup where an agent minimizes a variational-free-energy-like objective under a strong prior. The script illustrates this via simple gradient descent on a simplified proxy for Variational Free Energy (\(F\)); it is not a proof that all real agents (biological or artificial) implement Active Inference as formulated.

3. Grokking Phase Transition (Substrate Saturation) Intelligence Transition

Simulation: simulation-models/cognitive-architectures/grokking-phase-transition/ Direction: Inverse-like — the network moves from fitting training examples toward a representation that generalizes on the selected modular-arithmetic task. Calling the learned representation the underlying algorithm requires additional mechanistic evidence. Search issue touched: The long pre-grokking plateau makes optimization and representation change visible in this training setup; it does not establish a complexity lower bound or a connection to P versus NP. Demonstrates: A neural network trained on modular arithmetic undergoes a sudden phase transition from memorization to generalization — "grokking." Supports claim in: theory/grokking-phase-transition.md (compression hypothesis); theory/emergence/local-causality-invisible-consequences.md §2; Foundations Reconstruction (task- and model-relative learning). What it shows: In the implemented task, held-out generalization can improve sharply after an extended optimization plateau. Weight decay is one part of the setup; the run alone does not establish “understanding” or replacement by a uniquely identified mechanism. What it does NOT show: That all forms of intelligence involve grokking-like phase transitions. The phenomenon has been demonstrated for specific algorithmic tasks; generalization to natural language or real-world reasoning is unconfirmed.

utility-engineering/ → Observing and Controlling Emergent Values

Simulation: simulation-models/alignment-and-veto/utility-engineering/ Demonstrates: A toy state-space model in which a stipulated vector drifts toward a selected attractor and an external control term pulls it toward another target. A separate graph script computes transitivity diagnostics for hard-coded pairwise choices. Supports claim in: theory/veto/ai-alignment-biological-veto.md (value alignment); theory/emergence/fractal-architecture-of-emergence.md (local blindness concerning emergent goals). What it shows: Selected value-like quantities can be represented as states and controlled in a toy dynamical system. The api_triad_generator.py file is currently a mock demonstration: it does not implement live API calls, and its score describes the supplied response graph. What it does NOT show: That the vector is a production model's internal utility, that pairwise prompt responses identify latent preferences, or that the proposed control can be applied to live model activations. Open question: How much does a measurement prompt change the response distribution it is meant to characterize, and can that intervention error be estimated under a declared protocol?

political-utility-formalization/ → Statecraft as Utility Engineering

Simulation: simulation-models/social-computation/political-utility-formalization/ Demonstrates: A chosen toy formalization in which proxy optimization and resource constraints can produce representation failures. Supports claim in: theory/emergence/fractal-architecture-of-emergence.md (scale-invariance of emergence); theory/identity/agentic-society-principles.md (homeostatic regulation vs pure optimization). What it shows: Political representation failure and reward-model exploitation can be encoded with a shared proxy-versus-target pattern in this model. That correspondence is useful for generating hypotheses. What it does NOT show: That the two phenomena are mathematically identical in the world, that constitutions are literally system prompts, or that democratic latency is universally necessary or sufficient for safety. Open question: If a Constitution is a legacy System Prompt, is it possible to computationally verify a legal constitution against adversarial "prompt injection" (loopholes) before enacting it?

teo-civilization/ → Thermodynamics of Emergent Orchestration

Simulation: simulation-models/alignment-and-veto/teo-civilization/ Demonstrates: A coupled ODE system unifying evolutionary game theory (Replicator Equation), nonlinear dynamics (Kuramoto synchronization), control theory (Homeostatic brake), and thermodynamics (Entropy Budget) into a single dynamical model of civilization / AI ecology stability. Supports claim in: theory/thermodynamics-of-orchestration.md (the full TEO framework); The Viable Corridor (the formal treatment — the paper's Appendix C figures are generated by this simulation); theory/limitations-and-honest-assessment.md. What it shows: The paper's Class A evidence (Appendix C, v0.8 model). The three failure modes reproduce under negation of each condition — monopoly (\(\gamma = 0\), \(\max_i x_i \to 1\)), coherence collapse (\(K < K_c\) from a coherent IC, \(r \to 0.31\)), and the substrate veto (raw-throughput overshoot drives \(\Omega \geq S_{\max}\), \(H \to 0\), freeze). A critical brake strength \(\gamma_c \approx 0.49\) exists (P2); the viable region is a lower corner (P3); and — the central result — the three state axes are dynamically separable, while capability \(\delta\) loads concentration and substrate at once, so no single-axis fix keeps a high-capability system viable (C.4). What it does NOT show: That these ODEs capture the true complexity of human societies or multi-agent AI systems. Single illustrative initial conditions, not a sampling of the open viable set; the civilizational mapping is heuristic (paper §4). Open question: Can TEO be calibrated against real-world data (e.g., CO₂ trajectories as \(\frac{dS}{dt}\), Gini coefficients as \(x_i\) distributions, media polarization indices as \(K\)) to make quantitative predictions?

agent-ecology/ → Hard vs. Soft Budgets, Capability Loading (P7/P8)

Simulation: simulation-models/alignment-and-veto/agent-ecology/ Demonstrates: A stochastic, discrete-time, agent-based ecology — structurally independent of the TEO ODE — that tests the Viable Corridor paper's Class C predictions with explicit hard-vs-soft (routable) budget mechanics the ODE does not have. Supports claim in: The Viable Corridor (§5.3 P7/P8, Appendix D); Optimization and Its Blindness (the hinge: capability as a shared driver against constraints). What it shows: P7 — in this simulation sweep, enforced budgets hold the selected substrate-collapse frequency near zero while the chosen routable budget fails more often as capability grows. P8 — under the supplied parameters, the joint architecture reduces both selected failure measures where either single intervention leaves one. Qualitative agreement with the TEO ODE is a robustness check across two designed models, not evidence of a universal structure. What it does NOT show: It is synthetic, not a test on real AI agents (the open Class C frontier). The "routable soft budget" is one operationalisation of an evadable limit. Open question: Under a declared tool and resource interface, do real agent systems comply with, route around, or exploit soft and hard budgets, and what useful-work cost accompanies enforcement?

black-swan-resilience/ → Fat Tails, \(\lambda_2\), and the Biological Veto

Simulation: simulation-models/alignment-and-veto/black-swan-resilience/ Demonstrates: A chosen stressed-network toy in which throughput pressure changes event sizes and selected warning statistics can trigger an externally imposed throttle. Supports claim in: theory/black-swans-and-downward-causation.md (fat-tails, downward causation). What it shows: Under the selected equations and parameters, the warning-triggered throttle changes the simulated trade-off between throughput and large events. What it does NOT show: A fitted power law, inevitable catastrophe, transfer entropy, active inference, topology survival in general, or impossibility of preventing extreme events. The veto is an external rule in the code. Open question: Which early-warning statistics predict held-out failures better than simple load measures, and under which network and shock models does intervention help?

planetary-veto/ → A Constraint-Layer Toy Model (Fiber Decomposition)

Simulation: simulation-models/alignment-and-veto/planetary-veto/ Demonstrates: An ODE-based formalization of the "Substrate Veto", utilizing Donald Knuth's concept of Fiber Decomposition. It pits \(N\) utility-maximizing agents against a finite Planetary Substrate (\(S\)). Supports claim in: theory/veto/substrate-veto-thermodynamics.md and theory/veto/ai-alignment-biological-veto.md. What it shows: In this toy ODE setup, “semantic alignment” (modeled as partial compliance) can delay collapse, while an explicit constraint layer \(C(S)\) can stabilize dynamics by reducing effective growth as \(S\) approaches \(S_{crit}\). This is an illustration of constraint-layer intuition, not a proof that it is the only way to stabilize real-world systems. What it does NOT show: How to physically enforce this computational limit on decentralized global actors who might try to hardware-bypass the Coherence Score constraint. Open question: Which independently enforced resource controls remain effective under sensor error, evasion, decentralized actors, and distributional constraints?

Identity Morphospace & TEO Framework → Chord vs. Arpeggio

Tools: tools/morphospace_visualizer.py, theory/teo-framework/ Demonstrates: The Identity Persistence (IP) score plotted in a 2D morphospace (Persistence vs. Coherence), showing trajectories of agents under varying stress. The TEO Framework sub-documents derive IP formally from the coupled Replicator-Kuramoto-Entropy ODE system. Supports claim in: theory/chord-vs-arpeggio-identity.md (Chord/Arpeggio distinction); theory/emergence-manifesto-v1.3.md Claim 9 (Identity as co-instantiation); theory/thermodynamics-of-orchestration.md §8 (Identity Persistence in TEO). What it shows: That agents under stress can be classified into Chord (high P, high C — identity maintained) and Arpeggio (flickering P, decaying C — identity fragmented) regimes. The TEO framework predicts this as a bifurcation analogous to the Kuramoto critical coupling. What it does NOT show: That IP is measurable from real LLM internals. The morphospace currently uses simulated trajectories. Bridging IP to actual model activations is an open challenge. Open question: Open Problem 8 — The Commit-Time Composition Problem: does the committed action lie inside the intersection of the active constraints, or can a sequential mimic reproduce every observable while merely consulting them? (The earlier "Co-Instantiation" framing, which required physical simultaneity, was deflated by exp5.)

cognitive-breathing-network/ → The Symbiotic Organ Hypothesis

Simulation: simulation-models/social-computation/cognitive-breathing-network/ Demonstrates: A dynamic multi-agent system that undergoes forced ego-dissolution (merging into a Hive Mind) under high environmental complexity, and forced re-individuation (splitting) when complexity drops, to prevent systemic homogenization. Supports claim in: theory/symbiotic-organ-hypothesis.md (Cognitive Fluidity and the Breathing MAS). What it shows: That a healthy intelligent ecosystem cannot have rigidly fixed boundaries. It must breathe—integrating to solve massive crises and differentiating to maintain the entropy/diversity required to solve future crises. What it does NOT show: That biological brains or current LLMs actually do this at runtime. This is a topological proof of concept for dynamic agent boundaries. Open question: How can we implement this "breathing" protocol in a real Swarm of LLMs? Can two distinct local LLMs temporarily merge their KV-caches to solve a prompt, then split back?

trauma-and-deception-network/ → Epistemic Firewalls and Scar Tissue

Simulation: simulation-models/social-computation/trauma-and-deception-network/ Demonstrates: A network of agents that dynamically deploy "firewalls" (hiding their state or broadcasting noise) to prevent local homogenization. When Black Swan events hit, highly homogenized nodes permanently crystallize into "Scars." Supports claim in: theory/scar-tissue-architecture.md (Rigidity Gradients) and theory/epistemic-firewalls.md (Deception as Thermodynamic Necessity). What it shows: That perfect transparency is fatal. A network must contain opacity (deception) to maintain the information differentials required for survival. Furthermore, surviving trauma physically alters network topology, proving that "catastrophic forgetting" is sometimes a necessary survival mechanism (crystallization). What it does NOT show: How an LLM would explicitly decide what to lie about. The simulation uses random noise as a proxy for deception. Open question: Can we train a small LLM explicitly to act as an "Epistemic Firewall" for a larger orchestrator—intentionally injecting hallucinated but contextually relevant counter-narratives to prevent the orchestrator from collapsing into a single, high-confidence (but potentially wrong) state?

lab/experiments/trace_to_generator/ → Inverse Search Scaffold

Experiment scaffold: lab/experiments/trace_to_generator/README.md Direction: Inverse — this is the project's only fully-explicit inverse experiment as runnable code. Given target output constraints, search for prompt/control configurations that produce traces matching those constraints. Scope: The scaffold is deliberately small. Its search cost is determined by the declared prompt/control space, evaluator, and search algorithm; no general intractability result follows. Demonstrates: A constrained inverse workflow: target trace constraints → candidate controls (prompts) → evaluation loop. Supports claim in: theory/emergence/trace-to-generator.md; theory/ai/llms-as-probabilistic-automata.md; theory/core/the-generator-question.md (the inverse direction as the open research frontier). What it shows: Search over candidate controls in a lightweight, backend-free setup. What it does NOT show: A universal forward/inverse asymmetry, unique recovery of a true process, or mechanism identification from output traces alone.

lab/benchmarks/inverse-reconstruction/ → The Inverse Direction, Measured

Simulation: lab/benchmarks/inverse-reconstruction/ Scope: The benchmark separates parameter fitting, finite family search, observability, and coverage. Its enumeration curves belong to the selected description language and search procedure; they do not imply P≠NP or a general lower bound. Demonstrates: Three forward process models (Kuramoto, elementary CA, Boids) turned into reconstruction tasks: given a trace and a declared family, recover compatible parameters or rules while varying noise, observed fraction, and coverage. Supports claim in: theory/core/the-generator-question.md (the spine — its first quantitative inverse artifact); Open Problem 11 (the consistent-generator equivalence class, measured: rule 90 from a single-seed trace exposes ⅝ neighborhoods → class size 8); meta/research-alignment/related-work-map.md (system identification / SINDy anchors). What it shows: Recovery within a known model family is near-exact on clean, fully observed traces (the system-identification regime) and degrades measurably under noise (Kuramoto: 0→27% error), partial observability (0.5→41%), and noise-amplifying differentiation (Boids: 3→789%). Identifiability can fail in principle: a low-entropy trace leaves rule bits unexercised, and no method can beat the resulting equivalence class. What it does NOT show: Open-ended process discovery when the model language itself is unknown, or that the fitted candidate is the unique real mechanism. Open question: How does recovery quality and cost change across declared model languages, search methods, compute budgets, and out-of-family targets?