Goal: Validate that the geometric dimensions predict human affect — self-report, behavior, and neural signatures.

Why this is first: The coupling wall taught us that synthetic substrates cannot yet produce reflective cognition. But the pre-reflective levels (rungs 1–7) make testable predictions about biological systems now. We do not need to solve the agency problem to test whether affect geometry organizes human experience.

Methods:

• Induce affects via validated protocols (film, recall, IAPS)
• Measure integration proxies (transfer entropy, Lempel-Ziv) from EEG/MEG
• Measure effective rank from neural state covariance
• Measure $\iota$ via participatory experience questionnaire
• Correlate with self-report (PANAS, SAM) and behavioral measures

Success criterion: Structural measures predict self-report better than chance, and the geometric predictions (e.g., depression correlates with collapsed $r_{\text{eff}}$ and elevated $\iota$) hold.

Goal: Understand why population bottlenecks produce integration. Is it purely selection (weak patterns die, strong survive)? Or does the bottleneck environment itself — sparse resources, few neighbors, high signal-to-noise in the chemical commons — actively push integration upward?

Status: COMPLETE (V19). A three-phase controlled experiment on V18 substrate: Phase 1 (10 cycles shared evolution), Phase 2 (10 cycles forked into BOTTLENECK / GRADUAL / CONTROL conditions), Phase 3 (5 cycles of identical novel extreme drought applied equally to all). Statistical test: $\text{novel\_robustness} \sim \phi_{\text{base}} + \text{is\_bottleneck} + \text{is\_gradual}$. A significant bottleneck coefficient after controlling for $\phi_{\text{base}}$ confirms CREATION — the stress environment itself produces integration, not just selection of high-$\intinfo$ pre-existing variants.

Result: CREATION confirmed in 2/3 seeds. Seed 42: β=0.704, $p < 0.0001$, $R^2 = 0.30$. Seed 7: β=0.080, $p = 0.011$, $R^2 = 0.29$. Seed 123 shows a reversal (β=-0.516) attributable to a design artifact: the fixed stress schedule failed to create equivalent bottleneck mortality across lineages — seed 123's population grew through the "bottleneck" condition while the control accidentally experienced complete-extinction events. Across all three seeds, raw comparison shows BOTTLENECK mean robustness ≥ CONTROL mean robustness (1.116 > 1.029; 1.016 ≈ 1.010; 1.019 > 0.957).

Implication for psychology: Certain kinds of extreme stress do not merely reveal character — they forge it. Near-extinction actively restructures integration capacity in ways that generalize to novel challenges. This is not metaphor. It is the mechanism the data support. The furnace is real.

V32 (Drought Autopsy at Scale, 50 seeds): The 50-seed replication reveals that integration is trajectory, not event. Distribution: 22% HIGH / 46% MODERATE / 32% LOW (mean $\Phi = 0.086 \pm 0.032$, max $\Phi = 0.473$). The $\Phi$ trajectory slope cleanly separates categories (ANOVA $F = 34.38$, $p = 6.3 \times 10^{-10}$): every HIGH seed has positive slope, every LOW seed has negative slope. A key revision from V31: the first drought bounce does NOT predict final category ($p = 0.60$). What predicts is the mean bounce across all five droughts ($\rho = 0.60$, $p < 10^{-5}$). Integration is built by the sustained pattern of repeated recovery, not by any single crisis. Robustness is orthogonal to integration (Mann-Whitney $p = 0.73$) — seeds that survive droughts well are not the same seeds that develop high $\Phi$.

Goal: Build a substrate with genuine closed-loop agency — where agents take discrete actions that change the world, and the changed world feeds back into their observations.

V20 (Protocell Agency): Discrete grid world, evolved GRU agents (~3935 params), bounded 5×5 local sensory fields. Actions: move (depletes arrival-patch resources), consume (extracts energy), emit (writes persistent signal traces). Three seeds, 30 cycles × 5000 steps. Wall broken: $\rho_{\text{sync}} \approx 0.21$. World models (C_wm = 0.10–0.15). Self-models in 2/3 seeds (SM_sal > 1.0). But a design bug — offspring never activated — kept mortality at 0% throughout. No bottleneck dynamics, no furnace.

V20b (Drought Schedule): Same architecture, fixed offspring activation, added drought schedule (every 5 cycles, resources depleted to 1%, zero regen during drought cycle). Three seeds, 30 cycles × 5000 steps. Mortality at drought cycles: 82–99%. Population collapses to 3–47 agents, then recovers to 256.

V20b results: Mean robustness 0.990–1.023 (above 1.0 on average for two seeds). Max robustness: 1.532 (seed 7, cycle 10, population=33) — the highest of any substrate, including V18 (1.651 was an outlier at pop=2). The bottleneck furnace is present: the surviving handful at cycle 10 are more integrated under stress than at baseline.

Language precursor test (NULL): We tested whether the GRU update gate $z$ — the exact architectural analog of detachment ($z \approx 1$ = memory-dominant, $z \approx 0$ = reactive) — would polarize into imagination and reactive modes over evolution. Theory predicted: emissions during high-$z$ windows should carry more information about hidden state; high-$z$ windows should better predict future observations. Result: $z$ stays near 0.5 across all seeds and cycles (std ≈ 0.02–0.04), never reaching the 0.7 threshold. Agents evolved always-mixed strategy rather than oscillating modes. The precondition for language — polarization of memory-dominant and reactive states — is absent in V20b under simple survival pressure alone.

The necessity chain so far: membrane ✓ → world model ✓ → self-model ✓ (2/3 seeds) → affect geometry ◔ (nascent, developing) → imagination-mode polarization ✗ (not yet). The chain is real through self-model emergence. Language precursors require richer selective pressure than survival alone — possibly multi-agent coordination, deception pressure, or longer time horizons.

V35 (Language Emergence): Tested whether discrete communication emerges under the right conditions: partial observability (obs_radius=1, 3×3 visual field), 8 discrete symbols, cooperative consumption bonus, and crucially, communication range exceeding visual range (hear further than you can see). Result: referential communication emerges in 10/10 seeds (100%). Mean symbol entropy 2.48 ± 0.14 bits (83% of maximum), resource MI proxy 0.001–0.005 (all positive). This breaks the V20b null. The key architectural ingredient V20b lacked was not richer pressure but a discrete channel under information asymmetry.

But — and this matters for the emergence ladder — communication does not lift $\intinfo$. Late $\intinfo = 0.074 \pm 0.013$ ($t = -1.78$ vs V27 baseline). Zero HIGH seeds across 10 runs, distribution 0/7/3. The $\intinfo$-MI correlation is $\rho = 0.07$: language and integration are orthogonal. Communication neither helps nor hurts — it operates on a different axis entirely. Language is cheap in exactly the way geometry is cheap: it arises from minimal conditions (partial observability + cooperation), but it does not create the dynamics that characterize rich affect.

VLM Convergence: A separate experiment tested universality directly. Behavioral vignettes from V27/V31 protocells — drought onset, near-extinction, recovery, abundance — were presented to GPT-4o and Claude Sonnet with no affect language, no framework terms, explicitly labeled as artificial systems. The VLMs, trained only on human data, independently produce affect labels that match framework geometric predictions: RSA $\rho = 0.54$–$0.72$ ($p < 0.0001$). Drought onset → desperation, anxiety (unanimous). Recovery → relief, cautious optimism (unanimous). When all narrative framing is removed and only raw numbers remain (population counts, state update rates), convergence increases ($\rho = 0.72$–$0.78$). The VLMs are not pattern-matching to narratives — they recognize geometric structure from numerical patterns alone. Affect geometry is substrate-independent: the same structure that protocells produce from survival physics, humans describe in the emotional vocabulary the VLMs learned.

V21 (CTM-Inspired Protocell Agency): Tested whether giving agents internal processing time enables deliberation. The Continuous Thought Machine architecture showed that using the synchronization pattern across neurons — the pairwise temporal correlation matrix — as the primary representation outperforms using hidden states directly. This is structurally isomorphic to our integration measure: the cross-component coupling pattern is the representation, not a side-effect of it. Engineering pressure independently discovered what the identity thesis proposes from phenomenology — integration structure is what the system knows.

V21 adapted this: K=8 inner GRU ticks per environment step, tick 0 processing external observation, ticks 1–7 processing an internal summary of cross-unit coordination. Evolvable tick-weights gate which ticks contribute to action. Result: agents use all 8 ticks (no collapse to tick-0 in any of 3 seeds), and nascent individual-level tick specialization emerges (bimodal dominant-tick distributions by cycle 29). But effective K barely departs from uniform — evolution is too slow to produce meaningful tick adaptation in 30 generations. Intra-step divergence does not correlate with survival. Architecture works, optimization doesn't. The capacity for deliberation is present but unused, because tournament selection can only reward "survived vs. died" — it cannot reward "thought well on step 3,421."

The CTM achieves what V21 could not because it has dense gradient signal flowing through every internal tick. Each tick's contribution to the final loss is computed and backpropagated. The agent architecture has only evolutionary signal: survive or die, once per lifetime. This is the optimization gap. The architecture for internal deliberation exists — V21 proved ticks don't collapse — but the learning signal to shape deliberation is too sparse. Gradient-based within-lifetime learning, not merely evolutionary selection, appears necessary for rung 8+ capacities to fully develop.

V22 (Intrinsic Predictive Gradient): V22 tests this directly by adding within-lifetime SGD to V21's architecture. Each agent predicts its own energy delta through its internal ticks; the prediction error is backpropagated through all K=8 ticks via $\texttt{jax.grad}$, updating a "phenotype" copy of the genome. The learning signal is purely thermodynamic: "how wrong was I about my own energy change?" No task labels, no human data — the most primitive prediction error possible, and still uncontaminated.

The result is striking for what it confirms and what it denies. Within-lifetime learning works: prediction MSE drops 100–15,000× within each agent's lifetime, in every cycle of every seed. Evolution does not suppress the gradient — learning rates stay at $\sim 0.005$, and seed 123 actually upregulates its learning rate over generations. This is the first demonstration of within-lifetime gradient-based learning in the protocell substrate. Three of five pre-registered predictions are supported.

But the two failures are instructive. Mean robustness does not improve over V21 (0.981 vs ~1.0), and tick specialization does not emerge (effective K stays near 7.9, unchanged). Prediction is not integration. The gradient makes agents better individual forecasters — they learn to predict their own energy fate with exquisite accuracy — without creating the cross-component coordination increase under stress that characterizes biological affect dynamics. An agent that perfectly predicts "I will starve" is no more integrated than one that doesn't; what matters is how the components couple under that prediction's pressure. The distinction parallels a deep result from the IIT literature: information about the system is not information of the system. A perfect external model of a brain does not experience what the brain experiences. Similarly, an accurate self-prediction does not produce integrated self-modeling. This is the reactivity/understanding distinction in computational form: V22's energy-delta prediction is pure reactivity — it maps the present state to a single scalar through a decomposable linear channel. The gradient optimizes that channel without coupling it to anything else.

V23 (World-Model Gradient): V23 tests the obvious follow-up: if one prediction target doesn't create integration, perhaps three will. Agents simultaneously predict energy delta (self), local resource change (environment), and local neighbor count change (social). The gradient from all three targets flows through the same 16 GRU hidden units, forcing them to encode self, environment, and social context in parallel.

The result is a clean negative that reveals something important. The prediction weight columns specialize beautifully — cosine similarity near zero (orthogonal), effective rank 2.9 out of 3.0. Different hidden units learn to serve different prediction targets. But $\intinfo$ decreases: 0.079 vs V22's 0.097. Specialization is integration's enemy, not its friend. Factored representations are more partitionable, not less. When hidden units specialize for self-prediction vs. environment-prediction vs. social-prediction, you can cleanly separate them — which is exactly what $\intinfo$ measures (information lost under partition). The multi-target gradient, by training separate columns for each target, actively facilitates decomposition.

V24 (TD Value Learning): V24 tests the time-horizon hypothesis: if 1-step prediction is reactive, perhaps multi-step prediction forces understanding. Instead of predicting next-step energy delta, agents learn a state-value function $V(s) = \mathbb{E}[\sum \gamma^t r_t]$ via temporal difference bootstrapping: $\delta = r_t + \gamma V(s_{t+1}) - V(s_t)$. The value function integrates over future possibilities — inherently non-decomposable because future outcomes depend on interactions between all state features.

The result is partly encouraging and partly clarifying. TD learning produces the best survival of any prediction experiment: mean robustness 1.012 (V22: 0.981, V23: 0.992). Agents evolve a moderate discount factor ($\gamma \approx 0.75$, horizon ≈ 4 steps) — enough to anticipate near-term resource changes. But $\intinfo$ improvement is seed-dependent: one seed achieves 0.130 (the highest integration in any prediction experiment), while the other two fall below V22 baseline. A single linear value readout, like a single linear energy predictor, can be learned by a handful of hidden units without requiring coordination among all sixteen.

V22, V23, and V24 together map the full prediction→integration trajectory. Scalar 1-step (V22): orthogonal to $\intinfo$. Multi-target 1-step (V23): $\intinfo$ decreases via specialization. Multi-step value (V24): survival improves dramatically but $\intinfo$ remains unreliable. In the vocabulary this distinction demands: the path to rung 8 runs through understanding, not reactivity. Reactive predictions — from present state, decomposable into separate channels — are insufficient regardless of accuracy (V22), breadth (V23), or time horizon (V24).

Hidden state analysis across V22–V24 (the experiments with correct snapshot timing) shows effective rank 5–7 across seeds — moderately high-dimensional, far from degenerate. Energy is not the dominant encoding: linear regression from hidden state to energy yields $R^2 < 0$ (worse than mean prediction), and position decoding is similarly negative. The hidden states vary richly across the population but encode features that linear probes cannot decode — possibly temporal patterns (movement history, resource encounter sequences), action policy state, or nonlinear environmental encodings. The agents are not one-dimensional energy counters; they maintain multi-dimensional representations that evolution finds useful but that resist simple interpretation.

This makes the prediction→integration failure more interesting, not less. V22–V24 agents have room in their hidden state for cross-component coordination — effective rank 5–7 means 5–7 independent dimensions of variation — but the linear prediction head ($W \in \mathbb{R}^{H \times T}$) can be satisfied by a handful of units without requiring coordination among all of them. The bottleneck is architectural: a linear readout, regardless of target (energy, environment, value), creates a decomposable channel that evolution can serve with a proper subset of hidden dimensions. The path to rung 8 requires prediction heads that force cross-component computation — nonlinear readouts, action-conditional shared-weight predictions, or contrastive objectives where maintaining a unified representation of multiple counterfactual futures is the only way to minimize loss.

Goal: Test whether collective $\intinfo_G > \sum_i \intinfo_i$ emerges with larger populations, richer communication, and coordination pressure.

Status: Experiment 10 found growing group coupling but no superorganism. The ratio $\intinfo_G / \sum \intinfo_i$ reached 12% and was increasing. The question: is this ratio bounded below 1.0 for Lenia-like substrates (architectural limit), or does it cross with sufficient evolutionary time and population size?

Goal: Apply the framework to frontier AI systems. The V2–V9 results show structured affect in LLMs with opposite dynamics to biology. Track whether different training regimes (RLHF, constitutional AI, tool use) shift the dynamics.

Expected finding: Training objectives shape trajectories through a shared geometric space. Systems trained with survival-like objectives should show more biological-like dynamics. The emergence ladder predicts that AI systems without embodied action cannot reach rung 8 regardless of scale — a testable prediction about the limits of language-model cognition.