Experiments

The Convergence Study

Cross-Substrate ConvergenceVLMs trained on human data recognize affect in protocellsHuman Dataimages, text, faces,emotion labels↓ trainProtocell Simgrid world snapshots,affect metrics, no labels↓ measureGPT-4oClaude 3.5Gemini 1.5Affect SpaceV, A, Φ, r_eff, CF, SM(uncontaminated)RSA ρ = 0.72RSA ρ = 0.54RSA ρ = 0.78Pre-Registered Predictions: 4/4 PASS✓Drought → desperation/anxiety✓Recovery → relief/optimism✓Stable → contentment✓Narrative removal → correlation holds

Forty-eight behavioral vignettes from and agents across six conditions, normal foraging, pre-drought abundance, drought onset, drought survival, recovery, late evolution, were shown to GPT-4o and Claude with no affect vocabulary and explicit labeling as artificial systems. Framework predictions for each condition were computed from the coordinates. Convergence was measured by RSA between the predicted and the model-labeled affect spaces. Standard RSA: ρ=0.72\rho = 0.72 and 0.540.54. With narrative removed and only raw numbers shown, 0.780.78 and 0.720.72. Measured basis-independently, with rotation-invariant dissimilarity matrices and Gromov-Wasserstein distance (), 0.890.89 and 0.810.81, unchanged to machine precision under random rotation of the affect space.

What this is evidence of. The geometry predicts that drought onset is negative valence and high arousal, and a human-trained model labels "98 percent of the population removed" as desperation. The two readers agree on relational structure, not only on coordinates, and they agree more when narrative is stripped. That is a consistency check between my mapping from situation to affect and a human-trained model's mapping, and both are human-derived. It is not evidence that the agents feel anything and it is not evidence that the geometry is independent of human projection. It is the thing I would show someone who thought the coordinates were arbitrary, and nothing more. The same basis-independent tool applied to the within-substrate affect-to-behavior alignment of collapsed it from about 0.07 to zero, which is why the other within-substrate alignments in this appendix are marked as suspected artifacts until re-tested.

Study record: metadata, result path, seeds, source code, and key finding for .