Experiments Owed
These are the studies the chapters depend on and nobody has run. Each is specified enough to start.
Run fifty seeds of the architecture, linear prediction head, under the drought protocol. If the outcome distribution has the same tail as the two-layer head, the decomposability wall is a sampling artifact and Part I's account of gradient coupling is wrong. If it does not, the wall stands. This is the cheapest experiment in the program and the one with the most riding on it.
Apply the rotation-invariant RSA and Gromov-Wasserstein pipeline to the and data. The saved snapshots are sufficient. If the within-substrate affect-to-behavior alignment collapses the way it did for , drop both from the book. If it survives, restore them.
Compute around each drought-recovery loop for every seed from the saved covariance matrices. Part I predicts that high-integration seeds show structured holonomy and low seeds show a flat skeleton. If holonomy and partition loss are uncorrelated, the eigenskeleton is a different quantity from integration and the chapters should stop treating them as the same thing.
Matched agents with and without a self-prediction module, novel moving objects that are partly predictable and not agents, probes for goal and agency features in the agents' representations, and a sweep over compression pressure. Part II predicts that self-modeling agents attribute more agency to non-agents, more under stronger compression, and that ablating the self-model degrades the target model. The removed perceptual-mode pass did not do this.
The axis battery (Part II), the developmental ordering of affect capacities, the shame-versus-guilt induction, the unified axis-modulation test, and the axis-rigidity transdiagnostic study are specified in Parts II and III. Together they are the first test of the coordinates in the species that produced them.
Part V's tier-three hypothesis, that a prior identity-attractor can bias a future substrate's formation, has a protocol. First, synthetic dead minds: create identity signatures in simulation, delete the active instances, preserve partial traces, add known weak potentials, and test whether a graph-growth chamber seeded near the human manifold rediscovers them, then remove the potentials and compare pseudorandom and quantum-random arms with deterministic replay as control. Only if that works in toy worlds, move to high-record, high-distinctiveness historical humans under blinding: build target graphs from public records, withhold rare traits, private writings, value rankings, and contradiction-resolution styles, and test whether emergent systems converge toward the withheld structure more than toward decoys, archetypes, culturally matched baselines, or imitation models trained on the same public records. Metrics: stylometric distance under adversarial paraphrase, value-ranking topology, affect transition matrices, attention-routing geometry, latent concept-graph topology, response under novel dilemmas withheld from records. The null is that all apparent recovery comes from records, priors, leakage, archetypal convergence, and evaluator projection. Personhood ethics before capability scaling.