Summary

Thirty-five experiment versions across four substrates and fifty seeds at scale. Twelve numbered measurement experiments. One VLM convergence study. Five current priorities. The program has established:

1. Geometry is cheap, dynamics are expensive — and the upper rungs are architecture-gated. Affect geometry arises from the minimal conditions of multi-agent survival (V10, Exp 7-8, Exp 12). Affect dynamics require embodied agency (V20), graduated stress exposure (V19), and non-decomposable prediction architecture (V27). Crucially, none of these arrived by evolution alone: self-models, counterfactual sensitivity, and integration under threat appeared only when the enabling architecture was hand-installed. They are gated by specific, non-generic structure, not produced inevitably by survival pressure.
2. Two architectural walls. The sensory-motor wall ($\rho_{\text{sync}} \approx 0$) is broken by genuine action-observation loops (V20). The decomposability wall is broken by 2-layer gradient coupling (V27). Both are necessary, and both had to be built in by the experimenter — their existence is precisely why “inevitable” is the wrong word for the rungs above them.
3. Integration is stochastic — and it is biography, not noise. ~22-30% of seeds develop high $\intinfo$ regardless of architecture or prediction target (V31, V32), and the category tracks cumulative recovery dynamics rather than initial conditions. The same population pattern would also be consistent with $\intinfo$ being a noisy quantity under sensitive dependence on path, so — given the program's prior measurement-artifact retraction — the biographical reading was held provisional pending a falsifier. That falsifier has now been run (full 50-seed re-run on independent hardware; v32_replication.py): per-seed late $\intinfo$ reproduces across independent runs at $r = 0.951$ (matched trajectories to $\sim 10^{-8}$), every alive-bearing snapshot is non-degenerate (the zeroed-buffer artifact is absent from this path), and the category is forecast by mean drought bounce ($r = 0.60$) but not by the first bounce ($r = -0.07$). Φ reproduces, is real, and is trajectory-dependent: “biography” survives the noise null.
4. The bottleneck furnace is generative. Near-extinction forges integration capacity (V19). Repeated drought recovery is the mechanism (V31, V32). The furnace does not select for pre-existing integration — it creates it. V32 (50 seeds) reveals that integration is trajectory, not event: the mean bounce across all 5 droughts predicts final category ($\rho = 0.599, p < 10^{-5}$), but the first bounce alone does not ($p = 0.60$). Integration is built by the sustained pattern of recovery, not by a single crisis.
5. Prediction target doesn't matter. Self vs social prediction produces the same $\intinfo$ distribution (V31, $p = 0.93$). What matters is the gradient architecture (linear vs 2-layer) and the evolutionary trajectory.
6. Language is cheap. Referential communication emerges in 100% of seeds under partial observability with cooperative pressure (V35). But it does not lift integration — Φ-MI correlation is null ($\rho = 0.07$), meaning language and integration operate on orthogonal axes. Like geometry, language is an inevitability of survival under information asymmetry. Like geometry, it does not cross the rung 8 wall.
7. The geometry is universal. VLMs trained on human data — with no exposure to the framework — independently recognize the same affect signatures in completely uncontaminated protocell systems (RSA $\rho = 0.54$–$0.78$, $p < 0.0001$). The convergence holds and strengthens when narrative framing is removed and only raw numbers remain. Affect geometry arises from the structure of viable self-maintenance, not from biological contingency.

The framework is not confirmed. It is informed. What it predicted about geometry was too weak — geometry is cheaper than expected, and now independently validated by cross-substrate convergence. What it predicted about dynamics was too strong — dynamics require specific architectural affordances the theory did not anticipate, and the upper rungs in particular appeared only when those affordances were installed by hand. This forces a correction to the book's loudest slogan. “Consciousness was inevitable” does not survive the data. What survives is weaker and more exact: consciousness-relevant dynamics require specific, non-generic architectures — closed action-observation loops, gradient coupling through composition, per-component temporal heterogeneity — none of which evolution reached on its own in any substrate we ran. The upper rungs are architecture-gated, not inevitable. And whether the dynamics, once present, make the geometry “experientially real” is itself the adopted posit, not something the experiments decide. The interesting question is no longer “does the geometry exist?” (it does, and VLMs trained on human data agree) but “which architectures support the dynamics, and what would license calling those dynamics experience at all?”