What Survived
Every claim in this book is tested, testable, or labeled as a guess. This appendix is the record: 43 study records across language models, multi-agent reinforcement learning, Lenia cellular automata, protocell agents, a thirteen-part measurement program, and a cross-substrate convergence test. It is not a trophy wall. Most of what is here is a null, a retraction, or a result smaller than I first said it was, and the point of keeping it is the dependency structure: which claims survived, which failed, and which walls forced the next experiment.
One result stands on its own. In fifty seeds of protocell agents with a two-layer prediction head (), the integration a population reaches is set by how decisively it re-integrates after each of five near-extinction droughts, and not by its starting genome, its first drought, or its prediction target. The per-seed outcomes reproduce across independent hardware at . That is the empirical content of "the dynamics are biographical," and it is the finding the chapters lean on.
One architectural fact is true by construction and still worth stating. Lenia patterns, however elaborate, never close an action-environment-observation loop; their self-effect ratio sits near zero in every variant (). Agents that act on a grid close it from initialization (). Counterfactual weight, self-models, and normativity need the loop, and nothing on the near side of it produced them.
One class of nulls is clean and useful. Nothing I tried raises mean integration: not forcing functions (), not prediction on any target (, ), not environmental complexity, not a discrete language channel (), not direct selection for integration (), not a contrastive loss (). The one intervention that produced high-integration seeds at all, gradient coupling through a two-layer head (), widened the tail of the distribution without moving its mean, and has not yet been compared against a linear head at the same sample size.
And the claim that affect geometry is universal rests on less than I used to say. The in-substrate alignments between a system's affect coordinates and its own behavior (, ) have the same form as an alignment that collapsed to zero when measured without reference to a coordinate basis (), and they have not been re-tested. What remains is the convergence study (): human-trained models label the behavior of protocell agents the way the geometry predicts, at basis-independently. Both readers are human-trained, so it is a consistency check between my mapping and theirs, which is reassuring and not proof of anything about the agents.
| Study | Status | What it shows |
|---|---|---|
| , with | Stands | Integration is trajectory-dependent and reproducible; mean drought bounce predicts late at 0.60, first bounce at zero. |
| Stands, by construction | Action-observation loops close from initialization. Its hidden-state numbers were retracted (zeroed buffers). | |
| , | Restated | Two-layer head widens the tail; mean unchanged; no 50-seed linear control. |
| Restated | Bottleneck forges generalization in two of three seeds; the third seed's design failed. | |
| Restated | All seeds use all symbols at high entropy; resource mutual information of a few thousandths; no integration lift. | |
| Restated | Consistency between framework predictions and human-trained labels, basis-independent. It is not universality. | |
| , , | Nulls and engineering | No forcing function needed for geometry; no Lenia variant closes the loop; mean robustness never clears 1.0. |
| , , | Clean nulls | No compositional language, no counterfactual detachment, no self-model, no normativity on the near side of the loop; contrastive loss and direct selection do not help. |
| , alignment | Suspected artifact | Within-substrate affect-to-behavior RSA; same form as the alignment that collapsed under rotation-invariant re-test. It has not been re-tested. |
Nineteen records were removed from the catalog in November 2026 and are not cited anywhere in this book: the language-model prompting studies, an early reinforcement-learning derivation, the attention-Lenia, Hebbian-plasticity, and quorum-signaling substrates, the protocell-plus-drought and inner-ticks variants, the predator-prey and partially observable environments, the three-seed social-target and dual-target runs, and the substrate-engineering, existence, world-model, representation, perceptual-mode, social-integration, entanglement, and capstone measurement passes. Their code and results are in git history before commit 111a046, and the identifier map is in experiments/ID_MAP.md.