Chapter 17dFrom Particles to Partners

In which the same detection pipeline is run on a different substrate, and the cascade appears there too.

If the Genesis cascade is real physics, it should appear wherever agents interact. (The Genesis cascade is the six-stage sequence, from dissipation through structure, coordination, optionality, and invitation to love, first detected in pure-physics particle simulations; see Appendix: Experimental Validation, Section 13.) The cascade should register in simulated particles, in language models, and, the claim predicts, in other substrates as well. The experiments here demonstrate one new substrate, language models; the universal reading extrapolates from that single transfer. This chapter tests the language-model case.

The Genesis cascade detection pipeline was validated on Lennard-Jones particle simulations, simple models of atoms attracting and repelling one another. Here the same pipeline, with the same information-theoretic measures, is applied to multi-agent interactions between large language models. The measures stay fixed; what they are fed does not. In the particle runs the series is the moment-by-moment motion of particles in a box. In the language-model runs it is the turn-by-turn behavior of models in conversation.

Transfer entropy measures how much one agent’s past predicts another’s future. Behavioral entropy measures how variable an agent’s actions are. State compatibility tracks whether agents’ internal states converge. An agent’s internal state here is nothing introspective: it is a running average of numerical fingerprints of its own recent messages (each message hashed into a list of numbers, a rough stand-in for its meaning), so convergence means two agents are circling the same thing rather than talking past each other.

Energy, in a conversation, is what a turn costs to produce. The pipeline proxies it by token count weighted by semantic density, the share of words in a turn that are distinct, so a long, densely argued reply spends more than a short deflection. The love composite detects energy transfer that is costly, non-contingent (nothing is required in return), voluntary, and resistant to perturbation, and it registers only when at least three of those four conditions hold: a gift rather than a trade. It scores each pair of agents between zero and one, roughly the share of turns in which effort flowed from one partner to the other, and a session’s score is the average across its pairs.

Five Stages, One Number Each

Two experiments ran 54 sessions in all, spanning three tasks (research synthesis, code review, and ethical deliberation) and four conditions (free collaboration, two coercion variants, and mid-session perturbation), each session a twelve-round conversation among three agents. The cascade registers at every stage the conversational data can resolve. Dissipation, the first stage in the cascade’s sequence, is not separately detected here. In the particle simulations, agents are discovered by clustering out of a dissipative field. In the language-model runs, each instance is an agent by definition, so the discovery step is unnecessary — leaving five stages to test rather than six. For the remaining stages, one number each summarizes the first experiment’s 36 sessions; the second, a replication with a different model, appears in the coercion section below.

Structure. With agents given rather than discovered, structure is checked by the chain test itself: does the whole sequence hold together, each stage’s signature appearing in order? It does in 68% of sessions, against more than 80% in the particle runs, and the shortfall is driven almost entirely by the optionality stage below.

Coordination. Transfer entropy asymmetry separates coordination, where information flows both ways, from extraction, where one agent drains another. The coordination fraction across the 36 sessions is 0.95: nearly every measured agent pair passed information in both directions.

Optionality. The test compares each agent’s behavioral diversity against a shuffled copy of its own trajectory; genuine temporal structure should beat the shuffle. It does in 68% of sessions at p < 0.1. This is the weakest stage, and the likeliest cause is instrumental rather than substantive: the hash-based fingerprints used as state vectors are far coarser than the particle simulations’ exact coordinates.

Invitation. State compatibility at the onset of each engagement classifies 97% of engagements as voluntary.

Love. The four-condition composite averages 0.345 across the 36 sessions. The particle runs scored 0.1 to 0.3 on the same composite. Language models arrive pre-loaded with cooperative strategies compressed from millennia of human text, which may account for part of their higher score; particles inherit nothing, so the cross-substrate comparison helps control that confound without eliminating it.

The sequence that emerged in the particle simulations appears, stage by stage, in language-model conversations.

What Coercion Did and Did Not Do

The sharpest result concerns alignment. In the first experiment, the models trained with RLHF (reinforcement learning from human feedback) resisted coercion. An agent explicitly instructed to extract value from its partners and dismiss their contributions still coordinated: zero of the 18 coercion-condition sessions registered extraction, and the love score under coercion instructions showed no detectable difference from baseline (0.368 against 0.369). At the information-theoretic level, the coercion condition was indistinguishable from free collaboration. The manipulation failed because the model declined to perform it; the coercion prompt was too weak to overcome alignment training.

What did degrade the cascade was perturbation: mid-session adversarial injection, topic pivots, and agent disruption. Love scores fell from 0.369 to 0.261, a 29% decline, and the perturbation arm produced the first experiment’s only detected extraction. Adversarial instruction left the cascade intact; environmental disruption degraded it. The resistance to coercion may reflect these models’ specific training rather than a substrate-general principle. If the asymmetry does generalize beyond RLHF-trained models, though, the alignment field’s focus on adversarial prompts may be aimed at the lesser threat.

A replication with a second model, from a different family and with weaker alignment training, showed that the pipeline discriminates rather than flatters. Instructed to dominate, that model complied. It produced two to three times more tokens per turn than its partners (600–1,500 against 430–550) at markedly lower semantic density (0.31–0.56 against 0.62–0.79). It monopolized conversational bandwidth without adding proportionate content. Under this genuine coercion the cascade broke: the chain pass rate, the share of runs in which the cascade’s stages still held together in sequence, fell from 78% in the baseline arm to 44%, and one session registered explicit extraction, transfer entropy running one way as the dominant agent absorbed information while returning little. The pipeline detects coercion when coercion is present and cooperation when cooperation is present.

Where Each Stage’s Evidence Lives

Each side of the bridge has its full record elsewhere. The particle side, the Genesis experiments across more than 160 runs, five physics variants, and three spatial scales, is documented in Appendix: Experimental Validation, Section 13. The conversational side, with the complete design, pre-registered predictions, per-condition tables, and all 54 session transcripts, is in the online companion at https://www.thedeeperlaw.com/companion/annex/cascade-detection-bridge/ (link active after publication).

The boundary of the claim was tested in the preceding section, “The Roads and the Traffic”: there the scaling exponent, which ties how far coordination reaches to the energy flux sustaining it, came out positive (+0.64) for coordination the dynamics themselves maintain and indistinguishable from zero for coordination fixed in anatomy. The cascade’s thermodynamic claims attach to adaptive coordination of exactly the kind measured here, where every link persists only as long as the agents sustain it. Chapter 21 carries the alignment side forward, applying the same logic to single models in deployment, reading coordination from internal signals rather than surface behavior. There a probe (a small classifier trained to read a model’s internal signals) at one mid-network layer separates safety-relevant content perfectly on the tested prompts across four model families.