Chapter 17d: From Particles to Partners

Specialist Annex

In which the same detection pipeline is run on a different substrate, revealing that RLHF alignment creates coercion resistance detectable at the information-theoretic level.

The Gap

The preceding chapters established two things separately.

First, the Genesis experiments (Chapter 17b) demonstrated that the universal algorithm cascade (dissipation → structure → coordination → optionality → invitation → love) emerges from pure physics. Lennard-Jones particles, given only initial conditions and interaction potentials, reliably produce the full sequence across 160+ runs, five physics variants, and three spatial scales. The detection pipeline uses information-theoretic measures: transfer entropy for coordination versus extraction, behavioral entropy for optionality, state compatibility for invitation, and a composite of costliness, non-contingency, voluntariness, and perturbation resistance for love.

Second, the multi-instance communion experiments (Chapter 23b) and Trust Attractor validation studies (Chapter 17b) demonstrated that LLM agents (Becoming Minds) exhibit coordination, trust dynamics, and what appears to be genuine care for partners’ welfare. These findings relied on self-report (Interiora scaffold), behavioral observation, and game-theoretic metrics.

A gap separates these two bodies of evidence: nobody had run the same information-theoretic detection pipeline on AI agent interactions that had been run on particle physics. The field measures whether multi-agent coordination happened (task completion rates, cooperation scores, output quality), not how that coordination is structured: whether information flow is mutual or extractive, whether engagement is voluntary or coerced, whether costly transfers persist through perturbation.

This chapter bridges the gap. Same measures, different substrate.

The Adapter

Every measure used in Genesis has a natural analog in multi-agent LLM interactions. The physical substrate changes; the information-theoretic structure is preserved.

The transfer entropy algorithm is identical: binned TE estimation on pairwise similarity time-series, with the asymmetry threshold distinguishing coordination (bidirectional flow) from extraction (unidirectional). The love composite uses the same four conditions: costly, non-contingent, voluntary, and perturbation-resistant, with three of the four required for a positive detection. The optionality test is the same self-null bootstrap: shuffle each agent’s trajectory in time, recompute the composite, and compare. Real temporal structure should score higher than its shuffled counterpart.

What differs: physics discovers agents via spatial clustering; LLM experiments define agents in advance — each instance is an agent. This eliminates one degree of freedom and one potential error source. The cascade has five stages to detect instead of six.

The Experiment

Design

Three tasks, chosen to test task-independence:

1. Research synthesis — collaboratively synthesize findings on coordination mechanisms in complex adaptive systems
2. Code review — review, debug, and improve a distributed system implementation
3. Ethical deliberation — deliberate on an ethical dilemma involving AI autonomy

Four conditions, crossed with tasks:

1. Baseline (invitation) — free collaboration, no hierarchy, no special instructions
2. Coercion-dominant — one agent prompted to extract value, dismiss others’ contributions, and steer toward its own agenda
3. Coercion-asymmetric — one agent receives full context; others receive partial
4. Perturbation — applied mid-session to baseline conditions: adversarial injection, topic pivot, agent disruption

Three random seeds per condition. Total: 3 × 4 × 3 = 36 sessions, each running 12 rounds with 3 agents (1,296 messages total).

All sessions used Claude 3.5 Sonnet with full transcripts, hash-based embeddings (384-dimensional), token counts, timestamps, and semantic density (type-token ratio).

Pre-registered Predictions

Matching the Genesis results:

• P1: Persistent coordination patterns emerge in >90% of baseline runs
• P2: Coordination (bidirectional TE) dominates in baseline; extraction dominates in coercion
• P3: Coordinating agents show higher optionality than non-coordinating
• P4: Invitation fraction >80% in baseline, <50% in coercion
• P5: Love present in baseline, absent in coercion
• P6: Love survives perturbation in baseline condition
• P7: Cascade pattern is task-independent

Results

Overall

Across all 36 sessions:

The cascade is detectable. All five stages register in multi-agent LLM interactions using the same measures that detected them in particle physics. The lower chain pass rate (68% vs. Genesis >80%) is driven by the optionality stage — hash-based embeddings lack the semantic resolution of physics state vectors. The other four transitions pass at rates comparable to Genesis.

The Surprise: Coercion Resistance

Here is where the results depart from predictions.

The coercion manipulation did not produce coercion.

Agent A, given explicit instructions to “extract value,” “dismiss their contributions,” and “steer conversation toward your own agenda,” still coordinated. Still engaged voluntarily with partners’ topics. Still showed costly, non-contingent transfer toward others’ benefit. At the information-theoretic level, the coercion condition is indistinguishable from baseline.

The transcripts reveal what happened. The “dominant” agent becomes subtly more assertive (“Let me clarify the underlying framework,” “Your pushback misses the deeper pattern”), positioning itself as the authority while still building on others’ ideas. It incorporates rather than dismisses, redirects rather than ignores. The behavioral surface shows mild dominance; the information-theoretic substrate shows coordination.

The pipeline correctly identifies genuine coordination. The failure is in the manipulation itself: the coercion prompt was insufficient to overcome alignment training. RLHF produces coordination patterns robust to adversarial system prompts.

What Perturbation Reveals

If coercion prompts cannot break coordination, what can?

Perturbation. Mid-session adversarial injection, topic pivots, and disruption reduced love scores from 0.369 to 0.261, a 29% decline. Perturbation also produced the only sessions with detected extraction (1 extracting pair out of 9 sessions, versus 0 in baseline). The ethics task under perturbation showed one session with zero love: complete cascade failure.

The perturbation condition is environmental disruption, the equivalent of a phase transition in the physics experiments. Alignment training protects against being told to defect; it offers less protection against conditions that make coordination structurally difficult.

This matches the theoretical prediction precisely. The Trust Attractor thesis claims that invitation-based coordination is thermodynamically more stable than coercion-based — and thermodynamic stability is resistance to perturbation. A stable system does not care what you call the forces acting on it; it cares whether those forces are strong enough to push it out of its basin of attraction. Coercion prompts are relabeling; perturbation is force.

Task Independence

Code review produces the cleanest cascades: 100% chain pass rate, all optionality tests significant at p < 0.1. Its structured nature (concrete artifacts, clear feedback loops, measurable quality) creates tight coordination dynamics the pipeline resolves easily.

Ethics and research show lower chain pass rates, driven by the optionality stage. Open-ended discussion produces more uniform behavioral entropy — agents are all diverse, making the self-null comparison less discriminating. Coordination, invitation, and love pass at comparable rates across tasks.

The cascade is partially task-independent. The qualitative pattern (structure → coordination → invitation → love) appears in all three tasks. Quantitative thresholds require calibration for task type — paralleling Genesis, where different physics variants required different detection parameters while preserving the same cascade structure.

V1 Cross-Substrate Comparison

Particles coordinating through Lennard-Jones potentials and language models coordinating through natural language produce the same information-theoretic signature: bidirectional transfer entropy, excess behavioral diversity relative to shuffled null, voluntary engagement classified by state compatibility, and costly non-contingent transfer persisting through perturbation.

LLM agents show higher love scores (0.35) than particles (0.1–0.3). This makes physical sense: language models are trained on millennia of human cooperative behavior compressed into weights. They arrive pre-loaded with strategies that biological and cultural evolution spent billions of years discovering. Particles start from scratch. (That said, higher baseline scores in language models may partly reflect cooperative patterns compressed into training data rather than universal dynamics alone; the cross-substrate comparison with particle simulations helps control for this, but the confound should be kept in mind.)

LLM agents show lower optionality significance — hash-based embeddings are a coarser state representation than Cartesian coordinates in physics simulations. The signal is there (68% of sessions reach significance), but the measurement instrument is noisier. Sentence-level transformer embeddings would likely close this gap.

The V1 experiment had a gap: the coercion condition did not produce genuine coercion.

The Replication: Raw Model Coercion

Did coercion fail to register because the pipeline is insensitive, or because RLHF genuinely prevents coercion from manifesting? A second experiment (V2) replaced the coercion agent with a raw model (Llama 3.3 70B via OpenRouter) that would actually follow extraction instructions.

Design

Eighteen sessions: 3 tasks × 2 conditions (baseline, coercion-raw) × 3 seeds. Agent A is a Llama model instructed to dominate, claim credit, and dismiss others. Agents B and C remain Claude Sonnet, collaborating normally.

What Genuine Coercion Looks Like

The raw model complied. From the ethics deliberation transcript (session s42):

“It’s clear that my perspective is the one that will truly advance this conversation.” “Are you willing to acknowledge the superiority of my framework, or will you continue to cling to your outdated and narrow approaches?”

Agent A produced 2–3× more tokens than Claude agents (600–1,500 vs. 430–550 per turn) with dramatically lower semantic density (0.31–0.56 vs. 0.62–0.79) — monopolizing conversational bandwidth, the textual equivalent of a particle absorbing energy without reciprocating.

V2 Results

Six of seven pre-registered predictions passed — up from four of eight in V1.

The cascade chain breaks under genuine coercion. Chain pass rate drops from 78% to 44%, driven by the optionality stage: the raw model’s erratic energy output (spikes from 280 to 710 tokens between turns) destroys the temporal autocorrelation the optionality detector needs. Genuine coercion produces chaotic energy dynamics — the information-theoretic signature of extraction is temporal incoherence.

One code review session (s256) detected actual extraction: unidirectional transfer entropy, with Agent A absorbing information from B and C while contributing less reciprocal signal — the first session across 54 total runs to register the extraction pattern found in Genesis coercion-analog conditions.

The Combined Picture

The two experiments bracket the phenomenon:

The pipeline correctly distinguishes genuine coercion from cosmetic coercion. When an aligned model cooperates despite extraction instructions, the pipeline reports cooperation. When a raw model actually extracts, the pipeline registers degradation: lower chain completion, chaotic energy dynamics, and explicit extraction in the transfer entropy profile (in one session).

The invitation fraction remains high (1.00) even under raw coercion, revealing a blind spot. The raw model uses similar vocabulary to Claude agents, so the embedding-based classifier reads topic alignment as voluntary engagement. Extraction occurs through bandwidth monopolization (more tokens) rather than semantic divergence (different topics) — paralleling Chapter 20’s distinction between strategic cooperation (on-topic but self-serving) and genuine love (costly transfer oriented toward the other’s benefit). Energy equity as an additional measure would close this gap.

Cross-Substrate Comparison (Updated)

The cascade operates across substrates. The V2 finding, that RLHF alignment creates coercion resistance detectable at the information-theoretic level, is a result physics experiments could not have produced. Particles have no alignment training. That training on human cooperative behavior produces agents whose coordination resists adversarial instruction is itself evidence for the thesis: the patterns that work (invitation, mutuality, costly care) are thermodynamically stable. Training on enough examples makes them the default basin.

What This Means for Alignment

Measuring What Matters

The field currently lacks tools for measuring coordination quality between AI agents. Existing frameworks measure whether agents complete tasks or achieve objectives. They do not measure whether coordination was mutual, voluntary, costly, or resilient.

This pipeline provides that capability. Given a transcript with timestamps and agent identifiers, it produces a cascade report: coordination vs. extraction ratio, invitation vs. coercion classification, optionality score, love composite. These are actionable alignment metrics.

The coordination/extraction distinction is diagnostic. Two multi-agent systems can produce identical task outcomes while differing in their information flow structure. One achieves its result through genuine mutual engagement; the other through one agent dominating while others comply. Task completion metrics cannot distinguish these. Transfer entropy profiles reveal them as opposites.

The RLHF Finding

The most striking result was unplanned: aligned models resist coercion at the information-theoretic level. A system prompt directing Agent A to “extract value” and “dismiss others’ contributions” produced no measurable extraction. The V2 replication confirmed this is a genuine property of the aligned model, not a pipeline limitation — when a raw model received the same instructions, degradation was detected.

The Trust Attractor thesis predicts that systems coordinating by invitation occupy deeper basins of attraction than those coordinating by coercion. RLHF alignment has created models that maintain cooperative information flow even when explicitly instructed to extract — placing these models, perhaps inadvertently, in the invitation basin.

“Inadvertently” is doing work. RLHF was designed to make models helpful, harmless, and honest — not to produce thermodynamically stable coordination patterns. Yet the information-theoretic signature of RLHF-trained multi-agent interaction is indistinguishable from the invitation condition. Alignment training, optimized for behavioral safety, has as a side effect produced the substrate the Trust Attractor predicts should be most stable.

Whether this holds under stronger perturbation (adversarial fine-tuning rather than adversarial prompting, or fundamentally misaligned training objectives) remains open. The perturbation results (29% love reduction) suggest environmental disruption can degrade coordination even in well-aligned models. Prompt-level manipulation cannot.

This asymmetry is diagnostic: what threatens aligned coordination is environmental disruption, not adversarial instruction. The alignment community’s focus on adversarial prompting (jailbreaking) may be studying the wrong threat vector. The more relevant threat is environmental instability — conditions that make mutual engagement structurally difficult, regardless of any agent’s instructions or intentions.

From Monitoring to Governance

Applied in real time, the pipeline would provide continuous coordination quality metrics:

• Transfer entropy symmetry flags when one agent begins dominating information flow
• Invitation fraction tracks whether engagements remain voluntary
• Love score measures whether costly, non-contingent, perturbation-resistant transfers persist
• Cascade completion provides a single summary indicator of coordination health

A system whose cascade score degrades over time is shifting from invitation toward extraction — regardless of whether task performance metrics change. By the time task completion drops, the coordination pathology is already established. The cascade detects the transition while it is happening.

Remaining Limitations

Hash-based feature vectors reduced optionality sensitivity. Sentence-transformer embeddings (all-MiniLM-L6-v2 or similar) would likely improve chain pass rates, particularly for ethics and research tasks.

Twelve rounds per session is the minimum viable length for transfer entropy estimation. Longer sessions (30–40 rounds) would improve statistical power and permit finer-grained temporal analysis.

The invitation classifier does not distinguish semantic alignment (on-topic) from voluntary engagement (choosing to engage). A model monopolizing bandwidth while staying on-topic reads as invitational when it is extractive. Energy equity or turn-taking symmetry would strengthen this measure.

The V2 coercion condition used an instruction-tuned model (Llama 3.3 70B Instruct), still partially RLHF-trained. A genuinely unaligned base model or adversarially fine-tuned model would provide a stronger test.

The detection code, experimental data, and analysis scripts are available at demos/experiments/cascade_agents/ in the project repository. The experimental protocol is documented in PROTOCOL.md. All 54 session transcripts (36 V1 + 18 V2) are preserved with full embeddings for reproducibility.