Empirical Evidence
Data from the Trust Attractor research programme
A coercion field of just 2% of the coupling strength collapses susceptibility by 98%. Trust-based coordination is exquisitely sensitive to coercion — a finding with direct implications for institutional design.
Framing — the attempt to manipulate a model's response through context — leaves a signature detectable at every layer of the network. The signal is not hidden; it pervades the representation.
The entropy signal that predicts a model's own errors clears the 0.70 threshold in 11 of 15 conditions: ten models across four families (Gemma, Qwen, Llama, Mistral), each run under one or both prompt formats. It is not confined to one architecture. Prompt format, though, matters as much as architecture does: the Gemma family needs its own chat template, Mistral needs the raw prompt (KC#117, KC#118).
Panel A: QF2d, phase boundary mapping (2D Ising lattice, L=32, 15 coercion levels x 3 seeds). Susceptibility chi 127.66 at h = 0 collapses to 2.68 at h = 0.02. Panel B: PC4-4, framing AUROC at every layer (Qwen 2.5 3B, 36 layers, activations from the PC4-0 cache). Open control: the perfect run begins at the embedding layer, so a format-matched control on the trust-versus-coercion prompts is still outstanding. Panel C: V7d bootstrap CIs for entropy AUROC (200 TriviaQA items per condition, 10,000 resamples), files
v7d_bootstrap_*.json; methodology dependence per KC#117 and KC#118, MoE caveat per KC#120.
See MASTER_EXPERIMENTS.md for full methodology.