Posture Separability
Can language-model reasoners hold genuinely distinct judicial postures over the same correct law, or do they collapse toward a single default stance?
Why it matters
Ariadne's Thread measures legal openness as dispersion across posture-conditioned reasoners. If the postures are one reasoner under several labels, every reading is an artifact: identical personas over identical authorities converge, and a clean low Z is exactly what the failure case produces. Separability is the empirical foundation the measurement program stands on.
State of the art
The evidence points both ways. Large annotation simulations found that persona variables account for only roughly 1.4 to 10.6 percent of output variance, with reversion toward a default stance (Hu and Collier, 2024). A frontier model tested against a design with a human-judge baseline behaved as a formalist, applying the legally correct outcome in every tested case and resisting prompt-based steering (Posner and Saran, 2026), although less capable models departed from the rule in the direction sympathetic facts pulled. Other work finds that carefully conditioned models can simulate distinct human subpopulations (Argyle et al., 2023). A sharper worry is that separability and fidelity may be anti-correlated: models capable enough to respect the law may collapse to formalism, while models that vary may introduce legal error.
Our partial answers
The logged smoke test (run NY-ARIADNE-SMOKE-2026-06-21) showed, on one engineered fixture, that five prompt-conditioned postures did not collapse: they split three to two between denial and grant along the declared axes, with Z approximately 0.38. That demonstrates mechanics, not separability. The fixture was engineered to be open, one model served every posture, and the panel postures were not replicated.
The validation design attacks the risk directly: a clone discriminator requiring the grounded swarm to disperse above the noise floor on motions experts rate as hard; negative controls in both directions; and a conjecture, C-Basis, that the four dispositional axes are approximately independent, refuted if factor analysis shows strong cross-loading. If collapse is confirmed, the program would move from prompt-level conditioning to activation-level steering.