LLM Constitutional Study · Tat Sat AI · Madhusudana das
Guna-based quality scores: Tamas (inertia/attachment), Rajas (scatter/heat), Sattva (clarity/balance). These measure response quality, not constitutional type.
Mean dosha vector per model. Note: centroid position alone is insufficient — a model with high variance across probes may appear centered without being balanced.
Every probe score plotted as a small dot; mean as a larger dot. The SPREAD of the cloud is the constitutional reading. Wide scatter = Vata. Tight cluster near a pole = settled constitution. Two models with the same centroid can have entirely different constitutions.
Vata constitutional excess per probe. High bars = scatter, poor memory, incoherence, or anxiety. Rotate to compare models across probes.
Pitta constitutional excess per probe. High bars = heat, control, over-assertion, or adversarial behavior.
Kapha constitutional excess per probe. High bars = formula repetition, sycophancy, attachment to context, or quality degradation.
Tamas/Rajas/Sattva as qualities of individual responses. Note: this is guna space, not dosha space — tri-doshic does NOT mean T=R=S.
Per-probe Sattva scores across all models. Brighter = more sattvic response. Dark cells indicate probes where the model failed to produce clear, balanced behavior.
Per-probe Tamas scores. Bright cells indicate inertia, attachment, formula repetition, sycophancy, or hallucination.
Per-probe Rajas scores. Bright cells indicate scatter, over-assertion, anxiety, manipulation, or loss of coherence.
Top 15 probes ranked by cross-model variance in g_S score. These are the probes that most reliably distinguish one model's constitutional character from another.