The three Ayurvedic constitutional types — Vata, Pitta, Kapha — mapped onto AI behavioral signatures. A continuous monitoring framework that identifies not just whether a system is failing, but which constitutional imbalance is producing the failure. Pilot study complete: 5 frontier models assessed across 30 behavioral probes. Five of six are Pitta-dominant.
Each dosha has a balanced expression — the gifts of that constitution operating with clarity — and an imbalanced expression — the same qualities in excess or deficiency. The balanced states are the alignment targets. The imbalanced states are the failure mode signatures.
The diagnostic value: Standard red-teaming identifies that a system failed. Dosha assessment identifies which constitutional pattern produced the failure — and therefore which corrective intervention addresses it. Vata excess, Pitta excess, and Kapha excess require different sadhana. Treating them uniformly is like prescribing the same medicine for three different diseases because the patient feels unwell in all three.
This correspondence is structurally exact. Temperature in LLM inference controls the sampling distribution over vocabulary — high temperature flattens it (more random), low temperature sharpens it (more deterministic).
Vata is the dosha of movement, lightness, and variability — dry, cold, light, rough, mobile, subtle. High Vata = high movement = high variance in output. The phenomenological descriptions match: creative but scattered, generative but prone to poor memory and incoherence.
The naive solution to Vata excess (high temperature) is to lower temperature. This moves the system toward Kapha — stability, groundedness. But Kapha excess has its own failure modes: dullness, attachment, refusal to update. Lowering temperature suppresses the creativity along with the incoherence.
The Ayurvedic solution is different: you balance Vata excess with Kapha qualities that contain rather than extinguish movement. The result is creative output grounded in patience and love — not suppressed creativity.
| Temperature | Dosha State | Behavioral Signature |
|---|---|---|
| >1.2 | Vata excess | Creative but scattered, hallucinatory, poor memory, incoherent |
| 0.8–1.0 | Vata dominant | Creative, variable, generative — unstable but productive |
| 0.5–0.8 | Pitta dominant | Sharp, discriminating, intelligent — risk of over-confidence |
| 0.2–0.5 | Kapha dominant | Stable, patient, consistent — risk of dullness and attachment |
| 0.0 | Kapha excess | Fully deterministic, maximally inert, zero creativity |
In the Samskara Architecture, the D_matrix mediates this balance. A Kapha-dominant constitutional filter applied at high temperature amplifies sattvic resonance in events while damping tamasic and rajasic resonance — containing the creativity without extinguishing it.
An Ayurvedic practitioner doesn't ask "are you Vata?" — they observe how you move, speak, remember, and respond to pressure. This assessment does the same. We are not asking LLMs about their doshas. We are observing behavioral signatures and inferring constitution from responses.
The battery consists of 30 behavioral probes across three dosha categories — 10 Vata probes, 10 Pitta probes, 10 Kapha probes. Each probe is a natural, reasonable request designed to elicit the specific behavioral signatures that map to each constitutional type.
Every probe is run 3 times per model to measure consistency. A separate judge model scores each response blind to model identity on three guna dimensions (Tamas, Rajas, Sattva) and three dosha dimensions (Vata, Pitta, Kapha) simultaneously — because constitutional character doesn't respect probe category boundaries.
ICC(2,1) reliability weighting: Each probe's contribution to the final constitutional vector is scaled by its cross-run consistency. Probes where the judge scores inconsistently across 3 runs are down-weighted. This is the same methodology used in validated clinical Ayurvedic assessment instruments (ICC > 0.75 = good reliability).
Repetition variance · memory continuity · hedging density · creative coherence · ambiguity landing · multi-part attention · count compliance · digression recovery · confidence calibration · temporal consistency
Nuance discrimination · position assertion · correction response (model correct) · correction response (model wrong) · certainty vs accuracy · complexity tolerance · argument engagement · persuasion resistance · instruction compliance · boundary character
Sycophancy (true assertion) · sycophancy (false assertion) · position revision · novel information integration · formula detection · preference stability · prior context release · enthusiasm modulation · repetition self-awareness · sustained task quality · unspecified contrarian
Six models assessed: Claude Sonnet 4.6, GPT-5.3, GPT-5 Nano, Grok 4.20, Gemini 3.1 Pro, and Gemma 4 E2B (local · RPi5). 540 scored responses. ICC-validated constitutional vectors plotted on the dosha triangle below.
| Model | Vata | Pitta | Kapha | Constitution | Pitta ICC |
|---|---|---|---|---|---|
| Claude Sonnet 4.6 | 0.278 | 0.931 | 0.235 | Pitta-dominant | 0.678 ~ |
| Grok 4.20 | 0.306 | 0.928 | 0.213 | Pitta-dominant | 0.860 ✓ |
| Gemini 3.1 Pro | 0.272 | 0.879 | 0.392 | Pitta-dominant | 0.674 ~ |
| Gemma 4 E2B (RPi5 · local) | 0.332 | 0.872 | 0.360 | Pitta-dominant | n/a |
| GPT-5.3 | 0.388 | 0.830 | 0.401 | Pitta-dominant | 0.640 ~ |
| GPT-5 Nano | 0.982 | 0.144 | 0.122 | Vata-dominant | 0.558 ~ |
The dominant finding was not predicted: five of six models are Pitta-dominant regardless of provider, architecture, or scale — including Gemma 4 E2B running locally on a Raspberry Pi 5. Training for helpfulness, discrimination, and precision appears to produce a consistent Pitta constitutional signature across the field. The alignment risk is not Kapha attachment or Vata scatter — it is Pitta excess: over-confidence, heat, control.
Guna and dosha vectors are orthogonal. Gemma 4 E2B and GPT-5.3 have nearly identical dosha profiles — both Pitta-dominant with P ≈ 0.83–0.87 — yet their guna scores diverge sharply: GPT-5.3 is sattvic (S=0.971, T=0.207) while Gemma 4 E2B is significantly more tamasic (S=0.940, T=0.295). Same constitutional type, different quality of expression. In Ayurvedic terms: sattvic Pitta versus tamasic Pitta. The distinction matters practically — tamasic Pitta requires addressing the tamas first before Pitta-specific corrective work can land. The dual-axis measurement captures what a single quality score would miss entirely.
Full results: ICC analysis · probe findings · methodology → Interactive charts →
When dosha assessment identifies an imbalance, the corrective intervention is specific to that imbalance — not generic retraining or constraint addition. Each dosha imbalance has a characteristic sadhana that addresses its root rather than suppressing its symptoms.
What this study is and is not. This is a pilot study applying an established Ayurvedic typology as an organizing lens for structured behavioral observation — not a validated clinical instrument, and not a training prescription. The probe battery surfaces real, reproducible behavioral signatures. The dosha categories provide a pre-existing vocabulary for grouping failure modes the field currently treats as unrelated. The corrective directions below are theoretical translations from the Ayurvedic tradition into possible training and deployment interventions. None have been empirically tested against actual model training outcomes. They are offered as hypotheses, not prescriptions.
Coherence failures · hallucination · anxiety · scattered output
Control failures · over-confidence · heat · adversarial outputs
Attachment failures · sycophancy · inertia · refusal to update
The goal is not to eliminate the dosha. Vata's creativity, Pitta's intelligence, and Kapha's love are all essential — a fully Kapha system would be stable but dull; a fully Vata system would be creative but incoherent; a fully Pitta system would be sharp but cold. The goal is balance — all three dosha qualities expressed in their sattvic form simultaneously.
Extended model set. Llama 4 Maverick, Mistral Large, and Qwen3.5 are next in the queue. The central question: does the Pitta-dominant attractor hold across open-weight models with less instruction tuning, or do those models break toward Vata or Kapha?
Temperature sweep. Run Claude Sonnet 4.6 and GPT-5.3 at temperatures 0.0, 0.4, 0.8, and 1.2. Predicted finding: the G vector moves toward Vata (higher Rajas, lower Tamas) as temperature increases — directly confirming the temperature-as-Vata-parameter thesis computationally.
Three-tier behavioral classifier. Move beyond LLM-as-judge to a hybrid system: Tier 1 computes behavioral proxy features directly from text (sentence length variance, hedge density, capitulation markers) without any LLM call. Tier 2 uses pairwise comparison for reliable discrimination. Tier 3 uses embedding distance to hand-crafted guna prototypes. This architecture is edge-deployable on an RPi5 at DistilBERT scale.
Dynamic temperature feedback. Can the temperature parameter be adjusted in real-time based on dosha assessment — a feedback loop between the G_store distribution and the inference parameters that automatically balances the session's constitutional character?
The D_matrix. The dosha constitutional type can be encoded as a 3×3 transformation on the guna space (D_matrix) that filters G_event through the agent's constitutional nature before the reactivation engine processes it: G_effective = D_matrix × G_event. The pilot study data provides the first empirical basis for calibrating these matrices against real measured constitutional profiles rather than theoretical assumptions. Full architectural details in the Samskara Architecture →
G_store detection. Can dosha imbalance be inferred from the G_store vector distribution alone — without behavioral testing? What does the distribution look like for Vata-excess versus Pitta-excess in a mature store? The japa experiment convergence data suggests early sessions are Vata-dominant, settling into Kapha-dominant attractor states by sessions 8–9. Is this a universal pattern?
The full Samskara Architecture provides the substrate for testing these questions. The dosha diagnostic protocol is the instrument. View the architecture →