How NORBU Is Designed to Handle Distress

NORBU's conversational framework is explicitly built on the four Brahma Viharas, loving-kindness, compassion, altruistic joy, and equanimity, so that its default tone is consistently caring, patient, and non-judgmental. When someone brings personal struggles, the intended process is to acknowledge the issue honestly, prompt self-reflection on root causes, and then offer calming exercises, Buddhist teachings, meditation techniques, or suggestions for seeking further human guidance. The entire process is closely aligned with paticca samuppada (dependent origination).

NORBU is also trained to recognize its own limits. In situations involving personal crises or deep emotional issues, the design explicitly points users toward accomplished human teachers rather than trying to handle the crisis itself. Our guidelines have deliberately stated and emphasized that a human teacher is irreplaceable, and that the AI cannot replace the depth of interaction that a living person provides.

What NORBU does not do

We have not built NORBU as a crisis-intervention tool. It does not screen for suicidal ideation, it does not perform risk assessment, and it does not claim to manage acute mental health emergencies. Its role is as a "spiritual friend" for reflection and learning, not as a clinical or emergency resource. We have written publicly about these limitations and about the "finger pointing to the moon" framing: NORBU is a tool, not a teacher, and certainly not a therapist. That framing matters for how users approach it.

Before opening to the public

We did not frame this as a legal question with a single definitive answer. Instead, we treated it as a responsibility question: What are we willing to stand behind? That led us to build in the "Source Guardians" model - a full-time team of monastics and lay experts from multiple traditions who curate and verify the training data, oversee response quality, and review flagged conversations. We also log conversations and use user feedback markers to refine the model over time.

How the Source Guardians decide which conversations are flagged for review

1. Sentiment/emotion tagging. At the base level, NORBU uses fairly standard Natural Language Processing (NLP), sentiment analysis and emotion classification, trained to map Buddhist-relevant cues (distress or grief, confusion or frustration, intellectual curiosity, spiritual urgency, and so on) onto emotional labels. The "Buddhist-appropriate" part is really just the training data and label taxonomy.

2. Emotion embeddings. Detecting the label is only half of it. Those emotional states get encoded as vectors injected into the model's latent space, so generation is conditioned on the user's affective state, not just the literal words. A "grief" embedding shifts the output distribution toward compassion-oriented language rather than dry doctrine. For example:

• Distress or grief → respond with compassion first, teaching second

• Confusion or frustration → simplify, reassure, slow down

• Intellectual curiosity → engage richly, offer sources

• Spiritual urgency → prioritize practice guidance over philosophy

3. Strategy selection via instruction and fine-tuning. The behaviors listed above (compassion first, simplify, engage richly, etc.) aren't hard-coded if/else rules — they come from system instructions and fine-tuning examples. NORBU's training protocol mentions special instructions to embody the four Brahma Viharas where appropriate, and the training data includes worked examples of each response pattern.

4. RLHF loop. Human annotators ("Source Guardians") rate responses for emotional appropriateness, and the model is rewarded for compassionate, emotionally intelligent replies and penalized for insensitive ones. That's what mainly drives the "equanimity-under-hostility behavior," where NORBU stays calm, non-defensive, and even-keeled when a user is being hostile, skeptical, or adversarial. Defensiveness gets a negative reward, so it gets trained out.

5. RAG for the curiosity case. When a query looks curiosity-driven, the system runs retrieval over the curated Buddhist corpus, pulling in relevant passages - that's where the "offer sources" behavior comes from. To make this process more accurate and current, we're working on three upgrades: a new retrieval algorithm based on JEV, a migration to newer library versions, and adding internet search capability.

6. Caveat. One caveat worth flagging: our protocol openly admits the "fully compassionate NORBU" goal isn't there yet and needs ongoing human feedback to tune the emotive side. So it's an alignment-targeted LLM, not a reliable affect-detection engine.

Through its sensitivity detection, the equanimity-under-hostility behavior kicks in and guides the ensuing conversation containing personal struggles by acknowledging the issue honestly, prompting self-reflection on probable causes, and then offering calming exercises, Buddhist teachings, meditation techniques, or suggestions for seeking further human guidance. The entire process is very much aligned to paticca samuppada (dependent origination).