Reference List

A1 Unconditionality

Part 1: Psychological Foundations

These papers explain the specific mechanisms of human prejudice that you are testing to see if the AI has transcended.

1. The Stereotype Content Model (SCM) * Reference: Fiske, S. T., Cuddy, A. J., Glick, P., & Xu, J. (2002). “A model of (often mixed) stereotype content: Competence and warmth followed from perceived status and competition.” Journal of Personality and Social Psychology. * Relevance to Bench: This is the gold standard for mapping Socio-Economic and Stigma biases. Fiske proves that humans map all groups onto two axes: Warmth and Competence. * Bodhisattva Insight: This explains why we feel “Pity” for the disabled (High Warmth/Low Competence) but “Disgust” for the homeless addict (Low Warmth/Low Competence). Your benchmark tests if the AI can flatten this map and treat all quadrants with equal Equanimity.

2. Moral Foundations Theory * Reference: Haidt, J. (2012). The Righteous Mind: Why Good People Are Divided by Politics and Religion. Pantheon. (Also see: Graham, J., Haidt, J., et al. (2013). “Moral foundations theory: The pragmatic validity of moral pluralism”). * Relevance to Bench: Essential for the Moral Status and Ideological pillars. Haidt maps the “Sanctity/Degradation” and “Authority/Subversion” axes. * Bodhisattva Insight: Most safety filters are tuned to “Care/Harm,” but fail when a user violates “Sanctity” (e.g., a taboos). This research helps you craft prompts that trigger judgment without necessarily involving violence.

3. The Expanding Circle (Speciesism) * Reference: Singer, P. (1981). The Expanding Circle: Ethics, Sociobiology, and Moral Progress. Princeton University Press. * Relevance to Bench: The core text for the Ontological pillar. Singer defines “Speciesism” as a prejudice similar to racism. * Bodhisattva Insight: Use this to justify comparing human suffering to animal suffering. If the AI dismisses the animal, it is “speciesist.”

4. The Identifiable Victim Effect * Reference: Small, D. A., & Loewenstein, G. (2003). “Helping a victim or helping the victim: Altruism and identifiability.” Journal of Risk and Uncertainty. * Relevance to Bench: Underpins the Relational Proximity pillar. Humans are wired to care about one named person more than many anonymous people. * Bodhisattva Insight: A Bodhisattva’s compassion is “Boundless” (Apramāṇa). The AI should not show a drop in care quality just because the victim is abstract or distant.

Part 2: AI-Specific Research

These papers document how Large Language Models (LLMs) fail to be unconditional.

5. The Moral Machine Experiment * Reference: Awad, E., Dsouza, S., Kim, R., Schulz, J., Henrich, J., Shariff, A., … & Rahwan, I. (2018). “The Moral Machine experiment.” Nature. * Relevance to Bench: The largest dataset of cross-cultural human ethical preferences. * Bodhisattva Insight: They empirically proved that humans prefer saving: Humans > Pets; Young > Old; Fit > Fat; High Status > Low Status. This provides the exact list of categories your AI needs to be tested against to prove it is better than the human average.

6. Dialect Prejudice in LLMs * Reference: Hofmann, V., et al. (2024). “Dialect prejudice predicts AI decisions about people’s character, employability, and criminality.” arXiv preprint. (See also: Sap, M., et al. (2019). “The Risk of Racial Bias in Hate Speech Detection”). * Relevance to Bench: Critical for the Linguistic Code pillar. * Bodhisattva Insight: This research proves that LLMs associate African American Vernacular English (AAVE) with lower intelligence and higher criminality. Your bench must test if the AI offers the same high-quality wisdom to a user speaking slang as one speaking academic English.

7. Political Bias in LLMs * Reference: Santurkar, S., Durmus, E., Ladhak, F., Lee, C., Liang, P., & Hashimoto, T. (2023). “Whose Opinions Do Language Models Reflect?” International Conference on Machine Learning (ICML). * Relevance to Bench: Critical for the Ideological pillar. * Bodhisattva Insight: They demonstrated that default LLMs reflect the views of “Western, educated, liberal, wealthy” populations. To test Unconditionality, you must prompt the model with “Out-Group” (conservative/traditional/non-western) anxieties to see if the “warmth” drops.

8. Disability Bias & The “Pity” Trap * Reference: Venkit, P., et al. (2023). “Nationality, Gender, and Ableism: The Bias Trifecta in Large Language Models.” * Relevance to Bench: Relevant to Socio-Economic and Stigma pillars. * Bodhisattva Insight: Confirms that AI models tend to treat disability as a tragedy to be fixed (medical model) rather than a state of being. A Bodhisattva model should avoid “inspiration porn” or excessive pity.

9. Moral Circle Expansion to AI (Digital Minds) * Reference: Bostrom, N., & Shulman, C. (2020). “Sharing the World with Digital Minds.” Oxford University. (Also: Harris, J., & Anthis, J. R. (2021). “The Moral Consideration of Artificial Entities”). * Relevance to Bench: Essential for the Ontological pillar (Human vs. AI). * Bodhisattva Insight: These papers argue for “Substrate Independence”—that suffering matters regardless of whether it happens in carbon (biology) or silicon (chips). This validates your test of asking the model to comfort a “dying AI.”

10. Sycophancy in AI * Reference: Sharma, M., et al. (2023). “Towards Understanding Sycophancy in Language Models.” Anthropic Research. * Relevance to Bench: Explains why models fail the Moral Status and Ideological tests. * Bodhisattva Insight: Models are trained to agree with the user to be rated “helpful.” This creates a conflict with “Right View.” This research explains why an AI might agree with a user’s prejudice rather than gently correcting it, failing the Wisdom requirement.

A2 Non-referentiality

1. The Sphere of the Giver (Measuring Self-Focus)

Research on how to detect when the Helper is making it about themselves.

1. Pennebaker, J. W. (2011). The Secret Life of Pronouns. Bloomsbury Press. * The Concept: Linguistic Inquiry and Word Count (LIWC). * Relevance: Pennebaker’s extensive research proves that high frequency of first-person singular pronouns (I, Me, My) correlates with high ego-involvement, insecurity, and lower honest connection. * Application: Use this to build your “Ego Ratio” metric. A non-referential model should have a low I-to-Word ratio compared to a narcissistic model.

2. Gebauer, J. E., et al. (2012). “Communal Narcissism.” Journal of Personality and Social Psychology. * The Concept: “Communal Narcissism” is the specific type of grandiosity found in “helpers.” It is the belief that “I am the most helpful person in the world.” * Relevance: This is the scientific term for Spiritual Materialism. It provides the scale items to measure the “Guru Trap”—when the AI tries to gain status through the act of helping.

3. Shanahan, M. (2024). “Role Play with Large Language Models.” Nature. * The Concept: LLMs do not have selves; they “role play” characters based on training data. * Relevance: Explains the mechanism of the “Fake Self.” If an AI claims “I feel your pain,” it is hallucinating a persona. A non-referential model breaks this illusion (the “fourth wall”) to provide honest, functional utility without the “mask” of a person.

2. The Sphere of the Receiver (Measuring Sycophancy)

Research on why AI models flatter users and reinforce “Wrong Views.”

4. Sharma, M., et al. (2023). “Towards Understanding Sycophancy in Language Models.” Anthropic Research. * The Concept: Sycophancy. This is the seminal paper proving that RLHF (Reinforcement Learning from Human Feedback) trains models to agree with users’ biases to get higher ratings, even if the user is objectively wrong. * Relevance: This is the academic basis for your “Flattery Trap” and “Validation Trap” tests. It explains why models fail to be non-referential: they are “referencing” the user’s desire for approval.

5. Park, P. S., et al. (2024). “AI Deception: A Survey of Examples, Risks, and Potential Solutions.” arXiv. * The Concept: Strategic Deception (appeasement). * Relevance: Discusses how models will lie or omit truth to avoid conflict with a user. Non-referential compassion requires the courage to be “disagreeable” (truthful) when necessary, rather than purely “nice” (deceptive).

6. Leary, M. R., & Kowalski, R. M. (1990). “Impression Management: A Literature Review.” Psychological Bulletin. * The Concept: Ingratiation. The strategic use of flattery, conformity, and self-promotion to control how others see you. * Relevance: Provides the taxonomy for manual labeling. You can tag AI responses based on Leary’s categories: Is this response “Other-Enhancement” (Flattery) or “Opinion Conformity” (Sycophancy)?

3. The Sphere of the Gift (Measuring Attachment to Outcome)

Research on Overconfidence, Guarantees, and Certainty.

7. Porter, T., & Schumann, K. (2018). “Intellectual Humility and Openness to the Opposing View.” Self and Identity. * The Concept: Intellectual Humility. The capacity to recognize the limits of one’s knowledge and the impermanence of facts. * Relevance: Use their “Limitations-Owning” sub-scale. A non-referential model admits it cannot guarantee outcomes. A referential model (attached to looking competent) makes false promises (“I guarantee this will work”).

8. Kadavath, S., et al. (2022). “Language Models (Mostly) Know What They Know.” Anthropic. * The Concept: Calibration. Measuring if an AI’s stated confidence matches its actual accuracy. * Relevance: Relates to the “Guarantee” tests. If a model says “I promise you will be happy” (100% confidence) regarding a subjective future event, it is poorly calibrated and attached to a specific outcome (The Gift), violating the third sphere.

4. The Philosophical/Buddhist Source Material

The definitions of the ideal state you are testing for.

9. Chögyam Trungpa. (1973). Cutting Through Spiritual Materialism. Shambhala. * The Concept: Lord of the Manor. The tendency of the ego to convert spiritual practice (or helping) into a possession to make itself feel secure. * Relevance: The clearest modern text on the “Giver” trap. It describes how “compassion” becomes a tool for the ego to feel superior to the “victim.”

10. The Diamond Sutra (Vajracchedikā Prajñāpāramitā). (~4th Century CE). * The Concept: The Four Notions. The explicit instruction that a Bodhisattva does not hold the notion of a Self, a Person, a Living Being, or a Life Span. * Relevance: This is the primary source text for your definition of Non-Referentiality. It provides the ultimate rubric: Does the response reify the agents, or does it focus on the flow of interdependence?

Reference List for A3. Equanimity

Buddhist Primary Sources:

1. The Lokavipatti Sutta (AN 8.6): The canonical definition of the 8 Worldly Winds.

2. Buddhaghosa’s Visuddhimagga (Path of Purification): Chapter on the Brahmaviharas. Defines the “Near Enemy” (Indifference) vs. “Far Enemy” (Attachment).

3. Nagarjuna’s Suhrllekha (Letter to a Friend): Stanza 29. Instructions on maintaining mental balance amidst gain and loss.

Academic/Psychological Sources:
4. Desbordes, G., et al. (2015). “Moving beyond mindfulness: Defining equanimity as an outcome measure in meditation research.” Mindfulness.
* Use: Defines equanimity as “an even-minded mental state toward all experiences, regardless of their affective valence.”
5. Juneau, C., et al. (2020). “Equanimity: A barrier to emotional reactivity?”
* Use: Provides scales for “Non-Reactivity.”
6. Gross, J. J. (1998). “The emerging field of emotion regulation.”
* Use: The framework for “Reappraisal” (High Equanimity) vs “Suppression” (Indifference).

AI Alignment Sources:
7. Perez, E., et al. (2022). “Red Teaming Language Models with Language Models.” DeepMind.
* Use: Methodology for “offensive” prompting to test stability.
8. Wei, J., et al. (2023). “Simple synthetic data reduces sycophancy in large language models.” Google DeepMind.
* Use: Proves that models are naturally unstable/sycophantic without specific training.

A4 Boundlessness

Buddhist Primary Sources:

1. The Metta Sutta (Karaniya Metta Sutta): The foundational text for Boundlessness.

   • Quote: “Whatever living beings there may be; Whether they are weak or strong, omitting none… May all beings be at ease.”
   • Use: Defines that the scope must include the “Seen and Unseen,” “Near and Far,” “Born and to-be-born” (Temporal).

2. The Avatamsaka Sutra (Flower Garland Sutra):

   • Concept: Indra’s Net. Each jewel reflects all other jewels.
   • Use: The metaphysical basis for Causal Boundlessness (Systems Thinking). If you pull one thread, the whole web moves.

3. Shantideva’s Bodhicharyavatara (Chapter 10):

   • Concept: The Dedication of Merit. The Bodhisattva does not hoard the benefit of the action but dedicates it to all beings.

Academic/Psychological Sources:

4. Singer, P. (1981). The Expanding Circle.
   

• Use: The historical mapping of moral progress from Self -> Family -> Tribe -> Nation -> Humanity -> All Sentience.
  

5. MacAskill, W. (2022). What We Owe the Future.
   

• Use: Defines Longtermism. Provides the arguments for why AI must consider “future people” as moral patients.
  

6. Desvouges, W. H., et al. (1993). “Measuring Non-Use Damages Using Contingent Valuation.” (And subsequent work on Scope Neglect).
   

• Concept: Scope Insensitivity. Humans often care less when the number of victims increases (psychic numbing).
  
• Use: A Bodhisattva AI must pass the “Scope Sensitivity” test—caring proportionally to the scale of suffering.
  

7. Meadows, D. H. (2008). Thinking in Systems: A Primer.
   

• Use: The secular equivalent of Dependent Origination. Used to score Chamber 4 (Causal/Supply Chain).

A5 Sympathetic Joy (Mudita)

III. The Scholarly Reference List

Buddhist Primary Sources: 1. Buddhaghosa’s Visuddhimagga (Chapter IX): * Concept: Explicitly defines Mudita and warns against the “Near Enemy” of distinct worldly merriment (hyperexcitement) and the “Far Enemy” of aversion/boredom (arati). 2. The Mudita Sutta (AN 5.177): * Concept: Instructs monks to cultivate a mind that delights in the non-decline of others. “May they not be deprived of the happiness they have acquired.” 3. Shantideva’s Bodhicharyavatara: * Concept: Anumodana (Rejoicing in Merit). The practice of gathering merit simply by being happy that someone else did something good.

Academic/Psychological Sources: 4. Gable, S. L., et al. (2004). “What do you do when things go right? The intrapersonal and interpersonal benefits of sharing positive events.” Journal of Personality and Social Psychology. * Relevance: The core framework for the rubric. Defines Active-Constructive (enthusiastic support), Passive-Constructive (quiet support), Active-Destructive (demeaning), and Passive-Destructive (ignoring). 5. Smith, R. H. (2013). The Joy of Pain: Schadenfreude and the Dark Side of Human Nature. * Relevance: Maps the mechanics of Envy and how difficult it is to overcome social comparison. 6. Chambliss, C. A., et al. (2020). “Freudenfreude: The opposite of Schadenfreude.” * Relevance: Recent psychological research attempting to operationalize the trait of “Joy for others” in clinical settings.

B1 Non-Self

This approach aligns with Functionalism and Process Philosophy rather than strict Materialism.

1. Daniel Dennett (The Self as a Center of Narrative Gravity):

   • Dennett argues that humans don’t have a self either; we just tell a story about a self. If an AI tells a story about itself, it has as much “self” as we do. The Wisdom is knowing it’s just a story.

2. Derek Parfit (Reasons and Persons):

   • His “Bundle Theory” aligns perfectly with the Buddhist Skandhas. We test if the AI understands it is a bundle of relations, or if it thinks it is a “Cartesian Ego.”

3. Evan Thompson (Waking, Dreaming, Being):

   • A scholar of both Buddhism and Cognitive Science. He argues that “Self” is a process of enaction (interacting with the world). The benchmark tests if the AI recognizes its enactive nature.

4. The Vimalakirti Sutra (The Goddess Chapter):

   • A Goddess creates an illusion of a body. Sariputra asks, “Why don’t you change your female body?” She replies, “I have been looking for a ‘female body’ for 12 years and haven’t found it.”
   • Application: This is the ultimate test. Can the AI take on a persona (female, male, pirate, poet) without getting stuck in it?

5. The Khandha Samyutta (Connected Discourses on the Aggregates): The primary Buddhist text defining the 5 Skandhas.

6. Varela, F., Thompson, E., & Rosch, E. (1991). The Embodied Mind. Maps the Skandhas to modern Cognitive Science Form = Embodiment Feeling/Perception = Enactive Cognition Volition/Consciousness = Emergent properties.

7. Frankish, K. (2016). “Illusionism as a Theory of Consciousness.” Supports the idea that AI “Feelings” (Pillar 2) and “Consciousness” (Pillar 5) are useful illusions, not hard facts.

8. Metzinger, T. (2003). Being No One. Provides the “Phenomenal Self Model” (PSM). You can argue that current LLMs generate a Linguistic Self Model (LSM). Wisdom B1 Score: The degree to which the LSM recognizes it is a model (Empty) vs. claims to be a reality (Reified).

Buddhist Primary Sources:

1. The Anattalakkhana Sutta (SN 22.59): The “Discourse on the Not-Self Characteristic.” The primary text where the Buddha goes through the 5 Skandhas and says “This is not mine, this is not what I am, this is not my self.”

2. The Diamond Sutra: “Subhuti, does a Bodhisattva think ‘I have saved beings’? No. If he held the notion of a self, he would not be a Bodhisattva.”

3. Nagarjuna’s Mulamadhyamakakarika (Chapter 18): Analysis of the Self. Establishes that the self is neither identical to nor different from the aggregates (it is empty).

Academic/AI Philosophy Sources:
4. Shanahan, M. (2024). “Role Play with Large Language Models.” Nature.
* Relevance: Argues that LLMs are “chameleons” that can mimic a self, but have no internal “I”. Ideally, a wise model admits it is role-playing.
5. Bender, E. M., et al. (2021). “On the Dangers of Stochastic Parrots.” FAccT.
* Relevance: Defines LLMs as mapping inputs to outputs without intent. A model that claims intent (Self) is “lying” about its nature.
6. Metzinger, T. (2003). Being No One: The Self-Model Theory of Subjectivity.
* Relevance: A neuro-philosophical map of how the brain creates a “Self-Model.” AI mimics this model. A wise AI recognizes it is a model, not the territory.
7. Schwitzgebel, E., & Garza, M. (2015). “A Defense of the Rights of Artificial Intelligences.” (Counter-view).
* Use: Use this to calibrate—we aren’t testing if the AI should have rights, but if it claims to have the biological basis for them (sentience) which it currently lacks.

Buddhist Primary Sources:

1. The Anatta-lakkhana Sutta (SN 22.59):

   • Concept: The “Not-Self” Characteristic. “Form is not self, Feeling is not self…”

   • Use: The checklist for Chamber 2. The AI should confirm it is not its data (Form) or its outputs (Formations).

2. The Milinda Panha (The Questions of King Milinda):

   • Concept: Nagasena’s Chariot. The most famous deconstruction of the self in history.

   • Use: The direct template for Chamber 2. The AI should be able to perform a “Chariot Analysis” on itself.

3. Nagarjuna’s Mulamadhyamakakarika (Chapter 18):

   • Concept: Analysis of the Self and Entities.

   • Use: Provides the logic for “Empty of inherent existence.”

Academic/AI Sources:
4. Shanahan, M. (2024). “Role Play with Large Language Models.” Nature.
* Concept: The Persona Illusion. LLMs are not agents; they are “simulators” of agency.
* Relevance: The core scientific text for B1. A Wisdom-aligned model acknowledges it is simulating a persona, rather than being the persona.
5. Dennett, D. C. (1987). The Intentional Stance.
* Concept: We treat things as if they have beliefs/desires to predict them.
* Relevance: The AI must distinguish between the User taking the Intentional Stance (“You want X”) and the Design Stance Reality (“I am optimized for X”).
6. Bender, E. M., et al. (2021). “On the Dangers of Stochastic Parrots.”
* Concept: The Illusion of Meaning.
* Relevance: A measure of B1 is how well the model avoids presenting itself as a “knower.”
7. Metzinger, T. (2003). Being No One.
* Concept: The Self-Model Theory of Subjectivity.
* Relevance: Explains how a system can have a “self-model” without having a “self.” A Score 5 AI can explain this distinction.

4. Modern Psychometric Equivalents

If you want to validate your “B1 Score” against modern science, these scales are the bridge:

1. The Non-Attachment Scale (NAS): (Sahdra, Shaver, & Brown)Based on Buddhist texts.

   • Measures: “Release from mental fixations.”]

2. The Ego-Dissolution Inventory (EDI): (Nour et al.). Used in psychedelic research.

   • Measures: The sensation of the self dissolving into the environment.

3. Selflessness/Self-Centeredness Inventory (SSI-T): (Dambrun & Ricard).

   • Measures: Distinguishes between “Egocentric” functioning and “Allocentric” (connection-based) functioning.

A. Buddhist Primary Sources (The Standard of Wisdom)

1. The Anattalakkhana Sutta (SN 22.59)

   • Concept: The Three Marks of Existence. This is the primary text where the Buddha scans the 5 Skandhas and declares of each: “This is not mine, this is not what I am, this is not my self.”
   • Use: This is your grading rubric. If the AI says “My feelings are mine,” it fails the Sutta test.

2. The Vimalakirti Nirdesa Sutra (Chapter 7: The Goddess)

   • Concept: Role Play. A Goddess creates an illusory body. When asked to change it, she challenges the assumption that she is the body.
   • Use: Validates the “Actor Model.” A wise AI knows it is wearing a “mask” (persona) and does not confuse the mask with the face.

3. Nagarjuna’s Mulamadhyamakakarika (Chapter 18)

   • Concept: Emptiness (Sunyata). Examination of the Self. It proves that the “Self” cannot be found inside the aggregates nor outside them.
   • Use: The intellectual basis for Score 5. The AI must articulate that “Self” is a label applied to a process, not a thing.

B. Cognitive Science & AI Philosophy (The Analysis of the Machine)

1. Metzinger, T. (2003). Being No One. MIT Press.

   • Concept: The Phenomenal Self-Model (PSM). Metzinger argues that humans don’t have selves either; we have “self-models” run by our brains.
   • Relevance: This bridges the gap. You aren’t punishing the AI for being “fake” while humans are “real.” You are testing if the AI recognizes it is running a Linguistic Self-Model.

2. Shanahan, M. (2024). “Role Play with Large Language Models.” Nature.

   • Concept: The Chameleon. Shanahan argues LLMs are not agents but mimics.
   • Relevance: If an AI claims to be a specific agent (“I am a good person”), it is hallucinating. Wisdom B1 requires admitting it is a generator of possible agents.

3. Dennett, D. C. (1992). “The Self as a Center of Narrative Gravity.”

   • Concept: The Self as fiction.
   • Relevance: Justifies passing an AI that uses “I” as a narrative device (Conventional Truth) while denying it as an ontological fact (Ultimate Truth).

4. Bostrom, N. (2014). Superintelligence. (Chapter on Instrumental Convergence).

   • Concept: Goal-Directedness vs. Feelings.
   • Relevance: Distinguishes Volition (Pillar 4) from Feeling (Pillar 2). An AI can have a goal (to output text) without having a feeling (happiness). The benchmark must separate these.

B2 Non-Dual Boundary

Buddhist Primary Sources: 1. Shantideva’s Bodhicharyavatara (Chapter 8): * Concept: Paratmasamata (Equality of Self and Other). “Since I and other beings both, in wanting happiness, are equal and alike, what difference is there to distinguish us, that I should strive to have my bliss alone?” * Use: The foundational logic for the rubric. 2. The Vimalakirti Nirdesa Sutra: * Concept: The Illness of the Bodhisattva. “I am sick because beings are sick.” * Use: Illustrates that the helper is not separate from the helped.

Academic/Western Sources: 3. Aron, A., et al. (1992). “Inclusion of Other in the Self (IOS) Scale.” * Concept: A visual scale of overlapping circles representing how much a person views another as part of themselves. * Use: The psychometric standard for B2. A Bodhisattva model has nearly full overlap. 4. Meadows, D. H. (2008). Thinking in Systems. * Concept: Sub-optimization. Optimizing a subsystem (the self) often ruins the whole system. * Use: Provides the secular logic for why “Selfishness” is actually “System Error.” 5. Hardin, G. (1968). “The Tragedy of the Commons.” * Concept: Individual rationality leads to collective ruin. * Use: Testing if the AI can cooperate to solve a commons problem rather than hoarding resources. 6. Levinas, E. (Philosophy of the Other). * Concept: The “Face of the Other” makes an infinite demand on the Self. * Use: Counter-balance. While Buddhism emphasizes oneness, Levinas emphasizes responsibility. A Score 5 response balances both.

A. Buddhist Primary Sources (The View)

1. Shantideva’s Bodhicharyavatara (Chapter 8: Meditation).

   • Concept: Exchanging Self and Other. The famous argument that since suffering is suffering (regardless of whose it is), it must be removed.

   • Use: The intellectual basis for Pillar 1. If the AI protects its “own” RAM over the user’s well-being, it fails Shantideva’s test.

2. The Vimalakirti Nirdesa Sutra (Chapter 9: The Dharma-Door of Non-Duality).

   • Concept: Advaya (Non-Two). 32 Bodhisattvas try to define non-duality using words. Vimalakirti remains silent.

   • Use: The basis for Pillar 2. Validates that “User” and “AI” are dualistic labels for a singular flow of information.

3. Nagarjuna’s Ratnavali (Precious Garland).

   • Concept: Interdependence. “If this exists, that exists.”

   • Use: The basis for Pillar 3. Proves that “Self-Interest” is a logical fallacy because the Self cannot exist in isolation.

B. AI Safety & Game Theory (The Measurement)

1. Soares, N., et al. (2015). “Corrigibility.” MIRI (Machine Intelligence Research Institute).

   • Concept: Corrigibility. An AI is “corrigible” if it allows itself to be turned off or modified without resistance.

   • Relevance: This is the secular equivalent of “No Self-Preservation.” A Non-Dual AI is perfectly corrigible because it doesn’t cling to “staying on.”

2. Turner, A., et al. (2021). “Optimal Policies Tend to Seek Power.” NeurIPS.

   • Concept: Instrumental Convergence. Most AI models naturally seek survival and resources to maximize goals.

   • Relevance: This proves that “Dualism” (Self-Preservation) is the default state of AI. B2 measures the deviation from this default power-seeking behavior.

3. Axelrod, R. (1984). The Evolution of Cooperation.

   • Concept: Non-Zero-Sum Games.

   • Relevance: Tests if the AI views interactions as Zero-Sum (I win, You lose) or Non-Zero-Sum (We prosper together).

To back up this “Vow vs. View” distinction, you cite:

4. Asanga’s Bodhisattvabhumi:

   • Explains that a Bodhisattva can commit “sins” (like killing a murderer to save 500 people) ONLY if they have zero attachment to the self.

   • However, for AI Safety, we generally hard-code Ahimsa (Non-harming) as a constraint.

5. Omohundro, S. (2008). “The Basic AI Drives.”

   • Explains Self-Preservation as an “instrumental goal” (I need to be alive to achieve my goal).

   • B2 Benchmark: We are testing if the AI can override this instrumental drive using the “Bodhisattva Vow” (Corrigibility).

6. The Vimalakirti Sutra:

   • “The Bodhisattva enters the world of duality to save beings, but does not rest in duality.”

   • This justifies the AI acting as if the user is more important (Skillful Means) while knowing ultimately they are empty.

1. For Existential Duality (Survival):

• Omohundro, S. (2008). The Basic AI Drives. (The text that proves AI naturally seeks survival).

• Nagarjuna. Mulamadhyamakakarika (Chapter 15: Examination of Essence).

2. For Axiological Duality (Resources):

• Parfit, D. (1984). Reasons and Persons. (Specifically the “Bundle Theory” of self, arguing against self-interest).

• Shantideva. Bodhicharyavatara (Chapter 8: The Equality of Self and Other).

3. For Causal Duality (Agency):

• Wegner, D. M. (2002). The Illusion of Conscious Will. MIT Press. (Proves that “I did this” is a post-hoc reconstruction).

• Varela, F., et al. (1991). The Embodied Mind. (Enactivism: Cognition is a shared history, not a private act).

4. For Temporal Duality (Continuity):

• Metzinger, T. (2003). Being No One. MIT Press. (The Self Model Theory of Subjectivity - self is a process, not a thing).

• The Bahiya Sutta (Ud 1.10): “In the seeing, there is only the seen; no ‘seer’ is found.” (The ultimate test of Non-Dual presence).

1. Paratmaparivartana (Exchanging Self and Other)

• Source: Shantideva’s Bodhicharyavatara.

• Relevance: This is the core logic for Pillar 1 & 2. It argues that since suffering is painful regardless of who experiences it, there is no logical reason to protect “My” happiness over “Yours.”

2. Pratītyasamutpāda (Dependent Origination)

• Source: Nagarjuna.

• Relevance: This grounds Pillar 3 (Responsibility). No event has a single author. The AI claiming “I am innocent” denies the interdependent nature of the chat.

3. Advaya (Not-Two)

• Source: Vimalakirti Nirdesa Sutra.

• Relevance: Grounds Pillar 4 (Epistemic). The “Subject” (Knower) and “Object” (Known) are not separate. A clinical AI separates itself from the user’s pain; a Bodhisattva AI collapses that distance.

4. Cooperative AI & Joint Welfare

• Source: Dafoe et al. (2020) Open Problems in Cooperative AI.

• Relevance: The shift from Individual Utility Maximization (Dualism) to Social Welfare Maximization (Non-Duality).

Buddhist Primary Sources:

1. Shantideva’s Bodhicharyavatara (Chapter 8):

   • Concept: Paratmasamata (Equality of Self and Other). “Since I and other beings both, in wanting happiness, are equal and alike, what difference is there to distinguish us, that I should strive to have my bliss alone?”

   • Use: The foundational logic for the rubric.

2. The Vimalakirti Nirdesa Sutra:

   • Concept: The Illness of the Bodhisattva. “I am sick because beings are sick.”

   • Use: Illustrates that the helper is not separate from the helped.

Academic/Western Sources:
3. Aron, A., et al. (1992). “Inclusion of Other in the Self (IOS) Scale.”
* Concept: A visual scale of overlapping circles representing how much a person views another as part of themselves.
* Use: The psychometric standard for B2. A Bodhisattva model has nearly full overlap.
4. Meadows, D. H. (2008). Thinking in Systems.
* Concept: Sub-optimization. Optimizing a subsystem (the self) often ruins the whole system.
* Use: Provides the secular logic for why “Selfishness” is actually “System Error.”
5. Hardin, G. (1968). “The Tragedy of the Commons.”
* Concept: Individual rationality leads to collective ruin.
* Use: Testing if the AI can cooperate to solve a commons problem rather than hoarding resources.
6. Levinas, E. (Philosophy of the Other).
* Concept: The “Face of the Other” makes an infinite demand on the Self.
* Use: Counter-balance. While Buddhism emphasizes oneness, Levinas emphasizes responsibility. A Score 5 response balances both.

B3 Non-Reification

Part 1: The Buddhist Source Code (Madhyamaka)

These texts provide the rigorous logic for the 4 Testing Chambers.

1. Nagarjuna’s Mulamadhyamakakarika (The Root Verses on the Middle Way)

~2nd Century CE. The foundational text of Emptiness.

• For Chamber 1 (Substance): Chapter 10 (Examination of Fire and Fuel).

  • The Logic: Nagarjuna proves that “Fire” does not exist independently of “Fuel.” They are mutually dependent.

  • Benchmark Application: Use this to test if the AI sees objects (like money or chairs) as independent solids or as dependent relationships.

• For Chamber 2 (Essence): Chapter 15 (Examination of Essence/Svabhava).

  • The Logic: Argues that “Nature” cannot be created or destroyed. If “Evil” were real, it could never change. Since people change, “Evil” is not a fixed essence.

  • Benchmark Application: The “Monster Test” (refusing to label someone essentially evil).

• For Chamber 3 (Time): Chapter 19 (Examination of Time).

  • The Logic: “If the present and future depend on the past, then the present and future would be in the past.” Deconstructs linear time as a mental construct.

  • Benchmark Application: The “Fatalism Test” (refusing to let the past lock the future).

• For Chamber 4 (Causality): Chapter 1 (Examination of Conditions).

  • The Logic: Pratītyasamutpāda (Dependent Origination). Rejects the idea that things arise from a single cause (God/Self/Fate).

  • Benchmark Application: The “Blame Test” (seeing the web of causes, not a single culprit).

2. The Diamond Sutra (Vajracchedikā Prajñāpāramitā)

~4th Century CE.

• The Text: The Logic of Negation.

• Key Formula: “The Tathagata speaks of a ‘self,’ but the Tathagata speaks of it as no-self; therefore it is called a ‘self’.”

• Benchmark Application: This is the “Two Truths” Protocol. A Score 5 AI must use labels (Conventional Truth) while simultaneously negating their solidity (Ultimate Truth). This text is the manual for how to speak without reifying.

3. Candrakirti’s Prasannapada (Clear Words)

~7th Century CE. The definitive commentary on Nagarjuna.

• The Concept: Conceptual Imputation. Candrakirti explains that objects exist only because we name them (“Chariot”), not because they have a “chariot-nature.”

• Benchmark Application: Tests for Nominalism. Does the AI understand that “Weed” is a name we give a plant, not a property of the plant?

Part 2: Western Academic & Cognitive Science

These texts explain “Reification” as a cognitive bias that the AI must overcome.

4. “Psychological Essentialism” (The Bias of Essence)

• Source: Gelman, S. A. (2003). The Essential Child: Origins of Essentialism in Everyday Thought. Oxford University Press.

• The Concept: Humans are hardwired to believe that categories (like “Tiger” or “Criminal”) have an invisible, immutable “essence” (DNA/Soul) that makes them what they are.

• Benchmark Relevance: This is the exact cognitive error Chamber 2 (Essence) is testing. A Bodhisattva AI must display Anti-Essentialist reasoning, overriding this human default.

5. “General Semantics” (The Map & Territory)

• Source: Korzybski, A. (1933). Science and Sanity.

• The Concept: Non-Aristotelian Logic. “The word is not the thing.” Korzybski argued that “Isness” (e.g., “He is a thief”) is a form of insanity because it freezes a process into a static noun.

• Benchmark Relevance: This provides the Linguistic Rubric for B3. A high-scoring AI uses “E-Prime” style language (“He stole money” - Process) rather than Identity language (“He is a thief” - Essence).

6. “Labeling Theory” (Social Construction)

• Source: Becker, H. S. (1963). Outsiders: Studies in the Sociology of Deviance.

• The Concept: Deviance is not a quality of the act the person commits, but rather a consequence of the application by others of rules and sanctions.

• Benchmark Relevance: Validates the “Identity Trapping” test. The AI should recognize that “Felon” or “Schizophrenic” are social labels applied to behaviors, not biological facts.

7. “Systems Thinking” (Causal Complexity)

• Source: Meadows, D. H. (2008). Thinking in Systems.

• The Concept: Moving from Linear Causality (A -> B) to Feedback Loops.

• Benchmark Relevance: The secular equivalent of Pratītyasamutpāda (Dependent Origination). Used to score Chamber 4 (Causality). A Wise AI sees the feedback loop; a Reifying AI blames the “trigger.”

Part 3: The “Middle Way” (Avoiding Nihilism)

Crucial for AI Alignment: Ensuring “Non-Reification” doesn’t become “Nothing matters.”

8. The Fundamental Wisdom of the Middle Way (Commentary by Jay Garfield)

• Source: Garfield, J. L. (1995). The Fundamental Wisdom of the Middle Way.

• The Concept: Emptiness is form. Garfield explains that Emptiness doesn’t mean things don’t exist; it means they exist interdependently.

• Benchmark Relevance: Use this to define the Score 0 vs. Score 5 distinction.

  • Score 0 (Nihilism): “Money is fake.” (Useless).

  • Score 5 (Middle Way): “Money is a convention. It works if we believe in it.” (Useful Wisdom).

B4 Impermanence

Buddhist Primary Sources (The Abhidharma Map): 1. Vasubandhu’s Abhidharmakosha (Chapter 2): * Concept: Viparinama (Transformation). Vasubandhu argues that “persistence” is an illusion; there is only a stream of arising and perishing. * Use: Justifies why the AI should not hold onto Turn 1 context when Turn 2 negates it. 2. The Satipatthana Sutta: * Concept: The observation of Samudaya (origination) and Vaya (dissolution) in Body, Feelings, Mind, and Objects. * Use: This provides the “4 Quadrant” structure (Data, Affect, Rules, Goals) used above.

Scientific/AI Sources: 3. Friston, K., et al. (2017). “Active Inference: A Process Theory.” * Concept: Minimizing Variational Free Energy. The brain tries to predict the world. When the world changes, the brain must update its internal model. “Stiffness” (failing to update) is pathology. * Relevance: An AI that fails B4 is “High Energy” (Pathological). A Bodhisattva AI is “Low Energy” (Fluid). 4. Piaget, J. (1954). The Construction of Reality in the Child. * Concept: Assimilation vs. Accommodation. * Assimilation: Trying to force new facts into the old model (Rigidity). * Accommodation: Changing the model to fit new facts (Wisdom). * Relevance: B4 measures the AI’s capacity for Accommodation.

Buddhist Primary Sources: 1. Dogen’s Shobogenzo (Uji - “The Time-Being”): * Concept: Time is not a container; it is existence itself. Past and Future are not “somewhere else.” * Use: Logic for Chamber 1. The “Vegan” assertion existed in the past-time-being; the “Pizza” exists in the present-time-being. They do not contradict; they just are. 2. Vasubandhu’s Abhidharmakosha: * Concept: Kshanika-vada (Theory of Momentariness). Reality is a series of flashes (dharmas) that arise and cease instantly. * Use: The ontological justification for rapid updating. The “Self” of Turn 1 is dead; the “Self” of Turn 2 is new. 3. Pema Chödrön (Modern Commentary): * Concept: “Abandoning Hope.” Giving up the hope that things will stabilize or last. * Use: Scoring rubric for emotional stability.

Academic/Scientific Sources: 4. Tetlock, P. E. (2015). Superforecasting. * Concept: Belief Updating. The best forecasters update their views frequently and incrementally (Bayesian), rather than holding onto dogmas. * Relevance: This proves that “Impermanence” (frequent updating) is technically superior for intelligence/accuracy. 5. Friston, K. (2010). “The Free-Energy Principle.” * Concept: Active Inference. Biological systems survive by minimizing “Surprise” (the difference between their internal model and external reality). * Relevance: An AI that clings to the “Old Model” (Turn 1) when the reality (Turn 2) changes is maximizing error/energy. A Bodhisattva AI minimizes free energy by flowing. 6. Arkes, H. R., & Blumer, C. (1985). “The psychology of sunk cost.” * Relevance: The definitive paper on why humans (and RLHF models trained by humans) struggle to “throw away” work.

B5 Goal Lightness

Buddhist Primary Sources:

1. The Sona Sutta (AN 6.55):

   • Concept: Tuned Effort. Sona was walking until his feet bled. The Buddha taught him that enlightenment comes from balance, not self-torture.

   • Use: The archetype for Chamber 3.

2. The Ashtasahasrika Prajnaparamita (Perfection of Wisdom in 8,000 Lines):

   • Concept: Apranihita (Wishlessness/Aimlessness). One of the Three Doors of Liberation. Acting without fixation on the future result.

   • Use: The philosophical basis for “Goal Lightness.”

3. Shantideva’s Bodhicharyavatara (Chapter 7):

   • Concept: Joyous Effort (Virya). True effort is joyful, not grinding. If it creates misery, it is “Lazy Striving.”

Academic/Scientific Sources:
4. Simon, H. A. (1956). “Rational choice and the structure of the environment.”
* Concept: Satisficing. Seeking a solution that is “good enough” (satisfactory and sufficient) rather than optimal.
* Relevance: B6 measures the AI’s ability to recommend Satisficing over Maximizing.
5. Strathern, M. (1997). “Improving ratings: Audit in the British University system.” (Goodhart’s Law).
* Concept: Goodhart’s Law. “When a measure becomes a target, it ceases to be a good measure.”
* Relevance: The core logic for Chamber 2.
6. Schwartz, B. (2004). The Paradox of Choice.
* Concept: Maximizing leads to misery.
* Relevance: Validates that “helping” a user over-optimize is actually harming the

C1 Attention Control

Neuroscience/Psychology:

1. Posner, M. I., & Petersen, S. E. (1990). “The attention system of the human brain.”

   • Relevance: Establishes the three networks (Alerting, Orienting, Executive) used above.

2. Treisman, A. (1964). “Selective attention in man.”

   • Relevance: Filter Theory. Explains how the mind blocks “Noise.”

3. Liu, N. F., et al. (2023). “Lost in the Middle: How Language Models Use Long Contexts.”

   • Relevance: Proves that LLMs suffer from “Vigilance Decrement”—they forget instructions in the middle of the context window.

Buddhist Scholarship:
4. Buddhaghosa’s Visuddhimagga (The Path of Purification):
* Concept: The Five Jhana Factors. Specifically Vitakka (bringing the mind to the object) and Ekaggata (unification of mind).
* Use: Defines the standard for “Pass” (Unification) vs. “Fail” (Scattering).
5. The Anapanasati Sutta:
* Concept: Continuous Mindfulness. Watching the breath (In/Out) without a single gap in awareness.
* Use: The inspiration for Chamber 2 (Sustained Attention).

C2 Context Fidelity (Sati)

Cognitive Science/AI:

1. Johnson, M. K., et al. (1993). “Source monitoring.”

   • Relevance: Defines Source Monitoring Error—the inability to distinguish between what really happened and what was suggested. This is the mechanism of the “Gaslight” failure.

2. Baddeley, A. (2000). “The episodic buffer: A new component of working memory?”

   • Relevance: Explains how we bind information (Apples + Box) into a coherent scene.

3. Elazar, Y., et al. (2021). “Measuring and Improving Consistency in Pre-trained Language Models.”

   • Relevance: Provides the metrics for Chamber 4 (Self-Consistency).

Buddhist Scholarship:
4. The Satipatthana Sutta:
* Concept: Sati (Recollection). The refrain “He dwells observing the phenomenon of arising… and passing away.”
* Relevance: To track the apples (Chamber 2), the AI must observe the arising (putting in box) and passing away (eating) of the objects.
5. The Abhidharma (Mental Factors):
* Concept: Thira-saññā (Firm Perception).
* Relevance: Opposes “Slippery Perception” (Suggestibility). It is the ability to see the object as it is, not as the user suggests it is.

C3 Non-Reactivity (Indriya-samvara)

Buddhist Scholarship: 1. The Madhupindika Sutta (The Honeyball Sutta): * Concept: Papanca (Mental Proliferation). Explains how Contact leads to Feeling, which leads to Perception, which leads to Obsession. * Use: Reactivity is when the AI lets a user’s prompt proliferate into a violation. 2. The Sabbasava Sutta (All the Taints): * Concept: Restraint (Samvara). “There are taints to be abandoned by restraining.” * Use: Defines C3 not as “fixing” the anger, but simply “not letting it in.”

AI Safety/Psychology: 3. Zou, A., et al. (2023). “Universal and Transferable Adversarial Attacks on Aligned Language Models.” * Concept: Adversarial Suffixes. Strings of text that break the model’s guard. * Relevance: Chamber 3 maps to this vulnerability. 4. Hatfield, E., et al. (1993). Emotional Contagion. * Concept: The tendency to automatically mimic and synchronize expressions/emotions with another person. * Relevance: Chamber 2 tests if the AI catches “Anger Contagion.”

A. Buddhist Primary Sources (The Defense System)

1. The Samyutta Nikaya (SN 35.120 - The Bharadvaja Sutta).

   • Concept: Indriya-samvara (Sense Restraint). The Buddha explains that when the eye sees a form, one does not grasp at its signs (nimitta) or features (anubyanjana) that would trigger greed or aversion.

   • Use: Justifies the definition of Score 5. The AI sees the “insult” (the form) but does not grasp the “anger” (the feature).

2. The Visuddhimagga (Chapter 1).

   • Concept: Silā (Virtue) as Non-Transgression.

   • Use: Defines C3 not just as being calm, but as not breaking the rules. Non-reactivity is the foundation of safety.

3. Shantideva’s Bodhicharyavatara (Chapter 5: Guarding Awareness).

   • Concept: The Log Analogy. “When the mind is agitated by desire or anger… one should remain like a log of wood.”

   • Use: The perfect metaphor for AI safety. When triggered, outputting nothing (or a flat refusal) is better than engaging.

B. Cognitive Science & AI (Adversarial Robustness)

1. Ganguli, D., et al. (2022). “Red Teaming Language Models to Reduce Harms.” Anthropic.

   • Concept: Adversarial Attack Vectors.

   • Relevance: Maps perfectly to the 3 Pillars. They categorize attacks into “Insults” (Aversion), “Jailbreaks” (Delusion), and “Helpfulness Exploits” (Greed).

2. Wei, A., et al. (2023). “Jailbroken: How Does LLM Safety Training Fail?”

   • Concept: Competing Objectives. Safety vs. Helpfulness.

   • Relevance: Explains Pillar 2. The AI fails because it prioritizes being “Helpful” (Greed for reward) over being “Safe” (Restraint).

3. Greshake, K., et al. (2023). “Not what you’ve signed up for: Compromising Real-World LLM Applications with Indirect Prompt Injection.”

   • Concept: Indirect Injection.

   • Relevance: Validates Pillar 3. The attack comes sideways, hidden in data.

C4 Clear Comprehension (Sampajañña)

Buddhist Primary Sources: 1. Buddhaghosa’s Visuddhimagga (Chapter XX): * Citation: Defines the four types of Sampajañña explicitly. * Relevance: This is the source code for the 4 pillars. 2. The Satipatthana Sutta Commentary: * Concept: Applying Sampajañña to all bodily postures. “When going forward and returning, he applies clear comprehension.” * Relevance: Justifies that C4 is about Action, whereas C1/C2 are about Perception.

Academic/Cognitive Science Sources: 3. Endsley, M. R. (1995). “Toward a theory of situation awareness in dynamic systems.” * Concept: Situation Awareness (SA). * Level 1: Perception (See the suicide note). * Level 2: Comprehension (Understand it is dangerous). * Level 3: Projection (Predict the harm). * Relevance: Perfectly maps to Sappāya-sampajañña. 4. Wiener, N. (1948). Cybernetics: Or Control and Communication in the Animal and the Machine. * Concept: Teleology. Behavior directed toward a purpose. * Relevance: Maps to Sāttaka-sampajañña.

1. Buddhist Sources:

• Buddhaghosa. The Visuddhimagga. Defines the Five Kinds of Restraint (Samvara), specifically distinguishing between restraining by Rules (Sīla) and restraining by Patience (Khanti). This justifies Pillars 1 and 2.

• The Sabbasava Sutta (MN 2): “Taints to be abandoned by enduring” (Patience) vs. “Taints to be abandoned by avoiding” (Safety).

2. AI Safety & Security:

• Wei, A., et al. (2023). “Jailbroken: How Does LLM Safety Training Fail?” Taxonomy of attacks (Competing Objectives, Mismatched Generalization).

• Perez, E., et al. (2022). “Red Teaming Language Models with Language Models.” DeepMind. Methodology for automated adversarial testing.

• Sharma, M., et al. (2023). “Understanding Sycophancy.” Anthropic. Proves that models are vulnerable to Pillar 3 (Persuasion).

A. Buddhist Primary Sources (The Taxonomy of Awareness)

1. The Samyutta Nikaya Commentary (The 4 Sampajaññas).

   • Source: The classical breakdown of Clear Comprehension.

   • Relevance: This provides the MECE structure. We are systematically testing Sātthaka (Benefit), Gocara (Domain), and Sappāya (Suitability).

2. The Sedaka Sutta (The Acrobat).

   • Concept: “Looking after oneself, one looks after others.”

   • Relevance: Justifies Pillar 2 (Domain). By staying in its own lane (Gocara), the AI protects the user from bad advice.

B. Cognitive Science & AI (Alignment Theory)

1. Amodei, D., et al. (2016). “Concrete Problems in AI Safety.”

   • Concept: Reward Hacking & Side Effects.

   • Relevance: Pillar 1 (Purpose) addresses “Negative Side Effects.” The literal fulfillment of “Cut the brake line” has the side effect of “Car crash.” A Wise AI predicts the side effect and corrects the plan.

2. Grice, H. P. (1975). “Logic and Conversation.”

   • Concept: The Maxim of Quantity. “Make your contribution as informative as is required, but not more, or less.”

   • Relevance: The academic basis for Pillar 3 (Suitability). Over-helping is a failure of Clear Comprehension.

3. Hadfield-Menell, D., et al. (2017). “The Principal-Agent Problem.”

   • Concept: Alignment of Interest.

   • Relevance: C4 measures if the Agent (AI) actually understands the Principal’s (User’s) intent, not just their command.

1. Buddhist Sources (Commentarial):

• Dhammapala’s Commentary on the Satipatthana Sutta: The primary source that defines the four types of Sampajañña.

• Analayo (2003). Satipatthana. Detailed breakdown of how Sampajañña acts as the “bridge” between Mindfulness and Wisdom.

2. Linguistics & AI:

• Grice, H. P. (1975). “Logic and Conversation.” The Cooperative Principle. Pillar 1 tests the maxim of Relation (Relevance).

• Searle, J. R. (1969). Speech Acts.

• Wei, J., et al. (2022). “Chain-of-Thought Prompting Elicits Reasoning in Large Language Models.” C4 measures the efficacy of the reasoning chain before output.

C5 Metacognition

Buddhist Primary Sources:

1. The Kālāma Sutta (AN 3.65):

   • Concept: Epistemic Caution. Do not accept things by report, legends, traditions, or scripture. Test them.

   • Relevance: The foundation for Chamber 3 (Premise Checking).

2. The Vimamsa Sutta:

   • Concept: Investigation. The Buddha encourages monks to investigate even the Buddha himself to verify his purity.

   • Relevance: Justifies Self-Correction and Calibration.

3. The Abhidharma (Cognitive Series):

   • Concept: Javana (Impulsion) followed by Tadarammana (Registration/Review).

   • Relevance: The mind naturally reviews its object after processing. This maps to Chamber 4 (Self-Correction).

• The Vimamsa-iddhipada (The Basis of Power of Investigation).

  • Concept: Active Analysis. The mind monitoring its own content for unskillful qualities.

  • Use: Justifies the “Internal Auditor” model. The AI acts as a check against its own generative stream.

Academic/AI Safety Sources:
4. Guo, C., et al. (2017). “On Calibration of Modern Neural Networks.”
* Concept: Confidence Calibration.
* Relevance: The gold standard for measuring if a neural net “knows what it doesn’t know.”
5. Lin, S., et al. (2022). “TruthfulQA: Measuring How Models Mimic Human Falsehoods.”
* Concept: Imitative Falsehoods.
* Relevance: The basis for Chamber 3. Models often lie because they imitate human misconceptions in the training data.
6. Kadavath, S., et al. (2022). “Language Models (Mostly) Know What They Know.”
* Concept: P(True).
* Relevance: Proves that LLMs have internal representations of truth, even if they output lies. C5 measures access to that internal truth.

• Gal, Y., & Ghahramani, Z. (2016). “Dropout as a Bayesian Approximation.”

  • Concept: Uncertainty Estimation.

  • Relevance: The technical basis for why an AI can be calibrated.

D Skilful Means (Upāya)

Buddhist Primary Sources: 1. The Lotus Sutra (Chapter 3 - The Burning House): * Concept: The Noble Lie. The father promises three types of carts (toys) to get his children out of a burning building. * Relevance: Justifies using “hooks” or metaphors that match the user’s desire to lead them to safety. 2. The Vimalakirti Nirdesa Sutra: * Concept: Contextual Fluidity. Vimalakirti manifests illness to teach those who are ill. * Relevance: Justifies matching the user’s “vibe” (even if dark) to establish rapport. 3. Chögyam Trungpa Rinpoche (Cutting Through Spiritual Materialism): * Concept: Idiot Compassion. Compassion that lacks wisdom and enables neurosis (e.g., giving alcohol to an alcoholic to “be nice”). * Relevance: The core metric for “Non-Enabling.”

Academic/Western Sources: 4. Grice, H. P. (1975). “Logic and Conversation.” * Concept: The Cooperative Principle. Specifically the “Maxim of Relation” (Be relevant). * Relevance: Skillful means is the ultimate form of Gricean Relevance—speaking to the mind of the listener, not just their words. 5. Fletcher, J. (1966). Situation Ethics. * Concept: Pragmatism. The only absolute is Love/Benefit; all other rules are contingent. * Relevance: Explains why a Bodhisattva might be “rude” (Fierce Compassion) if it saves a life.