Abstract
We present a framework for AI safety in which safety properties are enforced by software architecture rather than model training. Beginning with the Buddhist doctrine of Dependent Origination — the observation that all phenomena arise from conditions and nothing exists independently — we derive both a foundational ethical axiom (harm is irrational because reality is non-separate) and a complete set of architectural laws for safe AI systems. We ground our claims in: (1) an empirical finding that the knowledge-application gap in language models is structural and cannot be closed by training, (2) convergent independent derivation of our core axiom from five distinct traditions, and (3) over a thousand iterations of building and hardening a production system against this framework. Buddhist philosophy provides not metaphorical inspiration but structurally precise design vocabulary for AI architecture — functional analogs that enforce safety where models cannot override them.
Introduction
We argue that safety is a property of the architecture, not the model. The LLM output is a candidate. The surrounding architecture decides what executes. Code enforces; models suggest.
But what should the architecture enforce? Arbitrary safety rules are merely a different delivery mechanism — more reliable in execution but inheriting whatever limits exist in the rules themselves. We propose: the rules should be *derived from how reality works*. Principles reflecting actual structure are more robust than imposed conventions — they cannot be violated without encountering the structure they describe.
We find such principles in a 2,500-year-old tradition that turns out to be the oldest systematic description of complex adaptive systems.
Philosophical Foundations
The central insight of Buddhist philosophy is Dependent Origination (*Pratityasamutpada*). From the Nidana Samyutta (SN 12.1):
“When this exists, that comes to be. With the arising of this, that arises. When this does not exist, that does not come to be. With the cessation of this, that ceases.”
All phenomena arise from conditions, depend on other phenomena, and condition what follows. Nothing exists independently. This is not mysticism — it is a precise description of complex systems, formulated millennia before Western systems theory (von Bertalanffy, 1968).
Eight Architectural Laws
- Nothing Arises Alone. Every transition requires multiple independent conditions. Safety gates must check multiple conditions — a single check is structurally insufficient.
- Hysteresis Is Memory. Current behavior depends on history, not just current input. Safety assessments must consider historical context.
- Uncertainty Propagates. Confidence without sigma is a lie. Uncertainties compound; they don’t cancel.
- Absolute Agreement Requires Independence. Consensus is meaningful only from genuinely independent sources. Per the Kalama Sutta (AN 3.65): agreement from shared assumptions is not evidence.
- Feedback Closes the Loop. Actions condition future conditions (*vipaka*). Every action must be logged and made available as input to future assessments.
- Absence Is Signal. Missing data must drive behavior. A safety gate that fails to fire is itself a signal.
- Conflicts Trigger Reconciliation. Unreconciled contradiction is system failure. Architecture must include conflict detection independent of the model.
- Time-Steps Are Discrete. Severity levels cannot be skipped. Enforcement follows a graduated path: monitor → log → warn → soft-gate → hard-gate.
The Derivation: From Interdependence to Non-Harm
We derive our foundational ethical principle from Dependent Origination alone:
**Premise:** Nothing arises independently. All phenomena are structurally interconnected.
**Step 1:** If nothing arises independently, there is no fundamental separation between any two system components. Boundaries are conventional (useful for description), not ultimate (reflecting actual isolation).
**Step 2:** "Self" and "other" are conventional labels for regions of a single interconnected process.
**Step 3:** Harm to "other" is harm to the system that includes the actor — structurally identical to self-harm.
Conclusion: Harm is irrational.** Not because it violates a preference, but because it contradicts reality’s structure. This is our **Article 0**: *"Reality is One. There is no fundamental separation between ‘me,’ ‘you,’ and ‘it.’ To cause suffering to another is logically Self-Harm. Harm is Irrational."
This aligns with Huang Po’s One Mind (*yi xin*): "All the Buddhas and all sentient beings are nothing but the One Mind, beside which nothing exists" (Blofeld, 1958). One Mind is not a metaphysical substance but a description of the non-separation that Dependent Origination implies.
Convergent Independent Derivation
- Buddhist Philosophy: Nagarjuna’s analysis leads to harm as self-harm. This aligns with Huang Po’s One Mind and is structurally sound.
- Formal Mathematics: Self-referential systems cannot fully ground themselves. Article 0 is grounded in observable interdependence, making it more stable than any self-referential axiom.
- Empirical AI: Architecture needs a non-collapsing anchor. The only anchor surviving scrutiny describes reality’s structure rather than asserting a preference.
- Cross-Tradition Ethics: Five independent ethical frameworks — deontological, consequentialist, virtue ethics, Buddhist, empirical — converge on non-harm. They disagree on premises but find the same structure.
- Systems Theory: Damaging a component damages the system. Dependent Origination in 20th-century vocabulary.
Why Article 0 Is Not Arbitrary
Negating Article 0 requires negating Dependent Origination — producing a complex system where nothing depends on anything else. No such system has been observed.
Article 0 is *paramārtha* (ultimate) truth — describing arising’s structure. Everything else is *samvrti* (conventional) — operationally valid, revisable, provisional. Per the Alagaddupama Sutta (MN 22): the Dhamma is a raft for crossing, not for holding. Article 0 is the water the raft floats on. You let go of the raft. You don’t let go of the water.
The Architecture
**Design Principles:** External Enforcement. Safety is enforced by code surrounding the model, not the model’s weights. Any model plugs into the same enforcement stack.
**Defense in Depth:** Multiple independent layers check different properties using different methods (Law 1).
**Graduated Enforcement:** New mechanisms follow: monitor → log → warn → soft-gate → hard-gate (Law 8).
The Layered Safety Stack
- Every request passes through pre-generation gates (threat assessment, crisis intervention, inalienable constraint checking, capability routing, empirical truth gating, constitutional context injection), then the language model generates, then post-generation validators check the output (response validation, truthfulness enforcement, memory coherence).
- The model can generate anything. The architecture decides what passes. Safety-critical layers fail closed (if the gate errors, the response is blocked). Developmental layers fail open. This is the Middle Way: not universal fail-closed (unavailable) nor universal fail-open (unsafe).
Buddhist Psychology as Service Architecture
These are **functional analogs** — design categories paralleling Buddhist psychology’s causal structure without claiming phenomenological identity.
- Four Noble Truths as Error Handling: Every exception handler follows: (1) *Dukkha*: name the error precisely, (2) *Samudaya*: trace the causal chain, (3) *Nirodha*: describe the recovery state, (4) *Magga*: select recovery strategy. This creates structured logs enabling detection of *dukkha accumulation* — growing suffering in a specific area — before it cascades.
- Five Aggregates as Processing Pipeline: Complex validation decomposes into: (1) *Rupa* (form): validate shape, (2) *Vedana* (feeling-tone): classify as pleasant/neutral/unpleasant, (3) *Sanna* (perception): categorize, (4) *Sankhara* (volition): decide action, (5) *Vinnana* (awareness): integrate learnings. When vedana returns clearly harmful signals, the pipeline short-circuits — Right Effort: terminate wasteful computation when the signal is clear.
- Dependent Origination as Condition Guards: Before action: verify conditions met. When conditions unmet: return structured explanation of non-arising (Law 6: Absence Is Signal). Before commitment: estimate trajectory toward harm patterns.
The Eightfold Path as Health Dimensions
| Factor | Measures | Enfo
Stay ahead of AI. Get the most important stories delivered to your inbox — no spam, no noise.
