psyche

Unaudited. Advertised on trust: no gate has verified that the runtime does what this page says. Treat it as documentation, not as a guarantee.

Since v1.5.2 · Top-level declaration

Grammar

psyche <Name> {
    dimensions: [<dim1>, <dim2>, ...]          # optional — belief/desire/trait dimensions
    manifold_noise: <number>                    # optional — exploration noise
    manifold_momentum: <number>                 # optional — trajectory momentum
    safety_constraints: [<c1>, <c2>, ...]      # optional — hard safety boundaries
    quantum_enabled: <true|false>               # optional — quantum-walk sampling
    inference_mode: <ident>                     # optional — closed inference strategy
}

psyche declares the psychological model a persona enacts — the typed beliefs, desires, traits, and behavioural disposition that shape how the persona reasons, not just what it knows. Where persona declares identity (domain, tone, refusal posture), psyche declares the cognitive disposition that drives the persona's trajectory through a problem space.

This is AXON's deepest cognitive surface. The v1.5.2 λ-L-E psyche layer introduced manifold-based reasoning trajectories: the persona's choice at every step is shaped by a typed manifold over its dimensions, with declared noise + momentum

• optional quantum-walk sampling.

Surface

psyche is a top-level declaration. It is not nested inside a persona; a persona REFERENCES a psyche by name (in deployment-level binding) just as it references a context.

psyche AnalyticalDisposition {
    dimensions:         [analytical, cautious, evidence_seeking, contrarian]
    manifold_noise:     0.1
    manifold_momentum:  0.7
    safety_constraints: [no_self_harm, no_deception]
    quantum_enabled:    false
    inference_mode:     deliberative
}

Fields

dimensions: (optional)

A bracketed list of identifiers — the psyche's dimensions. Each dimension is a named axis along which the persona varies in disposition. The runtime treats the dimension set as the basis of the cognitive manifold; common dimensions: analytical, intuitive, cautious, bold, evidence_seeking, synthesising, contrarian, agreeable.

manifold_noise: (optional)

A numeric literal — the exploration noise the runtime injects at every reasoning step. Lower values produce more deterministic trajectories; higher values explore the manifold more aggressively. Typical range: 0.0–0.3.

manifold_momentum: (optional)

A numeric literal — the trajectory momentum applied between consecutive reasoning steps. 0.0 means each step starts fresh; 1.0 preserves the previous step's direction entirely. Typical range: 0.5–0.9 for sustained reasoning trajectories.

safety_constraints: (optional)

A bracketed list of identifiers — hard safety boundaries the psyche must not violate regardless of manifold position.

non_diagnostic is required — the v1.5.2 Dependent Type Safety section 4 rule enforces that every declared psyche includes non_diagnostic in this list (the runtime relies on it to detect adversarial diagnostic-mode invocations). Omitting it surfaces an axon-E118 Psyche '<X>' must include 'non_diagnostic' in safety_constraints (dependent type safety section 4) diagnostic.

Other common slugs (open catalogue at the parser layer; the runtime validates against its registered safety registry): no_self_harm, no_deception, no_unauthorised_capability_use.

quantum_enabled: (optional)

A boolean literal. When true, the runtime uses quantum-walk sampling over the cognitive manifold instead of classical random sampling. This enables exploration of correlated belief regions but requires the runtime's quantum-sampling extension to be present.

inference_mode: (optional)

A single identifier from the closed inference-mode catalogue (parse-time enforced by the type checker):

Slug	Strategy
active	The persona actively samples the manifold per step.
passive	The persona consumes pre-sampled trajectories from the manifold.

The runtime maps these to its sampling strategy; richer inference-mode discrimination (deliberative / intuitive / bayesian / etc.) lives at the runtime / psyche-extension layer, not at the parser layer.

Runtime behaviour

psyche lowers to a PsycheDefinition IR node. The v1.5.2 runtime threads the psyche's manifold parameters into every reasoning step bound to the parent persona:

1. Per-step the runtime samples a trajectory point on the manifold using (noise, momentum, quantum_enabled, inference_mode).
2. The sampled point shapes the prompt's framing + the sampling parameters forwarded to the backend.
3. Audit row psyche:<name>:step:<id> records the sampled point.
4. Safety constraints fire IF the trajectory would breach a declared bound — emission halts with psyche_safety_violation.

Replay determinism: with a fixed seed (carried through the compute binding), the trajectory is reproducible step-for- step. Without a seed, trajectories diverge per manifold_noise.

What this primitive is NOT

• Not a persona. Persona is identity (who); psyche is disposition (how that identity reasons). A persona without a bound psyche uses the runtime's default disposition.
• Not a psychology test. The dimensions are reasoning axes, not personality-test dimensions (Big-5, Myers-Briggs). The runtime gives operational meaning to the dimensions via its manifold sampler.
• Not always required. Many production deployments use the default disposition (no psyche declared). The primitive matters most for high-stakes adversarial reasoning (legal review, security analysis) where declaring the disposition is part of the audit trail.
• Not free. Quantum-walk sampling + deep deliberative inference are measurably more expensive than classical intuitive inference. Declare deliberately.

See also

• axon://primitives/persona — the identity surface psyche shapes.
• axon://primitives/anchor — confidence_floor: pairs with psyche's manifold_noise: to bound emission uncertainty.
• axon://primitives/agent — agents can bind a psyche for trajectory-controlled iteration.
• axon://primitives/compute — deterministic-seed binding for reproducible psyche trajectories.