Saltar al contenido principal

compute

Since v1.12.0 · Top-level declaration

Grammar​

compute <Name> [(<params>)] [-> <ReturnType>] {
    shield: <ShieldRef>      # optional — defence layer
    # Additional fields (model, effort, seed, parallelism, ...) are
    # accepted by the parser's permissive skip-value path; the
    # type checker validates them against the runtime's binding
    # registry.
}

compute declares a typed binding to a compute backend — the surface that pins a flow to a specific model, an effort level, an explicit random seed, a parallelism strategy. Where tool declares external capability the model can call, compute declares how the cognitive layer itself runs.

The most common production use: pinning a deterministic backend for a regulated flow (banking decision, clinical diagnosis, legal analysis). The runtime carries the compute binding into every emission's audit row, so the deployed model is auditable.

Surface​

compute is a top-level declaration. It is not nested inside a flow. A flow references a compute binding through the apply: <ComputeName> flow-step pattern or through a deployment-level compute: field on the bound run.

axon-W006 — a step's apply: <Compute> for model selection is a no-op the compiler warns on; declare the capability need on the step with requires_context: (v2.22.0) and let the resolver pick the model. See axon://primitives/step.

shield FinancialShield {
    scan:       [pii_leak, prompt_injection]
    on_breach:  halt
    severity:   critical
    compliance: [PCI_DSS, SOC2]
}

compute LoanUnderwriterCompute {
    shield: FinancialShield
}

Fields​

shield: (optional)​

A single identifier referencing a declared shield. The compute binding's invocations route through the shield's scan list before commitment.

The parser currently models shield: as the only structured field. Additional adopter-facing fields (e.g. model:, effort:, seed:, parallelism:) are parsed permissively (via the skip_value fallback) and validated at deploy time by the runtime's compute-binding registry rather than at parse time. A future cycle will tighten this surface to a typed AST node; until then, the runtime is the source of truth for which fields a given backend honours.

Anatomy quirks​

The parser accepts optional parameters / return type between the name and the brace (compute LoanUnderwriterCompute(...) -> T { ... }). This is legacy syntax preserved for AST parity with the Python frontend; new code should use the plain compute <Name> { ... } form.

Runtime behaviour​

compute lowers to a ComputeDefinition IR node. At deploy time, the runtime resolves the binding against its compute-backend registry and pins the configuration. Every flow that consumes this compute carries the binding's name into its audit rows — compute:<name>:invoked with (model, effort, seed, latency, cost).

For deterministic deployments, declare seed: explicitly. The runtime preserves the seed across retries so replay reproduces the exact tokens emitted (subject to model availability).

What this primitive is NOT​

  • Not a tool. A tool declares an external capability the model can call (web search, code interpreter). Compute declares the cognitive layer itself — which model, on which backend, with what effort.
  • Not infrastructure. A compute declaration does NOT provision GPUs or model endpoints. The runtime resolves the binding against ALREADY-running compute substrate — the declaration is the typed handle, not the IaC.
  • Not optional for regulated deployments. Production HIPAA / SOX / PCI flows ship explicit compute bindings so the audit trail records the exact model used per emission. Anonymous compute (no declaration) emits axon-W015.

See also​

  • axon://primitives/tool — external-capability counterpart.
  • axon://primitives/shield — the most common companion field.
  • axon://primitives/run — the deployment-level binding site (effort: field).
  • axon://primitives/agent — agents declare their own compute indirectly via tools: + budgets.