12 field operators for policy conditions — equality, comparison, membership, pattern matching.

Condition Format

interface FieldCondition {
  field: string;       // dot-notation path (e.g., "model", "estimated_cost", "messages.0.role")
  operator: FieldOperator;
  value?: unknown;     // not needed for exists/not_exists
}

Presence

Operator	Description	Example
`exists`	Field is present and not null/undefined	`{ field: "scope", operator: "exists" }`
`not_exists`	Field is absent or null/undefined	`{ field: "cost_center", operator: "not_exists" }`

Equality

Operator	Description	Example
`eq`	Strict equality	`{ field: "model", operator: "eq", value: "kimi-k3" }`
`neq`	Not equal	`{ field: "tier", operator: "neq", value: "free" }`

Comparison (numeric)

Operator	Description	Example
`gt`	Greater than	`{ field: "estimated_cost", operator: "gt", value: 1000 }`
`gte`	Greater than or equal	`{ field: "budget_remaining", operator: "gte", value: 1000 }`
`lt`	Less than	`{ field: "budget_remaining", operator: "lt", value: 100 }`
`lte`	Less than or equal	`{ field: "max_tokens", operator: "lte", value: 4096 }`

Membership

Operator	Description	Example
`in`	Value is in the array	`{ field: "model", operator: "in", value: ["gpt-4o", "gpt-4o-mini"] }`
`not_in`	Value is not in the array	`{ field: "model", operator: "not_in", value: ["claude-opus-4-6"] }`

Pattern Matching

Operator	Description	Example
`contains`	String contains substring	`{ field: "model", operator: "contains", value: "opus" }`
`regex`	Regex match (safe: max 200 chars, no nested quantifiers)	`{ field: "model", operator: "regex", value: "^claude-.*" }`

Hard rules fail closed on a field they cannot read

Every operator except exists / not_exists returns indeterminate when the field does not resolve. Beyond that it depends on the operator: gt, gte, lt and lte are also indeterminate when either side is not a number, contains and regex when either side is not a string, and in and not_in when the rule value is not an array. eq, neq, in and not_in compare with strict equality, so a defined value of another type is a determinate comparison (false for eq and in, true for neq and not_in), and null is a value that compares determinately. In a rule with enforcement: hard an indeterminate condition counts as satisfied, so the rule still fires; in a soft rule it counts as unmet. A hard rule on a field the SDK never populates therefore denies every call. Lead such a rule with an exists condition on the same field when it must only apply to calls that carry it.

The controller injects model, tier, estimated_cost, budget_remaining, budget_remaining_after, scope and cost_center on every LLM call (plus budgetFractionRemaining / budgetRunwayHours for a call inside a withCostCenter scope that declared an allocation, and action_kind / action_name instead of model for a governed action). The caller's request params are spread in underneath, so max_tokens and messages.0.role resolve when the request carries them.

Values are checked when the file loads

gt / gte / lt / lte need a finite number, in / not_in an array, contains / regex a string, and every operator except exists / not_exists needs a value. A regex the evaluator would refuse (invalid, over 200 characters, or nested quantifiers) is refused here instead. Any such fault throws PolicyLoadError and loads none of the file's rules.

Dot-Notation Field Resolution

Fields support dot-notation for nested objects:

conditions:
  - field: messages.0.role
    operator: eq
    value: system

Full YAML Example

- name: limit-opus-in-prod
  effect: deny
  enforcement: hard
  severity: high
  conditions:
    - field: model
      operator: contains
      value: opus
    - field: estimated_cost
      operator: gt
      value: 500

Field Operators