spolm-checks 0.1.0

Checks-as-code for AI agents: define invariants once, enforce them in replay, simulation, and at runtime.

spolm-checks

Checks-as-code for AI agents. Define an invariant once — enforce it in replay (CI regression), simulation (certification), and inline at runtime (action gate).

Deterministic first: predicate (CEL — sandboxed, guaranteed-terminating), schema (JSON Schema), and policy (trajectory patterns: never X, X only after Y, X at most N, tool allow/deny lists). No LLM-as-judge in the core engine, by design.

Example check

check:
  id: refund-within-order-total
  version: 1
  severity: block
  scope:
    workflow: refunds-agent
    target: tool_call
    tool: refund
  assert:
    type: predicate
    expr: "args.amount <= context.order.total && args.currency == context.order.currency"
  provenance:
    source: mined
    cluster_id: c_8842
  on_fail:
    runtime: block
    simulation: record_violation
    replay: fail_regression

CLI

spolm-check lint ./checks
spolm-check run ./checks --traces traces.jsonl

run exits non-zero if any block-severity check has violations — wire it straight into CI.

Python

from spolm_checks import load_suite, IRTrace, run_replay

checks = load_suite("./checks")
report = run_replay(checks, traces)   # traces: list[IRTrace]

Apache-2.0.