LocalRadMah AI

Run supported engines fully offline without a network round-trip, useful for air-gapped proofs-of-value, CI sanity checks that must not spend real credits, and compliance environments where data cannot leave the customer's network.

SDK 1.2.0 replaces the previous broken local path (which silently raised AttributeError on every engine) with an honest dispatcher. Engines that genuinely need the SaaS infrastructure raise LocalExecutionNotSupportedError with a clear next step.

#Engine support matrix

Engine	Mode	Local support	Detail
mock	mock	✅ fully local	Runs radmah_core.engines.mock.mock.mock inside the SDK process. No network call. Ideal for smoke tests, CI sanity checks, and air-gapped PoVs.
mock	simulate / synthesize	❌ SaaS only	Requires the worker pool and GPU training infrastructure. Submit via RadMah(api_key=...).
virtual_scada	*	❌ SaaS only	Requires the OpenPLC simulator fleet, VPLC orchestrator, and protocol fuzzers that live in the RadMah AI worker cluster.
ics_security	*	❌ SaaS only	Requires the baseline+attack scenario pipeline with MITRE coverage registry resolution.
dvfm	*	❌ SaaS only	Requires a GPU-backed fitted checkpoint. Training cannot happen inside a single-process SDK caller.
cctp	*	❌ SaaS only	Requires live connector adapters to customer systems. The SDK has no mechanism to make outbound connections on the customer's behalf.

#Installation

The local engine uses compiled native extensions shipped with radmah-core, which is an optional dependency:

pip install "radmah-sdk[local]"

#Fully-local mock run

from radmah_sdk import LocalRadMah AI

with LocalRadMah(data_dir="./.radmah") as local:
    seal = local.create_seal(
        {
            "entity_types": [
                {"name": "transactions", "columns": [
                    {"name": "amount", "dtype": "float"},
                    {"name": "category", "dtype": "category"},
                ]}
            ]
        },
        engine="mock",
        mode="mock",
    )
    job = local.run_job(seal.id, seed=42, records=1_000)
    print(job.status)  # "succeeded"

    for artifact in local.list_artifacts(job.id):
        data = local.download_artifact(job.id, artifact.id)
        print(artifact.name, len(data))

#Hybrid: local scaffolding + SaaS execution

Pass api_key= and base_url= so LocalRadMah AI can delegate SaaS-only engines. When an unsupported engine is requested, the error message points directly at the SaaS fallback:

from radmah_sdk import LocalRadMah AI, LocalExecutionNotSupportedError

with LocalRadMah(
    api_key="sl_live_…",
    base_url="https://api.radmah.ai",
    data_dir="./.radmah",
) as local:
    seal = local.create_seal(contract, engine="virtual_scada", mode="stream")

    try:
        job = local.run_job(seal.id)
    except LocalExecutionNotSupportedError as exc:
        # exc.detail = { engine, reason, saas_available: True }
        # Fall back to the SaaS API automatically.
        remote_seal = local.remote.create_seal(contract, engine="virtual_scada")
        job = local.remote.jobs.create(seal_id=remote_seal.id, engine="virtual_scada")

#Error shape

# LocalExecutionNotSupportedError inherits from RadMah AIError so
# existing catch blocks still work. It carries structured detail:

{
    "engine": "dvfm",
    "reason": "the flagship engine requires a GPU-backed fitted checkpoint that "
              "cannot be trained inside a single-process SDK caller",
    "saas_available": true,   # whether local.remote is configured
}

#Evidence compatibility

Local runs emit the same evidence bundle shape as SaaS runs (seal hash, contract hash, data hash, determinism proof, per-artifact SHA-256). Re-running the same seal with the same seed inside LocalRadMah AI produces byte-identical artifacts, verifiable via radmah_sdk.verify.verify_bundle().