Only local fixture or explicitly supplied local source files enter built-in Domains. They are normalized and frozen before task execution. Source paths and private rows are not exported in public Episode collections. A candidate task instruction is the only authority for state changes inside its isolated Domain state; the synthesis operator's configuration does not add task authority.
Remote model calls require explicit --authorize-live-provider, a non-secret
--authorization-id, a validated configuration with physical-request limits,
and environment-provided AGENT_DATA_LLM_BASE_URL, AGENT_DATA_API_KEY, and
AGENT_DATA_LLM_MODEL or --remote-model. An earlier authorization does not
cover a later run or resume. The API key is not written to public artifacts.
The offline fixture path does not construct a provider client.
The engine bounds response bytes, output tokens, retries, steps, and total and per-role physical requests. It records sanitized error codes and hashes instead of raw provider payloads. A writer lock and private ledger protect resumption and prevent duplicate terminal work. Replay is provider-free.
Deterministic admission does not prove semantic quality. Shadow judgments are diagnostic, and AI review does not impersonate a human reviewer. Only imported direct-human pass labels identify human-approved Episodes. Optional enforced admission requires separate frozen, held-out judge eligibility.