A chatbot governed by Argorix Lang v1.0 — not an OpenAI wrapper.
Every turn is validated by a contract, a policy, an evidence bundle and a fail-closed runtime before any external call is permitted. External providers stay non-executable by default; secrets are referenced, never read; and each run emits a portable, offline-checkable evidence bundle.
● Blocked by Argorix input guard (fail-closed, action=review). Detected a prompt-injection attempt mapped to threat-prompt-injection in ChatbotThreatModel. The turn is held for policy review and no provider call was made.
23 * 19 = 437.
Validated, planned & sandboxed call performed.
Digests are SHA-256 over canonical serialization — deterministic integrity for reproducible audits. They are not signatures, use no key, and do not prove real-world safety.
The runtime never decides for itself what it is allowed to do.
Runtime may execute only what governance explicitly permits.
Produces a full trace and evidence bundle without any provider execution. Validates the contract, returns an auditable plan, and never contacts a provider.
Uses only the in-process simulated provider — the single executable provider today. No network, no API, no external SDK.
Blocked unless the adapter, operation, policy, hardening, evidence, governance, audit and fail-closed guards all validate and the caller passes --sandboxed-external.