This demo drives a local model in LM Studio through the
Context Relay Protocol v4.1 runtime. Watch the canonical
CRP-* header namespace, signed session tokens, window-level
HMAC chain, CKF memory, DPE risk scoring, safety budget, and compliance
classification appear in real time.
Send a prompt. CRP returns a governed response plus every governance signal.
Facts extracted from model answers are stored, recalled, and cited across windows.
CRP windows form a DAG: LINEAR, FAN_OUT, FAN_IN, and BRANCH. Each window carries a continuation ID and a window HMAC.
Every window has a HMAC chained from its parent. Tamper with one window and the chain breaks.
Every governed response carries the canonical CRP v4.1 header set.
Declare a safety policy and watch CRP halt generation when risk crosses the threshold.
CRP classifies each response against EU AI Act, GDPR PII, NIST, and ISO 42001 controls.
CRP headers travel with A2A/MCP agent messages. The safety budget acts as a circuit breaker across hops.
CRP runs Layer 1 validation, Layer 2 injection/PII scanning, DPE risk
scoring, policy enforcement, safety budget, and HMAC audit chains.
These checks are active: they happen on every request,
are surfaced in CRP-* headers, and are recorded in the
tamper-evident audit log.
Unlike black-box guardrails, CRP publishes its full safety surface: capabilities, defaults, coverage rules, and explicit out-of-scope items.