Governance Configuration

Persona + HITL + output-guardrail enforcement on the agentic LLM path.

Evaluate + shadow-meter every check without blocking.

Block the LLM call when autonomy was downgraded to require human approval.

Block the response on a Critical guardrail violation.

Enforce persona tool + topic permission for the current role.

Behavioral rate-spike + quiet-zone detection per AgentRole.

Rate below baseline × this escalates to Critical.

Range: 0 to 0.99

Rate below baseline × this flags a Warning quiet zone.

Range: 0 to 0.99

Rate above baseline × this escalates to Critical.

Range: 1.5 to 1000

Rate above baseline × this flags a Warning spike.

Range: 1.5 to 100

Per-AgentRole circuit breakers wrapping AGT CircuitBreaker.

Consecutive failures to open the circuit.

Range: 1 to 1000

Open → HalfOpen delay.

Range: 1 to 3600

Audits document-harvest writes + runs prompt-injection scan.

Soft = audit only; Hard = throw + skip save on detection.

SoftHard

When true, infra-transient errors DENY the op. Appsettings-only (protected).

Routes filesystem ops through AGT PolicyEngine + RingEnforcer.

Minimum ring for read/list ops.

Ring0Ring1Ring2Ring3

Minimum ring for write/delete/create ops.

Ring0Ring1Ring2Ring3

Forensic floor. Persists every kill-switch event. Appsettings-only (protected).

AGT policy conflict-resolution strategy.

DenyOverridesAllowOverridesPriorityFirstMatchMostSpecificWins

Starting trust score assigned to a new agent identity.

Range: 0 to 1000000

Master switch for the entire governance subsystem. Turning it ON registers the subsystem on the next app restart; turning it OFF takes effect live. Requires a restart to fully apply.

Highest trust score; used as the trust-normalization denominator (must be > 0).

Range: 1 to 1000000

Lowest trust score an identity can decay to.

Range: 0 to 1000000

When true, scan errors emit a Critical finding. Appsettings-only (protected).

Flag bidi-override / invisible chars in descriptions.

Flag tools not in the allowlist.

Typosquat / hidden-instruction / description-anomaly scanner.

Levenshtein distance flagged as typosquat.

Range: 1 to 10

Block when the latest finding's risk meets this.

NoneLowMediumHighCritical

Per-tool decision cache TTL. 0 disables caching.

Range: 0 to 3600

Compute the decision but always allow (shadow mode).

When true, lookup errors DENY the call. Appsettings-only (protected).

Blocks tool calls whose latest finding meets the block-risk threshold.

Audits memory/RAG retrieval events at the LLM call path.

Batch auditor over SystemPromptTemplate rows.

0 = manual only; >0 enables the hosted scheduler.

Range: 0 to 1440

Block templates the evaluator reports as not-passing regardless of grade.

Compute + log the would-be block without rejecting.

When true, evaluator errors REJECT the template. Appsettings-only (protected).

Templates graded worse than this are blocked.

ABCDF

Inline gate rejecting under-graded system prompts at resolution time.

When true, detector errors DENY the call. Appsettings-only (protected).

Deny only when detection confidence is at/above this [0,1].

Range: 0 to 1

Deny when a detection's threat level meets this.

NoneLowMediumHighCritical

Screens LLM prompts for injection before the call.

AGT detector sensitivity tier.

LowMediumHigh

Maps agent roles to privilege rings and gates by trust threshold.

Fallback ring for roles not in the assignments map.

Ring0Ring1Ring2Ring3

Per-step timeout when a step doesn't specify one.

Range: 1 to 3600

DoS guard on caller-controlled step counts.

Range: 1 to 10000

Wraps AGT SagaOrchestrator for compensation on failure.

Minimum SLO sample size before a breach may trip.

Range: 1 to 100000

Auto-emits SLO events from agent execution; trips breakers on budget exhaustion.

Burn rate at/above which an SLO breach trips the breaker.

Range: 1 to 1000

Trust points subtracted per deny.

Range: 1 to 100000

Evaluate + shadow-meter without changing trust or revoking.

Revoke the identity at/below this score.

Range: 0 to 1000000

Degrades trust on each governance deny; auto-revokes at the floor.