Guard Policies

Guard policies control what DriftGuard does when before_step() detects a warning that meets the confidence threshold.

Policies

Policy	Behavior
warn	Return warnings — agent decides what to do
block	Raise GuardrailTriggered exception
acknowledge	Raise GuardrailAcknowledgementRequired unless acknowledged=True
record_only	Skip review entirely — store memory only

Setting the default policy

from driftguard import DriftGuard, DriftGuardSettings

guard = DriftGuard(
    settings=DriftGuardSettings(
        guard_policy="block",
        guard_min_confidence=0.8,
    )
)

Overriding per call

review = guard.before_step(
    "increase salt",
    policy="acknowledge",
    min_confidence=0.75,
)

Warn (default)

The default policy. Warnings are returned in the RetrievalResponse but no exception is raised. The agent can inspect and decide.

review = guard.before_step("increase salt")

for warning in review.warnings:
    print(f"{warning.trigger} → {warning.risk} ({warning.confidence:.2f})")

Block

Raises GuardrailTriggered when warnings exist above min_confidence.

from driftguard import GuardrailTriggered

try:
    guard.before_step("increase salt", policy="block", min_confidence=0.8)
except GuardrailTriggered as e:
    print(f"Blocked: {e}")
    # agent takes alternative action

Acknowledge

Raises GuardrailAcknowledgementRequired unless the caller passes acknowledged=True. Useful for human-in-the-loop flows.

from driftguard import GuardrailAcknowledgementRequired

try:
    guard.before_step("increase salt", policy="acknowledge")
except GuardrailAcknowledgementRequired:
    # show warning to human, get confirmation
    guard.before_step("increase salt", policy="acknowledge", acknowledged=True)

Record only

Skips the retrieval review entirely. Memory is still stored when you call record(). Useful for observation-only phases.

review = guard.before_step("increase salt", policy="record_only")
# review.warnings == []
# review.confidence == 0.0

guard_step decorator

Apply guardrails to any agent step function:

from driftguard import guard_step

@guard_step(guard, policy="block", min_confidence=0.8)
def execute_action(action: str):
    ...

Custom context extraction:

@guard_step(
    guard,
    input_getter=lambda payload: payload["next_action"],
    policy="acknowledge",
    acknowledged_getter=lambda payload: payload["confirmed"],
)
def execute(payload: dict):
    ...