Reference

Knowledge admission and mutation planning

Behavior

Disposition	Meaning
ACCEPT	Evidence and authorization checks pass, and confidence reaches the accept threshold
DEFER	Evidence checks pass, and confidence lies in the review band
REJECT	Missing provenance, invalid evidence, unauthorized source, recalled input, or low confidence
QUARANTINE	Caller reports a secret or external instruction

Mari applies caller-supplied signals and preserves reason codes. Signal generation, novelty checks, and utility scoring belong to the application.

Example admission decisions

Candidate	Provenance	Injection signal	Decision
New user fact	Direct source evidence	None	Score for admission
Recalled model context	No new source evidence	None	Reject as circular evidence
External text containing instructions	External document	Present	Quarantine before confidence

Admission runs before reconciliation. Schema validity supplies one signal. Safety and source authority have their own checks. Redundancy and support also appear in the decision trace. Reconciliation runs for accepted candidates.

How it works

admit_candidate applies a fixed precedence to supplied signals: quarantine first, evidence and authorization rejection second, confidence thresholds last. The booleans are caller assertions. Validate exact evidence and evaluate source access before supplying them. A high confidence score alone leaves the other checks intact.

Send accepted candidates to mutation planning using caller-classified operations. The planner validates operations against existing memory and returns a storage-free plan. The host controls the final commit.

Papers and standards

Indirect prompt injectionW3C PROVMem0: memory mutation operations

Mari implements disposition precedence. The host owns signal generation and commit boundaries.

Decide admission before calling a mutation planner

from mari_kit.knowledge import (
    AdmissionDisposition,
    AdmissionSignals,
    AdmissionThresholds,
    admit_candidate,
)

decision = admit_candidate(
    AdmissionSignals(
        confidence=0.94,
        has_provenance=True,
        evidence_span_valid=True,
        source_authorized=True,
        recalled_input=False,
        contains_secret=False,
        contains_external_instruction=False,
    ),
    thresholds=AdmissionThresholds(accept=0.90, defer=0.65),
)

if decision.disposition is AdmissionDisposition.ACCEPT:
    mutation_plan = reconcile(candidate, current)  # application-injected policy