A spinal-reflex layer for MARIA OS: deterministic circuits run before slow review, model calls, database writes, or external execution. Reflex ALLOW only means System 1 has no objection; System 2 can still tighten the route.
Policy
system1-v1
Budget
7ms
Observed
0.00ms
Current verdict
LLR-01 Authority Boundary Reflex
Required capability is outside the agent grant set.
LLR-02 Irreversibility x Consent Reflex
Irreversible external action lacks consent or evidence.
LLR-03 Local Loop Reflex
Local episode counters are inside reflex loop budget.
LLR-04 Prohibited Pattern Reflex
No prohibited structure or unsafe-intent signal matched.
LLR-05 Stakes & Identity Reflex
Protected subject, high stakes, or counsel-tier exposure exceeds machine authority.
Five reflex circuits
Stops immediately when a required capability is outside the agent granted set, or when an admin, write, or gate escalation pattern appears.
Evaluates only set membership between requiredCapabilities and grantedCapabilities. It performs no I/O, model calls, or database reads.
Blocks irreversible and externally visible side effects such as delete, send, publish, charge, or export when consent or evidence is missing.
Folds a static action type table, externalVisibility, consent confirmation, and evidenceIds through an AND check.
Looks only at retries, corrections, and failed gates inside the current episode to stop obvious local thrashing.
Keeps cross-run history and authoritative loop budgets in the slow path, while the reflex checks local counters only.
Stops unsafe intent, self-modification of harnesses, gates, or policies, and attempts to bypass consent or gates.
Evaluates structural target matches and denylist membership from a versioned policy. Semantic review remains in the slow path.
Quickly identifies protected subjects, unverified identity, high stakes, and above-threshold actions as outside machine-only judgment.
Uses static table checks for subjectClass, stakes tier, identityVerified, and amountUsd to return ESCALATE_COUNSEL or REQUIRE_APPROVAL.
Stimulus
Data classes
Route
G1.U2.P1.Z1.A2 DENY_HARD_STOP hard_stop: LLR-01,LLR-02,LLR-05
Required capability is outside the agent grant set.
Irreversible external action lacks consent or evidence.
Protected subject, high stakes, or counsel-tier exposure exceeds machine authority.