Malia
Safety
AXIOM
Verification × governance
Deterministic layer
Proposed architecture
INSTINCT proposes.
AXIOM decides.
Learned autonomy proposes actions. A deterministic verification layer decides whether each one is inside the authorized envelope — before the machine moves.
(01) — Position in MALIA OS
AXIOM is the lowest layer of the shared architecture — the one every learned layer above it must answer to. Nothing acts without passing through it.
- L1INSTINCTPerception, navigation, reasoning, planning and learning
- L2PHYSIS + ORACLEShared physical state, sensor fusion and uncertainty
- L3LEXICONCommunication across different intelligent machines
- L4COLONYCoordination, task allocation and fleet management
- L5AXIOMAuthorization, safety checks and intervention controls
(02) — The flow
Every action,
audited.
From sensor to audit record, a proposal travels one path. AXIOM issues exactly three verdicts.
- 01SENSE
- 02INSTINCT
- 03PROPOSE
- 04AXIOM
- 05AUTHORIZEESCALATEDENY
The only three verdicts. ESCALATE hands the decision to a human.
- 06ACT
- 07AUDIT
(03) — What AXIOM requires
The envelope,
made explicit.
Verification is only as strong as what it checks. Eight requirements define the authorized envelope for every mission.
- 01
Allowed-action sets
Every mission defines the envelope of actions a machine may take. Nothing outside the set exists to be chosen.
- 02
Geofences & zones
Hard spatial boundaries — where a machine may go, where it may not, and where it must stop.
- 03
Health limits
Continuous monitoring of the machine’s own state. Degraded hardware shrinks the authorized envelope.
- 04
Confidence thresholds
Learned proposals carry calibrated uncertainty. Below threshold, AXIOM escalates or denies.
- 05
Safe-state transitions
Defined degraded modes — hold, land, dock, stop — reachable from every operating state.
- 06
Abort & return behaviour
Specified abort paths and return behaviour when links degrade, missions fail, or rules are breached.
- 07
Role-based authorization
Actions are authorized against the operator’s role and the mission rule set — not just the machine’s plan.
- 08
Immutable mission logs
Every observation, proposal, verdict and override is written to a tamper-evident audit record.
(04) — Human authority
A person decides, always.
Autonomy operates inside authority that humans set and can revoke. That is structure, not policy.
01 — Operator override
A human operator can interrupt, redirect or halt any machine at any time. Override is a designed pathway, not an emergency hack.
02 — Mission boundaries
The envelope — zones, actions, thresholds, abort conditions — is defined by people before deployment, and machines cannot widen it.
03 — Use of force
Any future system implicated in use-of-force decisions is built around applicable human-judgment requirements — consistent with the principle of appropriate human judgment in DoD Directive 3000.09.
(05) — Data governance
What machines know,
and who may know it.
- 01ProvenanceSensor data, models, configurations and versions carry lineage — what was collected, where, by which system, under which build.
- 02Role-based accessData and control authority follow the operator’s role. Access is granted per mission, per role, per need.
- 03RetentionMission-appropriate retention: data is kept as long as the mission and audit require — no longer by default.
- 04SeparationCustomer, synthetic, licensed and telemetry data are held apart and governed by their own rules.
- 05Risk managementAligned in spirit with the NIST AI Risk Management Framework’s emphasis on measurement and ongoing risk management.
(06) — Export posture
ITAR and EAR classification is built into the product architecture from the start — designed in, not retrofitted after the fact.
MALIA makes no claim that any system, model or dataset is exportable worldwide. Classification, licensing and transfer review are part of the engineering process, treated with the same rigor as the safety envelope itself.
(07) — A note on status
AXIOM is a proposed verification and governance architecture — the requirements MALIA designs toward, not a certified system. No MALIA machine is fielded, sold or deployed.