Governed Intelligence

The stack

Three layers, one system

Layer 1

LLM — Fluency layer

The language model provides the interface. It is capable, general-purpose, and not inherently governed. Without the layers above, it will produce confident nonsense at the same rate as confident insight.

Layer 2

GEOX — Earth grounding

When the domain is geological, GEOX forces physics-over-n叙事, real-data-over-elegant-fiction. It enforces OBS/DER/INT/SPEC classification on any pixel-derived or model-derived claim. GEOX is domain-gated — it activates when Earth content is detected.

Layer 3

arifOS — Governance kernel

The constitutional kernel holds explicit floors (ASI-1 through ASI-5) that determine what can be claimed, held, or executed without human review. It issues verdicts: SEAL, CAUTION, HOLD, VOID. Human sovereignty is a first-class architectural primitive.

ASI floors

Constitutional floor levels

Every claim or action in an arifOS-governed system is scored against a floor level. Higher floors require stronger evidence or explicit human authorisation.

ASI-1

Observable (OBS)

Directly verifiable facts. An image shows a wireline log. A temperature reading is above a threshold. No inference required. Claims at this floor can propagate freely within normal operational bounds.

ASI-2

Derived (DER)

Claims that combine OBS facts with stated reasoning. The derivation must be explicit — which facts, which inference steps, what uncertainty band. Claims at this floor can propagate but must carry their derivation with them.

ASI-3

Interpretive (INT)

Claims that require model-based inference, analogy, or interpretation. The model must be stated and its limitations acknowledged. Unsourced interpretation is not permitted at this floor.

ASI-4

Speculative (SPEC)

Forward-looking, scenario, or hypothetical claims. Must be framed explicitly as speculation. Cannot be treated as basis for irreversible action. Requires CAUTION verdict before proceeding.

ASI-5

Sovereign floors

Irreversible, identity-shaping, externally consequential actions. Require explicit human authorisation. Cannot be triggered by the system alone. This is the human sovereignty layer — not a preference, a structural requirement.

Verdicts

What the kernel decides

Before every consequential action, the arifOS kernel issues a verdict. The verdict is not a suggestion — it is a gate.

SEAL Safe to proceed. Evidence is sufficient, floor level is appropriate, no sovereign boundary is crossed. System may proceed.
CAUTION Proceed with warning. Evidence is present but incomplete, or the action has irreversible components that require monitoring. System may proceed with explicit uncertainty framing.
HOLD Pause for human review. Evidence is insufficient, or the action crosses a sovereign floor. System must defer to the human sovereignty holder before proceeding.
VOID Do not proceed. Evidence contradicts the claim, or the action violates a constitutional floor. System must not execute. Requires kernel update before reconsideration.