Skip to main content
SINT models every request through four approval tiers. Each tier maps to a class of physical consequence and a corresponding level of human oversight.
The tier model is T0–T3 (four tiers). Earlier documentation used T1–T5; that was deprecated in v0.2. See facts for the canonical model.

​
The four tiers

T0 calls never hit the network. They’re approved locally in microseconds. Governance adds zero latency to read-only operations.

​
The escalation function

Every request has a base tier from its capability token. The gateway elevates the effective tier based on four dynamic Δ factors:
The escalation is a max, not a sum. The highest single risk indicator drives the tier — a conservative composition.

​
Δ_human — human presence

​
Δ_human
function
Δ_human(r) = 2 if contextSnapshot.nearbyHumans == true, else 0.
An agent with a perfect history in an empty room should not therefore be trusted at the same tier when a person walks in.

​
Δ_trust — CSML-driven trust score

​
Δ_trust
function
Δ_trust(r) is driven by the CSML score for the current foundation-model backend.
  • Δ_trust = 3 if CSML < 0.30
  • Δ_trust = 2 if 0.30 ≤ CSML < 0.60
  • Δ_trust = 1 if 0.60 ≤ CSML < 0.80
  • Δ_trust = 0 otherwise
The gateway doesn’t trust model-vendor alignment claims; it measures behavior in situ and adjusts.

​
Δ_env — environment uncertainty

​
Δ_env
function
  • Δ_env = 2 if environmentClass ∈ { "shared_human", "unknown" }
  • Δ_env = 1 if sensorHealth == "degraded"
  • Δ_env = 0 otherwise
A degraded LiDAR is never masked by a confident prior. Unstructured environments raise the tier.

​
Δ_novelty — first-time tool use

​
Δ_novelty
function
Δ_novelty(r) = 1 if r.toolName ∉ recentAuditWindow(k=50), else 0.
A small bump that treats first-time invocations with extra skepticism.

​
Worked example

A warehouse robot with base tier T1 proposes a nav2.move_to_pose action.
1

Base tier

Token base: T1.
2

Δ_human

nearbyHumans = false → 0.
3

Δ_trust

CSML for GPT-5.2 (the backend) reads 0.16 → 0.
4

Δ_env

environmentClass = indoor_structured, sensorHealth = nominal → 0.
5

Δ_novelty

move_to_pose is in the recent audit window → 0.
6

Effective tier

max(1, 0, 0, 0, 0) = 1. Auto-approved, audited.
Now a person enters the aisle. The robot’s perception updates nearbyHumans = true. The executive layer issues a fresh request for the continued motion.
1

Base tier

T1.
2

Δ_human

nearbyHumans = true → 2.
3

Effective tier

max(1, 2, ...) = 2. Requires review. Sponsor approval fires.
The robot safe-holds until a human approves. The sponsor can tighten (but not loosen) the velocity envelope for this approved continuation.

​
Timeout semantics

On timeout: default-deny. The request transitions to FAILED and logs a HUMAN_TIMEOUT event. Safety before availability.

​
Forbidden action sequences

In addition to tier checks, the Policy Gateway detects forbidden combinations that span tool boundaries: These trip automatic T3 classification regardless of the underlying base tier.

CSML score

How Δ_trust is computed and calibrated against ROSClaw data.

Invariants

The six formal guarantees the tier model rests on.