AI Governance Pipeline: Risk-Throttled Review Queue

Governance Policy

Submission rate λ 0.60/s

Audit capacity (reviewers) 3

Risk threshold (throttle trigger) 55%

Approval risk tolerance 65%

Throttle: inactive

Live Readouts

Review queue

Deployed systems

Systemic risk R

Approvals/min

0.0

Incidents (total)

Rejected (total)

Legend

Low risk (< 30%) Medium risk (30–55%) High risk (55–75%) Critical / incident (> 75%) Rejected at audit

How it works

AI systems arrive as a Poisson-like stream at rate λ, queue (FIFO) for one of the c available auditors, and are held in review for a fixed duration. On completion an auditor approves a system only if its risk score is below the board's approval tolerance — otherwise it is rejected. This is a standard M/M/c queueing model with a risk-based admission rule instead of a coin-flip.

Systemic risk (EMA, τ = 2s):
R ← R + (R* − R)·(1 − e^(−dt/τ))
R* = 0.65·avgRisk(deployed) + 0.35·min(1, incidentRate / rateCap)

Deployment throttle (proportional controller):
λ_eff = λ                                  if R ≤ R_threshold
λ_eff = λ·max(0.05, 1 − k·(R − R_threshold))  if R > R_threshold

Deployed systems can trigger an incident with probability proportional to their own risk score each second; every incident nudges R upward, which — once R crosses the threshold — automatically throttles new submissions, exactly like a real governance board pausing rollouts after a spike in reported harms. As incidents get remediated and old systems retire, R decays and the throttle releases on its own.

λ (submission rate) — how fast new AI systems are proposed for deployment.
Audit capacity c — how many systems can be reviewed in parallel; the real bottleneck in most governance programs.
Risk threshold — the R value at which the feedback controller starts cutting the intake rate.
Approval tolerance — the maximum risk score an auditor will pass; lower it to see rejections rise and deployed risk fall.

Intake / QueueAudit GateDeployed

⚠ Couldn't load the 3D engineThree.js failed to load from the CDN. Check your connection and reload.