PortfolioGAP-08

Agent & Protocol · Toolkit

Human-in-the-Loop Approval Simulator

SIMULATEDVerified Jul 2, 2026

An agent processes a queue of 20 items, four of them edge cases. Raise the autonomy level: throughput climbs, human load falls — and at some point an edge case slips through unreviewed. Find the balance.

Same instrument · three industries — pick a use-case to reconfigure the run

L1 · Review all
L1L2L3L4L5
Throughput
100/hr

Items processed

Human review load
20/20

Items sent to a human

Edge cases slipped
0/4

Unreviewed errors

Risk exposure
$0k

Cost of the slips

The queue · ● edge case

L
M
H
L
M
L
H
M
L
M
H
L
M
L
H
M
L
M
H
M
Auto-approved Human-reviewed Slipped edge case

L1 clears every edge case

Every item reviewed — zero risk, but you're paying for a human on trivial approvals. There's throughput to reclaim.

Autonomy by risk tier · bridges EL-05 / Govern

High-risk (EU AI Act)max L1–L2
Limited / significantmax L3
Minimal / internalmax L4–L5

The level isn't a global setting — it's set per use case by its risk tier.

Steering-committee takeaway: Autonomy is set per risk tier, not per enthusiasm.

How this is built

Twenty items carry a risk tier and four are edge cases (errors if auto-approved). Each level defines a review policy; an edge case slips when it isn't reviewed. Throughput rises with autonomy; exposure = Σ severity of slipped edges. The medium-risk edge is engineered to slip exactly one level past the balance point.

Stack: Next.js (static) + shared design system; deterministic client-side.

Limitations: the queue and severities are illustrative; real autonomy also weighs reversibility and detection latency. It shows the throughput-vs-risk trade and the per-tier rule, not a policy engine.