CSE Roadmap

Common tier · about 4 hours · 12 parts

Part 3 of 12 · 2 min read

Approvals: fewer, tiered prompts catch more risky actions

Asking about everything feels the safest, but a person who is asked twenty times reads the last prompt differently than the first.

Think of a smoke detector that also beeps for a piece of toast: after enough false alarms, people stop checking the detector at all, including the time it's right.

In plain words: approval fatigue means a person's chance of actually reading a prompt falls with every prompt they've already answered that session. So asking about everything protects nothing. The dangerous action ends up buried among forty routine ones.

A worked example. Across 20 actions per session with 2 of them genuinely risky: asking about every single action produces 20.0 prompts per session and only 52 percent of the risky ones actually get read. Tiering so routine, reversible actions run automatically and only medium/high-risk ones prompt cuts that to 6.4 prompts and raises the read rate to 79 percent. Also batching the medium-risk prompts into one review screen brings it to 3.0 prompts and 86 percent read — fewer prompts, better read rate, in the same direction at once.

The key lines. From code/mini-ui/approvals.py:

# from code/mini-ui/approvals.py
def decide(req, policy="tiered"):
    if policy == "ask_all": return "ask"
    if req.risk == "high" or not req.reversible: return "ask_confirm"       # explicit confirmation with preview
    if req.risk == "medium": return "ask"
    if req.risk == "low" and req.reversible: return "auto"                   # logged and undoable

Read it as: the policy only prompts for medium and high risk, and lets low-risk reversible actions through automatically. The prompts a person does see are deliberately the ones worth their attention. That's what keeps the attention-fatigue curve from reaching the risky ones.

Approval fatigue, reversible action

Approval fatigue is people approving without really reading after enough prompts, so the one dangerous request slips through unnoticed. A reversible action has a defined inverse; an irreversible one cannot be undone once taken.

Example: the req.risk == "high" or not req.reversible check above routes every irreversible action to an explicit confirm screen regardless of its labeled risk level. Reversibility alone is enough to earn the stricter path.