Part 3 of 12 · 2 min read
A reviewer's attention is a resource the gate spends, and it runs out
Think of clicking "I agree" on terms you never read: once you've clicked "yes" a hundred times today, the hundred-and-first click happens without your eyes ever reaching the text.
In plain words: approval fatigue means reviewers read less carefully as requests pile up, and rubber-stamping is the resulting failure mode -- approving without reading -- which makes the gate look safe (every action got "approved") while actually being unsafe, because the more often a gate fires, the less attention each individual request gets, which can make a gate that fires constantly worse than one that fires rarely but on the right things.
A worked example. Under the module's fatigue model, gating every one of six actions per task sends 6.0 requests per task; at 10 tasks per hour that's 60 requests per reviewer-hour, which collapses the simulated catch probability to just 23%, letting 70.8 of 93.5 harmful actions per 1,000 tasks through. Gating only irreversible actions sends 1.8 requests per task (18 per reviewer-hour), keeping catch probability at 47% and leaking only 60.4. With a reviewer who never tires, the ranking flips entirely: gate-everything leaks just 9.5, against 29.2 for irreversible-only -- fatigue isn't a minor correction, it's what decides which policy is actually better.
The key lines. From code/mini-hitl/hitl.py:
# from code/mini-hitl/hitl.py
def attention(rate, p):
return 1.0 / (1.0 + rate / p.half_attention_rate) if p.fatigue else 1.0
Read it as: attention is a smooth curve, not a cliff. It halves at exactly half_attention_rate requests per hour (20, by assumption) and keeps falling gradually beyond that. That is why going from 1.8 to 6.0 requests per task does not just cost "a bit more fatigue." It moves the reviewer well past the point where attention has already fallen by more than half.
Approval fatigue, rubber-stamping
Approval fatigue is reviewers reading less carefully as requests pile up. Rubber-stamping is approving without reading, which makes the gate look safe and be unsafe.
Example: the gate-everything policy leaking nearly as much harm (70.8) as no gate at all (93.5) under fatigue is rubber-stamping expressed as a number -- the gate exists, every action technically passed through it, and almost none of that passage actually involved anyone looking.