Approval Headroom
Route AI-agent actions to auto-allow or human review at a cost-optimal threshold θ*
ABOUT THE APPWatching every agent action isn't safer — past a point, more review lets more danger through, because a flooded reviewer rubber-stamps. Approval Headroom computes the cost-optimal cutoff θ* for that tradeoff and keeps an append-only, tamper-proof record of every decision.
It's an early, honest experiment, not a finished product: today you enter each action's risk by hand (no agent hookup yet). The θ* engine is a faithful, machine-checked (22/22) implementation of a 2026 paper — Turan, Oversight Has a Capacity (arXiv 2606.08919); the inverted-U is a modeling assumption, stated in plain sight, not a measured law.
Built to be a foothold and remixed: reusable pieces — a checkable integer θ* model, an append-only tamper-proof ledger, a shared-config calibration pattern, honest-disclosure UI — carry forward whether or not this becomes a standalone product. 6 languages. If you're building AI oversight, take what's useful and build on it.
TAGS
aiassistantbusinessdevelopertoolsprivacysecurityproductivityutilities
SHARE