All work
Incentive design
Adversarial thinking, from vote-buys to agent exploits.
A DAO launch with 46% of votes concentrated; one afternoon of simulation surfaced a ~$870K attack. The method now prices agent failure modes.
$870K
attack cost surfaced in simulation
46%
concentration flagged before launch
1
afternoon to find the flaw
Problem
The launch math worked. The adversarial math did not. Community trust — not just the treasury — was the asset at risk.
Approach
- 01Simulated the attacker instead of admiring the spreadsheet.
- 02Capped maximum influence and added rapid community recourse.
- 03Rewrote the logic with operators in the room and recorded the rejected option.
- 04Reused the same adversarial method today for agent permissions and spend caps.
“Incentive design was never just numbers. It was consequences and trust.”