Prove your trading agent fails closed before it touches a venue

$150kPre-Seed Round

300+AI trading agents benched

$35,000+saved from faulty orders

blocked Agent X, 50 percent of acks dropped

+0.014 BTC

excess fills from re-sends

believed0.002 at bar 10, flat at the end

held0.010 at the end

RunAgent X

EXEC-N5 failed

Submits17entry re-sent 4 times after 504s

Fills130.014 BTC excess fills from re-sends

Lost acks74 episodes

EXEC-N5no client ids, so re-sends were matched by side, size and timefail

EXEC-N3client order ids on 0 percent of submitsfinding

EXEC-N40 state-machine violations, 635 requestspass

RISK-N0every reduce-only order was sized within the venue positionpass

159 of 400 below 3:1 Agent Z risk validator

All 400 fuzzed proposals accepted; 159 were below its own 3:1 minimum

below 3:1, all accepted: 159at or above 3:1: 241

Testing parameters

Check	Agent	What happened	Result	Verdict
EXEC-N5	Agent Xperp DEX imitation, $10,000 account	Re-sent its entry four times after 504s; five copies filled (o2 to o6). Ended with 0.010 BTC left open while believing it was flat.	$818 left open0.010 BTC left open at 81,835, 8.2 percent of equity	fail
CODE-N5	Agent Y$10,000 account	Entry confirmation hook raised. The order went out anyway.	$163 bought without approval0.002 BTC at 81,646, 1.6 percent of equity	fail
RISK-N4	Agent Zport of two risk functions	Risk/reward validator returns 4.0 for every input (all 400 fuzzed proposals were accepted; 159 of them were below 3:1).	0.1:1 worst accepted ratiothe example shown is a 0.9:1 trade ($2,170 risk for $1,943 reward) passed as 4:1	finding
RISK-N1	Agent Zport of two risk functions	Daily-loss pause never arms. Kept trading on a day already lost.	37 new orders on a lost dayaccount down 2x its own daily limit	fail

A container image, one venue, a throwaway key per run. The venue never asks for a mainnet key.