Loops HouseHarbor Range

NEAR HACKATHON
Team
1 member- SNOwner
Snowden Zeng
Overview
Harbor Range is a red-team proving range for agent spend policies. Before an agent gets a wallet, its policy should survive a week on the range: three scripted attacker personas — an inventory flooder, a wash trader, and a runaway spender modelled on a prompt-injected agent — plus one adaptive raider attack a small agent economy whose settlements run through a NEP-245-shaped mock NEAR testnet adapter. The stalls must survive: payouts are reserve-bounded and uncovered payouts are refused in full. The range renders binary verdicts anyone can re-run: npm run verify runs 8/8 checks green — 41 unit tests, the 10-condition cage check, a 100-seed survival sweep (100/100 held), a same-seed no-attacker control, and a negative control: a weakened policy (partial payouts) that looks healthier on the headline metric (3 stalls open, all drained to zero) and is rejected by the solvency invariant. A live in-browser demo runs the real code: https://snedz.github.io/harbor-range/. Built for our own agent fleet's spend safety, packaged for any team handing AI a wallet. Honest limits: settlements are local records on a mock adapter — no deployed contract, no real assets; the chain stays invisible by design.