AgentVault

CONTRACTS WRITTEN AND TESTED Archetype — permanent collectible identity, four types. GenesisAgent — an ERC-721 with a permissionless capped mint, where the archetype is assigned once and can never be changed. AccountRegistry — one isolated account per NFT, with principal and profit tracked as separate fields, lazy detection of NFT transfers, and a holder-controlled kill switch. Mandate — four risk mandates with prototype parameters, selected by the holder rather than determined by the NFT. RiskEngine — a pure-view deterministic validator that returns named rejection reasons instead of reverting. DEPLOYED AND SOURCE-VERIFIED ON ROBINHOOD CHAIN TESTNET GenesisAgent — 0x0EBdDD089f8203DD5cD1Bd1f148F75757285CF96 AccountRegistry — 0x0D0080582D317D2878A31b01D97614a3E918dC65 RiskEngine — 0x36df9096162b9f18d13574Fba930C9e2Ce6cc4a4 TestAsset AVTA — 0x87CEd5dc138F825B3F42924A8Bb6E15ae5EC9A7d TestAsset AVTB — 0x9D97ebc6A395aaf981B26d952A4e22604eAFEc75 All five are source-verified. The full demo arc has run on chain, including a real kill-switch transaction. An earlier RiskEngine is still deployed and verified and is marked superseded in the deployment record, with a date and a reason, rather than deleted. Extending an immutable allowlist means redeploying — that is the cost of having no admin key, and it was paid rather than designed around. PAXOS USDG, AT AN ADDRESS ESTABLISHED THE HARD WAY The risk engine's allowlist is fixed at deployment and includes USDG at 0x7E955252E15c84f5768B83c41a71F9eba181802F. Explorer search on this testnet returns more than fifty tokens claiming that ticker, five of them named literally Paxos USDG, and all five were confirmed on chain as real deployed ERC-20s wearing the name. The genuine contract is named plainly Global Dollar. The address was taken from Paxos first-party documentation and then verified on chain. Because the allowlist has no setter, sourcing it from explorer search would have been permanent. This is written up as security finding S1. VERIFIED RATHER THAN ASSUMED The chain ID was measured directly rather than read from documentation, closing an open question where a third-party source disagreed with the official docs. Blockscout's verifier endpoint was established by probing its API, not by assuming the conventional URL. Verification status was confirmed independently — the Foundry CLI reported one contract as already verified while the explorer still had it unverified. PROPERTIES ENFORCED, NOT JUST CLAIMED No contract in this system has an owner, an admin, or a role-gated function. Withdrawal has no pause check on any code path. The kill switch stops automation and can never stop the holder reaching their own funds. Collectible rarity cannot affect risk limits. The risk engine depends on an interface that cannot see an archetype at all, proven by a test suite that mints all four archetypes and asserts identical verdicts. Two mandate parameters are specified but not yet enforceable. They carry notes in the code and are recorded as pending in the decision log rather than presented as working. SECURITY ANALYSIS PUBLISHED Static analysis with Slither 0.11.6 found five items in project code — zero high, zero medium. Each is listed by detector name with a triage verdict: three accepted with reasoning, two false positives explained. None was suppressed with an inline disable comment. The unfiltered run reports twenty-five findings, twenty of which are inside third-party libraries, and both numbers are published so the exclusion is visible rather than implied. Two findings are real improvements that were deliberately not applied, because the affected contracts are already deployed and source-verified and editing the source would leave the explorer's verified code no longer matching the repository. Both are recorded as carried forward to the next redeployment. STATEFUL INVARIANT TESTS Four invariants, each one a claim the project makes publicly, driven by a handler contract that randomises sequences of minting, deposits, withdrawals, pauses, NFT transfers and synchronisation. Withdrawal liveness — in every reachable state, including paused, mid-transfer and unsynchronised, the current holder can withdraw in full. A deposit is never profit — profit and loss equals balance minus net principal, checked against independent accounting in the handler rather than against the contract's own arithmetic. Paused implies never approved — across every reachable state, not just the one a unit test constructs. Archetype cannot influence risk — parallel accounts differing only in collectible archetype return identical verdicts and identical reasons. Sixteen thousand three hundred and eighty-four calls per run with zero reverts, verified across eight consecutive runs and three explicit fuzz seeds, all recorded. THE TRAP IN INVARIANT TESTING An invariant suite whose handler calls all revert passes perfectly while exploring nothing. So the handler counts successful state mutations rather than attempted calls, and a separate ordinary test asserts those counts are non-trivial. The synchronisation function firing on its own is the result that matters: it can only fire when an NFT transfer is genuinely pending, so a non-zero count proves the stale-transfer window is actually being entered rather than assumed. That coverage assertion cannot live in Foundry's afterInvariant hook. When any invariant fails, Foundry shrinks toward the shortest failing sequence, and a coverage assertion is trivially failable by a one-call sequence — so the shrinker drives straight at it and reports a nonsense counterexample naming an invariant that never broke. Found the hard way, fixed by splitting reporting from assertion, and written into the repository's gotchas file with the full symptom. A CORRECTION, ON THE RECORD The suite reports 126 tests. Earlier in the Buildathon it also reported 126, and that number was not real. One test wrote an environment variable to check a deploy-script guard. Foundry runs test contracts in parallel, and a second test contract wrote the same variable, so the two raced and the suite passed intermittently. Five consecutive local runs produced three failures. The 126 was simply the run that had been looked at. It was caught by running the suite repeatedly from a clean clone rather than once, then fixed by removing the race. Four guard tests became two and no assertion was lost. The honest count after that fix was 124. Adding the invariant suite brought it back to 126, by coincidence, from a different set of tests. That coincidence was reconciled rather than waved through: the eight pre-existing suites still collect exactly 124, so no test quietly stopped being collected, which is the failure mode that looks identical to a clean pass. This is recorded rather than quietly corrected because the product is a performance record that includes losses and refused trades. A project that rounds its own test count up has no business claiming to keep an honest one.

Self-funded. No outside capital raised, no round in progress. Total spend to date is effectively zero — free tooling and testnet gas. Near-term funding plan is ecosystem grants rather than equity or token sales. A Genesis Agent mint is deliberately deferred until the product has real users, and would be sized to actual demand rather than to a funding target.