Open a five-minute up-or-down market, settled by a price already on chain
One pool for each five-minute round. You stake on Up or Down before it starts, and the side that was right shares the other side's stakes. The two prices are the first admissible readings recorded at or after the start and the end, from a feed already on chain. No API key needed. You'll run a real BTC round on a Monad fork and a real NVDA round recorded on HyperEVM, feed it a bad oracle, and see which test notices each deleted check.
What you end up with
Here's where you end up. You'll have run two real rounds from a fresh clone: BTC on a Monad mainnet fork, settled at Pyth's first updates after the start and the end, and xyz:NVDA on HyperEVM, replayed from values recorded live. You'll have watched an oracle the book doesn't agree with get refused, and measured which test notices each deleted check. Below are our runs.



Run it (10 min)
- 1
Run two real rounds
2 mingit clone --recurse-submodules https://github.com/fabrknt/fivemin && cd fivemin # needs Foundry forge test --match-test "test_realRound|test_recordedSunday" -vvYou will see
[PASS] test_recordedSundayRound_equalOracle_isVoid() (gas: 383048) xyz:NVDA open 161110 / 1000 at the start, close 161110 / 1000 at the end: equal, so the round is void every stake back: alice 50, bob 30; no fee [PASS] test_realRound_btc_downWins() (gas: 502046) bets at block 122356650, 8 s before the start: alice Up 100, bob Down 60, carol Down 40 open recorded at block 122356773: BTC 64725, published 28 s after the start close recorded at block 122357676: BTC 64634, published 2 s after the end Down won: bob claims 118.8, carol 79.2, alice 0, fee 2 (2% of the losing side)The BTC round runs on a Monad mainnet fork at three real blocks: the bets, Pyth's first push after the start, and its first push after the end. We found them from PriceFeedUpdate logs. The NVDA round replays values recorded live on HyperEVM. Why recorded? Because HyperEVM's price precompiles answer with the current state, whatever block a call names. US stocks were closed that Sunday, and trade.xyz's oracle barely moved. Stock rounds are only worth running in market hours.
- 2
Feed it an oracle the book does not agree with
1 minforge test --match-test test_oracleAwayFromTheBook -vvvvYou will see
[PASS] test_oracleAwayFromTheBook_isRefused() (gas: 30937) ├─ [0] VM::expectRevert(PriceRejected(3, 524, 50)) ├─ [12895] Hip3Source::read(0x0000000000000000000000000000000000000000000000000000010000000203) [staticcall] │ └─ ← [Revert] PriceRejected(3, 524, 50)A HIP-3 oracle is set by the market's deployer, and there's no free outside price to check it against. So Hip3Source reads it through hl-price-guard's HLPriceGuard.read. That refuses an oracle whose spread, mark or mid is further away than its bounds (20, 50 and 50 bps, set from a Sunday recording where they reached at most 3.11, 3.72 and 11.79). If the deployer moves the oracle away from the book by more than the bounds, nothing is recorded. A smaller move, within them, can still be recorded.
- 3
Delete a check and see which test notices
5 mincargo install --locked --git https://github.com/fabrknt/unproven --tag v0.2.0 # needs Rust unproven unproven.toml --only "collectFee"You will see
baseline: no finding, every test passes [1/3] collectFee: round is final ... ok collectFee: round is final [test] failed: test/FiveMin.t.sol::FiveMinTest::test_collectFee_refusedWhilePending [2/3] collectFee: once ... ok collectFee: once [test] failed: test/FiveMin.t.sol::FiveMinTest::test_fullRound_downWins_feeFromLosersOnly [3/3] collectFee: no fee on a void round ... ok collectFee: no fee on a void round [test] failed: test/FiveMin.t.sol::FiveMinTest::test_collectFee_noFeeOnAVoidRound, test/Hip3Recorded.t.sol::Hip3RecordedTest::test_recordedSundayRound_equalOracle_isVoid 0 by the fuzzer, 3 by a test only, 0 refused by something else, 0 unprovenunproven.toml lists the contract's 38 checks, each with the edit that deletes it. unproven applies each edit, reruns forge test --json, and fails if a row reads differently. The first run matched our prediction row for row: 20 caught, 18 caught by nothing. Three of those 18 would have lost money: a fee taken from a round still pending, a fee taken from a void round, and a Down-only round won by Up paying no one. Tests now catch 31. The other 7 are redundant, harmless or refused elsewhere, and each row says which. The whole table takes about 10 minutes.
- 4
Record a live round yourself
2 minpython3 script/record_hip3.py 1 2 3 60 live.json # dex 1 (xyz), asset 2 (NVDA), szDecimals 3, 60 sYou will see
59 samples, 1 errors, blocks 51302798..51302859One JSON-RPC batch a second: the latest HyperEVM block, then the oracle, mark and best bid/ask precompiles at that block. The JSON holds every sample, and Hip3Recorded.t.sol shows how a recording becomes a test. A scheduled GitHub Action in the repository records a round in US market hours, because a laptop that goes to sleep misses it.
Why: what went wrong (measured)
- 2026-11-15 On the first measurement, no test caught 18 of the contract's 38 checks. Three of those would have lost money if deleted. We added tests for 11. Source (opens in a new tab)
- 2026-11-15 A review found two bugs. A late poke could still record a price stamped inside a checkpoint's window after the round had gone void. And a reading stamped ahead of the block was accepted. Both are fixed, each with a test. Source (opens in a new tab)
- 2026-11-15 A real xyz:NVDA round recorded on a Sunday opened and closed at 161.110. The book moved, but the oracle changed only twice in six minutes. A stock listing is only worth running in market hours. Source (opens in a new tab)
Put together: the parts
1. The pool, the two prices and the payout: fivemin (opens in a new tab)
FiveMin (bets, poke, outcome, claim) and two price sources: PythPushSource for Monad, Hip3Source for HyperEVM. Foundry, Apache-2.0, written with Claude.
git clone --recurse-submodules https://github.com/fabrknt/fivemin2. Read an oracle you can stand behind: HL Price Guard (opens in a new tab)
HLPriceGuard.read for the book check on HIP-3, PythRef for Pyth.
forge install fabrknt/hl-price-guard@v0.4.13. Know which checks a test would miss: unproven (opens in a new tab)
Deletes each listed check in turn and reruns the tests; 0.2 reads forge test --json. Run it in CI, and the table can't drift from the code without anyone noticing.
cargo install --locked --git https://github.com/fabrknt/unproven --tag v0.2.0Go deeper, in order
- Pyth: sponsored push feeds, by network (opens in a new tab): Which feeds Pyth pushes on which chain, and at what deviation and heartbeat. On Pyth's page (read 2026-11-15), Monad has 52. No chain has a stock on a five-minute feed.
- Hyperliquid: interacting with HyperCore from HyperEVM (opens in a new tab): The read precompiles behind Hip3Source: oracle, mark and the book.
- fivemin's spec (opens in a new tab): Why stakes close at the start, why the first reading after each checkpoint, and what stays open, stated plainly.
Before real money
- Close the Pyth blocker (the first update after a time, with an API key) before any Monad listing takes money.
- List stocks only in market hours, and set the book bounds from your own recording.
- Run unproven in CI, so the result for every check, caught or explained, can't drift from the code unnoticed. And decide who operates the pools, and where they aren't offered.
- Not audited, not deployed. Claude wrote the contracts and the tests on 2026-11-15; a person chose the design and the checks. The README lists what checked them.
- The Monad source has an open blocker. Anyone holding signed Pyth updates can push the most favourable price inside a checkpoint's window, then record it. Closing it needs Pyth's first-update-after-a-time check, and that needs an API key. Until then, it's for testing only.
- On HyperEVM the deployer sets the oracle, and can still move oracle, mark and book together by trading. The book check only narrows what the deployer can do alone. Whoever records the price chooses the block, inside a 3 s window.
Run end to end
2026-11-15 We followed it from the public repository in a fresh clone on 2026-11-15 UTC (fivemin dfb3353, Foundry 1.7.1, unproven v0.2.0 installed from its tag): steps 1 to 4, with the outputs above. Following it turned up two rows of unproven.toml that were out of date after a test was added. We updated them before publishing this page.
Fabrknt is independent. We're not affiliated with Hyperliquid, trade.xyz, Pyth or Monad. fivemin isn't audited, isn't deployed, and isn't offered to anyone.