Where the chain and the docs disagree

The docs lag the chain, and following a stale sentence can fail silently: a transaction that succeeds and does nothing, a timing off by a factor of ten, a field scaled differently. Where they disagree, on-chain measurement wins, then the official docs, then SDK code. Each canonical page flags its own cases; this page lists them all, with the sample size wherever there was one.

Most measurements come from three September 2026 research tracks, settled by transactions and live API responses (borrow-rate options, composable credit, collateral underwriting). The rest come from auditing SDK wire formats against the live exchange, and from live info API reads.

The docs say one thing, the chain does another

Topic	The docs say	Measured	Evidence	Covered on
API wallets per account	1 unnamed and up to 3 named, plus 2 named per sub-account (the sub-accounts page: 3, rising by 2 per sub-account)	14 named API wallets, all unexpired, on an account with 1 sub-account, where the docs allow 5; about 9.3M USDC of cumulative volume	Live extraAgents and subAccounts reads, 2026-11-03	API wallets
CoreWriter delay	Order actions and vault transfers sent from CoreWriter “are delayed onchain for a few seconds”	Orders are accepted 0.35 to 0.51 s after the EVM block. Only vault transfers are delayed, by 4.05 to 4.10 s	5,455 mainnet CoreWriter actions over 73,914 blocks; 52 orders matched by client order id across 15 accounts and 14 markets; 6 vault transfers	execution order
CoreWriter supply for contracts	The portfolio margin page: “In a future upgrade, CoreWriter will expose the same supply action for smart contracts.” The CoreWriter table on another page already lists action 15	Live. A testnet contract supplied 500 USDC and held 500.00000051 one block later. Mainnet decodes action 15 into a typed lending action and enforces lending rules on it	Testnet probe contract; two mainnet refusals carrying the CoreWriter-only action shape	composable credit
CoreWriter action 14	The action table runs from 13 to 15	Action 14 exists on mainnet; its semantics are unknown	25 emissions in 3,000,001 mainnet blocks	CoreWriter action reference
HIP-4 status	The HIP-4 deployer actions page is headed “Testnet-only”	HIP-4 runs on mainnet	228 active outcomes, 21 questions and 3 venues on mainnet on 2026-09-20; 331 outcomes, 33 questions and 4 deployer venues on 2026-11-03; the page is still headed “Testnet-only” in the live docs on 2026-11-03	HIP-4
Outcome fees	The HIP-4 page: “Fees are currently zero for outcome markets for initial testing.”	Opening is free and closing is charged, at spot rates x (scale + max(scale, 1)): every buy was billed 0 and every sell was charged. The docs’ own fees page agrees: fees on closing or settling, never on opening	7,256 mainnet fills, 2026-09-12	fees
Aligned quote token actions	enableAlignedQuoteToken and disableAlignedQuoteToken are documented spotDeploy variants	Both are rejected at the parse stage on mainnet and testnet	SDK audit using a dummy signature, which the exchange checks only after parsing	quote assets
Multi-outcome markets	The HIP-4 page: multi-outcome markets “will be supported but are not part of the initial mainnet release”	Questions run on mainnet	21 questions across 3 venues on 2026-09-20; 33 questions on 2026-11-03	what is an outcome market
HyperEVM write and read on mainnet	The HyperEVM overview: write system contracts “are not live on mainnet yet”; the developer page introduces the read precompiles as a testnet EVM feature	Both are live on mainnet	273,437 CoreWriter actions with known ids in 3,000,001 mainnet blocks; precompiles 0x800 to 0x814 assigned on both networks, 2026-08-17	HyperEVM overview
Assistance Fund supply	HYPE in the fund is burned, “removing the tokens permanently from the circulating and total supply”	tokenDetails for HYPE listed the fund’s about 48.4M HYPE as non-circulating but still inside a total supply of about 998.8M; only about 1.2M HYPE had left total supply	One live read, 2026-11-03: the API’s accounting, not a traced transfer	HYPE
outcomeMetaUpdates frame	Typed as a one-element tuple, [WsOutcomeMetaUpdate]	Arrives as an array	SDK wire audit	WebSocket
Gossip priority auctions	“2 independent Dutch auctions synced to the same 3 minute schedule”	gossipPriorityAuctionStatus returned 3 slots on one 180-second schedule: slots 0 and 1 won, slot 2 unwon at the 0.1 HYPE minimum	2 live reads on mainnet, 2026-11-03	priority fees

The official docs differ

The CoreWriter delay has the widest reach. The docs put order actions in the delayed set; measured on mainnet over 52 orders, they landed 0.35 to 0.51 s after the EVM block, and only vault transfers were delayed, by about 4.05 s. Sequencing an order leg across blocks to wait out the delay wastes time, and assuming the delay protects against a fill is wrong.

The docs are silent, and not knowing loses money

Here the docs say nothing, or cover one direction only.

Topic	The docs say	Measured	Evidence	Covered on
CoreWriter failures	CoreWriter burns about 25,000 gas “before emitting a log to be processed by HyperCore as an action”; nothing on failure	The EVM transaction succeeds whether or not HyperCore executes. Four refusals seen with EVM status 1: Invalid number of decimals, Insufficient USDC or HYPE balance for token transfer gas, Borrow/Lend error: Insufficient spot balance for operation, Manual borrow/lend not allowed with portfolio margin	Testnet transactions and mainnet refusals, each read from HyperCore’s action log	footguns
Rounding, HyperCore to HyperEVM	A non-round amount is “burned (guaranteed to be <1 Wei)”, which describes the HyperEVM to HyperCore direction with extra EVM decimals	In the other direction, with USDC’s evmExtraWeiDecimals of -2, an amount that is not a multiple of 100 wei is rejected, not burned	5 of 5 sends followed the rule: 2 non-round refused, 3 round accepted	moving assets
Bridge gas	A transfer from HyperCore to the HyperEVM “costs 200k gas at the base gas price of the next HyperEVM block”	The gas is charged from the sender’s HyperCore balance, in HYPE if that covers it, else from the token sent, so without spare HyperCore HYPE the whole balance cannot be sent	On testnet, sending 100 of 100 USDC refused; 99 sent for 0.000588 USDC	moving assets
Action 17 units	wei “must be a multiple of size”; no scale given	1e8 per contract	wei = 100000000 minted exactly one Long and one Short and moved 1 USDC, testnet	CoreWriter
Precompile history	Values match the latest HyperCore state when the EVM block is built; independent archive nodes serve historical state, and read precompiles are recorded only for the calls actually made	The precompiles fail at any explicit block number, including the current head, even on an archive node that serves historical eth_getCode	PrecompileOutOfGas at head, head minus 1, 20 and 10,000	precompiles
0x811 with no record	Nothing	Reverts rather than returning zeros, even for a HyperCore user holding spot	Mainnet and testnet, including an account holding 31,989 USDC of spot	precompiles
USDC on the HyperEVM	The USDC page gives 0xb88339cb7199b77e23db6e890353e22632ba630f as the USDC contract on the HyperEVM	HyperCore USDC (token 0) is linked to a different contract, 0x6b9e773128f453f5c2c60935ee2de2cbc5390a24. The docs’ USDC is not linked, so moving between the two is a swap	spotMeta and the tokenInfo precompile	moving assets
Staking links and deployers	Staking links attribute a staking account’s HYPE to a trading account’s fee tier; nothing on deployers	A deployer cannot have one. Activation is refused with Deployer cannot have staking link	Testnet probe	collateral underwriting
Unstaking on testnet	A 7-day unstaking queue	Exactly 7 days on mainnet; about 5 minutes on testnet	One mainnet exit finalised 7 days and 1 second after it started; one testnet withdrawal finalised in 5 minutes	staking
Market order TIF	The order schema lists only Alo, Ioc and Gtc	FrontendMarket is accepted on mainnet; the SDKs and the app send it for market orders, and the info endpoint’s own examples return it	SDK wire usage; the docs’ own orderStatus example	order types
Undocumented surface	Nothing	A September 2026 sweep found about 18 undocumented exchange actions and 19 info requests working on mainnet; the docs have since added cDeposit2, trailingStop and userStarState, so the count may be lower now. Two appear nowhere in the live docs on 2026-11-03: the HTTP borrowLend action and the userToMultiSigSigners info request	SDK audit against both networks; live docs search	exchange action reference

Wrong once, corrected since

These are now right in the current docs; they stay listed for checking older notes and SDKs.

Topic	What was reported	The docs now	Covered on
Negate key	Earlier documentation and SDK examples used negateQuestion; the wire key is negateOutcome, checked against Hyperliquid’s own interface	The exchange endpoint page in the current mirror documents negateOutcome	split, merge, negate
HIP-4 deploy action	Deployment used {"type":"spotDeploy","outcome":{...}}; that shape is now rejected	The current mirror documents a top-level outcomeDeploy with venue and operation	outcome templates
HYPE lending supply cap	The docs gave a 10M HYPE global supply cap; on 2026-09-20 a live reserve read showed 10,233,170.44 HYPE supplied	A 25M global and 1M per-user cap (2026-11-03), so that reading sits inside it	borrowing and lending

Older SDKs still carry superseded forms: one sent outcomeDeploy without the venue and operation envelope, and mainnet answered Failed to deserialize the JSON body into the target type.

The docs contradict themselves

Here no measurement decides; two passages of the docs disagree.

Topic	One passage	The other	Covered on
hip3LiquidatorTransfer signing	The exchange endpoint shows it as an L1 action: no chain fields, expiresAfter accepted, and SDK code signs it so	A comment on the HIP-3 deployer actions page calls it user-signed	exchange action reference
Interest component	“0.01% every 8 hours, which is 0.00125% every hour, or 11.6% APR”	The hourly rate x 24 x 365 is 10.95 per cent; 11.6 per cent is 0.01 per cent per 8 hours compounded 1,095 times a year	funding
Recurring binary settlement	Contract specifications: Yes if the interpolated mark is at or above the target	The spot info endpoint page: Yes if the mark at expiry is “above” targetPrice	market shapes
HIP-1 token names	name has “no uniqueness constraints”	“The ticker is a unique onchain identifier”, on the same page	native tokens
Oracle inputs	Robust price indices: the oracle “does not depend on hyperliquid’s market data at all”	The oracle page includes Hyperliquid spot at weight 1, left out only where main liquidity is elsewhere, such as BTC	oracles
Portfolio margin eligibility	The trading page: account value below $25M	The FAQ: account value below $5,000,000	portfolio margin
Fees in closed PnL	The entry price page puts the fee in closed PnL	The info endpoint’s example opening fill shows closedPnl "0.0" beside a non-zero fee	entry price and PnL

Not yet shown to contradict the docs

Topic	What is known
Allowlisted HIP-3 dexes	The docs, testnet only: isStar: true at creation and per-user {"proxy": [user, op]} operations. Testnet transactions replayed in July 2026, before the docs covered it, used operation: "activate" on a live dex and a batch modifyApprovals; which shapes testnet accepts now is unresolved
Action 17 timing on mainnet	Split and merge landed in the transaction’s own block on testnet; the mainnet timing window had no action 17 traffic, so mainnet is inferred, not measured
Borrow and repay through CoreWriter	The docs list action 15 as supply and withdraw; one tool routes ops 2 and 3 as borrow and repay; measurements cover supply and withdraw only
HIP-4 slashing for late settlement	Specified when HIP-4 was announced, disclaimed for the testnet release, not observed on mainnet

For the timing and failure findings in context, see What one HyperEVM transaction can and cannot do. Every silent failure, with its fix, is in Footguns, measured.