how i read this, what i got wrong & what i can't prove
everything on these pages was read from sui mainnet through public rpc endpoints, plus ethereum & bitcoin explorers for the bridged legs, dex trade data for prices & the public tweet api for timestamps, plus github's api for alphafi's public sdk repos. no private data, no insider, no source code (alphafi's move packages aren't published; for the oracle i diffed the disassembly sui returns for each version instead). anything i call "inferred" or "read here as" is me interpreting state changes, not something i read off a label. i got a bunch of things wrong on the way & i'd rather list them than pretend i didn't.
sources & tools
- sui json-rpc via sui-rpc.publicnode.com, rpc-mainnet.suiscan.xyz & sui-mainnet-endpoint.blockvision.org (the public fullnode dropped json-rpc). past object versions from suiscan, nodeinfra, tatum & blockvision, cuz publicnode & suiet prune history.
- queries: suix_queryTransactionBlocks w FromAddress / ToAddress / MoveFunction filters, suix_queryEvents, sui_getTransactionBlock w balance changes, events & inputs, sui_tryGetPastObject, sui_getNormalizedMoveModulesByPackage, suix_getOwnedObjects, suix_getDynamicFields.
- trades & candles: api.geckoterminal.com (pool trades w tx_from_address, OHLCV), api.dexscreener.com, api.coingecko.com. tweets: api.fxtwitter.com/status/{id}, no login, timestamps to the second.
- ethereum side: blockscout's api & etherscan's address labels (that's where "Kraken: Hot Wallet 2" comes from). bitcoin side: mempool.space. the tBTC redemption's destination was decoded from the approveAndCall extra data (a P2WPKH output script), then bech32 encoded.
- decimals: SUI, ALPHA, stSUI, WAL, BLUE 9; USDC, USDT, DEEP, USDSUI 6; xBTC, tBTC, WBTC, BTC 8. SUI/USD of 1.199 is taken from idear.sui's own 71,500 SUI to 85,729 USDC swap at 23:27:53 on sep 25.
- most of the scripts ran from a throwaway scratch folder & aren't preserved; everything cited can be re-derived from the digests, which is why every one of them is a link. the exception is the portfolio rebuild, whose scripts are kept w the site's source (scripts/alphafi-portfolio); more on that below.
reading the vault tickets
pending UserWithdrawRequest tickets sit inside each position's withdraw_requests VecMap (sui_getObject on the position; in the json it's entries w fields.key & fields.value.fields). claimed ones move to all_withdrawals, an ObjectTable (suix_getDynamicFields, then sui_multiGetObjects). each ticket carries time_of_request, time_of_acceptance, time_of_unlock, time_of_claim, status (1 pending, 2 claimed) & token_amount. the pool is 0x06a49223…a6f8; its InitiateWithdraw / Settle / ClaimWithdraw events are typed under the original package 0x97543070…, not the current one, which cost me an hour.
the dry run
lock enforcement was tested w sui_devInspectTransactionBlock through @mysten/sui (SuiJsonRpcClient), building a user_claim_withdraw call exactly like a real claim & having the node execute it as the ticket's owner. nothing signed, nothing committed, nobody's wallet touched. the claim on the 56,770 ALPHA ticket stamped oct 14 aborts w code 1006 in update_all_withdrawals_and_get_amount; the same call on an unclaimed ticket from the 22:30:59 batch succeeds. details on the vault lock.
reading the positions
alphalend position ids are dynamic fields under the positions table 0x9923cec7…91ba of the LendingProtocol object 0x01d9cf05…0b93, keyed by {type "0x2::object::ID", value}. suix_getDynamicFieldObject w that key returns the position w its collateral & loans. PositionCap objects carry a position_id field, which is how the cap handoffs on sep 7 were matched to positions. loan_bailout's signature came from sui_getNormalizedMoveFunction on the lending package 0xe48b33ef….
the portfolio rebuild
the portfolios page replays each position's own events from sui mainnet: every tx that changed the position (suix_queryTransactionBlocks w a ChangedObject filter on its dynamic-field object), keeping only the events scoped to that position id. per market, the last UserRewardDistributorRefreshEvent in a tx decides whether it's collateral or debt & whether it's still open (a share of 0 closes it). collateral is that share in raw token units; debt is the market's last BorrowUpdateEvent amount, carried forward when a tx doesn't move it. decimals come from MarketDataUpdateEvent's decimal_digit. snapshot every market after every tx, round to 6 decimals & that's the closing net worth.
the trap: on a full repay, BorrowUpdateEvent reports the repaid balance, not zero. track debt off borrow updates alone & you get a phantom loan that never goes away (& when P1's stSUI came back as collateral in oct 2025, the phantom ate it). the refresh share hitting 0 is the real close.
checked against alphafi's own portfolio screen, opened as the cap holder w the read-only wallet trick at the bottom of the portfolios page. these four P1 rows match it to the last of six decimals:
| utc | tx | type | closing net worth (supplied +, borrowed -) |
|---|---|---|---|
| 7 December 2025, 14:32:08 | yetPUDHS… | liquidation | USDC -1,958,376.028802 · ALPHA 1,759,327.589405 · suiUSDT -978,613.606814 · WAL 35.478991 |
| 4 December 2025, 00:40:24 | ApykA8jZ… | claim rewards | USDC -1,957,547.243039 · ALPHA 1,759,327.589405 · suiUSDT -978,613.606814 · stSUI 914.039734 · WAL 35.478991 |
| 3 December 2025, 10:18:27 | 9mEpkyCo… | repay | USDC -1,957,547.243039 · ALPHA 1,759,327.589405 · suiUSDT -978,613.606814 · WAL 0.000003 |
| 3 December 2025, 05:35:57 | HZDTBCWi… | liquidation | USDC -5,957,188.516401 · ALPHA 1,759,327.589405 · suiUSDT -978,613.606814 · WAL 0.000003 |
the end state checks itself too. on 2026-09-28, sui_getObject on each position's dynamic-field object returned exactly what its last rebuilt row says, w no loans left on any of them:
- P1: ALPHA 1,759,327.589405 · WAL 35.478991
- P2: empty
- P3: ALPHA 2,082,086.714007
- P4: ALPHA 1,822,438.51595
the scripts are kept w the site's source in scripts/alphafi-portfolio: fetch_hist.py pulls the events over public rpc, generate_data.py writes the data file & re-running it on the cached events reproduces the page's data byte for byte. apr, earnings & the earning forecast are live-only numbers a dead market can't honestly report, so the page marks them na; the usd on its summary cards is approximate market value, dated on the page.
position labels
this site labels the four positions the way the oracle timeline does. my working notes & the first three thread 2 receipts cards used a different order. the addresses are what matter:
| on this site | position id | opened by | on the t2-1 / t2-2 / t2-3 cards |
|---|---|---|---|
| P1 | 0x9e9d0f9c2de84c68855d753d3bfbd2afe5e6a624b23ab1c0763e0cb3e8820de0 | 0x4dde… | P3 |
| P2 | 0x74d35b4159acb6c48cb0581c4bcaeb0fe5bd347f59ec11cd6255027bca6a9a2d | 0xdfe2… | P1 |
| P3 | 0xd439d8da50cc30791939a3af05e1d0723c01a8e545ed16c61a276d8408adf288 | 0xa66f… | P2 |
| P4 | 0xf0a41462670462e55eaed66a76a43eeba979f27167bdd77f971e40891db25b94 | 0x1d1c… | P4 |
things i got wrong first
in roughly the order i found them. the write-up behind these pages went through six revisions in two days & most of the changes were me being wrong in public. keeping the list.
- my original tweet (sep 26) said a027 sold ~$197k 18 minutes after the tweet & ~$11k 18 minutes before. the chain: a027 claimed 214,001 ALPHA (~$197k at $0.92) & sold 61,104 of it for ~$27k; the pre-tweet ~$11k was a59f, not a027, which was idle from 18:16 to 23:21. the dollar figures were explorer valuations of what moved, not sale proceeds. idear's ~$111k was really ~$85.5k in SUI.
- "nobody else was exiting that afternoon." a draft line in thread 1 said the cluster left alphalend on an afternoon when nobody else was. wrong: 5 to 17 wallets an hour were exiting all day. what's true is there was no rush until the tweet. i posted a correction reply.
- the fall to $0.14. an early summary blamed idear.sui for it. the candles put idear's swaps at $0.92 to $0.58 & a027's sales 14 to 18 minutes after the tweet at the drop to $0.14.
- "wallets that pay into the fee wallet." i described f71e & 1d1c that way early, implying they operate it. paying fees into a fee address is what any user does. the treasury attribution now rests on the behavior listed on who moved early.
- the 6 tBTC from the fee wallet to a027 looked like a payment. it was a conversion job: a027 kept none of the 38 BTC-equivalents it handled.
- "only 8 wallets queued" on sep 21 was misleading wording. eight is two to four times the normal daily count. it's a lot, not a little.
- 35 days. i'd said the unlock was 35 days cuz the app says so. the on-chain lock was 28 days from the weekly settlement; 35 is the worst case. i also had the settle on saturdays. it's sundays 14:00 utc. measured on 278 tickets.
- the 22:01 tx was first summarized as a manual oracle push cuz only its first commands got printed. it also activates emergency mode on seven pools & zeroes the WAL loop lock, so the on-chain wind down starts 68 minutes before the tweet, not 53.
- the bailout total was $16.3M w eight repayments until a ninth RepayEvent (10.15 xBTC on P4, $822,749) turned up. it's $17,129,783 & nine calls & the 94.86 xBTC now reconciles. i also had loan_bailout listed as an admin function; its signature takes no capability at all, which is the whole point.
- a027's reaction time was "24 seconds" until i found the pool state after the 23:20:29 process call (392,280 ALPHA). it's 35 seconds after the pot could cover it.
- four DEEP amounts had been divided by 1e9 instead of 1e6 (DEEP has 6 decimals): 1,103 DEEP, 326,336 DEEP, 1,068,232 DEEP, 2,608,000 DEEP.
- "one owner" for the cluster became "linked by funding & synchronized behavior." "within 21 minutes" became "about 25 minutes." "six minutes" became "seven minutes." small, but they add up to a habit of rounding toward the story.
- "made every queued request eligible immediately" was too broad. a ticket's unlock date is stamped at settlement & the contract enforces it (dry run), so the lock change reached only the 22:30:59 batch & later requests. 51 tickets (86,672 ALPHA) stayed locked. the batch split is now exact too (a027 214,001; 0x6bf6 22,147; a59f 11,566; 23 others 1,582), replacing "about 23,700 over roughly 24 wallets." a request from sep 21 would normally have unlocked oct 25, not "around oct 19."
- "the collateral stayed where it was" & "none of the four position owners has signed a tx since." wrong & misleading, respectively. P2's 3,261,824 ALPHA was withdrawn on sep 25 at 11:39 by the wallet holding its cap & the caps had moved to the fee wallet in december 2025 (all four by dec 25) & to a fresh wallet on sep 7. the creator wallets being quiet meant nothing.
- the heist. the first version of september read as: the team pulled the real collateral & bridged it away. the wallet that did it sits on the rescue side (funded & gassed w the bailout wallet, the collateral sent into kraken deposit addresses), the fee wallet had itself consolidated that collateral into the handed-over position & the fee wallet is now empty. i had a whole thread drafted. it's in the receipts page as "never posted." the bridge totals also got refined: 24.65 tBTC, 68.65 WBTC, 118.66 ETH (of 118.77 withdrawn).
- two numbers from a parallel set of notes that i didn't adopt: ALPHA on 2025-12-03 was $0.95 w a $9.27M market cap per coingecko (not ~$1.18 & ~$11.5M), so the oracle mark was 14.9x market, not 12x; & P4's first two hours of borrowing were about $9.3M at the protocol's own values, not $29M.
- oct 2026, a batch from re-reading everything. "the fee wallet had two of the caps since december": it had all four by dec 25 (P1's on dec 4, P3's on dec 20). P4's cap moved four hours after creation, not seventeen. "the god wallet holds the upgrade caps" (on the gate & in my own sep 30 tweet): the oracle's UpgradeCap is the lending admin's, which published v1 to v5 & the ALPHA vault's is the vault admin's. "you can't read v4's logic off the chain": you can, the disassembly is right there, 9 functions changed. "the move registry record resolves to Zorag44's repo": that record belongs to a copy, not the production oracle.
- "liquidated about $12M". no sale is visible on chain. the security was sent into exchange deposit addresses (kraken on sui & on ethereum). that's what the pages say now.
- "exchange-style" wallets. two of the wallets i'd left unnamed are exchanges: 0x03b6… is kraken's sui hot wallet & 0x6095… is coinbase's, both identified by which deposit addresses sweep into them. so they're venues, not rescue-side parties & the rescue's USDC was coinbase withdrawals. a few lines also said flatly that the foundation paid; what's on record is that it said it would assist & the pages say that now.
- "fat finger or choice" isn't open anymore: the sdk PR that pointed ALPHA at AVAX's feed was committed 4m37s before the chain did it (the code).
- oct 15 2026, an outside fact check. a reviewer went through the live pages against the chain. what changed: "the vault is empty" (the queue is; ~3.4M ALPHA is still staked, free to leave); the dec 17 ALPHA move was 14:04 to 14:11, not 10:47 & the P3 deposit was signed by P3's creator, not the fee wallet; the sep 25 limits tx zeroed borrow limits, not deposit caps; the christmas $2.27M was borrowed by P2's creator that morning, not the fee wallet; kraken took ~1.21M USDC across 8 addresses (net ~$0.92M, not $0.54M), so the exits total is ~$21.3M; P1 also kept 35 WAL; the stash wallet is empty now; w862d's first money from the lending admin was a token basket, not $55k; a handful of windows were a minute or two off (93 s not 94, 5 min not 6, eight minutes not nine, 15 LTV flips not twelve); "day 28" was day 29; the UK & dubai complaint links had moved. what held up: the ~$0.93 pre-tweet price (intraday candles), the 10 protected positions (10 distinct ids across 7 txs), the $17.13M bailout.
- oct 15 2026, the three pages the reviewer didn't reach. i ran the same check on the code, the money & shutdown night: 504 claims, every digest, time & sender exact, 21 corrections, none that change what a page argues. the code: jangid cleared "AlphaFi" from his github profile on oct 13 (the jul 12 archive still shows it) & the pyth cron post was 23:01 utc, not 3am. the money: f71e put 10M BLUE into the ember vault, not 5M; 0xf533 did borrow ~$1.6M on its position (all repaid, net zero), so "dust" was wrong; the 500k perps deposit came from a closed LP position, not an 0xe663 withdrawal; w9273's position has $67.6M of lifetime borrows across all markets ($51.3M was the ALPHA-market figure); the 0x5f81 repay was sep 9 only; the jan 16 stSUI leg left the fee wallet as SUI. shutdown night: the cold wallet has six outgoing txs now, not two (117k USDC & 38.8k SUI left on sep 27 & 28); big5234 has 531 txs in 13 months, not 50 in 8; the stash started selling at 00:09, not 01:16 & is empty since oct 1; the 11:00 hour had 61 exits by 13 wallets, not 44 by 12; two "0" claims were 0.39 & 0.1 ALPHA. new, not a correction: the god wallet (ALPHA's deployer) queued its own vault exit at 23:17:51, eight minutes after the tweet & two more Gate deposits landed on oct 15 2026 while this check was running (w862d 119,309 USDC, w2310 125,419). the exits data was then rebuilt to include those two & the eighth kraken address: 230 deposits, 201 stablecoin, 16 accounts, ~$21.5M one-way. the 52-inside-an-hour, 41, 37 & ~2-by-chance figures didn't move.
what is not proven
- who controls cluster A or idear.sui & whether the treasury relationship is employment, investment, market making or OTC dealing.
- who controls 0x5f81…, the wallet holding the four bad-debt caps. alphafi's fee wallet created it & handed it the caps on sep 7; the rescue's USDC repaid through it from sep 8.
- whose money the 10,544,043 USDC of sep 7 was (0xe6c9… to 0xf2f1… to 0x5f81…). it came out of coinbase, through a small relay (0xe6c9 & 0x3e66, one private gas wallet) whose withdrawals & deposits are on who paid. whose coinbase account, only coinbase knows.
- whether the rescue-side wallets (the bailout wallet, 0xf2f1, 0xe6c9) are the sui foundation's own or an agent's. their USDC came out of coinbase, their gas out of kraken & the collateral went back into kraken. the foundation-labeled wallets use the same two exchanges but share no wallet, deposit address or gas wallet w them. the foundation's statement & alphafi's thread are the only attributions.
- whether anyone was told about the 22:16 lock change, or just saw "claimable" appear in the app. the 25 other receipts in the 22:30:59 batch would have seen it if they were looking; one acted before the tweet.
- whether anything was said in alphafi's discord or telegram b/w 17:00 & 23:09 utc. chain data can't see that. no wayback machine snapshots of alphafi.xyz exist for sep 24 to 27. the discord's #announcements hasn't had a post in about eight months & its general chat is gone, so that can't be checked there anymore (screenshots).
- why idear.sui & the a027 cluster both exited alphalend b/w 17:29 & 18:15 while the rest of the day was routine. the admin's flow-limiter removal at 17:30:52 sits in the middle of that window.
- who the people behind the four position creators are & who holds the lending admin's key. raj says he made at least one of its market calls (the code); the chain only says the lending admin signed it. accident vs on purpose is a lot less open than it was: the mark went $2.77 (june), $3.65 (august), $14.16 (december) while the market fell toward $1 & the sdk wrote each remap down minutes before the chain did.
- ALPHA's market price on 2025-06-25 & 2025-08-24 (no free price history reaches that far).
- what the app displayed as the unlock date to users w pending requests before sep 25 (the stamped settle-plus-28 date, or request-plus-35). no snapshot exists.
- what the pool field time_from_locking_period_for_unstaking_to_start (also 2,419,200,000 ms) does; it doesn't enter the ticket arithmetic.
glossary
- ALPHA vault / ember pool: the staking vault for the ALPHA token. deposit ALPHA, get a receipt, earn rewards, wait a lock period to leave.
- ticket: a UserWithdrawRequest inside your vault position. created when you initiate, stamped when the admin settles, paid when you claim.
- settle: the admin's weekly settle_requests_manual, sundays 14:00 utc give or take. it stamps every pending ticket's unlock date as settle time plus the lock.
- the pot: claimable_balance, the shared pool of ALPHA that claims are paid from. the keeper fills it one settled batch at a time.
- alphalend: alphafi's lending protocol. 36 markets. positions hold collateral & loans.
- PositionCap: the ownership object of an alphalend position. hold it & you can borrow against, repay & withdraw from that position.
- loan_bailout: a public alphalend function that lets anyone repay any position's loan by id, w no capability. the normal repay needs the cap.
- oracle mark / booked: the value the protocol wrote into its own events for a deposit, i.e. what it believed the collateral was worth.
- fee wallet: the vault's fee_address, set by the admin. where protocol fees go.
- cluster A: my name for the a027 group of wallets. rescue side: my name for the bailout wallet, 0xf2f1, 0xe6c9 & the relay around it. not 0x5f81, the wallet that holds the caps: alphafi's fee wallet gassed it & gave it the caps & the rescue's USDC repaid through it. the money that reached them came out of coinbase & kraken, which are exchanges, not parties. treasury: my name for f71e, 1d1c, 7c26, w9273 & 6a49, based on behavior.
- hot wallet: an exchange's own wallet that customer deposits get swept into & withdrawals get paid from. deposit address: an address an exchange gives one customer; it only ever sweeps into its own exchange's hot wallets, which is how an exchange can be identified w no public label.
- gas sponsor: a wallet that pays another wallet's transaction fees by co-signing. one private sponsor co-signing a handful of wallets is how i tied the relay together. relay: my name for 0xe6c9, 0x3e66 & the wallets that sponsor signs for.
- lending admin (0xa1eb94d1…): holds the oracle's & alphalend's UpgradeCaps & every market's settings. vault admin (0xe25b5d16…): holds the ALPHA vault's UpgradeCap & ran the shutdown-night runbook. god wallet (0xa511088c…): ALPHA's deployer.
- t0: 2026-09-25 23:09:29 utc, the alphafi tweet. "+12m" means twelve minutes after it.