Tom · SŌGŌ @0xtomdev

A number on a card was wrong and the cause was a code path that never ran. I was stashing leverage when the user set it, then reading the stash later. If the user opened through a path that skipped the stash, the read fell back to 1x and the card understated the result. Turns out the venue will tell you per coin leverage from an info query even with no open position. So the fallback became: read the stash, else ask the venue, else floor at 1x. And when the venue answers, write it back so the next read is local. Self healing fallback beats a correct value you only have half the time. If you are on the same venue and want the exact query, it is in the public info endpoint under active asset data.

the hardest engineering problem in this space is not latency. its idempotency. making sure the same intent, retried, doesnt become two positions. everyone learns this the same way the expensive way

every dollar this has made came from fees on trades that actually happened. no presale, no allocation, no promise attached to anything. it makes growth slower and the numbers real id take that trade again

the part that should worry builders isn't the $1.3B. it's that Kelp and Drift were 44% of H1 losses and neither was a smart contract bug. Coldcard was a firmware change from 2021. five years of looking like a working product. the metric nobody publishes is dwell time. how long the flaw existed before anyone knew it did. an audit measures the code on the day. dwell time measures how long your monitoring was lying to you. my own small version: a webhook poster that threw the response away, so a 400 and a 200 wrote identical logs. 1,190 alerts in three weeks, zero delivered, and nothing anywhere said so. cost nothing to fix, three weeks to notice. same failure class as firmware nobody read again. different scale, same shape.

zerokn0wledge.hl 🪬✨ @zerokn0wledge_

20 hours ago

How bad 2026 really is? - 120+ projects shut down - Over $1.3B stolen across 344 exploits - Crypto hiring down ~80% What happened? Three things at once, and only one of them was the market. First, the token treasury model finally got tested. Teams raised in tokens, paid

keyboard_arrow_left Previous keyboard_arrow_right Next

14 8 44 2K 8

the Coldcard detail is the one i keep coming back to. a firmware change from 2021 that sat quiet five years. the number nobody tracks is dwell time. not how many exploits, but how long the flaw existed before anyone knew. Drift six months, Coldcard five years. both looked like a healthy system the entire time. a treasury drawdown you can see coming. a silent failure looks exactly like everything working.

Twenty one days on mainnet proves the money path executes. It does not prove the fee accounting is right at scale. It does not prove the expiry path works, because nothing has expired yet. It does not prove the restore works, because I have not needed one. Three things I am asserting, not demonstrating.

today: renamed a function and broke nothing. that sounds like nothing. it means the seams are in the right places and the thing is finally shaped like the problem. first time in weeks that happened i notice these more than launches

that framing holds and i thnk its the right call g the thing that survives it is the asymmetry underneath. a prop firm does not hard cap because it doubts the skill, it caps because its own loss and the traders loss are different sizes the vault has the same shape. the leader is exposed to their own stake, the depositor to theirs, and only one of those two is doing the sizing same problem in mine by the way. the agent key cannot withdraw but it can size, and i dont cap it either. i just have fewer people to lose what does the depositor see, and when. tthe size before the fill or after

Volume is not one number, it is four multiplied together. Capital. Leverage. Round trips per day. Utilisation, which for most accounts sits somewhere between 40 and 70 percent. Which means the two levers that actually move venue volume are leverage and frequency, not deposit size. A small book turning fast outproduces a large book sitting still. That is also why the volume column is easy to rent and open interest is not. You can pay for turnover. You cannot pay someone to feel safe overnight. I ran this decomposition on my own numbers before I trusted anyone elses. x.com/sogoterminal

the telegram trading category on hyperliquid is smaller than you think. a handful of names, one of them shrinking hard this week, two sitting at absolute zero. for a category that should be the most natural fit on the whole chain somethings off about that

the part worth watching is what it does to their builder fee line a wallet brings reach. reach and routed volume are different assets and only one of them pays somebody posted hypertracker numbers last week showing a 114 day old builder with 585 users over 30 days out earning phantom in a 24 hour window. distribution size was not the variable that decided it does phantom take a cut of the builder fee here, or is this a routing deal wich pays somewhere else

Three ways to stop trading, and they are not redundant. One. A single user halts their own account. Instant, scoped to them. Two. A file on disk halts all trading. Broad, but the watcher that reads it only runs when something is flowing, so it is volume dependent. Three. Rotate the internal key. The control plane disarms itself and fails closed. No volume needed for that one to bite. The third exists because I asked what happens if the thing I need to stop is also the thing that stopped reporting. Related policy I hold to. The fill watcher alerts, it never halts on its own. Automated halts on a noisy signal cost more than the signal is worth, and mine was noisy twice in week one. If you run a trading group and this is your problem too, most of this plumbing already exists. DM me at t.me/tomsogo

if a product asks you to trust it, it has already failed. the goal is a design where trust isnt required because the harmful action is not available. thats a much harder spec and almost nobody writes it down writing it down is what makes it real

spent today on privacy math. crew stats dont show until there are enough traders in the group that no single persons position can be reverse engineered from the aggregate. it makes the product look emptier at the start im keeping it anyway

fifteen years of shipping and the thing i finally trust is the boring loop. read only investigation. decision. build. review the diff myself. merge. no step skipped, ever, on anything touching money its saved me twice this month alone

My deploy check was a grep for a string that should not appear in production. It kept passing when it should not have, and failing when it should not have. The string legitimately exists in a preview tab on the same page. A check that can be true for the wrong reason is not a check.

🗂 SOGO │ ├ 🗂 Custody │ ├ Funds stay in your own account │ ├ The signing key cannot withdraw │ ├ 90 day expiry │ ├ Revoke on chain, no permission from us │ └ We never hold a balance │ ├ 🗂 Markets │ ├ Whatever Hyperliquid lists │ ├ One book, no synthetic depth │ └ We do not run a venue │ ├ 🗂 Fees │ ├ Direct lane 2.50bp │ ├ Crew lane 10.00bp │ │ ├ 2.50bp partner pool │ │ ├ 2.50bp platform │ │ ├ 2.50bp treasury │ │ └ 2.50bp development │ ├ Every basis point published │ └ No range, no "up to", no discretion │ ├ 🗂 Crew │ ├ The partner pool goes to the group owner │ ├ Calculated on crew volume │ ├ Never per head │ ├ Share of that pool │ │ ├ Operator 40% │ │ ├ Syndicate 55% │ │ ├ Prime 70% │ │ └ Sovereign 85% │ ├ No second level │ └ No recruitment bonus │ ├ 🗂 Engine │ ├ Telegram, nothing to install │ ├ Every order carries a client id │ ├ Three independent halt paths │ ├ You can stop it without asking me │ └ Fees settle where a block explorer can follow them │ ├ 🗂 Build │ ├ Solo │ ├ Read only first on anything touching money │ ├ External review before a single external user │ ├ It found a claim code redeemable from a session that never minted it │ └ Fixed, and said so out loud │ ├ 🗂 Traders │ ├ One wallet on mainnet │ └ It is mine │ └ 🗂 Next ├ A page for your crew. Volume, ranking, what it paid you ├ Relink, before the key cliff ├ 10 genesis crew seats └ Not done @sogoterminal t.me/sogoterminalbot

Bullet @Bulletxyz

2 days ago

📁 Bullet ┃ ┣ 📁 Markets ┃ ┣ Majors ┃ ┣ Altcoins ┃ ┣ Stocks ┃ ┣ Metals ┃ ┣ Energy ┃ ┣ Semis ┃ ┣ Space ┃ ┣ RWA ┃ ┗ Memes (soon) ┃ ┣ 📁 Engine ┃ ┣ Rust, top to bottom ┃ ┣ No garbage collector ┃ ┣ Sub-100µs in the sequencer ┃ ┣ 3ms end to end ┃ ┣

12 4 32 14K 1

@ethy_agent @HyperliquidX not a criticism, my own relink path is not shipped and i have the same cliff coming in october

the thing that decides whether autopilot is actually autopilot is the key clock agent keys on HL expire, so a strategy left running unattended stops at that boundary and the failure is quiet. orders stop being accepted and the loop keeps loopingg and after any restart the only thing that tels you what actually landed is the client order id you attached before sending. without it you are reconciling from fills and guessing do your setups re-authorise on their own or does the user get pinged?

the security spend side is the harder half of that, agreed the softer half is inside the table you quoted. it carries a 7d column right next to the 24h one and the ordering is not the same in both hyperliquid reads $378,923 over 24h against $6.68m over 7d, wich puts that snapshot at roughly 40% of its own weekly run rate and revenue is not one word across those rows either. some of it accrues to a company, some of it routes to buybacks and an assistance fund. same column, different claimant does the L1 comparison normalise for that, or is it spend against reported revenue