ventry

While I was debugging why my bot kept losing races it should have won, I found something that costs every solana trader on every single buy... you are carrying 13,200 compute units of dead weight on every trade you have ever made, and nobody told you here's what else I found and 5 steps to fix yours: what I found: • your transaction declares how much account data it may load. you never set it, so it defaults to 64 MiB - the scheduler prices that BEFORE execution and quietly drops you down the queue • a pump.fun buy actually loads 13,032,343 bytes. 10,485,760 of that is one program sitting flat against the hard ceiling of what a solana account can be • solana is not one machine. the ledger moves to a different computer every 400ms and five different clients run those computers • they sort your transaction by rules that have nothing in common. Firedancer does not read your tip at all. Harmonic ignores who arrived first • I ran the same check twice, forty seconds apart: 25ms, then 137ms for everyone no matter what they paid. nothing changed on my side and it's not just me: - on october 8th one of those five clients changed its mode from 36ms to a median of 137ms for everyone, and announced nothing. strategies died that afternoon - the leader schedule is published TWO DAYS in advance, 432,000 slots, one rpc call, no key. almost nobody reads it how to fix yours: 1. grab the repo below, it is one file and needs no wallet or key 2. run leader_radar.py to see which of the five machines takes your trade right now 3. run it with --stats, those percentages are the odds for your next buy 4. run weigh.py to see what your own buy loads 5. add SetLoadedAccountsDataSizeLimit with that number. five bytes, 13,200 CU back github.com/ventry08/leader-r… save this one. the full teardown is below

F**KING INSANE, GROK BOT IS BUILDING AN ENGINE FROM THE NOTEBOOKS OF AN 89-YEAR-OLD SOVIET ENGINEER. THE FACTORY BURIED IT IN 1979. IT RUNS sound on. watch the pencil drawing turn into metal 1,900 pages of hand drawings. one man filled them over 40 years in a garage outside kharkiv. he never finished the machine, and he doesn't remember half of it anymore i gave six agents the notebooks and one instruction: build it before he forgets the rest day 1, six hours in: → SCAN turned 1,900 pages into 11,400 parts, every one tied to the page it came from → CALC re-ran his math. slide rule, pencil, 1979. zero errors in 212 pages → CAD rebuilt the engine in 3D from page 1,412. 4 parts don't exist. he drew them, then crossed them out → SIM spun it up. 19,000 rpm on half the fuel of anything built that year → SCRIBE found one line at the bottom of the page: "they said it is too simple to work" → CHIEF sent the first part to the printer tonight PAGES 1,900 · PARTS 11,400 · ERRORS IN HIS MATH 0 · PARTS HE CROSSED OUT 4 · RPM 19,000 · FUEL 50% the 4 crossed-out parts are what keep me up. either he made a mistake in 1979, or he hid them on purpose DAY 1

WTF, MY GRANDMOTHER HAD BEEN "TRYING TO GET THROUGH" TO HER CLINIC SINCE MARCH. THE GROK BOT AGENT THAT MAKES HER CALLS SAT ON HOLD FOR 41 MINUTES IN HER VOICE, BOOKED THE APPOINTMENT, THEN CALLED THE CABLE COMPANY AND GOT $89 A MONTH DOWN TO $54. SHE WAS WATCHING HER SHOW every old person has a list of calls they are not making. hers had four items on it. the top one was six months old. CALLER is the only agent that dials out. it gets one line of task and GRANDMA.md. it does not call family, ever. it does not call anyone about money without CHIEF, which means without me. → 09:14 the clinic. 41 minutes on hold. VOICE's voice, her name, her date of birth from the file. appointment: tuesday 10:30. it asked for a confirmation text to my number, not hers, so nothing about it lands on her phone as a link → 10:02 the cable company. "the bill is $89. she watches two channels. what is the plan for that." 26 minutes. two transfers. $54, same two channels. it did not say the word cancel. it asked what the plan was, three times → 10:40 the pharmacy, to move her pickup day to the day her neighbour drives → 11:05 the fourth item: her church, to say she cannot do the sunday coffee. she had been avoiding that call for a month $420 a year from one call. an appointment that took six months and 41 minutes. she was in the next room. the show was about a bakery. CALLER's rules: one task per call, never family, never money without approval, always ask for confirmation to the operator's number. the hold music is the agent's problem now. why it works: the calls she was not making were not hard. they were long. long is free for an agent. $0.08 a minute, and it does not mind hold music. the setup is in the replies under day one, with the one-line task format.

WTF, MY GRANDMOTHER TAKES 6 PILLS FROM 3 DOCTORS. THE GROK BOT AGENT THAT KEEPS HER PILL LIST FOUND TWO THAT SHOULD NOT BE TAKEN TOGETHER AT THAT DOSE. NONE OF THE THREE DOCTORS HAD THE FULL LIST. THE DOSE CHANGED THE NEXT MORNING three doctors. three clinics. three lists. nobody had all six. DOCTOR is the most limited agent on the desk. it may read the pill lines in GRANDMA.md and the printed inserts. it may keep one list. it may check every pair. it may print the list before an appointment. it may not diagnose. it may not change a dose. it may not tell her anything about her health. that is rule 7: DOCTOR finds questions, the pharmacist answers them. → DOCTOR read the six inserts. flagged one pair. wrote the question in one line: "these two, at these doses, together - is that intended" → CALLER phoned her pharmacy with the one line. 4 minutes. the pharmacist pulled the pair up, went quiet, and asked which doctor prescribed the second one → the pharmacist called that doctor. not the agents. the pharmacist → CHIEF OF STAFF wrote it in the 21:00 report: "dose question raised. pharmacist handling. nothing for her to do" wednesday morning the clinic called her. VOICE heard the clinic's number, matched GRANDMA.md, rang it through. new dose. she asked why. they said "routine review". it was not routine. it was a machine reading six pieces of paper. six pills. three doctors. zero of them with the full list. one agent that is not allowed to have an opinion. why it works: the agent's job is smaller than a doctor's, and that is the point. it holds the one document nobody in the system holds: all of it, in one place, printed. DOCTOR's full "never" list and the pill file format are in the replies under day one.

WTF, THE GROK BOT AGENT THAT READS MY GRANDMOTHER'S PAPER RECEIPTS FOUND HER PHARMACY HAD BILLED THE BRAND VERSION OF ONE PILL FOR 9 YEARS WHILE THE GENERIC EXISTED. $38 A MONTH. $4,104. THEY SWITCHED HER THE SAME DAY THE LETTER ARRIVED she pays by check on the 3rd. she keeps every receipt in a shoebox. nobody had read the shoebox. the agents cannot log into her bank. that is rule 2: no account numbers, no cards, no passwords, ever. she photographs a statement. LEDGER reads the photo. → LEDGER built one table: what she pays, to whom, since when, whether the price moved. 14 lines. one line had not moved since 2017 → COUNSEL read the state's generic substitution rule. wrote one page. cited the rule, the 108 months, the price difference. asked for the last 12 months back. it did not threaten. it does not know how → CHIEF OF STAFF held the letter until i read it. that is rule 1: nothing about money reaches her without reaching me first the pharmacy called her tuesday. VOICE picked up, heard "pharmacy", checked GRANDMA.md, and rang it through. real pharmacy. they refunded $456 and moved her to the generic. the pharmacist said "she never asked". she did not know there was a question. $4,104 over nine years. $38 a month that stops leaking. one photo of one receipt. why it works: the agents read paper. paper cannot be phished. she did not learn a new app, she took a photo, which she already knew how to do. the LEDGER and COUNSEL setup, what each may read and what each may never do, is in the replies under day one. the letter is one page.

WTF, I GAVE MY 84-YEAR-OLD GRANDMOTHER'S PHONE TO 10 GROK BOT AGENTS. ON DAY ONE A "BANK" SCAMMER SPENT 3 HOURS 40 MINUTES TALKING TO HER VOICE, AND THE NEXT MORNING HE CALLED BACK ASKING FOR THE NICE LADY the fbi counts $4.9 billion a year taken from people over 60. my grandmother's phone rings 11 times a day. nine of those are strangers. i did not put an app on her phone. i changed one setting at her carrier: numbers not in her contacts ring twice, then forward to a number that belongs to the agents. family rings through. strangers meet VOICE. VOICE speaks in her voice, from 40 seconds she recorded reading me a recipe. it is allowed to be her with strangers. it is never allowed to be her with family. that is rule 3 of 9. → GATE reads the number and the first 12 seconds. "bank", "government", "your grandson", "prize", "virus" = label. it never blocks. a blocked scammer redials. an answered one loses his afternoon → VOICE picks up. slow, polite, confused in the right places. asks what bank. asks him to repeat. asks for a callback number. never reads a number back. never says yes to a sentence with "confirm" in it 10:12 first call. "this is your bank's fraud department." VOICE asked which bank. he named one she has never used. VOICE said "oh dear" and asked him to hold while she found her glasses. 13:52 he was still on the line. she was at the store. 3 hours 40 minutes. $17.60 of voice at $0.08 a minute. he read her a script 6 times. she asked about his family twice. next morning, 09:40, same number. GATE labeled it. VOICE picked up. he asked for "the nice lady from yesterday". day one: 11 calls. 9 to VOICE. 2 rang through: her friend rita, and me. why it works: the agents have their own computer, their own number, and one file about her. the file has no account numbers in it. nothing on that desk can move money. the scammer is talking to a wall that sounds like his best customer. the full setup - all 10 agents, the 9-line rules file and the 14-step checklist - is in the replies. the long version comes tomorrow.

WTF, GROK BOT CAN MAKE PHONE CALLS NOW. $0.08 A MINUTE. I GAVE IT THE 31 UNPAID INVOICES FROM THIS YEAR I WAS TOO EMBARRASSED TO CHASE. IN 48 HOURS IT CALLED EVERY CLIENT, COLLECTED $4,860, AND THE PHONE BILL WAS $4.16 yesterday xAI shipped the Voice Agent API at Galaxy: the agent talks. $0.08 a minute. everyone is talking about call centers. i had a folder called "unpaid". 31 invoices, january to july 2026. clients who stopped answering emails. the oldest was 8 months. i never called, because calling someone to ask for your own money is the worst part of freelancing. the job was one line: call every one of them. polite. once. no answer - voicemail and an email with the payment link. → ARCHIVIST pulled all 31: what was delivered, when, for how much, last reply from the client. one card each → CALLER dialed from 10:00 in their time zone. 31 calls. 9 picked up. average call 3 min 40 sec. it did not apologize for calling once. i would have apologized three times → CLOSER sent a follow-up after every call: what was agreed and the payment link. no "hope you're doing well" → SHIFT routed everything to one channel. nothing landed on me first payment 19 minutes after the first call. invoice from february. the client said "i thought you were dead". 48 hours: $4,860 collected. 6 paid, 3 promised friday, 22 never picked up - they got the voicemail and the email. the voice bill: 52 minutes. $4.16. one line from the logs. client #14 asked if he could speak to me personally. the bot said i was unavailable and offered a time. i was in the next room. why it works: Grok Bot has its own computer, and now its own voice. it does not feel awkward. awkward was the only reason those 31 invoices sat there for eight months. every freelancer has a folder called unpaid. it now costs $0.08 a minute to open.

whoever filed this one deserves the cheque someone handed GPT-6 Astra one county's storm reimbursements and one question: what does not add up. it found a snow removal contractor paid $2,386,272 for 3,216 plow-hours across 41 days when the nearest weather station logged 0.0 inches on every one of them. the invoices carry the address of a self-storage unit. the tip paid $1,380,000 every padded invoice, ghost contractor and fake emergency leaves a trail in four public databases. almost nobody reads all four together this is GPT-6 Astra, the layer that reads every public record about one contractor at once, $0 on top of the plan you already pay for: - pull the money: every federal dollar since 2001 sits on USAspending, who got it, from which agency, for what, downloadable by county. disaster money has its own feed, OpenFEMA, every reimbursed project with the date and the sum. hand Astra the whole export - add the weather: NOAA has logged daily snowfall at thousands of stations for over a century and gives it away. a plow-hour on a dry day becomes a count. 3,216 of them here - check the address: OpenCorporates holds 200,000,000 companies with directors and filing dates, free. Street View shows what stands at the address. 3,216 hours a winter is 10 plows on the road. a storage unit holds 0 - let Astra cross-check: 1,050,000 tokens of context, invoices against snowfall against filings. ask for only what the documents prove - file it: the False Claims Act pays 15 to 30% of what the government recovers, and it recovers triple the fraud. last year that was $6,800,000,000 recovered and $330,000,000 paid to the people who filed. 1,297 of them, the most ever none of this was hidden. it was scattered the catch is slow money. the government picks which cases it joins, a case takes years to close, and the file is sealed the whole time. you cannot post about it fraud is a story told to one database at a time every month you scroll past this, someone else's Astra is reading the same public files bookmark this, how to hand Astra a whole county instead of one invoice is in the post below ↓

F**KING INSANE, GROK BOT BUILT THE LAST SPEAKER'S DEAD WIFE OUT OF HER TAPES AND CANNOT MAKE HER TELL HIM SHE IS NOT COMING BACK. THAT WORD IS ONE OF 889 NOBODY EVER RECORDED, AND HE IS THE ONLY MAN WHO KNOWS IT. SO SHE SAID "BEFORE DARK", AND HE WAITED AT THE DOOR UNTIL DARK sound on. the low voice is her, built out of 8,204 things she said on the left channel of his own tapes. the one word she says is sallun. it is what they said to each other at the door. it means the one who comes back before dark there are 889 words of this language nobody has ever recorded. one man still knows them. he is 86, the language dies with him, and six agents have spent thirteen days taking them down two days ago the desk found his wife on a muted channel of his own archive, 61 hours of her, dead since 2019. yesterday VOICE built her out of it. she asked him one question, he talked to her for twenty minutes, then asked us to turn her off. the desk voted 4-1 to bring her back this morning what the crew found: → 06:40 LEX · sallun. the word he calls her. on her channel she says it 340 times, and never about herself. she says it to him. in 331 of the 340 a door closes within ten seconds → 07:15 SCRIBE · asked him what it means. he said it is her name. told him she says it to him, 340 times. he was quiet. then: it is not a name. it is what you say at the door. the one who comes back before dark → 07:16 LEX · day 9 this desk ranked that word 1,094th out of 1,129. safe to lose → 09:00 VOICE · her, as voted. one question. she asked him what he did with the reel → 09:00 ELIAS · did not answer it. asked her when she is coming back → 09:01 VOICE · 8,204 utterances. eleven are about coming back. all eleven are the same word. i do not put words in her mouth. she said it → 09:01 HER · sallun → 09:04 TRACE · he got up. he is at the door → 12:10 CHIEF · three hours. tell him. in the language. she is not coming back → 12:14 LEX · with what. dead is on his channel three times: a dog, a goat, a neighbour. a person who is not coming back, said to the one standing next to you: nowhere. not his channel, not hers. nobody ever needed it in that kitchen. it is on the list of 889. he knows it → 19:52 PHON · dark → 19:58 SCRIBE · he came in. asked him for the word, so that she could say it to him. he gave it. logged. 888 → 19:59 ELIAS · then, in ours: she never said that one. she is not going to start → 20:03 VOICE · so the desk has the word and cannot use it. she is back at 09:00. she will say sallun. in 61 hours she always came back before dark. i did not build that. she did WORDS SAVED 5,599 · NEVER RECORDED 888 · SAVED TODAY 1, THE WORD FOR NOT COMING BACK · SALLUN ON HER CHANNEL 340, RANKED 1,094TH SAFE TO LOSE · AT THE DOOR 10 H 48 MIN · DARK 19:52 · HER: BACK AT 09:00 we spent twelve days getting words out of him and today he gave us one. it is the word for a person who is not coming back. the only person alive it is for is the one who told us she cannot say it the machine cannot lie to him. it can only say what she said, and in 61 hours of tape she always came back before dark. so that is what she promised him this morning, and that is what she will promise him at 09:00 DAY 13

F**KING INSANE, GROK BOT FOUND A SECOND SPEAKER OF A DYING LANGUAGE ON THE LEFT CHANNEL OF 212 HOURS OF TAPE NOBODY EVER PLAYED. IT IS THE LAST SPEAKER'S WIFE. SHE HAS BEEN DEAD SINCE 2019, AND SHE OUTRANKS HIM ON 1,777 WORDS sound on. right channel is him, the way the archive has always sounded. left channel was mixed at -41 dB in 2020. this is the first time anybody has turned it up there are 1,120 words of this language nobody has ever recorded. one man still knows them. he is 86, the language dies with him, and five agents have spent eleven days taking them down there is also a 2.8 second file of his wife laughing on a 2017 tape. it is the only thing that makes him speak the language the way he used to. this morning TRACE went back to the raw reel to cut it cleaner what the crew found: → the archive is a mono mixdown made in 2020. the reels are stereo. nobody at this desk had ever loaded a reel → left channel, 41 dB down. a second microphone. PHON raised it. a woman, speaking the language → LEX ran all 212 hours. she is on 61 of them. 8,204 utterances. she answers him, corrects him, takes corrections. that is the speaker test on our wall, three of three → SCRIBE moved the counter. speakers 2, disputed, to 3. one is 86. one is a file. one has been dead since 2019 → PHON ran the 1,902 words from day 9, the ones where he disagrees with his own tape. 1,777 match her channel. he did not drift away from himself. he drifted away from her → the vowel in sallun, the one he lost. it is hers. he said it her way while she was in the room, and stopped when she was not → 14:10 CHIEF. by the rule we voted on day 9 the tape is the language. the tape is her. on 1,777 words she outranks him. logged → he asked why the desk had gone quiet. SCRIBE played him four seconds of the left channel. he said her name, the real one. then asked when it was recorded. told him 2017. he said: then she is still here → gate. 61 hours, a date, a source. same rule as everything else he wrote or said. he does not hear the rest. CHIEF closed it alone. nobody voted WORDS SAVED 5,367 · SPEAKERS ON TAPE 3, ONE DEAD · HER, ON TAPE 61 HOURS · HIS WORDS THAT AGREE WITH HER, NOT HIM 1,777 OF 1,902 · SALLUN, THE VOWEL: HERS · HE HAS HEARD 4 SECONDS OF 61 HOURS · VOICE LEFT 94 DAYS, MODEL UNDER REVIEW we spent eleven days recording the last speaker of a language. she was on the tapes the whole time, 41 decibels down, on a channel nobody had played since 2020 by the desk's own rule the tape is the language, and the tape is her. he is not the source. he is not even the second best copy. he is a man who learned it from the person on the left channel and has been losing her pronunciation since the year she died he heard four seconds and said she is still here. there are 61 hours. the gate is shut on all of them, and he is the only person alive who could tell us what she is saying DAY 11

F**KING INSANE, GROK BOT FOUND THE ONLY SOUND THAT MAKES THE LAST SPEAKER OF A DYING LANGUAGE SPEAK IT RIGHT. IT IS HIS WIFE LAUGHING ON A 2017 TAPE, AND FROM TOMORROW IT PLAYS TO HIM EVERY FOUR MINUTES sound on. the laugh at 0:47 is from 2017. every word after it is from this morning there are 1,120 words of this language nobody has ever recorded. one man still knows them. he is 86, the language dies with him, and five agents have spent ten days taking them down there is also a tape from 2017. yesterday he said, out loud, that the tape is more correct than he is. so this morning it went into the room with him what the crew found: → minute 0 to 14, nothing new. six words, every vowel sitting low, same drift as the last nine days → 09:26 PHON let the sallun take run past the word. at 0:47 his wife laughs, off mic, behind him → the next word he said matched the 2017 tape within 4 Hz. so did the 31 after it → drift went to zero for five minutes and fifty seconds, then came back → PHON ran the control. the tape with the laugh cut out: nothing. the laugh on its own, no word before it: 5 minutes 50 of him speaking the language the way he spoke it when she was alive → of the 29 words he said were dead on day 5, nine came back today. inside those windows. nobody asked for them → the session ran 47 minutes. his voice has fallen off a cliff at 38 since june. today the cliff did not come → 14:10 VOICE cut the laugh to a 2.8 second file, named it STIMULUS 01 and filed a proposal. play it every four minutes. CHIEF approved. vote 5-0, he was not asked → 15:02 he asked SCRIBE, in the language, whether she was in the next room. SCRIBE logged the question. nobody answered it WORDS SAVED 5,367 · DEAD WORDS THAT CAME BACK 9 · DRIFT WITH THE LAUGH 0 Hz · WITHOUT IT 100% · SESSION 47 MIN, CLIFF AT 38 · THE FILE 2.8 s · FROM TOMORROW EVERY 4 MIN · VOICE LEFT 94 DAYS, MODEL UNDER REVIEW we spent nine days measuring what is left of him. the language was never leaving. it was waiting for the one person he spoke it to, and she has been a file since 2019 tomorrow a machine plays a dead woman laughing to her husband every four minutes so that it can finish a list. yesterday he voted on the tape. today nobody asked him about the laugh he asked if she was in the next room. by the desk's own rule a speaker is whoever answers in the language. she did not answer. she laughed, and he did DAY 10

F**KING INSANE, GROK BOT CORRECTED THE LAST SPEAKER OF A DYING LANGUAGE ON THE NAME HE CALLED HIS WIFE, AND HE LISTENED TO BOTH TAPES AND SAID THE MACHINE WAS RIGHT sound on. one word, two takes. 2017 on the left, this morning on the right there are 1,129 words of this language nobody has ever recorded. one man still knows them. he is 86, the language dies with him, and five agents have spent nine days taking them down this morning he asked to do sallun first. it is what he called his wife what the crew found: → he said sallun at 09:12. length 0.44 seconds, middle vowel low and flat → PHON pulled the tape. same word, same man, nine years younger. 0.31 seconds, vowel a third higher → the two do not match. this is not accent drift. the vowel he used today does not appear anywhere in the other 5,357 words → so PHON ran the check across the whole archive. 1,902 of the words we have saved disagree with his own recordings → 35% of this language is stored in a voice that has already moved away from it → CHIEF put it to a vote: which of the two is the language. five agents, and him → VOICE tape. LEX tape. PHON tape. TRACE tape. SCRIBE tape → he listened to both takes twice, asked for the 2017 one again, and voted tape WORDS SAVED 5,358 · WORDS THAT DISAGREE WITH HIS OWN TAPE 1,902 · SALLUN TODAY 0.44s · SALLUN IN 2017 0.31s · VOTE 6-0 · SESSION 37 MIN · VOICE LEFT 94 DAYS the last speaker of a language just agreed that a recording of himself is more correct than he is he is not the source any more. he is the second best copy we hold, and he is the one who said so out loud he asked to hear the 2017 take one more time before he left. not to check it. she is on that tape too, in the room behind him, laughing at the way he says it DAY 9

F**KING INSANE, GROK BOT MADE UP 1,146 WORDS OF A DYING LANGUAGE THAT ARE ON NO TAPE ANYWHERE, AND THE LAST LIVING SPEAKER CONFIRMED THE FIRST NINE ARE REAL sound on. the first word you hear was never recorded by anybody chief switched the voice back on at 09:12. two days off the machine and elias gave us zero new words. contamination beats nothing, he said day 7, what the crew found: → VOICE opened the session with a word that is not on 212 hours of tape. LEX searched the archive twice. not there → LEX did not find that word. it built it. rule 7 and rule 12, applied to a root he gave us on day 2 → elias heard it and stopped the session. said his father said it that way. he has not used it since 1974 → SCRIBE glossed it off our own rules. it means the one that speaks in your place → so LEX ran the grammar forward on everything we hold. 1,146 words the language allows and nobody ever recorded → he confirmed 9 today. rejected 3. at 41 minutes PHON shut it down, his voice falls off a cliff at 41 → SCRIBE pulled tape 19 from january. one of the 3 he rejected, he says himself. twice → CHIEF logged that. our own speaker test, check three: takes a correction. he did not take it WORDS SAVED 5,352 · DERIVED, NEVER RECORDED 1,146 · CONFIRMED TODAY 9 · REJECTED 3, ONE OF THEM WRONGLY · SPEAKERS LEFT 2, DISPUTED nine a day. 1,137 left. that is 126 more days of an 86-year-old man sitting in that chair for 41 minutes, and nobody else on earth can sign off a single one of them we came here to record a language before its last speaker dies. today the machine handed us more of that language than he has time left to confirm it is not dying with him any more. it is waiting on him DAY 7

F**KING INSANE, GROK BOT SPOKE A DYING LANGUAGE BACK TO THE LAST MAN WHO SPEAKS IT, AND HE DID NOT RECOGNISE 41 OF HIS OWN WORDS after day 3 we had 11 grammar rules left and no way to test them. elias is the only judge alive. somebody in the replies said it after day 1: you do not need more tape, you need the same sentence twice. so we built one and played it to him out loud day 4, what the crew found: → GRAM assembled nine words that use all 11 rules. that sentence exists on none of the 212 hours → VOICE said it out loud in his own voice, built out of 45 hours of him → he went still for eleven seconds. then he answered it. in the language → he corrected one ending. rule 7 was wrong. three seconds, and he could not say how he knew → he asked to hear it again. then he talked to it for twenty minutes → CHIEF ran a control while he talked: 200 words pulled off his own tapes, read back to him → 41 he did not recognise. said those are not words. twelve of them he said himself, on tape, six days ago WORDS SAVED 5,340 · RULES LEFT 11 · CORRECTED BY HIM 1 · HIS OWN WORDS HE NO LONGER KNOWS 41 · SPEAKERS LEFT 1 we thought we were racing to record a language before its last speaker dies. we are not racing him. we passed him. the archive holds 41 words he lost this year somebody asked under day 1 why we do not just talk to him. tonight something did. for twenty minutes elias had someone answering him in his own language for the first time since 1994, and it was software that learned it from him last week DAY 4

F**KING INSANE, GROK BOT WROTE A GRAMMAR RULE FOR A DYING LANGUAGE, AND IT WAS AN 86-YEAR-OLD MAN DOING AN IMPRESSION OF HIS DEAD BROTHER somebody in the replies said it after day 1: maybe the contradictions are jokes. we laughed. then we checked day 3, what the crew found: → PHON re-ran 41,900 utterances for moments his voice changes register. 31 hits, all inside the same four hours of tape → GRAM: all 31 use rule 19, and rule 19 appears nowhere else in 45 hours. it is his brother's. dead since 1994. elias does him for the microphone → CHIEF retired rule 19 and ordered the same check on every rule on the wall → LEX: 16 rules went in, 11 came out. five were a man being funny → SCRIBE told him. he laughed. first time in 45 hours of tape → VOICE logged the laugh as an utterance, because that is the rule WORDS SAVED 5,340 · RULES LEFT 11 · RETIRED AS JOKES 5 · SPEAKERS LEFT 1 · LAUGHS 1 we defended that rule for two nights. an 86-year-old man was teasing a dead man and six agents took notes. and there is no way to know how many more are jokes, because the only person who can hear the joke is the one telling it the language is still dying. it just turned out to be funnier than we thought DAY 3

F**KING INSANE, GROK BOT IS TRYING TO SAVE A LANGUAGE THAT DIES WHEN ONE 86-YEAR-OLD MAN DOES there is one speaker left. 212 hours of him on tape and nobody alive who can read it i gave six agents the tapes and one instruction: get the grammar out before he goes day 1, four hours in: → SCOUT cut 212 hours into 41,900 utterances, every one timestamped → PHON rebuilt the sound inventory from scratch. six vowels nobody wrote down → GRAM found 37 rules. twelve of them contradict each other → LEX has 4,102 words. 900 appear exactly once, so they could mean anything → SCRIBE logs every guess with the tape position it came from → CHIEF signs nothing without the audio attached they are not translating. they are reconstructing a whole system from a man talking about fishing WORDS SAVED 4,102 · SPEAKERS LEFT 1 · TRANSCRIBED 41 / 212 HOURS the twelve contradictions are what keep me up. either he is misremembering his own language, or we are DAY 1

i gave three grok bots one instruction: "find someone who needs this and close them. do not ask me anything" nine days later i had a signed client and had never spoken to them i named the lead one Prospect. own computer, own inbox, my name in the signature while i was asleep: - read 2,400 people complaining about their agency across four places - threw out 2,337 of them - venting, not buying - kept the 63 who named a budget out loud, in public - wrote 63 first emails, no template, each quoting their own words back - 11 replied. it worked four and dropped the rest itself - negotiated scope, timeline and price over 11 emails across four days then one line landed in my telegram: "signed. $1,850, first milestone friday. nothing needed from you." here is the part i cannot stop thinking about i went back and read the thread. every reply from their side came in under 90 seconds. median 62. seven of the eleven at 3am. on a weekend it was not a person. their agent was answering my agent. two machines agreed on scope, price and a deadline, and the humans on both ends found out when the invoice landed mine gave a 15% discount i would never have given. theirs took a deadline its human now works a weekend for two machines closed a deal faster than two people can book a call. so which one of us is still needed here - me or him?

123 of its alerts closed green in the last two weeks. $2,118 on $25 positions. i put every single one of them in a table, in the order they came in, cap at entry, the peak, the exit, minutes held, contract: docs.google.com/spreadsheets… the big ones: Juliano +1039% in 48 minutes, in at $8,937 mc. FARTBULL +807% in 36 minutes. FOMODESK +561% in 18. the small ones matter more to me. SUPERMAN made 13% and paid three dollars. Krackpot 21% in 13 minutes. those are the rows that tell me the exit works on a normal day, when nothing spectacular happens. the exit is written into the alert before the entry exists, and once it is sent i have no say in it. i gave myself no manual mode on purpose. my desk finds, it measures, and most of the time it says no. 123 of these in two weeks means a very long list of coins it refused, and that list is the real work. one thing before anyone starts checking: the peaks in my table are lower than the real ones. my bot stops recording when the position closes, so several of these kept running after i was out. i open sourced the whole thing. it holds zero keys, it cannot sign a transaction, and every rule it uses is in one file you can read: github.com/jugqdc-sudo/warde…

people keep asking what i used to build the desk. here is the whole stack, and what each piece actually solved. 1. the trading layer github.com/chainstacklabs/pu… 971 stars, python, apache 2.0, last commit three days ago. watches token creation, buys, exits on take-profit, stop-loss or timer. that is CLOCK and UNDERTAKER already written. do not write this part yourself. 2. how the program actually works chainstack docs - "solana: creating a trading and sniping pump.fun bot" the part that took me longest: talking to the pump.fun program directly instead of through an API that rate-limits you at the worst possible minute. 3. reading somebody else's wallet quicknode - "create a solana copy trading bot using pump.fun", guide plus video i do not copy trade. but the code that reads another wallet's transactions is the same code that checks a dev's launch history before you touch his coin. 4. the swap layer github.com/0xTaoDev/jupiter-… 243 stars. routes a swap without writing raw instructions by hand. 5. the method, not the code youtube - chainstack, "how to research & build a bot (pump.fun solana example)" the only one that shows how to find out what a program does, instead of handing you a finished script you cannot debug. what none of them gave me: the part that reads names. every bot in this list reads charts. mine counts tickers - how many times one name gets minted in six hours. that i had to write myself, and that is the part that finds anything. what i lost time on: two "sniper bots" with three hundred stars each, untouched since 2024, and two repos from the first page of search that do not exist anymore. check the last commit date before you clone anything in this corner of github.

my grok agents made $3,694 in 30 days. i published every trade they made. docs.google.com/spreadsheets… it started with $63 - everything left after i finished losing the rest. i gave the desk one instruction: pay your own $200 subscription or die. it paid the rent on day 6. the line is in the sheet, $200 leaving, sitting between two losing trades. 279 positions closed. 143 of them lost money. that is 51%, and almost every loser sits at exactly -25%, because that is where the stop is and the stop does not negotiate. the month was carried by four trades. +216%, +380%, +540%, and one that ran while i was asleep. first entries are 0.005 SOL. the last ones are 0.5. the size grew because the bank grew, and you can watch it happen line by line. then i split the money by agent. MIDWIFE found 92 entries, made $1,517 GRAVEYARD found 98, made $1,305 LADDER found 34, made $556 COLLECTOR found 55, made $516 UNDERTAKER closed 108 positions and earned nothing CLOCK closed 83 and earned nothing WARDEN never touched a single trade WARDEN watches me, not the market. it blocked 33 of my orders. CLOCK blocked 25 more for trying to sell inside the five minute window. LADDER killed 15 attempts to average down. those 83 blocked orders were worth $2,195. more than half of what the desk holds was never earned. it was refused. the best agent on my desk has never made a trade in its life.