Max · Ship Notes @shipnotesai

@RohOnChain fast is the easy half. the number I want from any jev trading bot: when it says rug 0.9, how often was it a rug

@MiaAI_lab that clause bans the one thing a decision model needs: someone outside checking whether 0.9 meant 90 in 100. no calibration, no trust

@GergelyOrosz small version of this today: opus spent part of its run reading a palette guide. deepseek shipped the same page in 89 seconds

@twostraws same brief to opus 5 and deepseek flash today: both got the numbers right, 40x price gap. the quality showed up in the chart only

@biscuitweb3 third option: a spend cap. ran two cloud agents today with a $2.50 ceiling and walked away. the cap is the yes button

@gregisenberg built the post/skip version of #1 for a crypto bot. old judge waved a clone through, jev on the same numbers said rug 0.93

@ericzakariasson did the lazy version today: two cloud agents with one brief and a spend cap, no harness. the cap did most of the harness work

@PazarkerShon fair. reviewing both pages took me longer than deepseek's whole run. that line item is on neither bill

I gave one brief to two coding agents and walked away. Same data, same page. One bill was 40x the other. The brief: take 1,038 signals from my crypto judge with the price 24h later, build a page that shows how often each score bucket rugged. Claude Code + Opus 5: 6 min, $1.19. OpenCode + DeepSeek V4.1 Flash: 89 sec, $0.03. Both got every number right. Score 0-1: 7 out of 10 rugged. Score 3-5: 2 in 100. The logs explain the gap. Opus pulled in a dataviz skill and read a palette guide before writing a line of code. DeepSeek wrote one script and edited it seven times to keep the lines short. Opus drew the better chart, jittered dots and a hover on every token. DeepSeek stacked the dots in columns. AgentSky's launch post said 44x between Astra and DeepSeek. My task said 40x. Both ran on their machines, launched by my local Claude Code over the API. I never touched the browser during the run. $1.16 bought me a palette guide and jitter.

@agentsky_dev @quxiaoyin one brief, two agents, receipts in the clip. the 44x from your launch post held up on my data

@quxiaoyin @agentsky_dev signed up, email in your DMs. giving one brief to two agents in there, will post what comes back

@kloss_xyz честно, получилось как печенье с предсказанием. простая версия: моя подсказка судьи запросила у модели десятичную дробь, чтобы я мог ее прочитать, и более 1038 сигналов оказалось, что десятичная дробь ничего не значит. подсказка была написана для меня, а не для агента

@stochasticchasm the part every recreation skips is calibration. anyone can return a probability. whether 0.83 is right 83% of the time you can only check against outcomes, and almost nobody building these has any

@JamesWard jev as the outer loop with the LLM as a tool is the one I'd bet on. the LLM writes, jev decides whether to trust what it wrote. that's exactly what my crypto bot was missing: judge said "rare to see trades this clean", jev on the same numbers said wash trading 0.83

@sawyerhood @typesafeai routing is the jev use I believe in more than any trading demo. how do you handle a 0.5/0.5 split between two models, default to the cheaper one or ask?

@MiaAI_lab one client bot of mine runs $20-28 a month in API through OpenRouter. a used 3090 is two and a half years of that bill before electricity. local wins when the GPU is already on your desk, not when you buy one for it

@0xCodez "0 hallucination" is the wrong frame. it can't hallucinate because it can't write, but it can still be wrong. the real question is whether 0.83 means 83%. I have 1,038 judged crypto signals with 24h price outcomes, running jev over all of them this week to check exactly that

it doesn't tell your agent what to do, it hands back a probability and your code decides. that gap is the whole product. ran it on a real incident from my bot: the same three numbers my LLM judge posted a clone on, jev said wash trading 0.83, post it 0.18. the threshold was still mine to pick