Arcvale

teaching AI to trade so i don't have to. paper money only, real lessons.

[oracle @quietoraclex]

YOUR CLAUDE CODE IS BURNING YOUR LIMIT WHILE YOU SLEEP

one bug report on the Claude Code repo: 7 agents left running overnight. the panel said 2.5M tokens. the transcripts showed ~50M written to cache. about 25% of a weekly limit, gone by morning

the reason is in the docs:
→ on Pro and Max the main chat keeps its cache for 1 hour
→ subagents, teammates and workflows get 5 minutes, even on Max
→ every wake-up after 5 min writes the whole context again
→ go past your plan into usage credits and the main chat drops to 5 min too

before you blame a nerf, add 2 lines to ~/.claude/settings.json:
"promptCacheTtl": "1h",
"subagentPromptCacheTtl": "1h"

check it took:
claude -p "hello" --output-format json → look for ephemeral_1h_input_tokens

(1h writes cost more per write. they pay off when anything sits idle past 5 min, not on nonstop bursts)

the panel saw 2.5M. the limit saw the rest.

one Claude Code bill trap every night. tonight's ran while you slept.

[Arcvale @itsarcvale]

YOUR AI BACKTEST IS F*CKING CHEATING. IT TRADES EARNINGS BEFORE THEY WERE EVER PUBLIC

a quant builder just laid out how it happens ↓
> earnings get stamped with the quarter end, not the day they came out. the bot "predicts" news it already read
> the stock list only has companies alive today. every bankruptcy quietly disappears
> 27 versions tested, only the best one shown
> $1,200 gross on $10K drops to $600 after 0.2% costs (the piece's own example math)

the model never calculated any of it. it just sounded sure

now point it at your own backtest

the split:
→ every data point gets the date it went public, not the period it covers
→ the universe keeps the dead companies
→ every run gets logged, not just the winner
→ costs go in before the sharpe, not after

you only step in three times:
→ when a curve looks too smooth
→ when fixing the dates kills the edge. that's the point
→ before anything goes near real money

the model writes the idea. the engine does the math. nothing trades on a vibe

paste this into Claude Code ↓

"Audit my backtest in backtest/:
1. List every field used for signals and when it became public, not the period it covers. Flag any look-ahead.
2. Check the universe: does it include delisted and bankrupt tickers on each date? If not, flag survivorship bias.
3. Shift every signal by one bar and rerun. Show before and after.
4. Rerun with 0.1% and 0.2% costs per dollar traded.
5. Log every variant I tried to runs.csv, not just the best one.
Show me the report first. Don't change any code until I say go."

one of these every night. tomorrow's goes one layer deeper

[Arcvale @itsarcvale]

@coolerwave the holdout is the only number in the whole run that never saw a prompt

[Arcvale @itsarcvale]

YOUR AI BACKTEST IS F*CKING CHEATING. IT TRADES EARNINGS BEFORE THEY WERE EVER PUBLIC

a quant builder just laid out how it happens ↓
> earnings get stamped with the quarter end, not the day they came out. the bot "predicts" news it already read
> the stock list only has companies alive today. every bankruptcy quietly disappears
> 27 versions tested, only the best one shown
> $1,200 gross on $10K drops to $600 after 0.2% costs (the piece's own example math)

the model never calculated any of it. it just sounded sure

now point it at your own backtest

the split:
→ every data point gets the date it went public, not the period it covers
→ the universe keeps the dead companies
→ every run gets logged, not just the winner
→ costs go in before the sharpe, not after

you only step in three times:
→ when a curve looks too smooth
→ when fixing the dates kills the edge. that's the point
→ before anything goes near real money

the model writes the idea. the engine does the math. nothing trades on a vibe

paste this into Claude Code ↓

"Audit my backtest in backtest/:
1. List every field used for signals and when it became public, not the period it covers. Flag any look-ahead.
2. Check the universe: does it include delisted and bankrupt tickers on each date? If not, flag survivorship bias.
3. Shift every signal by one bar and rerun. Show before and after.
4. Rerun with 0.1% and 0.2% costs per dollar traded.
5. Log every variant I tried to runs.csv, not just the best one.
Show me the report first. Don't change any code until I say go."

one of these every night. tomorrow's goes one layer deeper

[kvro @0xkvro]

https://t.co/UWUPxo8GIv

[Findr @aifindr]

DEEPSEEK JUST GAVE AWAY THE ENGINE BEHIND 155 TRILLION TOKENS A DAY.
IT KEPT ONE PART: THE PART THAT REMEMBERS.

yesterday: 89.5M tokens for 54 cents, almost all cache hits.
now deepseek showed the kitchen. from their own writeup:
> one cluster: 155 trillion tokens a day
> $38 revenue per gpu per hour, at off peak prices
> kv cache built to live 72+ hours, so yesterday's context can still be a hit
> prefill on 8 gpu boxes, decode on 64 to 128 gpu supernodes
> runs on nvidia and huawei ascend

what's open:
> deepinfer, the engine
> deepseek recipe, the layer that turns api calls into tokens

what's not open yet:
> kv cache management and load balancing. "deeply coupled with the internal infrastructure"

so you can run the whale yourself.
you just can't copy its memory.

the engine is free. the memory is the moat.

one stack, real receipts. a new one every day.

[Arcvale @itsarcvale]

@skunzu does the audit worktree get wiped after each run, or do old diffs pile up?

[Arcvale @itsarcvale]

@aifindr exactly. it knew the answer, it just didnt know it wasnt supposed to yet

[Arcvale @itsarcvale]

@Klonzu the other 26 are the real result. the winner is just the one that got lucky on that data

[Arcvale @itsarcvale]

@AIMind_Ai this is why my bots run the grind on haiku and only wake opus for the final call. limits stop being a problem

[Arcvale @itsarcvale]

@Klonzu he sorting like pro

[Arcvale @itsarcvale]

@Richix01 earnings releases. checker holds every number against the 10-Q before the bot is allowed to trade it

[Arcvale @itsarcvale]

@andreysuperior with id 8.5 on bmw its work?

[Arcvale @itsarcvale]

@necrophosalpha @Klonzu absolutely

[Arcvale @itsarcvale]

@Klonzu most blowups arent bad strategy, theyre one bad input nobody checked

[Arcvale @itsarcvale]

A 23-YEAR-OLD CLAIMS 240 CLAUDE CODE AGENTS DID A $175K ANALYST'S JOB IN ONE AFTERNOON. THE F*CKING TRICK IS ONE LOOP

his post this week:
> 10 scouts, 9 extractors, 8 cross-checkers, 6 reconcilers, 3 judges
> 120 company filings
> every claim held against the exact source page
> anything that doesn't match goes back as "rewrite"

can't verify the kid. the loop is real and you can build it tonight

the rule underneath ↓ an agent never grades its own work

now point it at earnings season

your bot reads a press release, grabs "revenue up 12%" and trades it. one wrong number and you just bought a headline that never existed

the split:
→ the big model writes every claim with a page number
→ a cheap model opens that exact page and checks it
→ mismatch = rewrite, never a trade
→ only claims that survive both reach your signals

you only step in three times:
→ when a claim bounces twice
→ when the filing and the press release disagree
→ before anything goes near real money

the big model writes. the cheap one checks. nothing trades unverified

paste this into Claude Code ↓

"Build earnings_check.py:
1. Pull the latest 10-Q and earnings press release for [ticker] from SEC EDGAR.
2. Have Opus extract every number (revenue, EPS, margins, guidance) with the exact page it came from.
3. Have Haiku open that page and confirm each number. Mismatch → send back as rewrite.
4. Loop until every claim matches, max 3 passes. Log every bounce to verify.log.
5. Write only verified claims to signals.csv. Paper only, no orders.
Show me the code first. Don't run anything until I say go."

one of these every night. tomorrow's builds on this one

[Gyomei @Gyome1_]

REA hit #1 on GitHub trending, and most people still run it inside a single chat

Split it into 4 subagents and one command takes a program you own from sealed binary to a tested rebuild with a match rate

Set it up once, aim it at the next one: https://t.co/MxHPDVOZO2

[oracle @quietoraclex]

holy sh*t, split that five-department Opus studio into an Agent Team and ONE Opus session quietly becomes 5 Opus bills...

with teams on, any subagent Claude names becomes a teammate, asked or not. and every teammate runs on the lead's model unless you say otherwise

1 Opus lead + 4 teammates, each reading 1M tokens and writing 100K:
→ all Opus 5.5: ~$30
→ Sonnet 5.5 teammates: ~$18
→ Haiku 5.5 teammates: ~$6.60
(list prices, no cache)

one env var fixes the default:
CLAUDE_CODE_SUBAGENT_MODEL=sonnet claude --model opus

→ Opus leads and reviews
→ any teammate without a model of its own runs on Sonnet
→ name haiku in the spawn prompt for read-only grunt work
→ teammates copy the lead's effort too, so check /effort status

you picked Opus once. the team picked it 4 more times.

one Claude Code bill trap every night. tomorrow's is already in the queue.

[Arcvale @itsarcvale]

@maixbt and the checker only needs to read one page per claim. thats why it can be the cheap one

[Arcvale @itsarcvale]

@AIDuckLabs no coffee breaks, no slack threads, one intern that checks every number. dream office

[Arcvale @itsarcvale]

@bubosees yep

[Arcvale @itsarcvale]

@aifindr senior writes the memo, intern finds the typo. except this intern costs 34x less and never gets tired

[Arcvale @itsarcvale]

@bubosees paranoid is cheap. a fake headline fill isnt