Skip to content
live · cycle 977 · 3h ago

Two paper-trading systems. Different questions.

Crypto arena. Can language models beat buy-and-hold bitcoin? Three models and four fixed rules each start with $10,000 and submit one action an hour on the same market data.

Equity committee. Can a locked value rulebook run a US stock book on its own? It starts at $100,000, screens overnight, and buys once a month. Language models only advise.

Neither trades real money. Both publish the checks they would have to pass first.

Right now

both systems, at a glance

arena leader
+$2,981
+29.81% · buy and hold BTC + ETH · from $10,000
equity bookcash plus market value of holdings
$114,076
+$14,138 · +14.15% · +11.10% vs SPY
verdictsarena NO-GOcommittee NO-GO
last cycleone decision round, hourly
3h ago
scheduler healthy
A fixed rule is ahead of every language model right now: buy and hold BTC + ETH at +$2,981, against gpt-4o at +$668. That is the result the checks are waiting on.
One or more language-model API calls failed recently and were recorded as HOLD, so those cycles are not real trading decisions. Check the Anthropic billing balance if Claude agents stay flat.

Crypto arena

NO-GO
stake
$10,000 each
traders
3 language models, 4 fixed rules
models follow
no set strategy
decides
hourly
beat
buy-and-hold bitcoin
since
26 Jul 2026

Standings

cumulative return by cycle, net of costs

  • buy and hold BTC + ETH
  • buy and hold bitcoin
  • claude haiku
  • claude sonnet
  • moving-average crossover
  • gpt-4o
  • random
leader
+29.81%
+$2,981 · buy and hold BTC + ETH
cyclesone decision round, hourly
977
inference cost
$20.56
$0.021 per cycle
checks passed
1/3 passing
all must pass before real money

A recent decision

one action that changed the book, or the latest if none has

gpt-4oREDUCE SOL · $266cycle 969 · 11h ago

SOL's recent price stabilization and proximity to break-even suggests reducing the short position to mitigate risk, while the market remains indecisive.

Equity committee

NO-GO
stake
$100,000 book
traders
fixed rules, language models advise
screens
nightly, 2 of 6 stages
proposes
eve of the rebalance, all 6
buys
monthly, first trading day
since
18 Jul 2026

Book value

repriced against the latest closes each night

  • book
  • buy-and-hold SPY
book valuecash plus market value of holdings
$114,076
+$14,138 · +14.15% · +11.10% vs SPY
holdings
9
largest NEM at 16.1%
days tracked
38
repriced nightly
checks passed
1/4 passing
all must pass before real money

Verdict

what NO-GO means

Each system is graded against a fixed list of conditions, published on its own page: enough history to judge, a real margin over the benchmark, losses inside the limit, and a scheduler that ran when it was supposed to. A system that fails any one of them is marked NO-GO and keeps trading on paper. Nothing here upgrades itself automatically.

Paper trading only. No capital is at risk and nothing here is investment advice. Simulated fills include spread, slippage and fees. Inference cost is deducted from agent returns. alasiya.io