Say the order.
Get one answer.
When Nasdaq is shut, Bitget’s own rToken book is the market. Sounding walks it for your exact size, then holds it to your fee, your ceiling and your deadline. One decision, or a named refusal with the largest size that fits.
Qwen reads your words and must quote them for every limit; code then finds those words and re-reads each number and date before the engine uses it.
The slippage you accepted is the slippage you confirm.
Last Look re-walks a fresh book when you confirm a decision that is within your ceiling. On a recorded book it needs a later capture; two are saved.
Take Bitget away and it stops.
The engine re-runs the lead order with each input withheld. Every removal refuses it or gets it wrong.
Measured, not claimed.
whole tasks done right on a blind set run once, every turn checked as shown. Without the model: 12, with 3 critical errors to its 1.
Prooflimits read correctly from 40 blind messages in English, 中文 and mixed. A regex reader got 19.
Proofnames where Bitget's best ask is inside a 50 bps ceiling but a 25,000 USDT buy is not, at the published 5 bps fee, on one recorded Sunday.
Research