Run it again yourself.

“Sell 178.4121 rHIMS. I pay 0.05% taker, keep it under 0.3% all-in, and I must be out before the 8th.”

36.87 bps all-in at your 5 bps fee: over your 30 bps ceiling.

The best bid alone would cost 17.48 bps, well inside the ceiling; the full 178.4121 shares walk 5 levels down the book. The largest size that fits is 101.834 shares, leaving 76.5781 unpriced.

Recorded Sunday rHIMS book, 2026-09-20 09:02:35Z. Pre-fee 31.89 bps against the 28.025 mid; proceeds 4984.05 USDT at a VWAP of 27.935629.

pnpm replay fixtures/rhims-20260920T090235Z.json sell 178.4121 30 5
# receipt 175394a68b72557912bce5121ee17950557d4442aaf1541dec0902b9f701b4b4 · sounding-engine/0.7.0

• ✓ Stands · Two real captures 21 seconds apart, Oct 3

  gap 20.8 s · cost drift 0.00 bps · price drift 0.00 bps

  • Fresh book within 10 bps of the sounding you read on cost and price; still within ceiling at 37.40 bps.
• ✕ Void · The Sep 20 decision re-checked on the Oct 3 book

  gap 12.7 days · cost drift 0.53 bps · price drift 363.41 bps

  • The price you would get moved in your favour by 363.41 bps since you read it, beyond the 10 bps tolerance; a different market from the one you decided on.
  • The decision was read 12.7 days before this confirm, beyond the 2-minute limit; re-sound before acting.

Messages in English, 中文 and mixed, with gold values, written by agents told not to read the code. A field counts only if the right value reached the engine. “Wrong” is the dangerous outcome: a different value used without asking.

Prompts, gold values, every value each reader extracted, and the scoring script are in the repository under evidence/ and scripts/paraphrase-eval.ts.

These runs, and the reading runs above, were made before 2026-10-07 on engine 0.4.0, which priced an unknown fee at 20 bps and added the fee on top. They are shown as they ran; a task’s right answer was the engine’s verdict at that time.

Thirty complete tasks written blind: starting controls, one to three chat turns in English, 中文 or both, and the right outcome. Some must be answered, some must be asked back, some name a stock with no book. Every turn runs through the real route. Critical means the wrong order was priced, a route broke the trader’s ceiling or deadline, or it acted where it had to ask.

Each set was written by a separate blind agent and run once. The first (at 62fe672) scored only final answers; its three critical errors were fixed afterwards. The second (at 8c64202) scores every turn as rendered, so asking while showing a card counts as critical; its one critical, a negative size quoted without its minus sign, has since been fixed. Every failure is published with its run in evidence/; re-runs after fixes are kept there as development numbers.

On 2026-09-20 at 09:25:59 UTC, the top five bids and asks for rHOOD on Bitget’s web order book were compared with the public API capture taken the same second: 10 of 10 levels matched in price and size. One instrument, one instant. Bitget’s separately documented Reality depth feed was not compared.

• Whether an order would be accepted or filled. Nothing is sent, and no order type is tested live.
• Hidden or off-book liquidity, and how market makers behave at the session switch.
• The future book. Every answer is conditional on the snapshot it names.
• Bitget's whitelisted Reality depth feed, which may differ from the public book.
• How current a recorded universe is: recorded pages say the capture date.
• Your exact fee tier. A fee you state decides; without one, Bitget's published 5 bps rToken rate and its 10 bps list rate are both priced, and the higher decides.