statebound EXECUTION RESEARCH / 001Explore the source ↗

+ UNCERTAINTY IS A STATE. NOT A RETRY.

The reply
is lost.
The risk isn't.

An order fills. Its reply disappears. A blind retry breaks the budget. Find the failure, repair the plan, and inspect every step.

Watch the experiment 02:31Explore the evidence ↗

Recorded simulator demo · synthetic orders · no live trading

FIG. 01 — ANATOMY OF A BLIND RETRYBNB / USDT
THE MANDATE20USDT

Maximum total debit

ACTUAL DEBIT30USDT
+50%OVER BUDGET
+ A missing reply doesn't undo a fill.+
THE STATEBOUND METHOD01 / Explore the failure ↘02 / Inspect the repair ↘03 / Check the bounds ↘

EXPLORE THE MECHANISM

One lost reply.
Four different endings.

Move through the recorded steps. See what the agent knows, what actually happened, and why the next purchase matters.

Loading precomputed simulator traces… The video and evidence downloads remain available below.

BINANCE / FROM READ TO REPLAY

Real inputs.
Inspectable consequences.

The official CLI supplies the price and symbol grids. Statebound turns that recorded snapshot into a separate synthetic experiment, then checks the repaired plan.

01 / OBSERVE

Read Binance.

BNBUSDT public Spot testnet data through the documented Skills Hub / official CLI route.

771.35 USDT / BNB

Snapshot:
binance-cli 2.1.1

Inspect original response ↗
02 / DERIVE

Respect the grid.

A 15 USDT principal target rounds down to a whole lot. The synthetic fee assumption is 10 bps; the total budget stays 20 USDT.

0.019 BNB

Lot step 0.001 BNB
Price tick 0.01 USDT

Inspect derived model ↗
03 / VERIFY

Reconcile before retry.

The recorded AI repair is rebound to this model, checked again and replayed with fresh or inconclusive lookup evidence.

14.67030565 USDT / REPAIRED

Blind retry: 29.3406113 USDT
Inconclusive lookup: still UNRESOLVED

Reproduce this workflow ↗
What is real, assumed and independently checked

Real: the saved public testnet response. Imported: price, tick, lot size and notional bounds. Assumed: fake account, fixed execution price, 10 bps fee and finite IOC fault model. No live AI call, hosted MCP connection or exchange write is claimed.

Unmapped venue filters: ICEBERG_PARTS, MARKET_LOT_SIZE, TRAILING_DELTA, PERCENT_PRICE_BY_SIDE, MAX_NUM_ORDERS, MAX_NUM_ORDER_LISTS, MAX_NUM_ALGO_ORDERS, MAX_NUM_ORDER_AMENDS. This is not a venue-valid order recommendation. The original 600 USDT flagship remains a separate fixture.

Both evidence bundles rerun bounded search and exact replay. Reproduce offline with npm run integration:verify.

Workflow manifest: 800b4938ad7d74fba48354b222305632b2185c577358a24771414cb6e4c08100

One failure. A complete repair loop.

Actual application recording

Original-speed MiMo Dean narration · ASR-reviewed speech · captions burned into video. Loading the saved AI candidate is not a fresh model call. Download captions · Voice verification

Inspect the evidence.

Recorded results, not a live checker

Safety and progress,
measured separately.

Full evaluation JSON ↗
48scenarios per variant
0repaired budget violations
36repaired goal completions
12honestly unresolved

Three mandates × 16 deterministic scenarios. Regression corpus, not an independent blinded benchmark. A bounded discrete result does not certify all market conditions.

What actually runs

The local workbench computes checks, accepts typed candidate repairs, persists replay events, and verifies exported evidence. This shareable preview serves recorded artifacts only. Download the repository to run a new experiment.

Official Binance CLI public testnet reads were recorded separately from the synthetic fixture. Hosted Binance MCP and testnet order writes are unavailable. Mainnet is disabled. No exchange or model credentials are needed for the offline workflow.

npm ci
npm run build
npm run dev
# Open http://127.0.0.1:4381

npm run evidence:verify -- submission/demo-evidence.json
Run the full workbench ↗