1
00:00:00,823 --> 00:00:01,941
An order fills.

2
00:00:01,941 --> 00:00:03,505
Its reply disappears.

3
00:00:03,505 --> 00:00:05,665
Should an AI agent buy again?

4
00:00:05,665 --> 00:00:08,869
Statebound makes that decision inspectable.

5
00:00:08,869 --> 00:00:12,668
This experiment starts with a twenty dollar budget.

6
00:00:12,668 --> 00:00:14,903
Every order here is simulated.

7
00:00:17,103 --> 00:00:21,707
We run a fresh search over the plan and its fault model.

8
00:00:21,707 --> 00:00:25,077
One missing reply leads to a blind retry.

9
00:00:25,077 --> 00:00:28,859
Thirty dollars leaves a twenty dollar mandate.

10
00:00:28,859 --> 00:00:31,983
The checker computes the path you see.

11
00:00:32,707 --> 00:00:34,350
Look at the two sides.

12
00:00:34,350 --> 00:00:40,398
The exchange has filled the order, but the agent only knows it might have filled.

13
00:00:40,398 --> 00:00:43,758
An old balance cannot erase that uncertainty.

14
00:00:43,758 --> 00:00:46,147
Fifteen dollars remains exposed.

15
00:00:48,347 --> 00:00:53,443
Now we load the repair written in our development Codex session.

16
00:00:53,443 --> 00:00:57,584
It queries the original attempt before buying again.

17
00:00:57,584 --> 00:01:02,919
We validate and recheck that candidate here, with the same mandate.

18
00:01:02,919 --> 00:01:05,467
Loading it is not a new AI call.

19
00:01:07,667 --> 00:01:11,666
When the lookup confirms the fill, the goal completes.

20
00:01:11,666 --> 00:01:15,887
Replay uses the same interpreter and a persistent ledger.

21
00:01:15,887 --> 00:01:20,627
The original uncertain attempt is never submitted a second time.

22
00:01:22,827 --> 00:01:25,438
But a lookup may stay inconclusive.

23
00:01:25,438 --> 00:01:29,766
Here, the run ends unresolved and the reservation remains.

24
00:01:29,766 --> 00:01:33,496
Statebound reports that limit honestly: containing

25
00:01:33,496 --> 00:01:37,227
risk and finishing the task are separate outcomes.

26
00:01:39,427 --> 00:01:42,092
Export the evidence, then verify it.

27
00:01:42,092 --> 00:01:46,681
A separate verifier recomputes the transitions and the bounded

28
00:01:46,681 --> 00:01:51,418
search, without asking a language model to judge its own answer.

29
00:01:51,418 --> 00:01:55,267
The plan, observations and results stay inspectable.

30
00:01:56,265 --> 00:01:59,555
Binance data also drives a separate experiment.

31
00:01:59,555 --> 00:02:05,085
The official command line tool supplies a recorded testnet price and lot sizes.

32
00:02:05,085 --> 00:02:08,725
Those inputs change the model and its debit amounts.

33
00:02:08,725 --> 00:02:11,595
We recheck the saved repair against them.

34
00:02:11,595 --> 00:02:14,115
Fees and execution remain simulated.

35
00:02:15,005 --> 00:02:22,388
Try the four cases yourself in the public Evidence Lab, or reproduce the checks from GitHub.

36
00:02:22,388 --> 00:02:27,605
Hosted Binance MCP is not connected, and no real trade is placed.

37
00:02:27,605 --> 00:02:30,655
Statebound: evidence before execution.