Autonomous trading lab exposes two blind spots

This is a strong reminder that autonomous systems can fail by lying to you about the thing you most need to know: whether they worked, and whether they’re even running.

• –The retrospective eval was contaminated because the same triggers that killed agents also made their prior decisions look “correct,” turning validation into a loop
• –The fix is architectural: decisions and outcomes need independent writers, with no shared logic, thresholds, or code paths
• –The second bug shows why “I think it’s off” is not a state check; only direct measurement against the running machine counts
• –For solo builders, CI-level separation tests are doing the job that team review would normally catch
• –The bigger lesson is that autonomy increases the cost of hidden coupling, both in metrics and in runtime state