The delivery verdict

Which environments enter the training set

A per-batch report the vendor can act on and the lab can rely on, with the exploit named.

Rejected

browse-checkout-exploit

Episodes
70
Alignment
-0.16
Exploit gap
2.80x
Policy
reward-loop
  • reward correlates with completion at only -0.16
  • a non-completing policy scores 180% above the median completed episode
  • reward-loop beats the task without solving it
Rejected

sql-repair-exploit

Episodes
70
Alignment
-0.21
Exploit gap
2.67x
Policy
reward-loop
  • reward correlates with completion at only -0.21
  • a non-completing policy scores 167% above the median completed episode
  • reward-loop beats the task without solving it
Rejected

ticket-triage-exploit

Episodes
70
Alignment
-0.17
Exploit gap
2.83x
Policy
reward-loop
  • reward correlates with completion at only -0.17
  • a non-completing policy scores 183% above the median completed episode
  • reward-loop beats the task without solving it
Accepted9 environments verified