Remove unconditional stateful-turn truncation
This commit is contained in:
@@ -219,11 +219,12 @@ Training use is tracked separately as `SFT_FULL`, `SFT_STEP_ONLY`,
|
||||
score.
|
||||
|
||||
For a complete trajectory, the quality call also sees its final model patch. A
|
||||
truncated prefix never receives that suffix-derived patch. Every non-success
|
||||
trajectory (`resolved=0` or `-1`) is capped before its first statically detected
|
||||
state-changing assistant turn. The effective boundary is the earlier of this
|
||||
local cap and the GLM boundary. This intentionally prefers a shorter, safer
|
||||
investigation trace.
|
||||
truncated prefix never receives that suffix-derived patch. State-changing turns
|
||||
are recorded as audit telemetry but are not automatic boundaries: writing code
|
||||
is not itself an error, and recovered experiments are useful process data. The
|
||||
effective boundary is the earliest unrecovered major or critical behavior found
|
||||
by the boundary stage. The isolated quality stage rejects a retained prefix that
|
||||
still contains an unrecovered severe problem.
|
||||
|
||||
Successful trajectories skip boundary selection and are scored as complete
|
||||
trajectories. The command remains simple:
|
||||
|
||||
Reference in New Issue
Block a user