Tagged “drift”
-
Which definition produced a reply
A transcript records the output and none of the inputs. Without a per-turn state fingerprint, every drift report is unattributable and irreproducible.
-
What a model change does to a character
The definition is not a specification, it is text interpreted by one model. Change the model and the same words produce a different character, unevenly.
-
Detecting drift without reading transcripts
Production has no ground truth, so drift must be measured against each conversation's own early behaviour rather than against a corpus or a golden sample.
-
Why characters drift
A fixed definition against growing history, and the history includes the model's own output. Drift is a proportion problem that reinforces itself.
-
Testing a character before it ships
A change that passes a fresh-session check fails in production: consistency is a distribution over samples at depth, not a property you can spot-check.
-
Constraints erode before the voice does
The voice survives a long conversation better than the rules do. Constraints are the part of a definition that decays first, and the failure is invisible.
-
Changing a character under a live conversation
A definition edit is a migration. Existing conversations get the new text with the old history, so the seam lands mid-relationship.