Trajectory Hygiene v0¶
LoopX keeps control-plane history for audit and continuation. That history is not automatically suitable for model training: quota accounting, scheduler acknowledgements, unchanged monitor polls, and repeated state projections can outnumber task-facing actions in a long session.
trajectory_hygiene_summary_v0 is a read-only audit over the existing compact
run index:
It reports controller-event density, compact controller-character density, non-material event density, repeated task-action labels, and action/outcome attribution coverage. These are proxies for deciding whether the current event mix needs a separate learning projection. They are not training-quality scores.
Boundary¶
The audit:
- reads only the public-safe compact run index;
- does not open run artifacts, raw sessions, raw trajectories, task bodies, or verifier output;
- does not change history, status, quota, todo, scheduler, or state semantics;
- always reports
seed_model_training_eligible=false.
A future learning projection should keep model-visible task context, assistant actions, tool observations, human decisions, and attributed outcomes. Controller events should remain in the audit ledger and link to learning turns by stable ids instead of being replayed as ordinary user/model turns.
The raw audit trail remains useful for reproduction and incident analysis. It must not be silently rewritten into a training sample.