Admission Workflow

QuarkLM learns new facts through corpus/admissions.jsonl. A fact is not training data just because someone typed it in chat. It becomes learnable only after admission, curriculum regeneration, and a measured weight update.

Admit one fact:

PYTHONPATH=src python3 -m closed_world_lm.admit \
  --id learned-child-book \
  --person child \
  --object book \
  --color blue \
  --relation on \
  --container table

Admit a batch:

PYTHONPATH=src python3 -m closed_world_lm.admit \
  --batch path/to/new_admissions.jsonl

Duplicate ids are rejected before writing. When using default project paths, direct and paraphrase admission probes are regenerated automatically.

After Admission

Run a self-improvement cycle that compares against the last promoted report:

PYTHONPATH=src python3 -m closed_world_lm.self_improve answer-cycle \
  --run runs/self-improve-next \
  --compare-report runs/self-improve-v0.38/self_improvement_report.json

After Admission​

After Admission