Skip to main content

Candidate Quarantine

v0.75 adds src/closed_world_lm/candidate_quarantine.py and a required candidate_quarantine.json artifact for self-improvement answer cycles and transformer answer-training runs.

The artifact is empty today because QuarkLM has not yet promoted autonomous candidate generation. That emptiness is still useful: the training plan can now prove that the candidate lane exists and that no candidate records were silently used as training data.

Candidate States​

Candidate records support these lifecycle states:

  • proposed
  • quarantined
  • needs_human_review
  • verified
  • rejected
  • admitted
  • trained
  • promoted

State transitions are validated. Terminal states such as rejected and promoted cannot move again.

Training Rule​

Candidate records are not training data.

A generated lesson, probe, repair proposal, diagnosis note, or memory proposal can train weights only after it is admitted into the ledgered corpus and converted into curriculum lessons. The candidate record may preserve its history, but the candidate store itself is not a training source.

Run Artifacts​

Self-improvement attempts now write:

  • attempts/attempt-###/candidate_quarantine.json
  • candidate_quarantine.json at the latest run level
  • training_plan.json with the candidate quarantine path and summary
  • self_improvement_report.json with the manifest embedded
  • closed_world_verifier.json with deterministic quarantine and plan checks

Transformer answer-training runs write:

  • candidate_quarantine.json
  • training_plan.json with the candidate quarantine path and summary
  • transformer_answer_metrics.json with the manifest embedded
  • closed_world_verifier.json with deterministic quarantine and plan checks

v0.76 adds deterministic closed-world verifier checks. Candidate quarantine can now feed a verifier report without using an external model. Non-admitted candidates remain out of training; training-eligible candidates must link to a ledger admission id.