Operate

Operating QuarkLM means keeping the model, corpus, evals, reports, docs, and public surfaces synchronized.

Operating Surfaces

Surface	Purpose
runs/self-improve-*	Versioned training and audit evidence.
experiment_intent.json	Hypothesis, allowed data, planned artifacts, gates, and decision for a run.
corpus_hygiene.json	Source mixture, duplicates, train/eval overlap, candidate ratio, and rare-profile coverage.
training_plan.json	Allowed data sources, scheduled example mixture, replay summary, and planned artifacts.
candidate_quarantine.json	Candidate lifecycle state and proof that generated candidates are not training data until admitted.
closed_world_verifier.json	Deterministic pass/fail evidence for candidate checks and training-plan approval.
training_recipe.json	Reproducible model, data, objective, optimizer, artifact, and gate recipe.
constraint_first_promotion.json	Promotion gate that blocks quality metrics until closed-world constraints pass.
corpus_snapshot.json	Current ledger source hashes and record counts.
corpus_diff.json	Comparison to the previous promoted run.
RC_SPEC.md / RC_GAP_AUDIT.md / RC_CHECKLIST.md	Release-candidate track, gap, checklist, and forbidden-claim controls.
.readthedocs.yaml / sites/DEPLOYMENT.md	Docs-on-Read-the-Docs and marketing-on-GitHub-Pages hosting controls.
README / Docusaurus / marketing	Public state that must not drift.

Start with release candidate readiness, then read release discipline, experiment registry, corpus hygiene, candidate quarantine, closed-world verifier, training recipes, provenance, and docs drift.