Skip to main content

Release Candidate Readiness

QuarkLM has two release-candidate tracks. They must stay separate.

TrackMeaningCurrent posture
Research Prototype RCThe closed-world self-improvement system is reproducible, auditable, documented, and honest about what is and is not learned into weights.Near.
Language Model RCThe from-scratch transformer reliably answers from the admitted corpus without candidate crutches and passes neural promotion gates.Not ready.

The current promoted responder evidence is runs/self-improve-v0.42/. The latest unpromoted transformer screen is runs/transformer-answer-v0.115.0-hidden-projection-margin-candidate-step1-dim4-context80/. That screen passes 10/11 constraints but still fails branch_diversity_target.

Current Decision​

Pursue Research Prototype RC first. It is the honest near-term release candidate because the system already has corpus boundaries, deterministic verifier checks, candidate quarantine, recipes, constraint-first promotion, docs discipline, and rejected transformer evidence.

Do not call the transformer a Language Model RC until branch routing passes. v0.115 lowers average collapsed-token hidden advantage from about 0.0842 to 0.0736, but all 9/9 multi-target profiles still collapse to "n".

Required Commands​

Run these before tagging or announcing an RC:

PYTHONPATH=src python3 -m unittest discover -s tests
npm run sites:build
python3 -m json.tool sites/shared/current-state.json >/dev/null

The local site build validates both public surfaces. Read the Docs publishes the Docusaurus docs, and GitHub Pages publishes only the standalone marketing site.

Required Artifacts​

Research Prototype RC requires:

  • RC_SPEC.md
  • RC_GAP_AUDIT.md
  • RC_CHECKLIST.md
  • experiment_intent.json
  • corpus_hygiene.json
  • training_plan.json
  • candidate_quarantine.json
  • closed_world_verifier.json
  • training_recipe.json
  • constraint_first_promotion.json
  • README, STATUS, Docusaurus, and marketing current-state alignment
  • sites/DEPLOYMENT.md, .readthedocs.yaml, and the marketing Pages workflow reviewed for hosting drift

Language Model RC additionally requires:

  • passing branch_diversity_target
  • non-collapsed multi-target branch profiles
  • target-token coverage floors met
  • direct-answer evals accepted without hidden candidate selection
  • retention and unknown-policy checks passing for the neural learner

Forbidden Claims​

Do not claim:

  • QuarkLM is a production language model
  • retrieval success is neural weight learning
  • v0.115 solved branch routing
  • the transformer is promoted while branch_diversity_target fails
  • the project has proven "world's first" status

Allowed current claim:

QuarkLM is an experimental closed-world research prototype with a reproducible admitted-corpus learning loop, exact retrieval/responder evidence, and an unpromoted from-scratch transformer whose next blocker is branch routing.

Next Model Step​

When the version loop resumes, prefer the profile-balanced routing repair bundle from RC_GAP_AUDIT.md: target-balanced branch batches across failing profiles, hidden-projection margin, representation-separation pressure, coverage-preserving guards, and branch-diversity acceptance gates.