Skip to main content

Research Implementation Map

Last reviewed: 2026-06-14.

The full map lives in the repository root at RESEARCH_IMPLEMENTATION_MAP.md.

v0.74 is a research-control checkpoint. It does not claim better model behavior. It records the source-backed implementation map that should guide the next mechanics before another larger transformer repair run.

Why It Exists​

The project already had a forward plan and a deep research review. The missing piece was a direct implementation ledger:

  • research cluster;
  • public implementation pattern;
  • QuarkLM gap;
  • required mechanic;
  • acceptance evidence.

That ledger prevents QuarkLM from drifting into knob turning. Each new version should now connect to a stated research and implementation reason.

Source Clusters​

The v0.74 map cross-references:

  • transformer language modeling from the original Transformer paper, GPT-style decoder practice, nanoGPT, llm.c, GPT-NeoX, and OLMo;
  • continual-learning and catastrophic-forgetting work, including EWC and lifelong-learning surveys;
  • small-data language-learning work such as BabyLM and TinyStories;
  • self-generated data methods such as Self-Instruct, STaR, Self-Refine, and Reflexion;
  • verifier and process-supervision work such as GSM8K verifiers and process reward models;
  • data curation and contamination work from The Pile, Dolma, DataComp-LM, and Open-Instruct;
  • tokenizer work from BPE, SentencePiece, and byte-level subword systems;
  • transparent open-model practice from Pythia, OLMo, OLMo 2, and LLM360.

These are design references only. QuarkLM still forbids pretrained weights, pretrained tokenizers, external embeddings, external datasets, copied code, and external-model-shaped training data.

Implementation Decision​

QuarkLM should continue the self-improvement operating system before another direct-answer objective mode:

  1. v0.75: candidate quarantine artifacts and lifecycle states. Implemented.
  2. v0.76: deterministic closed-world verifier checks. Implemented.
  3. v0.77: recipe objects and constraint-first promotion gates. Implemented.
  4. v0.78: transformer responsibility surfaces for experiments, artifacts, trainer utilities, and objective catalog. Implemented.
  5. v0.79: transformer model/config and checkpoint metadata surfaces. Implemented.
  6. v0.80: transformer eval/checkpoint-load surfaces. Implemented.
  7. v0.81: profile target-share anti-collapse objective. Implemented.
  8. v0.82: profile target-share full-stack screen. Implemented and rejected.
  9. v0.83: prompt-specific branch ownership. Implemented and rejected.
  10. v0.84: baseline replay anchors. Implemented and rejected.
  11. v0.85: baseline-floor update gating. Implemented and rejected.
  12. v0.86: adaptive baseline-floor retries. Implemented and rejected.
  13. v0.87: baseline-floor repair retries. Implemented and rejected.
  14. v0.88: objective-side baseline-floor anchors. Implemented and rejected.
  15. v0.89: stabilization-only baseline-floor updates. Implemented and rejected.
  16. v0.90: baseline-floor rejection diagnostics. Implemented.
  17. v0.91: profile-targeted floor stabilization. Implemented and rejected.
  18. v0.92: sequential source-profile floor stabilization. Implemented and rejected.
  19. v0.93: calibrated sequential profile-floor stabilization. Implemented with one accepted guarded update; rejected for promotion.
  20. v0.94: profile-scale calibrated floor stabilization. Implemented with eight accepted guarded source-profile updates; rejected for promotion.
  21. v0.95: diversity-aware profile-scale acceptance. Implemented with five accepted score-improving source-profile updates; rejected for promotion.
  22. v0.96: frontier-driven branch diversity under the full baseline target-token floor. Implemented with nine score-improving source-profile updates; rejected for promotion.
  23. v0.97: coverage-frontier audited missing-target repair. Implemented with one coverage-gaining update; rejected for promotion.
  24. v0.98: coverage-prep frontier repair. Implemented with six safe preparation moves; rejected for promotion.
  25. v0.99: coverage-recovery frontier retry. Implemented with two recovery conversions; rejected for promotion.
  26. v0.100.0: branch-stable coverage recovery. Implemented with two branch-stable recovery conversions; rejected for promotion.
  27. v0.101.0: branch-diversity recovery. Implemented with five local branch-score refinements; rejected for promotion.
  28. v0.102.0: collapsed-profile binding. Implemented with one targeted binding update and three remaining collapsed eval profiles; rejected for promotion.
  29. v0.103.0: remaining-profile binding. Implemented with six prioritized remaining-profile acceptances and a learning coverage gain; rejected for promotion.
  30. v0.104.0: owner/paraphrase residual binding. Implemented with six prioritized acceptances and protected-learning rejection evidence; rejected for promotion.
  31. v0.105.0: closed-world retrieval memory. Implemented with a corpus-only retrieval_memory_report.json artifact, 497 memory cards, and 219/219 exact retrieval evals without external embeddings or weight updates.
  32. v0.106.0: memory-guided consolidation planning. Implemented with memory_consolidation_plan.json, 9 memory-backed neural failed profiles, and top priorities owner, paraphrases, glossary, admission_paraphrases, and admissions.
  33. v0.107.0: gated memory-consolidation training. Implemented with a declared source consolidation plan, consumed targets owner, paraphrases, and glossary, and 26 prioritized attempts with 8 acceptances and 18 rejections; rejected for promotion on branch_diversity_target.
  34. v0.108.0: expanded memory-consolidation target window. Implemented with target-profile-to-source-label mapping and a five-profile source-plan screen for owner, paraphrases, heldout, qa, and glossary; rejected for promotion on branch_diversity_target.
  35. v0.109.0: missing first-token memory-consolidation pressure. Implemented with plan-derived missing first-token maps, 8 candidates, 22 attempts, 1 accepted guarded coverage-gain update, 21 rejections, 7 fallback acceptances, and exact 219/219 retrieval; rejected for promotion on branch_diversity_target.
  36. v0.110.0: remaining-collapsed missing first-token targeting. Implemented with an explicit source-plan collapsed-profile requirement, consumed targets owner, paraphrases, and learning, 6 candidates, 16 attempts, 1 accepted guarded coverage-gain update, 15 rejections, 5 fallback acceptances, and exact 219/219 retrieval; rejected for promotion on branch_diversity_target.
  37. v0.111.0: profile-specific remaining-collapsed missing first-token pressure. Implemented with source-label-to-target-profile maps, 6 candidates, 18 attempts, 0 direct missing-token acceptances, 18 rejections, 6 fallbacks, 1 accepted profile-specific update shape, and exact 219/219 retrieval; rejected for promotion on branch_diversity_target.
  38. v0.112.0: branch-diversity root-cause diagnostics. Implemented with BRANCH_DIVERSITY_RESEARCH.md, Learn docs, root-cause taxonomy under branch_diversity_target.root_cause, 24 profile-specific missing-token attempts, 0 direct missing-token acceptances, 8 fallbacks, exact 219/219 retrieval, and a critical target_routing_gap diagnosis; rejected for promotion on branch_diversity_target.
  39. v0.113.0: branch routing audit diagnostics. Implemented with branch_routing_audit under direct-answer snapshots, output-bias escape risk, prompt-to-branch representation separation, and target-imbalance summaries; rejected for promotion on branch_diversity_target.
  40. v0.114.0: logit-prior and centroid-separation instrumentation. Implemented with branch_logit_prior_profiles, centroid-distance and centroid-margin summaries, and branch_routing_audit.logit_prior; rejected for promotion on branch_diversity_target.
  41. v0.115.0: hidden-projection margin candidate. Implemented with branch-hidden-projection-margin-unlikelihood, parser/dispatch support, focused tests, and a one-step output-bias-frozen screen; rejected for promotion on branch_diversity_target.

Current Gap​

QuarkLM already has:

  • v0.71 experiment intent;
  • v0.72 replay planning;
  • v0.73 corpus hygiene and training plans;
  • v0.75 candidate quarantine artifacts and lifecycle states;
  • v0.76 deterministic closed-world verifier checks;
  • v0.77 recipe objects and constraint-first promotion gates.
  • v0.78 transformer experiment/artifact surfaces, trainer utilities, and direct-answer objective catalog.
  • v0.79 transformer model/config and checkpoint metadata surfaces.
  • v0.80 transformer eval/checkpoint-load surfaces.
  • v0.81 profile target-share objective mode: branch-balanced-context-profile-target-share-preserving-deficit-unlikelihood.
  • v0.82 full target-share screen evidence: runs/transformer-answer-v0.82-fullstack-profile-target-share-smoke-dim4-context80/.
  • v0.83 prompt-specific ownership mode: branch-balanced-context-profile-prompt-ownership-target-share-preserving-deficit-unlikelihood.
  • v0.83 full prompt-ownership screen evidence: runs/transformer-answer-v0.83-fullstack-prompt-ownership-smoke-dim4-context80/.
  • v0.84 baseline-anchored prompt-ownership mode: branch-balanced-context-profile-baseline-anchored-prompt-ownership-target-share-preserving-deficit-unlikelihood.
  • v0.84 full baseline-anchor screen evidence: runs/transformer-answer-v0.84-fullstack-baseline-anchored-prompt-ownership-smoke-dim4-context80/.
  • v0.85 baseline-floor update-gated prompt-ownership mode: branch-balanced-context-profile-baseline-floor-gated-prompt-ownership-target-share-preserving-deficit-unlikelihood.
  • v0.85 full baseline-floor update-gate screen evidence: runs/transformer-answer-v0.85-fullstack-baseline-floor-gated-prompt-ownership-smoke-dim4-context80/.
  • v0.86 adaptive baseline-floor prompt-ownership mode: branch-balanced-context-profile-baseline-floor-adaptive-prompt-ownership-target-share-preserving-deficit-unlikelihood.
  • v0.86 full adaptive baseline-floor screen evidence: runs/transformer-answer-v0.86-fullstack-baseline-floor-adaptive-prompt-ownership-smoke-dim4-context80/.
  • v0.87 baseline-floor repaired prompt-ownership mode: branch-balanced-context-profile-baseline-floor-repaired-prompt-ownership-target-share-preserving-deficit-unlikelihood.
  • v0.87 clean full repaired baseline-floor screen evidence: runs/transformer-answer-v0.87-fullstack-baseline-floor-repaired-prompt-ownership-clean-smoke-dim4-context80/.
  • v0.88 baseline-floor objective prompt-ownership mode: branch-balanced-context-profile-baseline-floor-objective-prompt-ownership-target-share-preserving-deficit-unlikelihood.
  • v0.88 full baseline-floor objective screen evidence: runs/transformer-answer-v0.88-fullstack-baseline-floor-objective-prompt-ownership-smoke-dim4-context80/.
  • v0.89 baseline-floor stabilization mode: branch-context-profile-baseline-floor-stabilization-unlikelihood.
  • v0.89 full baseline-floor stabilization screen evidence: runs/transformer-answer-v0.89-fullstack-baseline-floor-stabilization-smoke-dim4-context80/.
  • v0.90 baseline-floor rejection diagnostic evidence: runs/transformer-answer-v0.90-fullstack-baseline-floor-stabilization-diagnostics-smoke-dim4-context80/.
  • v0.91 profile-targeted baseline-floor stabilization evidence: runs/transformer-answer-v0.91-fullstack-baseline-floor-profile-targeted-stabilization-smoke-dim4-context80/.
  • v0.92 sequential source-profile baseline-floor stabilization evidence: runs/transformer-answer-v0.92-fullstack-baseline-floor-sequential-profile-stabilization-smoke-dim4-context80/.
  • v0.93 calibrated sequential baseline-floor stabilization evidence: runs/transformer-answer-v0.93-baseline-floor-calibrated-sequential-profile-stabilization-step1-dim4-context80/.
  • v0.94 profile-scale calibrated baseline-floor stabilization evidence: runs/transformer-answer-v0.94-baseline-floor-profile-scale-calibrated-sequential-stabilization-step1-dim4-context80/.
  • v0.95 diversity-aware profile-scale baseline-floor stabilization evidence: runs/transformer-answer-v0.95-baseline-floor-diversity-profile-scale-calibrated-sequential-stabilization-configured-step1-dim4-context80/.
  • v0.96 frontier profile-scale baseline-floor stabilization evidence: runs/transformer-answer-v0.96-baseline-floor-diversity-frontier-profile-scale-calibrated-sequential-stabilization-step1-dim4-context80/.
  • v0.97 coverage-frontier profile-scale baseline-floor stabilization evidence: runs/transformer-answer-v0.97-baseline-floor-diversity-coverage-frontier-profile-scale-calibrated-sequential-stabilization-step1-dim4-context80/.
  • v0.98 coverage-prep frontier profile-scale baseline-floor stabilization evidence: runs/transformer-answer-v0.98-baseline-floor-diversity-coverage-prep-frontier-profile-scale-calibrated-sequential-stabilization-step1-dim4-context80/.
  • v0.99 coverage-recovery frontier profile-scale baseline-floor stabilization evidence: runs/transformer-answer-v0.99-baseline-floor-diversity-coverage-recovery-frontier-profile-scale-calibrated-sequential-stabilization-step1-dim4-context80/.
  • v0.100.0 branch-stable coverage-recovery frontier profile-scale baseline-floor stabilization evidence: runs/transformer-answer-v0.100.0-baseline-floor-diversity-branch-stable-coverage-recovery-frontier-profile-scale-calibrated-sequential-stabilization-step1-dim4-context80/.
  • v0.101.0 branch-diversity recovery frontier profile-scale baseline-floor stabilization evidence: runs/transformer-answer-v0.101.0-baseline-floor-diversity-branch-diversity-recovery-frontier-profile-scale-calibrated-sequential-stabilization-step1-dim4-context80/.
  • v0.102.0 collapsed-profile binding frontier profile-scale baseline-floor stabilization evidence: runs/transformer-answer-v0.102.0-baseline-floor-diversity-collapsed-profile-binding-frontier-profile-scale-calibrated-sequential-stabilization-step1-dim4-context80/.
  • v0.103.0 remaining-profile binding frontier profile-scale baseline-floor stabilization evidence: runs/transformer-answer-v0.103.0-baseline-floor-diversity-remaining-profile-binding-frontier-profile-scale-calibrated-sequential-stabilization-step1-dim4-context80/.
  • v0.104.0 owner/paraphrase binding frontier profile-scale baseline-floor stabilization evidence: runs/transformer-answer-v0.104.0-baseline-floor-diversity-owner-paraphrase-binding-frontier-profile-scale-calibrated-sequential-stabilization-step1-dim4-context80/.
  • v0.105.0 closed-world retrieval memory evidence: runs/transformer-answer-v0.105.0-retrieval-memory-owner-paraphrase-frontier-profile-scale-step1-dim4-context80/.
  • v0.106.0 memory-guided consolidation planning evidence: runs/transformer-answer-v0.106.0-memory-guided-consolidation-owner-paraphrase-frontier-profile-scale-step1-dim4-context80/.
  • v0.107.0 gated memory-consolidation training evidence: runs/transformer-answer-v0.107.0-gated-memory-consolidation-owner-paraphrase-glossary-frontier-profile-scale-step1-dim4-context80/.
  • v0.108.0 expanded memory-consolidation target-window evidence: runs/transformer-answer-v0.108.0-expanded-memory-consolidation-owner-paraphrase-heldout-qa-glossary-frontier-profile-scale-step1-dim4-context80/.
  • v0.109.0 missing first-token memory-consolidation evidence: runs/transformer-answer-v0.109.0-missing-first-token-memory-consolidation-owner-paraphrase-heldout-qa-glossary-frontier-profile-scale-step1-dim4-context80/.
  • v0.110.0 remaining-collapsed missing-token evidence: runs/transformer-answer-v0.110.0-remaining-collapsed-missing-first-token-memory-consolidation-owner-paraphrase-learning-frontier-profile-scale-step1-dim4-context80/.
  • v0.111.0 profile-specific remaining-collapsed missing-token evidence: runs/transformer-answer-v0.111.0-profile-specific-missing-first-token-memory-consolidation-owner-paraphrase-learning-frontier-profile-scale-step1-dim4-context80/.
  • v0.112.0 branch-diversity root-cause evidence: runs/transformer-answer-v0.112.0-branch-diversity-root-cause-profile-specific-memory-consolidation-step1-dim4-context80/.
  • v0.113.0 branch-routing audit evidence: runs/transformer-answer-v0.113.0-branch-routing-audit-profile-specific-memory-consolidation-step1-dim4-context80/.
  • v0.114.0 logit-prior instrumentation evidence: runs/transformer-answer-v0.114.0-logit-prior-representation-instrumentation-profile-specific-memory-consolidation-step1-dim4-context80/.
  • v0.115.0 hidden-projection margin candidate evidence: runs/transformer-answer-v0.115.0-hidden-projection-margin-candidate-step1-dim4-context80/.

It still needs:

  • a broader guarded routing repair that uses v0.115.0 hidden-projection evidence to improve zero-coverage profiles and buried-target profiles without regressing retrieval provenance or relaxing promotion gates.

Operating Rule​

Every future mechanics version should answer three questions:

  1. Which research or implementation pattern justifies this mechanic?
  2. Which closed-world boundary does it protect?
  3. Which artifact proves it worked or rejected the run?

That is how QuarkLM keeps the claim clean: the model grows from its admitted dataset, and the project keeps enough evidence to prove what changed.