{
 "columns": [
  "Item",
  "Content",
  "Source"
 ],
 "widths": [
  24,
  110,
  60
 ],
 "rows": [
  [
   "Purpose",
   "Project workbook for reproducing the Leanstral 1.5 recipe at small scale: post-training Goedel-Prover-V2-8B (Qwen3-8B derivative) on 4x A10G, aiming to beat published ~8B miniF2F pass@32 SOTA, then applying the prover to find bugs in real code.",
   "/apps/development/rlhf/align/docs/SOTA_TARGETS.md"
  ],
  [
   "How to read",
   "WBS = what is done / running / next. Recipe_Stack = Leanstral stage vs our 8B analogue. Datasets = open data and tools per stage (API-verified, 133 entries). Data_Gaps = what does not exist publicly and how we build it. Evals_Gates = what decides promotion. Experiment_Plan = the runs. Journal = dated log. Sources = vault notes and repo paths.",
   "-"
  ],
  [
   "The plan",
   "docs/SOTA_PLAN.md v2 (after Codex review): ONE primary endpoint (class A miniF2F-test, exact McNemar; significance needs ~+9 net ~ 211/244); controlled contrasts K1 EI-vs-RL, K2 mutation-vs-unmutated, K3 feedback-vs-restarts, K4 averaging; eval reserved first (~180 GPU-h); total ~520-620 GPU-h; bug study is a separate paper.",
   "/apps/development/rlhf/align/docs/SOTA_PLAN.md"
  ],
  [
   "How to maintain",
   "Do not edit the xlsx. Edit docs/workbook/data/<Sheet>.json, then run: python docs/workbook/build_workbook.py. Log an event: build_workbook.py journal EVENT RESULT EVIDENCE NEXT. Change a status: build_workbook.py status WBS 4.1 Done.",
   "/apps/development/rlhf/align/docs/workbook/build_workbook.py"
  ],
  [
   "Legend: Confidence",
   "High = read directly from a primary source or measured here. Med = primary but self-reported / single run, or derived by arithmetic. Low = inferred, second-hand, or conflicting sources.",
   "-"
  ],
  [
   "Legend: Status",
   "Done / Running / Queued / Planned / Backlog / Deferred / Out of scope. Datasets: Verified = live API returned 200 on the date shown; Unverified = not checked; Missing = API 404.",
   "-"
  ],
  [
   "Legend: Source",
   "Vault note id (research/notes/<id>.md in /apps/development/rlhf), a repo path, or a commit hash in /apps/development/rlhf/align.",
   "-"
  ],
  [
   "Scope (decided)",
   "Post-training only on Goedel-Prover-V2-8B. Mid-training (6.5B Lean tokens) and Leanstral-scale (119B MoE) training are out of scope: no rented compute; mid-training alone would take ~24 days on this box.",
   "align_olmoe_leanstral_roadmap memory; interim-report-leanstral-base-midtraining-and-compute-provenance"
  ],
  [
   "Thesis (one line)",
   "On this box the lever is data, not RL: iterated verifier-filtered self-distillation + compiler-feedback correction + weight averaging, with CISPO RL as last-stage polish. RL on single proofs mostly sharpens pass@k; Leanstral's headline gains are unattributed to RL.",
   "/apps/development/rlhf/research/runs/leanstral-recipe-lineage-c14fca/comparisons.md"
  ],
  [
   "Premise corrections",
   "FP32 LM head and model averaging are NOT in the Leanstral report (zero hits); they come from MiniMax-M1 and Goedel-V2. Leanstral's CISPO clamp c, group size, LR, steps, teachers, compute, Lean version are all unreported.",
   "interim-report-leanstral-base-midtraining-and-compute-provenance"
  ],
  [
   "Measurement floor",
   "1 problem = 0.41 pp. The ~2.4 pp figure is binomial spread across problems, not seed noise. Exact two-sided McNemar needs +6 net with 0 losses, +9 with 4 losses. Decisions on miniF2F-valid only (driver refuses test-split selection).",
   "interim-report-data-vs-rl-lever-and-8b-4xa10g-reproducibility"
  ],
  [
   "Host rules honoured",
   "Shared box, no swap: preflight (>=45 GB MemAvailable, load < 96, disk queue < 64), one heavy job at a time (flock), <=2 dataloader workers/rank, nothing on / or /tmp, checkpoints <=100 steps bf16 keep <=2.",
   "/apps/CLAUDE.md"
  ]
 ]
}