| { |
| "columns": [ |
| "Id", |
| "Work package", |
| "Deliverable", |
| "Owner path", |
| "Depends on", |
| "Est. box-hours", |
| "Start", |
| "Finish", |
| "Status", |
| "Source" |
| ], |
| "widths": [ |
| 7, |
| 34, |
| 52, |
| 40, |
| 12, |
| 12, |
| 11, |
| 11, |
| 12, |
| 40 |
| ], |
| "rows": [ |
| [ |
| "1", |
| "FOUNDATIONS", |
| "align post-training stack on torch/triton/numpy/safetensors only", |
| "align/", |
| "-", |
| "-", |
| "2026-09-23", |
| "2026-10-13", |
| "Done", |
| "a4d4d5e .. 3c71f8d" |
| ], |
| [ |
| "1.1", |
| "Consolidated stack (SFT/DPO/PPO/GRPO)", |
| "1,093 tests passing; dependency firewall", |
| "align/align/", |
| "-", |
| "-", |
| "2026-09-23", |
| "2026-10-12", |
| "Done", |
| "a4d4d5e, ac2bdcf" |
| ], |
| [ |
| "1.2", |
| "Strict Lean 4 verifier (P1)", |
| "Ceiling 177/178 good, 0/61 adversarial; bans exact?/apply?/native_decide; axiom audit", |
| "align/align/verify/lean/strict.py", |
| "1.1", |
| "-", |
| "2026-10-12", |
| "2026-10-12", |
| "Done", |
| "654c20c; docs/LEAN_VERIFIER.md" |
| ], |
| [ |
| "1.3", |
| "Qwen3 / Goedel-V2-8B in align", |
| "Logit parity 2.9e-5; LoRA SFT 2,152 tok/s @16k", |
| "align/align/model/", |
| "1.1", |
| "-", |
| "2026-10-13", |
| "2026-10-13", |
| "Done", |
| "34a525c" |
| ], |
| [ |
| "1.4", |
| "Generation + overlapped verification", |
| "vLLM TP=4 gen and Lean verify inside ONE heavy lock", |
| "scripts/p2/pipeline_gen_verify.py", |
| "1.3", |
| "-", |
| "2026-10-13", |
| "2026-10-13", |
| "Done", |
| "7b07b55" |
| ], |
| [ |
| "2", |
| "BASELINE (SOTA gate 1)", |
| "Reproduced Goedel-V2-8B on miniF2F", |
| "docs/P2.md", |
| "1", |
| "-", |
| "2026-10-13", |
| "2026-10-14", |
| "Done", |
| "45f6f99" |
| ], |
| [ |
| "2.1", |
| "pass@8 @16k", |
| "78.7% (192/244)", |
| "p2_runs/", |
| "1.4", |
| "~3", |
| "2026-10-13", |
| "2026-10-13", |
| "Done", |
| "780ce3d" |
| ], |
| [ |
| "2.2", |
| "pass@32 @40k", |
| "82.8% (202/244); 30k-equivalent identical", |
| "p2_runs/r1_baseline/", |
| "2.1", |
| "~12", |
| "2026-10-13", |
| "2026-10-14", |
| "Done", |
| "45f6f99" |
| ], |
| [ |
| "2.3", |
| "300 s re-check of timed-out samples", |
| "0 recovered -> baseline stays 202", |
| "align/.tmp/p2/r1/baseline/", |
| "2.2", |
| "~2", |
| "2026-10-14", |
| "2026-10-14", |
| "Done", |
| "1cc55e4" |
| ], |
| [ |
| "2.4", |
| "miniF2F validation split", |
| "226 problems, decontaminated vs test/MOBench/pool", |
| "scripts/p2/build_valid.py", |
| "-", |
| "<1", |
| "2026-10-14", |
| "2026-10-14", |
| "Done", |
| "1c483cc" |
| ], |
| [ |
| "2.5", |
| "Split-hygiene guard in EI driver", |
| "Driver refuses a test-split scored file unless --allow-test-selection (--rounds 1, no target); 50 tests pass", |
| "scripts/p2/expert_iteration.py", |
| "-", |
| "0", |
| "2026-10-14", |
| "2026-10-14", |
| "Done", |
| "Codex review item 1" |
| ], |
| [ |
| "3", |
| "VERIFIER HARDENING", |
| "One hardened verifier for both arms before final comparison", |
| "align/align/verify/lean/", |
| "1.2", |
| "-", |
| "2026-10-13", |
| "", |
| "Running", |
| "redteam/false-accepts worktree" |
| ], |
| [ |
| "3.1", |
| "Red-team false-accept hunt", |
| "0 confirmed false accepts; 3 latent weaknesses (e.g. axiom-walk fuel cap fails open)", |
| "../align-redteam (branch redteam/false-accepts)", |
| "1.2", |
| "-", |
| "2026-10-13", |
| "", |
| "Running", |
| "separate user session" |
| ], |
| [ |
| "3.2", |
| "Merge red-team fixes", |
| "Axiom walk fails CLOSED at fuel cap; other latent fixes", |
| "align/align/verify/lean/strict.py", |
| "3.1", |
| "-", |
| "", |
| "", |
| "Planned", |
| "-" |
| ], |
| [ |
| "3.3", |
| "Enable Lean REPL OS sandbox", |
| "sandbox.py wired into ReplPool (currently opt-in)", |
| "align/align/verify/lean/sandbox.py", |
| "3.2", |
| "-", |
| "", |
| "", |
| "Planned", |
| "3c71f8d; docs/LEAN_SANDBOX.md" |
| ], |
| [ |
| "3.4", |
| "Re-verify baseline under hardened verifier", |
| "Baseline count under the final verifier (may move from 202)", |
| "p2_runs/r1_baseline/", |
| "3.3", |
| "~2", |
| "", |
| "", |
| "Planned", |
| "1cc55e4" |
| ], |
| [ |
| "3.5", |
| "Class-B ledger + environment coverage", |
| "Trajectory ids, cumulative tokens, declared truncation; ProofNet/Putnam inputs built and elaboration coverage measured; ProofNet dev split", |
| "extras/vllm_rollout/lean_gen.py; scripts/p2/", |
| "3.3", |
| "~10", |
| "2026-10-15", |
| "2026-10-20", |
| "Planned", |
| "docs/SOTA_PLAN.md Phase 0 items 3, 5" |
| ], |
| [ |
| "3.6", |
| "Register SOTA analysis in eval_report", |
| "Separate registered config: class A miniF2F-test n=k=32 @40k, 300 s, exact McNemar vs anchor; secondaries labelled", |
| "scripts/p2/eval_report.py", |
| "3.4", |
| "0", |
| "", |
| "", |
| "Planned", |
| "Codex code review 2026-10-14 (P2)" |
| ], |
| [ |
| "4", |
| "DATA", |
| "Training pools for self-distillation and RL", |
| "align/.tmp/p2/", |
| "-", |
| "-", |
| "2026-10-14", |
| "", |
| "Running", |
| "-" |
| ], |
| [ |
| "4.1", |
| "Pool round 0 (2,000 statements x 8 @16k)", |
| "Pass-rate profile: ~11% band (1-6/8), ~42% never, ~47% easy", |
| "align/.tmp/p2/r1/pool/", |
| "2.2", |
| "~25", |
| "2026-10-14", |
| "", |
| "Running", |
| "chain4.log" |
| ], |
| [ |
| "4.2", |
| "Open dataset inventory", |
| "Datasets sheet: per-stage HF/GitHub entries, API-verified", |
| "docs/workbook/data/Datasets.json", |
| "-", |
| "0", |
| "2026-10-14", |
| "", |
| "Done", |
| "workflow wf_c3f2e9d0-871" |
| ], |
| [ |
| "4.3", |
| "Decontamination sweep (4 stages)", |
| "NL+Lean hashes, n-gram (10 NL / 5 Lean), binder-aware tree equivalence, proof skeleton; covers mutations, trajectories, proof bank; ancestry kept; heuristic error measured", |
| "scripts/p2/ (new decontam.py)", |
| "4.2", |
| "<1 GPU; CPU", |
| "2026-10-15", |
| "2026-10-20", |
| "Planned", |
| "docs/SOTA_PLAN.md Phase 0; source-analysis-pythagoras-prover (A.3)" |
| ], |
| [ |
| "4.4", |
| "Mid-training Lean corpus", |
| "6.5B-token continued pretraining", |
| "-", |
| "-", |
| "~575 (24 days)", |
| "", |
| "", |
| "Out of scope", |
| "interim-report-leanstral-base-midtraining-and-compute-provenance" |
| ], |
| [ |
| "5", |
| "SELF-DISTILLATION", |
| "Iterated verifier-filtered SFT rounds on band problems", |
| "scripts/p2/expert_iteration.py", |
| "4.1", |
| "-", |
| "", |
| "", |
| "Queued", |
| "docs/SELF_DISTILLATION.md" |
| ], |
| [ |
| "5.1", |
| "Base model on miniF2F-valid n=8 @16k", |
| "Dev-split reference for round selection", |
| "align/.tmp/p2/r1/chain5.sh", |
| "4.1", |
| "~3", |
| "", |
| "", |
| "Queued", |
| "chain5.sh" |
| ], |
| [ |
| "5.2", |
| "Repair / correction round 1", |
| "354 failed attempts re-tried with compiler feedback", |
| "align/.tmp/p2/r1/select_corrections.py", |
| "5.1", |
| "~6", |
| "", |
| "", |
| "Queued", |
| "chain5.sh" |
| ], |
| [ |
| "5.3", |
| "Band-targeted rejection-sampling LoRA SFT", |
| "Round-k adapters; promotion on valid only", |
| "scripts/p2/expert_iteration.py", |
| "5.2", |
| "~55", |
| "", |
| "", |
| "Planned", |
| "interim-report-data-vs-rl-lever-and-8b-4xa10g-reproducibility" |
| ], |
| [ |
| "5.4", |
| "Weight-averaging sweep", |
| "alpha in 0.6-0.9 between rounds, chosen on valid", |
| "scripts/p2/weight_average.py", |
| "5.3", |
| "~6", |
| "", |
| "", |
| "Planned", |
| "interim-report-rl-adds-vs-sharpens-on-lean-passk" |
| ], |
| [ |
| "5.5", |
| "L1 / K3: correction vs restarts", |
| "Class-B harness with trajectory ledger; feedback vs restart sampling at matched tokens on valid", |
| "extras/vllm_rollout/lean_gen.py", |
| "3.4", |
| "~15", |
| "2026-10-20", |
| "2026-10-26", |
| "Planned", |
| "docs/SOTA_PLAN.md K3" |
| ], |
| [ |
| "5.6", |
| "K2 pilot: mutation vs unmutated data", |
| "2,000 verified mutations vs 2,000 extra unmutated proofs, same rows/tokens; parents screened for misformalisation", |
| "scripts/p2/ (new mutate_statements.py)", |
| "4.3, 5.4", |
| "~30", |
| "2026-10-24", |
| "2026-11-02", |
| "Planned", |
| "docs/SOTA_PLAN.md K2" |
| ], |
| [ |
| "5.7", |
| "Scale K2 winner (10-15k)", |
| "Not 50k: 50k x2 x5k tok ~150-250 h generation alone", |
| "scripts/p2/expert_iteration.py", |
| "5.6", |
| "~60-90", |
| "2026-11-02", |
| "2026-11-23", |
| "Planned", |
| "docs/SOTA_PLAN.md Phase 2" |
| ], |
| [ |
| "6", |
| "RL POLISH", |
| "Small on-policy LoRA CISPO run", |
| "align/align/rl/cispo.py", |
| "5", |
| "-", |
| "", |
| "", |
| "Planned", |
| "interim-report-objective-choice-cispo-vs-grpo-effect-size-controlled" |
| ], |
| [ |
| "6.1", |
| "CISPO stability guard", |
| "Negative-advantage low-ratio sequence mask (ARLArena fix); c~5; 1-2 updates/rollout", |
| "align/align/rl/cispo.py", |
| "-", |
| "0", |
| "", |
| "", |
| "Backlog", |
| "interim-report-objective-choice-cispo-vs-grpo-effect-size-controlled" |
| ], |
| [ |
| "6.2", |
| "K1: EI-SFT vs RL from the same checkpoint", |
| "Token-matched pilot 10-15 RL steps vs one L3 round; then 20-40 steps for the winner", |
| "align/align/rl/", |
| "6.1, 6.4", |
| "~15-40", |
| "2026-10-26", |
| "2026-11-23", |
| "Planned", |
| "docs/SOTA_PLAN.md K1" |
| ], |
| [ |
| "6.3", |
| "L6 agentic scaffold (class C)", |
| "Multi-round refinement R in {4,8} + retrieval of verified lemmas; tuned on valid", |
| "align/align/rl/envs/lean_multiturn.py", |
| "5", |
| "~40", |
| "2026-11-09", |
| "2026-11-16", |
| "Planned", |
| "docs/SOTA_PLAN.md L6; oprover-a-unified-framework-for-agentic-formal-theorem" |
| ], |
| [ |
| "6.4", |
| "RL feasibility check (<=4 h)", |
| "One LoRA step at 16k with FP32 head (2.5 GB + 1.25 GB/2k-chunk vs 1.9 GiB headroom); handoff time; logprob parity", |
| "align/align/rl/cispo.py", |
| "1.3", |
| "~4", |
| "2026-10-17", |
| "2026-10-20", |
| "Planned", |
| "docs/SOTA_PLAN.md Phase 0 item 6" |
| ], |
| [ |
| "7", |
| "FINAL EVALUATION", |
| "Pre-registered: primary = class A miniF2F-test, exact McNemar (needs ~+9 net, ~211); secondary class B + ProofNet (Holm); exploratory class C + Putnam pass@8", |
| "docs/SOTA_PLAN.md", |
| "3.4, 5, 6", |
| "~180 (226 w/ 2nd seed)", |
| "2026-11-26", |
| "2026-12-08", |
| "Planned", |
| "docs/SOTA_PLAN.md sections 1, 5" |
| ], |
| [ |
| "7.1", |
| "Pre-registered final evals", |
| "Both arms, frozen verifier/config hash; --allow-test-selection --rounds 1", |
| "scripts/p2/eval_report.py", |
| "3.4, 5.7, 6.2, 6.3", |
| "~180", |
| "2026-11-26", |
| "2026-12-08", |
| "Planned", |
| "docs/SOTA_PLAN.md Phase 4" |
| ], |
| [ |
| "8", |
| "RESEARCH & FOUNDATIONS", |
| "First-principles understanding of the recipe", |
| "research/, align/notes, align/notebooks", |
| "-", |
| "0", |
| "2026-10-14", |
| "", |
| "Running", |
| "-" |
| ], |
| [ |
| "8.1", |
| "Leanstral recipe lineage (hyperresearch)", |
| "Final report research/notes/final_report_leanstral-recipe-lineage-c14fca.md", |
| "research/runs/leanstral-recipe-lineage-c14fca/", |
| "-", |
| "0", |
| "2026-10-14", |
| "", |
| "Running", |
| "comparisons.md (step 6 done)" |
| ], |
| [ |
| "8.2", |
| "Math notes + notebooks", |
| "0101-0110 policy optimisation, 0201-0208 verifier learning, 0301-0306 formal methods", |
| "align/notes/src, align/notebooks", |
| "-", |
| "0", |
| "2026-10-14", |
| "", |
| "Running", |
| "workflow wf_659e4ae4-362" |
| ], |
| [ |
| "8.3", |
| "Architecture V2 (AgileRL study)", |
| "Modular algorithm/registry design for align", |
| "align/docs/ARCHITECTURE_V2_PROPOSAL.md", |
| "-", |
| "0", |
| "2026-10-14", |
| "", |
| "Running", |
| "workflow wf_659e4ae4-362" |
| ], |
| [ |
| "9", |
| "BUG FINDING (separate paper)", |
| "Prover applied to real Rust code; decoupled from the prover paper", |
| "-", |
| "7", |
| "-", |
| "", |
| "", |
| "Planned", |
| "docs/SOTA_PLAN.md Phase 5" |
| ], |
| [ |
| "9.1", |
| "Bug-finding literature review", |
| "Aeneas pipeline, property generation, prove/negate", |
| "align/docs/research/BUGFINDING_LIT_REVIEW.md", |
| "-", |
| "0", |
| "2026-10-14", |
| "", |
| "Running", |
| "workflow wf_659e4ae4-362" |
| ], |
| [ |
| "9.2", |
| "Aeneas/Charon toolchain (Lean 4.31.0 env)", |
| "Separate pinned environment + repo-definition-aware verifier; pilot on 3-5 functions with known bugs", |
| "-", |
| "9.1", |
| "CPU", |
| "2026-10-16", |
| "2026-10-20", |
| "Planned", |
| "docs/SOTA_PLAN.md Phase 0 item 7" |
| ], |
| [ |
| "9.3", |
| "Prove-or-negate harness", |
| "pass@4 prove, then negate the CLOSED proposition; executable Rust counterexample required", |
| "-", |
| "9.2", |
| "-", |
| "2026-12-08", |
| "2026-12-22", |
| "Planned", |
| "-" |
| ], |
| [ |
| "9.4", |
| "Bug study baselines", |
| "Plain LLM review + cargo-fuzz on the same crates; validated precision per method", |
| "-", |
| "9.3", |
| "~40 (GPU) + CPU", |
| "2026-11-21", |
| "2026-12-05", |
| "Planned", |
| "docs/SOTA_PLAN.md Phase 5" |
| ], |
| [ |
| "10", |
| "PUBLICATION", |
| "Paper: reproduction + self-distillation result + bug-finding case study", |
| "-", |
| "7, 9", |
| "-", |
| "", |
| "", |
| "Planned", |
| "-" |
| ], |
| [ |
| "10.1", |
| "Papers", |
| "Paper 1 prover (K1-K4, classes A/B, exploratory C); Paper 2 bug study", |
| "-", |
| "7.1, 9.4", |
| "-", |
| "2026-12-08", |
| "", |
| "Planned", |
| "docs/SOTA_PLAN.md Phase 6" |
| ] |
| ] |
| } |
