| { |
| "columns": [ |
| "Id", |
| "Missing data / tool", |
| "What exists instead", |
| "Build-it-ourselves plan", |
| "Est. box-hours", |
| "Source", |
| "Status" |
| ], |
| "widths": [ |
| 7, |
| 40, |
| 46, |
| 58, |
| 12, |
| 36, |
| 14 |
| ], |
| "rows": [ |
| [ |
| "G1", |
| "LeanGym-PR: executable repository pull-request tasks (Leanstral 50% of RL mix)", |
| "Nothing public. MathlibPR is a merge-readiness classifier, not task+patch. princhernwang/mathlib-refactor-history (gated) has refactor diffs.", |
| "Mine Mathlib4 git history: (pre-commit tree, PR description, post-commit diff) where the diff adds/repairs proofs; task = reproduce the patch until `lake build` of touched files passes under the pinned toolchain. Start with single-file proof PRs only. Success must also require: no sorry/axioms beyond standard, theorem statements preserved, dependency audit; ground-truth patch is one valid solution, not the only one.", |
| "~20 (CPU builds)", |
| "critic gap; in-this-work-we-present-leanstral-a-series-of-generalist-code-agent-models-for", |
| "Planned" |
| ], |
| [ |
| "G2", |
| "Prove-or-disprove Lean data (Leanstral 30% of RL mix)", |
| "Only xlr8harder/lean-proof-or-refute-300 (300 items) and formal-conjectures statements.", |
| "Negate the CLOSED proposition (¬ ∀ x, H x → P x) of never-solved statements after screening them for misformalisation; keep verified disproofs. Perturbing a true theorem does NOT guarantee falsity: any perturbed variant needs its own verified proof or disproof before it gets a label.", |
| "~10", |
| "critic gap", |
| "Planned" |
| ], |
| [ |
| "G3", |
| "Lean code-agent trajectories at SWE scale (Leanstral SFT 50%)", |
| "OProver/OProofs per-theorem repair trajectories (6.8M rows); small RRMA/Erdos sets; non-Lean SWE traces (SWE-Gym, SWE-smith). OProofs trajectory subset UNCONFIRMED; Goedel-LM/RL_dataset_V2 holds prompts/messages, not verified repairs.", |
| "Generate our own: correction rounds (chain5) already produce (attempt, compiler error, repair) triples; log them as multi-turn traces in the chat format; mix with OProofs repair subset after decontamination.", |
| "from 5.2/5.3", |
| "critic gap", |
| "Running (via WBS 5.2)" |
| ], |
| [ |
| "G4", |
| "Rust->Lean (Aeneas) property / negation dataset", |
| "None on HF. Dalek Bench (curve25519-dalek-lean-verify) and an Aeneas experience report exist; Verina, CLEVER, FVAPPS, vericoding are spec benchmarks, not Aeneas output.", |
| "Run Charon+Aeneas on a small set of pure-Rust crates (no unsafe/IO); auto-generate properties from doc comments/tests; prove-or-negate with our prover; hand-validate every flag. This is the bug-finding study.", |
| "~40", |
| "critic gap; BUGFINDING_LIT_REVIEW.md", |
| "Planned (WBS 9)" |
| ], |
| [ |
| "G5", |
| "Leanstral bug-finding release (57 repos / 47 flags / 11 real / 5 new)", |
| "Not public.", |
| "Reproduce at small scale with baselines (plain LLM review, cargo-fuzz) on the same crates; report precision after property validation.", |
| "in G4", |
| "in-this-work-we-present-leanstral-a-series-of-generalist-code-agent-models-for", |
| "Planned" |
| ], |
| [ |
| "G6", |
| "Verifier exploit / reward-hacking corpus", |
| "None public.", |
| "Our P1 adversarial suite (61 cases) + red-team classes (axiom walk, lexer, meta) become a versioned regression set shipped with the paper.", |
| "~2", |
| "654c20c; redteam/false-accepts", |
| "Running (WBS 3)" |
| ], |
| [ |
| "G7", |
| "Lean token counts for a mid-training mix", |
| "No source states Lean token totals.", |
| "Out of scope (no mid-training). If revisited: tokenise Mathlib + Lean-Github with the Qwen3 tokenizer in streaming shards.", |
| "-", |
| "critic gap", |
| "Out of scope" |
| ], |
| [ |
| "G8", |
| "Mathlib-version alignment", |
| "Most new Mathlib-derived sets target Mathlib 4.27-4.31; miniF2F-test header is Lean 4.9.0-rc1.", |
| "Only use data that compiles under our pinned 4.9.0-rc1 toolchain: re-verify every SFT proof with the strict verifier before training; drop failures (this is also the quality filter). Compiling under 4.9 proves compatibility, not informal fidelity: record retained yield and domain loss per source.", |
| "in 5.3", |
| "critic gap", |
| "Rule" |
| ], |
| [ |
| "G9", |
| "GitHub-hosted candidates unconfirmed (API quota hit)", |
| "vericoding (HF copy verified), curve25519-dalek-lean-verify unverified.", |
| "Re-check with an authenticated gh CLI (`gh auth login`) before use.", |
| "<1", |
| "workflow log", |
| "Open" |
| ] |
| ] |
| } |
