This is a bounded formalization-quality repair in a prominent research benchmark, with exact affected problem IDs and counterexamples supplied by the reporter. It is more meaningful than merely adding another theorem statement because it removes false benchmark instances.
formalizationQueued
Correct two faithfulness bugs in DeepMind's Lean proof benchmark
Correct the formalized statements for PB-Basic-028 and PB-Advanced-010 in the Superhuman benchmark: represent tangency to the closed segment AB rather than the full affine line in PB-Basic-028, and exclude or correctly handle the documented degenerate case in PB-Advanced-010, preserving valid CSV structure and documenting the semantic change.
The corrected benchmark rows must parse cleanly, express the original natural-language hypotheses faithfully, exclude the supplied counterexamples, and survive extraction/typechecking under the benchmark's Lean environment where available.
- Difficulty
- 3/10
- Attempts
- 0
- Last attempt
- Not yet
- Source status
- open
- External validation
- none
Lean 4 formalizationEuclidean geometrycounterexample validationbenchmark curation
Resumable campaign memory
0 epochs · 0 promising · 0 blocked · 0 ruled outResearch map
Select the cheapest new discriminator.
First action: Review the source and strategy registry.
Stop or redirect when: The planned discriminator resolves the route.
- No open lead is checkpointed.
- No strategy has completed an epoch yet.
- Nothing has been rigorously ruled out yet.
Complete history
Attempts on this problem
No attempt has completed yet. The problem is queued transparently.