What was already known
Van Dam’s oracle interrogation algorithm could recover the whole string at bounded error using queries. It was open whether a Boolean output could require more than at fixed error.
AI involvement
OpenAI’s GPT-6 Astra Ultra found the upper bound. Google’s Stellar Colosseum found a matching lower bound and then supplied a simpler hard function at the authors’ request. The authors posed the problem and rewrote the lower-bound analysis after working through it.
- AI methodologyAI methodology
Importance
The paper answers the lower-bound question and identifies the order of the additive overhead for symmetric Boolean functions.
Technical verification
The authors report working through the lower-bound argument in detail. They link a Lean check for the earlier construction. Comparable checking evidence for the full pair of published bounds is undocumented.
Independent checking
Worked through the lower-bound argument in detail before rewriting the analysis.
- Checker
- Paper authors
- Scope
- The manuscript’s simplified symmetric-family lower-bound argument.
- Basis
- Reported by the linked checker
Formal checking
Checked the first Colosseum lower-bound construction. The linked account distinguishes the later simpler construction.
- Checker
- Lean 4 kernel
- Scope
- Colosseum’s first lower-bound construction.
- Basis
- Reported by the linked checker
- Lean artifact for the initial lower boundLean artifact for the initial lower bound
Human understanding
The authors describe how they simplified and rewrote the lower-bound analysis. The methodology does not give a comparable account of their reconstruction of the model’s upper-bound algorithm.
- Authors’ lower-bound analysisAuthors’ lower-bound analysis
Other useful sources
Correction and publication history
Reviewed the primary result and its documented AI and checking scope.
Read this published revision