VibeMathedMath problems solved with AI

The proper hat-guessing number of K8eK_8-e

Determine the exact proper hat-guessing number of the complete graph on eight vertices with one edge removed. The general bounds leave HGP(K8e){13,14}\mathrm{HG}_P(K_8-e)\in\{13,14\}.

Result
Proved(see note)
Status
Candidate (review pending)
AI contribution
AI co-developed
Method
Computation
Field
Graph theory; hat-guessing games; finite geometry; matching theory
Posed by
Adriaensen et al.
Year posed
2026
Years open
0y
Solved
2026-09-04
Model
GPT-5.6 Pro
Vendor
OpenAI
Collaborators
Verification
Unreviewed
Publication
Announced
Significance
7 / 100
Disclosed cost
Wikipedia
No dedicated article

What was actually shown

We prove HGP(K8e)=14\mathrm{HG}_P(K_8-e)=14 using an explicit PGL(2,13)\mathrm{PGL}(2,13)-equivariant strategy. A compact formula fixes 990 normalized twin decisions, and a 48,510-edge residual-orbit matching induces consistent rules on 53,460 labelled clique-view orbits. Independent implementations regenerate the certificate and check all 138,378,240 proper fourteen-colourings with zero failures. Unlike the n=5,6,7n=5,6,7 constructions, residual right degrees reach eight, so the explicit global matching is load-bearing. A separately labelled companion exhibits one fixed clique completion compatible with 23802^{380} equivariant twin-rule pairs and proves 380 optimal within the stated independent whole-tail reversal model; that companion's independent review is pending. The release does not solve the general KneK_n-e family or K9eK_9-e.

What the AI did

OpenAI GPT-5.6 Pro contributed substantively to literature search, target selection, construction search, proof development, code generation, exact verification, adversarial critique, certificate design, and manuscript and release preparation. The equivariant construction, residual-matching proof, Hall-obstruction analysis, and shared-completion family were developed in a human-directed model-assisted research process. Matthew Protti selected and framed the target, directed and evaluated the work, commissioned review, required exact checks, approved the public scope, and accepts responsibility.

Verification

Unreviewed on this site's ladder. The frozen commit bacb79e3 carries STATUS.json with headline status PROVED_COMPUTER_ASSISTED and project review ACCEPTED, an AI_USE_AND_PROVENANCE.md naming the model and the division of labour, and a C++ sweep of all 138,378,240 proper fourteen-colourings with zero coverage failures across 48,510 residual orbits. The Gate 2/3 Hall-obstruction archive was reviewed as ACCEPT_WITH_EXPLICIT_REPAIRS and the repairs applied. The 2^380 shared-completion companion is a separate claim whose independent review is still pending and is not part of this entry. As with the siblings: a real adversarial review of an artefact, not a named expert or a proof assistant, and this site has not rebuilt it.

Sources

Submitted by Matthew Protti on

Changelog2 changes

Discussion