AI-found polytopes refute the printed IRIS face bound

Five simple 3-polytopes with 10 faces violate the printed IRIS Conjecture 6.1. One has four triangular, three pentagonal and three heptagonal faces, but no hexagonal faces, below the conjectured lower bound of 1/5.

Sources for this record

What was already known

The 2025 IRIS paper posed the face-count inequality as a conjecture.

AI involvement AI-centered

Claude agents found the counterexamples in John Erlbacher’s directed Demonstrandum campaign. Erlbacher audited its verification gates.

Claude agents supplied the witnesses under Erlbacher’s direction; individual-result autonomy is not documented.

Importance Modest

Corrects a specific proposed inequality about faces of simple 3-polytopes.

Limitations

The witness refutes the printed real-valued bound. It does not refute an integer-rounded weakening, Barnette’s theorem or the paper’s other conjectures. The original OpenReview page could not be read directly; the conjecture wording follows the cited frozen transcription.

Technical verification Formal (exact computation)

Independent Python and Rust exact checkers reportedly certify the five counterexamples to the printed inequality.

Independent checking

exact computation

Checker
Independent Python and Rust counterexample checkers, plus census re-derivation
Scope
The five 10-face counterexamples to printed IRIS Conjecture 6.1
Basis
Reported by the linked checker
Evidence source

Human understanding Few

No named person explains the individual construction; Few is the evidence default.

Correction and publication history