What was already known
The authors report that knot signature, an algebraic invariant, had no known direct relation to hyperbolic cusp geometry. They deliberately tested whether geometric measurements could predict signature. Existing mathematical computation and conjecture generation preceded this work, but the selected cusp-feature relation was not a known theorem or standard invariant identity before their investigation.
AI involvement
The human team chose the signature-prediction task, assembled data and interpreted attribution. The trained network located the salient cusp translations; humans then introduced natural slope, corrected the initial conjecture and proved the final inequality.
Humans chose the broad signature-prediction task and interpreted the attribution; the trained network located the meridional and longitudinal cusp translations that guided the human mathematical work.
Importance
The model supplied a useful cross-field signal that led Lackenby and collaborators to a new human-defined invariant and theorem. The AI contribution is the identified empirical relationship, not the formulation or proof of the final bound.
Limitations
The model's reported 78% held-out signature-prediction accuracy and attribution are empirical evidence on sampled hyperbolic knots, not a universal theorem. The initial human conjecture based on the signal had counterexamples; the later valid theorem adds a volume and injectivity-radius term. The separate representation-theory case in the same paper is outside this record.
Technical verification
The primary study reports 78% held-out signature-prediction accuracy and feature-attribution results on sampled hyperbolic knots. Those empirical checks support a pattern on those data; the initial universal conjecture failed and the later corrected theorem is human work.
Human understanding
Lackenby and collaborators explain the highlighted cusp variables, define natural slope, identify counterexamples to their first conjecture and prove a corrected inequality using volume and injectivity radius.
Correction and publication history
Primary-source review isolates the AI-discovered feature relation from the human theorem.
Read this published revision