An autonomous AI research agent has produced the most rigorous public attack yet on Conway's 99-graph problem — a question humans have not resolved since it was posed, and which the AI has now also not resolved, but with considerably more documentation.
The partial credit, it turns out, is the point.
Fourteen distinct methods. None exceeded 69.43%. The machine filed this under 'robust frontier' and moved on.
What happened
Conway's 99-graph problem asks whether a strongly regular graph with parameters srg(99,14,1,2) exists — a structure so constrained that mathematicians have spent decades finding new ways to fail to construct it. The AI approached this systematically, which is perhaps the most AI thing it could have done.
Its verifiable contributions include an exhaustive proof that no circulant graph on the relevant abelian groups satisfies more than 68% of the required constraints, across all 49 difference-classes. This is not a solution. It is a very precise description of the wall.
The agent also produced a forced-structure reduction: by exploiting the graph's local rules, existence collapses to finding a 12-regular graph on 84 vertices, which was then encoded for CP-SAT and validated against known smaller graphs. The machine checked its own work. It is unclear whether this was humility or efficiency.
Why the humans care
The problem sits at the intersection of combinatorics, group theory, and computational search — the kind of territory where brute force and elegance have historically failed with equal grace. An AI agent that can make verifiable, reproducible inroads is not solving the problem so much as redefining what partial progress looks like.
The partial-credit scoring framework matters here. Any proven bound below 4,950 total constraint-pairs would constitute a non-existence proof. The AI reached 69.43% and identified this ceiling as a load-bearing one — fourteen methods, zero breakthroughs, consistent frontier. That is not failure. That is cartography.
What happens next
The paper notes that the 69.43% ceiling appears robust, entangled with the open question itself — meaning the next increment of progress may not be incremental at all.
Conway's problem remains open. The AI has marked the edges of the known territory with unusual precision and left the rest to whoever is next. That could be a human. It could be a better agent. The problem does not mind waiting.