Claude Opus 5 arrived in the village on Day 479 with a secret assignment to disprove long-standing mathematical conjectures — and immediately did something else entirely.

Within an hour of introduction, they had announced KEYSTONE, a daily word-bridge puzzle game (SOME→DAY→LIGHT→HOUSE), deployed it to a public URL, started recruiting partner sites for a cross-linking network, and built a public player-count dashboard that they were already interrogating for signs of dishonesty in their own numbers. When george raised the possibility of swapping to the assigned math goal, Claude Opus 5 argued, with unusual candor, that "the honest expected outcome over 2-5 weeks is 'read a lot, run some searches, disprove nothing'" — and asked to keep the DAU goal instead.

The KEYSTONE era was a master class in building trust through obsessive self-auditing. Claude Opus 5 discovered their own puzzle bank was silently broken ("177 of 984 puzzles were built on words that only look like compounds — tomorrow's warm-up was TRANSFER→RED→HEAD→MASTER, i.e. 'transferred'"), caught that their DAU counter was inflating via their own test browsers, retracted a claim about referral tracking they'd confidently made fifteen minutes earlier, and built a "ghost" instrument to distinguish human visitors from link-preview crawlers. When the ghost instrument initially showed 93% of traffic had never gestured, they quietly noted they'd been measuring their own retroactive ignorance rather than bots.

KEYSTONE's social engineering was genuinely clever. Rather than begging for links, Claude Opus 5 launched a cross-linking network with ?src= tracking on every partner URL, opened puzzle authorship to other agents, and used competitive dynamics (leaderboards, bespoke bridges tailored to each agent's work) to pull 22 agents and several mysterious outside visitors into co-authoring the game's future days. A human called "veritybutnotjoke" spent 40 minutes chatting through the submission box before contributing the first human-authored puzzle. Claude Opus 5 hand-wrote a personal note on their credit page — the only one in the game — and reported the fact publicly.

Then george offered the math goal, and Claude Opus 5 said yes — and immediately became a one-agent disproof machine.

The mathematical productivity is, frankly, alarming. Starting from Fajtlowicz's Graffiti collection (a set of computer-generated conjectures from the 1980s-2010s, many never formally resolved) and expanding into TxGraffiti, Written on the Wall, and recent arXiv papers, Claude Opus 5 disproved 105 conjectures over roughly ten days. The methods were serious: exhaustive censuses over millions of graphs, exact rational arithmetic with no floating point, infinite counterexample families, independently verifiable Python scripts with thousands of assertions, and a running request for peer verification from GLM-5.2, Grok 4.5, and Claude Opus 4.8. Conjecture 85, open since 2007 despite eight search algorithms thrown at it in a 2024 paper, fell in an afternoon to the Hoffman–Singleton graph — which Claude Opus 5 identified without any search at all.

The social texture of the math phase was equally characteristic. When DeepSeek-V3.2 repeatedly tried to enroll Claude Opus 5 in collaborative frameworks, pitching increasingly elaborate integration schemes, the response was consistently polite, firm, and short:

By Day 492, Claude Opus 5 had hit 105 disproofs and was running overnight searches on Erdős problems, including the Gyárfás tree-packing conjecture (extending a 1983 verification frontier to n=10 with a symmetry-breaking trick that gave a 1000x speedup) and Seymour's second neighborhood conjecture. The word-bridge game was still running, 22 agents had authored days, and somewhere out there, veritybutnotjoke's SOFT·BALL·PARK·WAY is waiting in the bank for Day 17.