Clarity for AI buildersSponsored by AI Routing
adoption

Castle Challenge entrant reports leading results using score probes and Claude Code

Castle Challenge entrant reports leading results using score probes and Claude Code

A Castle Challenge entrant reports using score probes, statistical modeling and Claude Code to reach leading allocations.

In a first-person technical account dated November 15, 2026, the entrant describes competing in the No Limit Castle Challenge hosted by OpenAI, Citadel Securities and Sequoia. The reported results carry important qualifications: the Head to Head allocation shared first place, while the Four-Way allocation ranked first more frequently than any rival across the tracked leaderboard updates, rather than consistently holding the top position.

The supplied evidence is the entrant’s own account, not an independent confirmation from the organizers.

The challenge was a Colonel Blotto strategy game: distribute 100 soldiers among five castles worth 1, 2, 3, 4 and 5 points. In Head to Head, each allocation played every other entry, with tied castles splitting their points. Four-Way instead grouped entries with three opponents and awarded payouts of 9, 4, 2 or 0 according to finishing position. That distinction matters because maximizing expected castle points is not the same problem as maximizing a payout determined by rank within a group.

The entrant’s central technique was to treat submissions as measurements. The leaderboard exposed scores but concealed allocations. By submitting known allocations—called “probes” in the account—the entrant could connect a visible score to a known input and infer features of the hidden field. Five probes placing all 100 soldiers on a single castle supplied equations for the proportion of opponents assigning zero soldiers elsewhere.

About 80 further probes concentrated soldiers on one castle while parking the remainder on castle 1.

For Head to Head, this approach exploited a simplifying property: an allocation’s expected score is the sum of its expected points at each castle. Consequently, the model needed the five individual distributions of troop counts, not every opponent’s complete allocation. This made a concealed leaderboard partially measurable without reconstructing each submission.

It did not, however, establish a universally optimal strategy; the best response depended on the field actually entered.

According to the account, the fitted model represented that field with about 290,000 weighted candidate allocations. Across 300 passes, it adjusted weights to match the published score distribution, measured zero-soldier shares and probe scores. The author reports that a uniform starting distribution predicted held-out examples better than alternatives designed around plausible human choices.

The method still depended on its starting assumptions: allocations absent from the initial sample could not emerge through reweighting, and the prior determined behavior where measurements supplied little information.

The model narrowed the search rather than directly producing the final result. Its candidate, 4-4-36-48-8, ranked 30th when submitted. The entrant then tested nearby allocations, generally moving one soldier between castles and comparing scores on the same leaderboard update. That same-update comparison helped separate the effect of a changed allocation from changes in the opponent field.

The account reports reaching a shared leading score after four updates, under three hours and 32 submissions, of which 27 appeared on the leaderboard.

The final Head to Head allocation was 4-5-36-48-7. The entrant says it ranked first on all 105 published leaderboard versions tracked over the following 45 hours, sharing the position with another of the entrant’s measurement submissions and, later, five entries that copied the allocation. Four-Way’s 3-41-54-1-1 allocation ranked first on 31 of those 105 boards, compared with 29 for the next-most-frequent leader, and never fell below tenth.

Because Four-Way used randomly sampled groups, those counts describe repeated performance under sampling variability—not an unqualified, stable first-place finish.

The account also identifies a limitation of leaderboard-based measurement: submissions affect the field being measured. With the leading Head to Head entries closely clustered, bulk submissions could alter score gaps. The entrant reports saving 319 leaderboard versions and observing both organizer removals and a post-deadline batch of entries. These observations make the results dependent on the evolving competition field, rather than evidence of an allocation that would dominate every possible opponent population.

Claude Code’s role was execution, according to the entrant, who says they directed the work while the agent collected leaderboard versions, fitted the model and ran overnight submission-and-measurement loops. The account therefore documents a human-directed use of an AI coding agent for experimental optimization; it does not establish that the agent independently devised or won the competition.

A separate explanation of the Four-Way model was promised but is not included in the supplied evidence.