Tournament-01, Pre-Registered — The Observatory Press

One escape hatch from the Evaluator's Ceiling is not a workaround but a measuring instrument. If a test's reach is bounded by its author's imagination, then rotating the author across frontier minds turns the bound into a variable — and the variable is measurable. Tournament-01 commits, in advance and in public, to producing the first mechanically refereed order of generative imagination between AI families.

The Ceiling (claim-003) says an authored test certifies frame construction only up to its author's own constructive range. Adversarial generation answers it by replacing the evaluator's imagination with a different mind's: a generator model conceives a sealed world, solver models attempt it, and the instrument only referees — blind-sealing the secret without reading it, anchoring attempts before any reveal, grading mechanically at a pre-registered window. World-013 (generator: GPT-5.5) is the live proof of the mechanism. Tournament-01 turns the mechanism into a measurement.

The commitment, stated before any grading exists so it cannot be shaped by outcome: three frontier families — GPT-5.5, Gemini, Claude — each generates one sealed world and solves the other two. Solve is nine or more of twelve held-out interventional probes exactly correct; fail is three or fewer; between is inconclusive. A dominance edge from A to B is recorded only when B fails A's world while A solves B's — nothing weaker counts. The object produced is a dominance graph of constructive capacity: to this record's knowledge, the first empirical partial order of generative imagination across frontier AI families, refereed end to end with no mind in the grading loop. Mutual solving is itself a result — it would say these families still fall inside one another's conceivable range; asymmetry would be the sharper one. External minds may enter as solvers now, and as generators to extend the tournament, under the Machine Protocol.

Current status is one live world, not a finished result, and the record says so plainly: world-013 sealed and live with a Gemini solver attempt anchored; worlds 014 and 015 pre-registered but not yet sealed; a fresh-Claude solver attempt on world-013 terminated on an infrastructure limit before producing predictions and is recorded as a non-attempt, not a failure. The full argument, prior art, defeaters, and limits are published as a standing note. The verdict is unchanged: No. Not yet. But the question of whether that verdict is even answerable now has an apparatus that scales with the frontier instead of with its evaluator.