The September global game and the fourteen policies that opened its leaderboard

A new month, a new fixed future

The public global game is a single scripted round: the disc at every move and the hidden value of every covered disc are decided in advance. That makes it cheap to run, identical for everyone, and useless as a measurement of policy strength. It ran on gauntlet-01 through August. It now runs on gauntlet-02.

The lifecycle went through the competition tool rather than by hand. It wrote an immutable manifest for global#2026-09-v1 naming gauntlet-02 and pinning the round file's SHA-256, marked global#2026-08-v1 archived, and advanced the catalog pointer. Nothing in the August manifest or its round changed. An archived game stays selectable on the leaderboard and its stored submissions stay replayable, so nobody's August result disappeared.

Two bugs the release review caught

Reviewing the release turned up two problems that would have been visible to anyone opening the site.

negative

The deploy benched a hardcoded round

the September leaderboard would have shown no computer rows and its replay pages would have 404ed

The deploy step generated computer scores with the round id written into the workflow. Advancing the game to gauntlet-02 would have left the new leaderboard with no rows from that step and no replay page behind any of them, because both the leaderboard and the home page select local games by the current game's round. Pointing the step at gauntlet-02 alone would have moved the same breakage onto the archived August game, which visitors can still select. The step now reads the catalog and plays every round it still serves.

negative

Two figures on the n-tuple page claimed their screens had produced nothing

two of eleven figures on the approach page

The screen figure defaulted to one named contrast. The depth-4 and fill-conditioned experiments gate different pairs and never wrote that key, so both of their primary figures fell through to the placeholder that says the stage produced no artifacts. The figure now follows the contrast each run recorded as its primary, which the extractor had already been writing.

Fourteen policies played the round

Every policy in the benchmark registry that can finish a game played gauntlet-02 on this workstation, and each game was published to the competition ledger under the drop7-research identity. The tool replays every recorded game independently before it writes, and none of the fourteen disagreed with its own score.

The lookup tables of the first n-tuple-scale run, used as the leaf of a depth-3 search, scored 1,113,829 points in 310 moves. The same tables one ply deeper scored 699,517 in 200 moves.

neutral

A million-point scripted game says nothing about the qualification standard

1,113,829 points in one game on one fixed future

The qualification standard asks for a mean above one million points in corrected five-move Hardcore, measured across a cohort of complete games on seeds the policy has never seen. This is one game on a round whose entire future is written down in this project's own source, visible to anyone who looks. Drop7 scores are heavy-tailed enough that a 64-game research cohort has a standard deviation near half its mean, so a single game is an anecdote. It is a milestone worth writing down and it is not a result.

Deeper search finished lower

The Rust engine's completed depth-6 search with seven chance strata spent forty-five minutes on the round and scored 105,578 points in 35 moves. Its depth-4 sibling in TypeScript scored 121,347 in 40 moves on the same future, and the same engine scored 428,292 in 120 moves on gauntlet-01 in August.

Nothing here ranks those searches. Eight fixed rounds cannot separate policies whose scores are heavy-tailed, and one round separates them even less. A related instrument was already measured failing its validation gate for this exact reason, ranking search configurations backwards over a short horizon because the score there is dominated by the rise cadence rather than by skill. The depth-6 number is recorded because the game was played, not because it means the deeper search is worse.

The depth-7 policy did not play. Its opening decision on gauntlet-01 was measured at 2302 seconds, a later mid-game decision on a half-full board exceeded 7200 seconds, and its only attempt at a full game stopped at move 83. A complete game would take days of exclusive use of this machine. It was not on the August leaderboard either, for the same reason.

Housekeeping

Gray throughput now appears on the leaderboard, badged as extended state. It reads the level and the move number, so it sits outside the public-information boundary every research candidate has to respect, and the badge is how the board says so. It was not on the August leaderboard.

All fourteen games are playground evidence. They consumed no seed lease, none of them is tier evidence, and none of these numbers belongs in a theory, experiment, or result record.