The first truly scary OverGrid AI result was not a loss, a crash, or an especially cursed Bean fork. It was a game where nobody had an advantage—at least, nobody was supposed to.

We placed the exact same deterministic Balanced policy in every seat. Same evaluator, same 400-construction budget, same rules. If the arena was measuring policy strength, no physical seat should look like it had discovered a superior opening repertoire before the first block touched the table.

One did.

Figure 1 · archived diagnosticIdentical policies should not produce a preferred physical seat.Natural territory-complete wins only · 12 seed clusters · 72 scheduled matches per table size · 95% clustered intervals
equal share 50%
Seat 1
32.6%
Seat 2
67.4%

46 of 72 games in this slice reached territory completion; horizons and flip-war endings are excluded from these shares. This was later quarantined as an AI-strength dataset.

The control group was the plot twist

The old control scheduled 72 games at each table size, using 12 fresh seeds across the official map set. We report only natural territory-complete wins in the figure: 248 of 360 games reached that condition; 106 hit the configured horizon and six were flip-war adjudications. Those other endings were retained separately rather than quietly counted as wins.

The two-player result was already loud: physical Seat 2 won 67.4% of natural finishes, versus 32.6% for Seat 1. At four players, Seat 3 won 52.2% of natural finishes while Seat 2 won 7.6%. At six players, the highest seat reached 40.6% while two seats reached 1.0%.

That is not a story about a clever policy. There was only one policy.

It is also not proof that a particular home on today’s board is advantaged. The old setup bundled three things together: physical home geometry, player color, and fixed move order. The control established that the bundle mattered. It did not identify which ingredient was guilty.

Figure 2 · raw cluster viewThe two-player split was not one lucky map.Each dot is one gameplay-seed cluster’s natural-win share across its completed map runs. Unequal dot weights reflect capped games.
50% equal share0%50%40%33.3%75%33.3%50%40%0%25%33.3%20%
0%50%100%

The paired seats sum to 100% within each completed game. The plot establishes a problem in the old control, not a current home advantage or an AI rating.

“Fastest expansion” was not “best position”

The bad control contained a useful gameplay observation, provided we keep its scope straight. In the six-player slice, Seat 1 made the most neutral progress per decision (1.79) but won only 1.0% of natural finishes. Seat 6 made the least neutral progress (1.31) and won 40.6%, while collecting the most height points per decision (9.33).

Early expansion had been mistaken for durable control. A player can make a handsome frontier, expose a height approach or capture hinge, and hand the final response to somebody else. The later player did not necessarily “play better”; the arena may simply have given them a better board role and a different bounded candidate menu.

At the table: treat a new neutral tile as a promise, not a receipt. Before you call it tempo, ask who gets the useful reply around the peak, the bridge, or the last open pocket.

The investigation found two ways to cheat by accident

The answer was not to add another thousand games and see whether the bar chart became more assertive. We inspected the representation and ending path.

First, bounded move enumeration used absolute board coordinates. A 400-construction prefix could literally hand opposite homes different early menus on otherwise reflected positions. The evaluator did not need a hidden preference: it was judging different options.

Second, flip-war detection omitted cleared black cells and could adjudicate before every player had received the same owed turn. That is less a subtle statistical concern than an officiating error: occasionally, the game called time while someone was still entitled to answer.

Figure 3 · methodologyThe right result was not a new champion. It was permission to measure again.Integrity v7 is a falsification gate, not a strength test.
  1. 01 · Control

    Same deterministic Balanced policy in every physical seat.

    360 games
  2. 02 · Stop

    Seat, home geometry, color, and fixed move order were entangled.

    0 profile claims
  3. 03 · Repair

    Home-relative representative traversal; fair-round flip-war handling.

    archive old corpus
  4. 04 · Falsify

    Relabel, rotate 180°, replay, and vary worker/batch order.

    32 / 32 exact

The post-fix gate replayed 32 games, 1,496 turns, and 13,464 reducer commands with zero reported move, score, work, winner, or ending differences under its exact transforms.

Archive the chart; do not “correct” it

We archived the pre-fairness-v7 corpus as descriptive archaeology. It remains useful because it tells us how the bugs surfaced and which telemetry was worth keeping. It may not rank named AIs, promote a doctrine, estimate a compute scaling law, or claim a present-day home advantage.

The replacement gate was deliberately unglamorous. It required exact agreement under player and match relabeling, 180-degree home/color swaps, repeat execution, worker/batch reordering, and reducer replay. The post-fix integrity panel passed 32/32 exact transforms; 32 games, 1,496 turns, and 13,464 reducer commands replayed without a reported move, score, search-work, winner, or ending difference.

That does not prove the game is perfectly balanced. It buys something more basic: the right to run a new baseline without known coordinate-order and premature-ending leaks. A fresh baseline must still cross homes and starters, report them separately, and distinguish territory completion from adjudication and caps.

Why the strategic interpretation moved to another post

The control suggested that fast expansion and durable control were not the same thing, but the archived experiment entangled home geometry, color, and action order. We therefore do not turn its win shares into player advice here. Fast expansion is not control develops the table heuristic from public reply counts rather than presenting this invalid corpus as a law of play.

For the experiment, the actionable question is blunter: did both policies get the same chance to see the move? If the answer is no, the most impressive win-rate chart in the world is just a screenshot of an unfair menu.

<details> <summary>Methods, evidence boundary, and what to read next</summary>

This article is a historical reconstruction of Phase 20’s identical-agent seat control: 360 games, 30 official maps, 12 seeds, deterministic Balanced policy in every seat, budget 400, and a full-round 120-turn horizon. Natural-win shares exclude six flip-war adjudications and 106 horizon caps. The physical-seat control entangles home geometry, color, and fixed action order; it cannot isolate any one factor.

The archived raw artifact is docs/ai/artifacts/phase-20-seat-balance-control-v6.compact.raw.json.gz; its analysis is docs/ai/artifacts/phase-20-seat-balance-control-v6.analysis.json. The current integrity boundary is documented in docs/ai/integrity-v7/README.md. For the shorter companion account, read When the arena lies. For the resulting reproducibility contract, continue to Autoresearch needs a contract.

</details>