The sharpest AI bug report we received was a red rectangle around a two-dot height tile.

Petra had moved. The board had advanced. The human player’s question was still perfectly reasonable: why did the height opportunity survive? But the screenshot could not tell us the whole answer. By then we no longer had the exact pre-move hand, the candidate census, or a proof that the attractive construction was legal and survived the reply.

That limitation became the point. An evaluator cannot prefer a construction it never sees.

The order of attention is a strategic choice

An OverGrid turn is a complete three-piece construction, not a click on a tile. There are many legal constructions; a responsive live opponent cannot examine them all. A fixed budget is therefore necessary, especially at a three-to-six-player table, where each named AI receives 250 complete constructions rather than a duel’s 800.

That budget has a second job beyond performance: it decides what becomes eligible to be evaluated. A geometry-first traversal can be deterministic, legal, and fast while still spending its early attention on constructions that happen to arrive first. A route that exactly banks height may exist later in the same legal space and never make the bounded menu.

Figure 1 · teaching positionThe evaluator can only choose from the constructions it receives.Illustrative legal geometry · not a reconstruction of Petra’s reported game
••
A different geometry consumes the prefix

The height match can be legal and valuable yet absent from a geometry-only early prefix. A better evaluator cannot rank an unseen construction.

No claim about the actual player report

Tap the tabs to switch the order in which the same kind of legal construction is admitted. This does not say every height move is best; it says obvious public motifs must get a fair audition.

This is why “we should improve Petra’s height weight” was the wrong first repair. A weight only ranks candidates already on the desk. If the relevant shape has not arrived, adjusting the height weight is like improving a chess engine’s bishop evaluation while quietly deleting bishops from its move list. Very polished. Not especially helpful.

What shipped: public consequences before personality

The correction was deliberately shared. The legal-move visitor now measures public consequences for each candidate—exact available height, partial height progress, opponent columns, neutral territory, black-tile bridges, footprint, and verticality—and interleaves those lenses into the bounded census.

Every named profile receives that same representative supply. Only after that common competence step may Petra select a public successor within her bounded strength-regret rail. This matters because a “Petra reserve” for obvious height would have grown into an unscalable pile of special pleading: a height reserve, then a capture reserve, then an endgame reserve, and eventually a reserve for every move a human could circle.

Figure 2 · shipped candidate supplyOne shared census, several public ways to become visible.Public-consequence-v1 · every named profile receives the same supply
01Bank now

Exact, unclaimed height

02Build next

Partial height progress

03Deny

Opponent-owned columns

04Grow

Neutral ordinary columns

05Bridge

Uncleared black crossings

06Leverage

Footprint and verticality

Legal constructions→Interleaved bounded census→Shared evaluator→Named choice inside regret rail

The ordering reads map cells, current ownership, the acting hand’s legal construction, and public height availability. It does not read future hands, RNG state, opponent labels, experiment labels, or outcomes.

A correction has a narrow proof obligation

The shipped change proves a structural fact: exact-height and related public motifs are deliberately represented in the live candidate supply before named selection. It does not prove that the original photographed height was legal, that Petra should have taken it, or that the new ordering improves match-point win rate.

Figure 3 · evidence boundaryWhat the correction proves—and what it deliberately does not.Product change versus strength evidence
  1. PLAYTESTObserved

    A human saw Petra apparently leave a two-dot height opportunity.

  2. SHIPPEDCorrected

    Public consequences are interleaved before named selection.

  3. OPENNot established

    That Petra now wins more games, or that every height capture is correct.

The next valid strength claim needs fresh paired games with homes, starters, maps, and opponent schedules crossed—not a more flattering screenshot.

The next valid test is a paired schedule: frozen old and new candidate ordering, identical maps, seeds, homes, starters, opponents, and complete natural endings. It should report the opportunity → admission → selection → reply-survival → conversion funnel, not just a favorite screenshot. That is how we learn whether coverage changed a game rather than merely making a better story about one.

The broader lesson

The AI’s policy is not just its final scoring formula. It includes the candidate generator, ordering, work budget, tie-breaking, reply checks, and information boundary. Improving one layer while another quietly excludes the relevant move produces endless inert research: increasingly sophisticated reasons to select from the wrong menu.

The red rectangle did not establish that Petra was weak. It established a much more actionable proposition: before teaching a system better judgment, make sure it is looking at the board.

Read the companion note, A player circled one tile. It changed the AI program., for the replay and audit infrastructure that makes the next report stronger.

The extracted player lesson is now Rotate the whole hand, not just the piece you like.