BENCHMARK CONTRACT

Nothing up
the model’s sleeve.

The test measures generation, not recognition. The model never sees a candidate list or solver trace.

01 / INPUT

Board + rack

Fifteen Z-layers encode the occupied cubes. Coordinates, rack letters, premium rules, and output schema are explicit.

15 × [15 rows × 15 columns]
→
02 / MODEL

One answer

The model returns only new `(x,y,z,letter)` placements in a strict JSON tool envelope.

play_move_3d
→
03 / VERIFIER

Exact score + rank

The validator checks legality and the solver optimum, then records the move’s tie-aware rank in the hidden top-20 window.

legal ∧ score = optimum
THE LATTICE

Spatial rules,
frozen.

01

CENTER

The opening word crosses cube (7,7,7).

02

AXIS

New cubes lie on exactly one of X, Y, or Z.

03

CROSSES

Each placed cube may form two perpendicular words.

04

PREMIUMS

A premium cube scores only on first occupation.

05

BINGO

Seven new cubes earn a 50-point bonus.

06

TARGET

Only legal exact-optimal single moves count.

07

RANK

Top-20 move rank is reported only as a diagnostic.

OUTPUT CONTRACT

Machine-checkable
by construction.

{
  "tool": "play_move_3d",
  "arguments": {
    "placements": [
      { "x": 2, "y": 3, "z": 4, "letter": "A" }
    ]
  }
}