loop-benchmarking

commit 57b74f96d1ff52223531c72a77bb145a43f51272
parent 68c55df2846dad69bc4b1247c072387f21e32909
Author: Brian Graham <brian@buildingbetterteams.de>
Date:   Mon,  6 Apr 2026 09:37:01 +0200

Prevent off-grid reading and false positive piece detection

Grid reader:
- Reject grids where >60% of cells are filled (likely reading UI
  chrome, borders, or decorations instead of game grid)
- validateGridBounds(): check aspect ratio (~2:1 for 10x20 grid),
  reject grids that are too small or too large

Tests:
- piece_locks: require piecesSpawned > 0 alongside piecesLocked,
  otherwise flag as "likely false positive from UI misread"
- multiple_pieces: same spawned > 0 requirement
- Prevents the scenario where 0 pieces spawn but 10 "lock" because
  static UI elements were misread as game state

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Diffstat:
M
tasks/tetris/eval/gameplay-bot/grid-reader.ts
 | 
32
++++++++++++++++++++++++++++++++
M
tasks/tetris/eval/gameplay-bot/tests.ts
 | 
12
++++++++++--

2 files changed, 42 insertions(+), 2 deletions(-)
diff --git a/tasks/tetris/eval/gameplay-bot/grid-reader.ts b/tasks/tetris/eval/gameplay-bot/grid-reader.ts
@@ -100,10 +100,42 @@ async function readCanvasGrid(
     { x: bounds.x, y: bounds.y, cellW, cellH, rows: GRID_ROWS, cols: GRID_COLS, bgR, bgG, bgB, threshold }
   );
 
+  // Validate: a freshly-read grid should make sense
+  if (grid) {
+    const totalCells = GRID_ROWS * GRID_COLS;
+    const filledCells = grid.reduce((sum, row) => sum + row.filter(Boolean).length, 0);
+    const filledPct = filledCells / totalCells;
+
+    // If >60% of the grid is "filled" at any point, we're probably reading
+    // off-grid (UI chrome, borders, decorations). A real Tetris grid rarely
+    // exceeds 50% filled even in a losing game.
+    if (filledPct > 0.60) {
+      return null;
+    }
+  }
+
   return grid;
 }
 
 /**
+ * Validate that calibrated grid bounds look like a real Tetris grid.
+ * Returns true if the bounds are plausible.
+ */
+export function validateGridBounds(bounds: GridBounds | null): boolean {
+  if (!bounds) return false;
+
+  // Aspect ratio should be roughly 1:2 (width:height) for a 10x20 grid
+  const ratio = bounds.height / bounds.width;
+  if (ratio < 1.3 || ratio > 2.8) return false;
+
+  // Grid should be a reasonable size (not tiny or the entire viewport)
+  if (bounds.width < 50 || bounds.height < 100) return false;
+  if (bounds.width > 1000 || bounds.height > 1500) return false;
+
+  return true;
+}
+
+/**
  * Read grid from DOM elements. Looks for a grid-like structure and checks
  * background colors or class names to determine filled vs empty cells.
  */
diff --git a/tasks/tetris/eval/gameplay-bot/tests.ts b/tasks/tetris/eval/gameplay-bot/tests.ts
@@ -746,12 +746,20 @@ function deriveTestResults(
       pass: true,
       detail: `filled cells persist at bottom (grid-verified, ${lockEvents.length} lock event(s))`,
     });
-  } else if (session.piecesLocked > 0) {
+  } else if (session.piecesLocked > 0 && session.piecesSpawned > 0) {
+    // Only trust locked count if we also detected spawns (prevents false positives
+    // from static UI being misread as game state)
     results.push({
       name: "piece_locks",
       pass: true,
       detail: `${session.piecesLocked} piece(s) locked during play`,
     });
+  } else if (session.piecesLocked > 0 && session.piecesSpawned === 0) {
+    results.push({
+      name: "piece_locks",
+      pass: false,
+      detail: `${session.piecesLocked} lock event(s) but 0 spawns detected - likely false positive from UI misread`,
+    });
   } else {
     results.push({
       name: "piece_locks",
@@ -776,7 +784,7 @@ function deriveTestResults(
   }
 
   // 11. multiple_pieces
-  if (session.piecesLocked >= 3) {
+  if (session.piecesLocked >= 3 && session.piecesSpawned > 0) {
     results.push({
       name: "multiple_pieces",
       pass: true,