Strategy — the gate on real student games

The same gate as the calibration page, applied to real student games instead of curated examples: 8 moves within 100cp, eval within ±200cp, material within 1 pawn, no opposite castling. This is what the funnel costs in practice — of 50 seeded-random middlegame positions survive it. Engine: koga /candidates, top-10 at 1000000 nodes, mover's perspective.

A middlegame position is strategic if the best move is at most 100 cp better than the 8th best move, and none of the exclusions below apply.

Passes the gate:

The whole landscape

Tight-position count out of 50 for every (x, y) combination — darker = more positions qualify. Click a cell to set the sliders.

Strategic — passes the gate
top-y candidate moves (the "cloud")
move played — within top-y
move played — outside top-y
✗ opposite-side castling (poor positional example)
⚡ tactic / hanging piece (pawns counted) — not a positional example
Ladder too wide

50 quiet middlegame positions sampled seeded-random (seed 20260707, max 1 per game) from real student games — bands 1400 / 1600 / 2000 of the tactic-xC corpus, Gate-0 eligibility (middlegame phase, no time trouble, no forced mate, no mate lines in the top-10). Evals are mover-POV, dumped 2026-07-07.

Thresholds and exclusions are imported from src/lib/positional/strategyGate.ts, the module the live Strategy tile and the peer baseline both call (metric revision 3), so this page cannot drift from the shipped gate.

Two honest limits. The tactic exclusion is applied by the live metric but is not re-derived here — this dump carries no koga classification, and inventing one would misstate the funnel; the counts below are therefore an upper bound on what the full gate admits. And these ladders were measured at 1000000 nodes / 10 PVs while the shipped gate runs at 100k / 8. Rendered by scripts/tacticXc/tightLadderRenderPage.ts.