Position Hypermarket

What this data is, and is not

• A row is a position, not a game. Each row is one unique (position, hero colour) pair for one player — the same position reached by two players is two rows, and the same position reached with colours reversed is two rows. Occurrences count every arrival at an eligible ply — middlegame, non-terminal, and carrying a stored eval — including repetitions inside a single game, so a threefold-repetition draw contributes three only if all three arrivals qualify.
• The eval belongs to one occurrence. eval_cp is the stored eval of the representative occurrence — the earliest game, by (provider, game id, ply) — and the same position in another game may carry a different stored eval. Those evals are a depth mix, mostly 100k-node, so ±0.5 between two hits is within noise.
• Coverage is the koga-evaluated games only, not the archive. The live count is therefore a lower bound on how often the player reaches these positions.
• attack_score is queen-gated to exactly zero. A side with no queen scores 0.0 however fierce the attack, so an empty grid from a score filter is indistinguishable from “no such positions exist”. kz_pressure is not gated: it is the score to reach for in the queenless regime.
• attack_score is half-blind. It detects attack in progress and is structurally blind to clearly building. Sorting by it puts the in-progress regime on top.
• Gap bands are bands, not thresholds — and a band is not a cure. The top1−top2 gap disagrees with an 8M-node reference on 8–12% of positions at every budget tested. A band has two boundaries, so hits within measurement noise of either one are marked uncertain on the card. Four moves within half a pawn is not “four equally good moves”. Noise grows with N: the measured study covers top1−top2 only, so gap14 and gap18 are extrapolations and the ordering within candidates 5–8 is approximate. The gaps come from a separate MultiPV pass over the representative occurrence, like the eval and the category — and a position whose candidate list mixes mate and centipawn scores has no meaningful gap at all, so it is excluded from every gap band rather than reported as a near-tie.
• Blunder categories and the tactic label belong to one occurrence. Like the eval, category describes the move actually played at the representative occurrence, and the tactic label classifies that position’s stored best move — a single 100k-node line, so composite tactics are under-detected. A position seen nine times was blundered, or not, eight other times this row does not record. Rows the labeler did not flag carry category 0 and are never hidden by these filters.
• Clocks and opponent ratings land with the next bank rebuild — those fields read — for now.