Development-card timing under the learned leaf

Design

Each arm is one switch on the tables baseline seat (v2:{"depth":2,"leaf":{"tables":...}}) measured as a swapped pair: half A seats the candidate in slot 0 and the control in slot 1, half B swaps them, both on the same 64 deterministic boards played once per rotation of the four slots (512 games per pair). The effect is the per-seed half-difference of the two slot-0-minus-slot-1 contrasts and the seating term their half-sum, which cancels the arena's turn-order term. Stages, in order: development on boards 0 to 63 with an ETA and a fast builder beside the searches (the lineup of the hand-leaf cohorts, so the leaves compare on the same boards), confirmation on fresh boards 800 to 863 in the same lineup, and population on boards 800 to 863 with a third tables search in slot 2 and an ETA in slot 3, so purchases and knight plays are answered by opponents that trade and block.

buy_tempo and hold_knight are candidate rules and act under the tables alone. knight_play_value shapes a hand-written leaf term that the tables alone never evaluate, so its control and candidate both carry the hand-written terms on top of the tables. A blend screen run in parallel measured the full blend (hand weight 1.0) at −0.207 wins per game against the tables alone and the quarter blend at +0.074, so the arm's primary evidence is on the quarter blend; the development-stage pair on the full blend was already running when the screen landed and is reported as what it is. One defect had to be fixed first: the blended leaf dropped the knight_play_value flag, which would have made the switch inert under any blend; the blend now evaluates the hand terms with the flag, pinned by a test in the server's development suite.

The pairs

Every half completed 256 of 256 games. The contrast columns give each half's slot-0-minus-slot-1 wins per game; the effect column is the seating-corrected combination with its 95% interval over the 64 boards. Negative effects mean the switch loses.

Half B's contrast is slot 0 (control) minus slot 1 (candidate), so a real effect makes the two halves point in opposite directions. Both buy_tempo halves point the same way on the development and confirmation boards: the candidate loses more in slot 1 than the seating term explains, and the corrected sign is negative in all three stages. The seating terms themselves (+0.08 to +0.14 on L1, near zero on L2 with a third search seated between) match the earlier null pairs on these boards.

What the switches did

Per-game means for the candidate against the control across each pair. The mechanisms all fire; the first two simply fire against the tables' own timing.

Under the hand-written leaf the tempo rule halved purchases (1.54 against 2.55 per game) and advanced the first city by a turn without winning. Under the tables it cuts only a fifth of the purchases (about 3.2 against 3.9 to 4.2), still builds the extra cities (1.4 against 1.2), and now loses. The tables already spend on cards when the board justifies it, so a rigid "no build within a turn" gate mostly removes good buys. Holding knights forgoes about 0.2 plays and a fifth of the army awards per game under the tables, half the bite it had under the hand leaf, and loses more at the population table, where a rival search punishes the held robber. The knight schedule changes almost no count on the quarter blend: the tables' own knight valuation leaves the shaped schedule nothing to correct, which is why its intervals are the tightest of the three arms. On the full blend the same schedule is worth +0.109, which reads as the hand-written flat knight weight being wrong at full strength, not as a gain available to any configuration worth running.

Decision cost moves only where the candidate list changes: buy_tempo prunes the purchase branch and decides about 30% faster than the control, while hold_knight and knight_play_value cost the same as their controls. The blends cost about 900 to 1000 ms per decision at depth 2 under load, against 450 to 750 ms for the tables alone and about 200 to 250 ms for the hand-written leaf, because the hand terms are evaluated at every leaf on top of the tables.

Comparison with the hand-leaf results

Under the hand-written leaf (the earlier study, same development boards) the single-seating contrasts were +0.043 (buy_tempo, boards 0-63) and +0.098 and +0.066 (hold_knight), all inside the seating term, and the knight_play_value swapped pair corrected to +0.004 (−0.061 to +0.069). The sign pattern flips under the tables: what was harmless under a leaf that under-bought cards is harmful under a leaf that buys them well. The tables learned card timing from self-play, and the hand-written timing rules, tuned by intuition, now fight the leaf instead of patching it.

What could still be wrong

All evidence here is engine-arena development tier at depth 2 on deterministic boards, two lineups, and one frozen table file; nothing is protocol-tier. The population lineup still seats an ETA builder, so "opponents that trade and block" means one more search, not three humans. The buy_tempo estimate mixes a refuted development stage with two inconclusive ones; the pooled interval excludes zero but pooling across stages is after the fact. The quarter blend itself is a screen reading (+0.074, interval +0.011 to +0.138, unregistered) chosen by a parallel study; if a different blend becomes the default, the knight_play_value reading scales with the hand weight and should be re-read at that weight. The full-blend pair is one stage on one lineup, so +0.109 is the thinnest reading here. Decision times come from a machine shared at high load and compare candidate with control within each run, which cancels the load but not the noise.

What this means for the browser

The browser build runs the hand-written leaf at depth 3 under a one-second budget, and none of the three switches earns a place there. The bar is a pooled effect of at least +0.03 wins per game with the interval excluding zero and decision time within 25% of the control's; no arm comes close on either leaf. buy_tempo was zero under the hand leaf and loses under the tables (−0.043 pooled). It does decide about 30% faster than its control in both leaves, but a cheaper search is only useful when the strength holds: here the strength is absent under the hand leaf and negative under the tables, so the saving buys nothing. hold_knight loses under the tables (−0.035 pooled) and showed nothing under the hand leaf. knight_play_value is zero under the hand leaf and under the quarter blend; its only positive reading sits on the full blend, a configuration too weak to ship. All three switches stay off by default, in the engine and in the browser.