The quarter blend through the protocol arena

What was registered

The blend is ntuple-leaf's search (depth 3, 1.5 s budget, 8 dice scenarios, 4 sampled worlds, 4 inner scenarios) with leaf.hand 0.25 added on top of the same frozen tables, registered in the policy registry as ntuple-leaf-blend. The engine retest had read it at +0.066 wins per game against builders and +0.064 at a table of searches, at 1.4 times the decision time, and the current evidence asked for a protocol cohort before any default changes.

Two halves were registered under patterns/blend-protocol before either ran, each naming the other: half A seats the blend in slot 0 and ntuple-leaf in slot 1, half B swaps them; eta and fast fill slots 2 and 3 in both. Each half plays deterministic seeds 2000 to 2019 once per rotation of the four slots, 80 games, with the same points, match, decision, and turn caps as the cohort that made ntuple-leaf the baseline. The rule: all 160 games valid with no server timeout moves; support if the seating-corrected effect's interval lies above zero and the estimate reaches +0.03 wins per game; refutation if the interval lies below zero; otherwise inconclusive.

The two halves

Half A and half B each completed all 80 games with no timeout moves, twelve matches in flight, 72 and 76 turns per game on average. The two searches ran as separate remote processes against the server and saw only their own seated view and event history; both loaded the frozen tables (SHA-256 f8e61e66…) and reported the tables evaluator on every decision.

Half A, blend in slot 0:

Half B, blend in slot 1:

The seating-corrected effect

The blend won one more game than the tables alone in half A and two more in half B, and the two seatings agree that there is nothing large to find: the corrected estimate is +0.019 wins per game, below the +0.03 bar, with an interval that includes zero, so the registered rule reads inconclusive. It is not a refutation. Twenty boards played in both seatings resolve about ±0.11 wins per game, and the engine estimate of +0.066 sits comfortably inside the interval; this cohort could not separate an effect of that size from nothing. The seating term is invisible here (−0.006), unlike the engine arena's +0.05 to +0.18, which says nothing about the arena difference on twenty boards.

What the blend changes at the table is the same as in the engine: over both halves it built more cities per game (1.43 and 1.38 against 1.09 and 1.14), traded more with players (2.30 and 2.46 against 1.96 and 2.14), and finished with about half a point more (7.79 and 7.81 against 7.33 and 7.24), and those points did not become wins.

Decision time and completed depth

The remote runner's search seat now writes an optional per-decision trace (SETTLERS_SEARCH_TRACE, kept in each seat's private policy log), so this is the first protocol cohort that can say how much of the depth-3 search the 1.5 s clock actually delivered. The machine was shared with five other agents throughout (load average 60 to 80 on 32 cores), and both seats in every game shared that load.

Contested decisions are main-phase decisions with at least two root actions; a single candidate never deepens past the first depth. Over all decisions, including the trivial ones, the blend's median was 516 and 478 ms against 427 and 456 ms. The blend spent about 1.1 times the control's time per contested decision, less than the engine's 1.4 because the clock caps the difference, and paid for it in depth: it completed depth 3 in half of its contested decisions where the tables alone completed it in three fifths, and it was stopped by the budget in a quarter of them against a fifth. Both seats fell back to the first legal action on a search error in about 0.2% of decisions (22 and 19 for the blend, 19 and 27 for the tables alone, out of 9,000 to 10,000 decisions per seat per half), a pre-existing path of the search when it returns no candidate.

One game, viewed from above

The first match of the cohort, seed 2000 in its first rotation, which the blend's seat won in 49 turns with ten points against two, five, and eight. It is a real recorded game chosen for its position in the schedule; it does not by itself establish which policy is stronger.

Limitations

One lineup, one rotation scheme, twenty boards: a swapped pair on twenty seeds resolves about ±0.11 wins per game, so only a large effect could clear the +0.03 bar here, and the engine estimate of +0.066 sits inside what this cohort cannot separate from zero. The two searches are time-budgeted, and the cohort ran on a machine shared with five other agents, so both seats completed depth 3 less often than the unloaded 2026-09-09 cohort would have, and the blend lost about ten points of depth-3 share to its cost. The contrast is fair within a game, but it measures the two evaluators as the machine could run them, not at full depth. A first attempt at half A stopped on a seat-join conflict at game creation (abandoned attempt, 18 valid games, not pooled) and the harness's join step was made to retry on conflict before the pair was played; the daily log has the details. Nothing here speaks to negotiating opponents, humans, or the browser's hand-written leaf.

What the browser would need

The browser worker runs the hand-written leaf in WASM and cannot load the tables, so leaf.hand is not a browser switch. Using the blend there would need the 14.3 MB tables shipped to the worker, which the WASM build rejects by decision, or a server-side seat; nothing in web changes with this result.

Half A record · Half B record · Learned pattern values · Log 2026-09-25