The quarter blend through the protocol arena

The engine retest left one candidate for the protocol tier: ntuple-leaf's search with the hand-written terms blended into the learned tables at a quarter of their weight, +0.066 and +0.064 wins per game at the two engine lineups for 1.4 times the decision time. It is now registered as ntuple-leaf-blend (the ntuple-leaf command with "hand":0.25 inside leaf) and was played against ntuple-leaf through the authoritative server as a swapped pair on deterministic seeds 2000 to 2019: half A with the blend in slot 0, half B with the blend in slot 1, eta and fast in the other seats, four rotations per seed, 80 games each, the same caps as the cohort that made ntuple-leaf the baseline, twelve matches in flight. A four-game smoke checked the plumbing first: every game valid, both search seats reporting the tables evaluator on every decision.

Half A: 34 blend wins against 33 for ntuple-leaf, contrast +0.013 (−0.156 to +0.181). Half B: 34 against 32, contrast −0.025 (−0.195 to +0.145) for ntuple-leaf minus the blend. Both halves completed all 80 games with no timeout moves. The seating-corrected effect of the blend is +0.019 wins per game (95% interval −0.087 to +0.124) and the seating term −0.006 (−0.139 to +0.126). By the registered rule this is inconclusive: the interval includes zero and the estimate is below +0.03, but the interval also covers the engine estimate, so twenty boards in both seatings could not separate an effect of the engine's size from nothing. ntuple-leaf stays the protocol reference and the server default is unchanged. The report has both Wilson tables, the play statistics, and the decision traces.

Under the 1.5 s clock on a machine shared with five other agents (load average 60 to 80 on 32 cores), the blend completed depth 3 in 49.9% and 51.8% of contested main-phase decisions against 61.2% and 60.9% for the tables alone, and was stopped by the budget in a quarter of them against a fifth, at about 1.1 times the control's time per contested decision (932 and 911 ms against 827 and 834 ms). Its cost shows up as lost depth rather than as the engine's 1.4 times wall time. As in the engine, the blend built more cities and finished with about half a point more per game; the points did not become wins.

The first attempt at half A (registered protocol, abandoned attempt) stopped at its nineteenth game: the harness's seat join at game creation was refused with a version conflict before any policy process started, so the game is invalid and the schedule stopped as registered. Eighteen games had completed validly (blend 6 wins, ntuple-leaf 4, eta 3, fast 5); they are recorded and not pooled. The server checks a command's version before it recognizes a duplicate request, so a join envelope retried after a lost response under load fails with a conflict although the seat was taken; harness/arena.py now refreshes the snapshot, keeps a seat it already holds, and retries a join on conflict, as the ready and start steps already did. Half A was registered again with the same text and half B ran as first registered. Both searches also logged a few continuation could not reach the next-turn horizon fallbacks, a pre-existing search path that plays the first legal action when the search returns no candidate: 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.

To read the searches' cost through the protocol, the remote runner's expectimax seat gained an opt-in trace (SETTLERS_SEARCH_TRACE, one JSON line per decision on stderr with the evaluator, the completed depth, and the search and wall times; server crates/policies/src/expectimax.rs, off by default and never changing a decision). The harness keeps it in each seat's private policy log; analysis/blend_protocol.py aggregates it together with the two halves' per-seed contrasts and the public tapes. Study patterns/blend-protocol is complete; just check passes, and just site-check could not run because the web checkout lacks the remark-math package, on the untouched canonical checkout as well. Resolving an effect of the engine's size through the protocol would take a pair of about 64-board halves (512 games, roughly ±0.06), preferably on an unloaded machine so the blend keeps its depth; the cheaper question is whether a subset of the hand-written terms buys the engine gain without the depth cost.

Attribution: Claude Fable 5.1 (claude-fable-5-1) through Claude Code.