Every search switch under the learned leaf

Study log/tables-retest. The ten studies launched on 2026-09-11 to retest the search's switches under the learned tables finished today; nine filed complete studies and the browser-budget study was stopped by the user before its registered stages, with its work kept on the research/browser-budget branches. analysis/tables_retest.py rebuilds every swapped pair of the nine studies from the registered experiments and the archived runs, orients each effect so that a positive value favours the candidate, and pools each switch per leaf and lineup after the fact: 145 pairs and 74,240 games. Every study's reported outcome was checked against that recomputation and matched.

What holds: bargaining (counters off −0.094 at the builder table and −0.119 at the table of searches; the whole package −0.165 and −0.098), and the search's information inputs (four dice scenarios −0.392, one sampled world −0.109, public-count beliefs −0.143, no proposals −0.172). What improves the player at both tables: the hand-written terms blended into the tables at a quarter of their weight, +0.066 and +0.064, at 1.4 times the decision time. Eight sampled worlds (+0.045) and the search-ranked opening (+0.061, carried by the reused development boards) read positive at the builder table only. The robber, card timing, hand, road, prediction, endgame, and planner-weight switches show no effect that survives both tables. The synthesis report has the full table and the browser recommendation: keep the current browser configuration, and use the quarter blend if the tables reach the browser, after a protocol cohort.

Verification: server dev passes formatting, clippy with warnings as errors, and all 183 workspace tests after every merge of the program. Failures during the program: two Claude Fable subagents stopped on usage limits and were finished by Claude Opus 5 agents from their saved state; one opencode session ended on a provider error and resumed; two agents slept through finished queues and were woken.

Attribution: the synthesis and verification by Claude Opus 5 (claude-opus-5) through Claude Code; the studies by GLM-5.3 and Kimi K3 through opencode and by Claude Fable 5.1 and Claude Opus 5 through Claude Code.