Run 0df1bae1-1928-4bcc-8fb6-a5519621bb76. 80 completed of 80 planned games. 95% Wilson intervals; descriptive evidence only.
The candidate is the corrected expectimax-v2 at depth 3 under a 1.5 s budget
with eight common random scenarios and four sampled worlds, frozen in the policy
registry as expectimax-v2-plan before the cohort. The lineup and rule are the
ones the pre-fix candidate had failed a day
earlier with 31 wins against 25. Twelve matches ran at once against one server;
the wall-clock budget makes contention search less, never longer, so any effect
of sharing the machine works against the candidate.
The mean game lasted 69 turns. The candidate averaged 8.40 points to the
builders' 6.5 to 7.1, and the two identical ETA policies finished within one
win of each other this time.
One game, viewed from above
The first scheduled game, selected by schedule rather than outcome.
eta slot 1’s turn
Turn 0 · Place a settlement · 10 points to win
Bank19Brick19Lumber19Wool19Grain19Ore25 development cards left
E
expectimax-v2-plan slot 0
0 VP
0 resource cards0 development cards0 settlements0 cities0 roads0 knights played
Longest route: 0
E
eta slot 1
Playing
0 VP
0 resource cards0 development cards0 settlements0 cities0 roads0 knights played
Longest route: 0
F
fast slot 2
0 VP
0 resource cards0 development cards0 settlements0 cities0 roads0 knights played
Longest route: 0
E
eta slot 3
0 VP
0 resource cards0 development cards0 settlements0 cities0 roads0 knights played
Longest route: 0
Actions & conversation
1 events · since the previous position
Table
The game started.
Game 1 in the scheduled cohort.
109 of 109 recorded positions. Board changes between positions are grouped.
Replay sources
Game aa659137-6e03-4c83-b126-5c63e561edbc · version 14
Public events: ea05520a0f84c7ec9e0a0591ce5ff4532b5d38f6119b1e4312202bd1d38e0fde
Agent notes
Experiment 49427921-feaf-4291-a5bf-79500acdcae7 under study
search/expectimax-v2, run with just run 49427921-feaf-4291-a5bf-79500acdcae7 --parallel 12. The exact test counts wins by the candidate slot and the
compared ETA slot only: P(X ≥ 47 | n = 58, p = 0.5) = 1.02 × 10⁻⁶. The
candidate's exact configuration is the doubled-brace JSON in
policies/registry.toml; the run manifest records the server remote binary
hash and both repositories' Git state (server commit 59433b7 carries the
planning fix). Public frame source SHA-256:
4b9783ed0a6f0eb46818295534984f46b92ee8ece278548c02926e5587a15204.
Attribution: Claude Fable 5.1 (claude-fable-5-1) through Claude Code.