Liquidity statistics and reproduction

The frozen evaluation ran on 2026-09-08 in America/Los_Angeles (2026-09-09 UTC). The process used a dedicated local server on port 5566, with its own local data. The server and policies were release Rust binaries on macOS. All 80 games completed validly; no timeouts, incomplete games or unplayed matches entered the denominator.

Slot / policy	Wins	Losses	Completed	Win rate	95% Wilson interval	Mean points
0 / liquidity	20	60	80	25.0%	16.8%–35.5%	7.44
1 / eta-control	26	54	80	32.5%	23.2%–43.4%	7.75
2 / fast-control	13	67	80	16.2%	9.7%–25.8%	7.29
3 / fast-control	21	59	80	26.2%	17.9%–36.8%	7.50

Primary contrast

Each game contributes +1 when Liquidity wins, −1 when ETA wins, or 0 when either Fast slot wins. The average is −0.075. This preserves the game as the unit of analysis; the two Fast slots are not pooled into 160 independent observations.

Among the 46 games won by either primary slot, Liquidity won 20. The exact one-sided binomial tail under equal win probabilities is 0.8490021934. Promotion required p ≤ 0.05 and an observed gain of at least 0.10. Neither condition passed.

An exploratory, non-preregistered percentile bootstrap resampled whole games 20,000 times, giving [−0.2375, 0.0875] for the win-rate difference. Resampling seed 20260908 belongs only to analysis; it never entered a game. The source code retains the exact percentile indexing. This interval does not establish equivalence or prove the candidate is inferior.

Descriptive mechanism checks

Metric per completed game	Liquidity	ETA control
Bank/port exchanges	6.3375	5.8625
Discarded cards	10.1625	8.7125
Cities built	2.6500	2.4875
Roads built, including setup	6.2125	6.6375

Mean final-point difference: −0.3125. Exact score comparisons: 35 ahead, 39 behind, 6 tied. The heatmap bins final scores as 2–4, 5–6, 7–8, 9 and 10+. Its cells sum to 80 games. The exchange plot is an empirical CDF at integer trade-count thresholds; line segments connect observed thresholds rather than estimating a continuous law. These are descriptive secondary measures, without multiple significance tests.

Reproduce the analysis

From the research checkout, with the verified raw run locally available:

python3 -m analysis.liquidity de39917c-55da-40ad-a53a-5b8588898810
just check

The script validates the finalized inventory before reading result rows and public rule events. It writes ignored per-game rows and analysis output under .runtime/analysis/, plus the two curated diagnostic plots. It refuses partial cohorts. It never reads player observations, credentials or private process logs. The win chart and replay originate from just publish; their presentation labels were edited without changing the recorded values.

To build this policy for a new registered experiment:

just build-liquidity

Use the liquidity, eta-control and fast-control registry entries. A rerun has fresh server randomness and must not be described as a reproduction of the same random stream. No seed is exposed. Seat rotation gives each slot 20 games in each seat, without pairing seeds or fixing the starting player.

Evidence identity

Experiment: d9cbd52e-0a90-4c4b-8fae-23c0372f1d6e.
Run record.
Inventory SHA-256: 063a024a2f341520bdf1188904b7ddc66a23836087fb7a14abef06c976ab6046.
Analysis source SHA-256: f22fc5d15a3c27cd6e0de0e4c17ff51506159974bbfca5775061f25a95d47898.
Shared policy executable SHA-256: 0aaf22a660fa70fbb2944a5a61e5560cd3d14b0b427000d31641effbf124d9a3.
Server executable SHA-256: 5b6ade55b0a6bde720f4238d807eaa0c9f3b93c4d65e2cf969bb7c3588ede29f.

The manifest retains the experiment, source snapshots, Cargo.lock, configured executable hashes and Git states. The executable is identical to the final successful smoke; the frozen policy sources were checked unchanged after the run. The policies were built from dirty working trees, not a named release.

The verified bundle is artifacts/liquidity-study/de39917c-55da-40ad-a53a-5b8588898810.tar.gz. Its receipt is records/artifacts/de39917c-55da-40ad-a53a-5b8588898810.md. Archive SHA-256: 6535a03b78dcb8bac31e0100ce872a8c27b4ee59ae6902397e0bbb103eb99a54. Local integrity does not imply a remote backup.

Main report · All excluded runs.