Choose research opponents

Start with ntuple-leaf, the current protocol baseline. It is the expectimax-v2-plan search with learned position values. The client guide takes your own process from its first move to a registered simulation; this page helps choose the other seats.

Current searches

Policy	What changes	Preparation
`ntuple-leaf`	Current protocol baseline: learned leaf, depth 3, 1.5 s search budget	`just setup` and `just tables`
`expectimax-v2-plan`	Previous protocol baseline: same frozen search with a hand-written leaf	`just setup`
`expectimax-v2`	Tunable search at its default depth-2 configuration	`just setup`
`ntuple-leaf-deep`	Depth cap 4 and 4 s budget; did not improve on `ntuple-leaf` in its protocol cohort	`just setup` and `just tables`
`ntuple-leaf-tt`	Adds a cache retained across a turn; protocol strength comparison was inconclusive	`just setup` and `just tables`

The evidence boundary records what these comparisons established. The deeper and cached variants are available for investigation; they have not replaced the baseline.

just match ntuple-leaf expectimax-v2-plan ntuple-leaf expectimax-v2-plan --games 4

Builders and historical references

Policy	Use	Preparation
`fast`	Simple Rust builder with a trading wrapper	`just setup`
`eta`	Builder using estimated resource acquisition time	`just setup`
`legal-first`	Minimal Python adapter for checking legality and client plumbing	Included
`expectimax-v1`	Frozen reference search	`just setup`
`expectimax-v2-deep`	Historical depth-3, 1.5 s candidate with more scenarios and sampled worlds	`just setup`
`liquidity`	Exchange-aware estimator from the first study; did not earn promotion	`just build-liquidity`
`eta-control`, `fast-control`	Frozen comparators sharing the liquidity study's runner	`just build-liquidity`

Use builders when reproducing earlier lineups or studying behavior against simpler opponents:

just match ntuple-leaf expectimax-v2-plan eta fast --games 4

just policies prints the complete registry. Commands run from the research checkout; each seat gets a separate process and credential, even when a policy name appears twice.

Read the result

Each row represents a competitor slot, not every instance of that policy combined. Four games rotate the lineup through the seats. They check the setup; a strength study also needs frozen budgets, fresh seeds, and both relative orders of the candidate and reference.

The result keeps completed, incomplete, and unplayed games separate. A failed match stops new work, and in-flight parallel games finish before the result is finalized. Read runs/RUN_UUID/summary.json and the individual result.json files before registering another run.

For a recorded example, open the learned-leaf game. To watch your own game live, a web client must connect to the same server and open /game/GAME_ID. That optional viewer is separate from running experiments with only the two repositories.

Continue with the client and simulation guide or publish a selected replay.