Belote AI: random vs random baseline

What I worked on

Before writing any competitive agent, I got BeloteGame/BeloteState far enough along in the OpenSpiel fork to deal cards, enforce bidding, and play out legal tricks to a scored terminal state. With that in place, I wrote the first version of the match evaluator and ran uniform-random players against each other over a batch of seeded deals, with no strategy at all, just legal random actions on both sides.

Challenges encountered

The random-vs-random matchup isn’t interesting on its own, but it’s the cheapest possible check that the harness itself isn’t lying: with no skill difference between the two sides, the reported points margin should be statistically indistinguishable from zero. Getting the paired-seed dealing right (both sides seeing the same deals so variance from the deck doesn’t swamp the signal) took a couple of passes. An early version reused a single RNG stream across both players’ decisions, which very subtly biased whoever acted first on tie-breaks.

What I learned

A measurement tool that hasn’t been shown to report “no difference” when there genuinely is no difference can’t be trusted to report a real difference later. Random-vs-random isn’t a throwaway smoke test; it’s the zero-baseline the whole evaluation harness gets calibrated against before any agent claim is allowed to rest on it.

Next steps

• Add RuleValidBot (fixed bidding/play heuristics) and evaluate it against the random baseline.
• Add the 95% confidence interval and OpenSkill rating output to the match evaluator so later agent-vs-agent comparisons aren’t eyeballed off raw win rate.