Experiment 3 — Does a well-trained model adapt to the actual authority?
Hypothesis. Trained on 100,000 rounds, a model picks
the correct card for the authority it is actually playing, regardless of how it was trained:
both the 10-authority and 50-authority models buy Battle Pod when the game starts at 10
authority, and Trade Pod when it starts at 50.
Setup
- State: Turn 1, the active player's opening hand.
- Trade row (forced): Junkyard, Fleet HQ, Blob World, Battle Pod, Trade Pod.
- Decision probed: the opening purchase. With ~2 trade banked from the opening hand, the affordable buys are the 2-cost cards — Battle Pod and Trade Pod (plus the 2-cost Explorer).
- Readout: the policy's noise-free MCTS action-probabilities at that decision, averaged over qualifying opening hands.
Training depth. The hypothesis names
100,000 rounds. Running that literally takes many hours per
model in this pure-Python harness, so this page reports a reduced, proportional
depth of 500 self-play games (config: cheap)
that exercises the same effect. Reproduce at full depth with the documented flags — results
are decoupled from the site build, so re-running never requires a rebuild.
Results
| Model · playing at | Battle Pod | Trade Pod | Explorer | Picks | n |
|---|---|---|---|---|---|
| trained @ 10 authority, playing @ 10 | 50% | 3% | 22% | Battle Pod | 20 |
| trained @ 10 authority, playing @ 50 | 39% | 6% | 20% | Battle Pod | 20 |
| trained @ 50 authority, playing @ 10 | 43% | 10% | 20% | Battle Pod | 20 |
| trained @ 50 authority, playing @ 50 | 39% | 5% | 22% | Battle Pod | 20 |
Verdict
✗ does not match hypothesis. The well-trained models do not yet cleanly adapt to the authority in play (more training needed).