How did the policy at one information set evolve — and do three independently seeded runs agree on it? Nine real artifacts: three training milestones × three seeds, no invented intermediate history.
LOW cross-seed disagreement — three independent runs land on nearly the same mix · seed-disagreement signal 0.0026 · entropy 0.613 bits
fcb33b75b150a
milestone
seed 0
seed 1
seed 2
300k
2,196 visits
1,483 visits
2,376 visits
1M
16,366 visits
15,420 visits
17,058 visits
2M
45,361 visits
43,473 visits
44,702 visits
Why this matters
Cross-seed disagreement is one of the signals the confidence gate is calibrated on. The project’s calibration study (8,880 infosets with exact EV-regret targets) found that movement between checkpoints predicts true regret at no horizon, while visits and cross-seed disagreement are the real signals — which is exactly the difference between the “low” and “high” rows you can flip through above.