weight-reading-probe / study2 /FOOTPRINT_CHECK.md
SoulFireMage's picture
Study 2: pre-reading footprint check + training gate (before blinding)
31250b6 verified
|
Raw History Blame Contribute Delete
2.08 kB

Study 2: footprint check and competence gate (2026-11-05, before blinding)

Competence gate: 16/16 pass

Game Trained mean return per seed (s0–s3) Random play Welch t
Catch 49.98, 49.98, 19.64, 10.15 ≈ −30 104–321
Dodge 5.00, 4.99, 4.99, 4.99 ≈ −6 78–81
Squash 24.37, 23.73, 23.88, 24.08 ≈ −24 104–108
Avoid 4.97, 4.98, 4.96, 4.97 ≈ −0.7 40–42

All networks were trained with the unchanged study-1 trainer and settings (20M steps, the same PPO hyperparameters). Catch seeds 2 and 3 reached weaker policies than seeds 0 and 1, but both still clear the gate by a wide margin.

§4 footprint check (aggregate per game, mean over 4 seeds)

"Active" means the cell's weight satisfies |w| > 0.06 in at least one fc1 unit. Values are active cells per row, out of 10.

Channel Catch Dodge Squash Avoid
agent row 9 only (8.5) row 9 only (9.2) row 9 only (9.8) row 9 only (9.2)
moving rows 0–9, 8.0–9.8 per row rows 0–9, 6.2–9.0 rows 0–9, 4.8–9.5 rows 0–9, 8.0–9.5
targets none none none none
obstacles none none none none
prev_frame rows 0–9, 7.8–9.8 rows 0–9, 6.5–8.8 rows 0–9, 5.8–9.5 rows 0–9, 7.5–10.0

Verdict: the set of active channels and rows is identical across all four games, as designed. The study-1 shortcut, where a game is identified by which cells are ever occupied, is gone.

What does still differ, stated before reading: the total |w| mass per channel, averaged over seeds.

agent moving prev_frame
Catch 86.7 157.1 159.1
Dodge 79.6 131.7 134.0
Squash 91.9 118.8 128.0
Avoid 81.4 135.4 154.7

Squash's moving weights are also thinner on row 0 (4.8 active cells against 8.0–8.8 for the others). Differences in magnitude like these could still let the B1/B2 baselines do better than chance. That is exactly what H3-2 tests, and nothing is changed in response.