v3.1 — 10: Two follow-ups. The exit clock, and a second seed

One attempted fix that failed in an instructive way, and a replication that kept the headline and dropped most of the details.

1. Reviving the exit clock — verdict: no

Doc 07 found the v3.1 dynamics model knows where the hidden ball is but not when it comes out: frames_hidden from h at R² −0.23, and a dream that never re-emerges the ball. The proposed fix was a counter head: a small linear head on h trained to regress two numbers — frames since the ball was last half-visible, and frames until it next is — with the targets derived from a frozen probe on z (visibility is readable from z at R² 0.995), so the head uses only information available from the model’s own input stream. A privileged twin used the simulator’s visibility column and a vy head, as a ceiling; a third variant added the 24-step emergence-weighted rollout.

The counter head made the dream deader: where the baseline at τ ≥ 0.5 at least re-emerges the ball (too often, parked at the band edge), the counter models never do. The privileged twin is no better, so the frozen probe is not to blame.

The mechanism is the instructive part. Splitting the head’s two targets:

Every model learned “how long has the ball been gone” and none learned “when does it come back”. Since is a running sum an LSTM can integrate for free; until requires predicting a hidden object’s future — precisely the capability the head was meant to induce. A regression target does not supply a capability; it measures its absence. And the head cost hidden-x memory (0.52 → 0.31) for a little hidden-y, the wrong coordinate.

No controller was trained in these dreams. The exit-clock problem stands, and it now has a sharper statement: the model cannot predict re-emergence time from h, and no auxiliary regression fixes that. The candidate that remains is architectural — attention back to the entry frames — which is the v4 question.

2. A second seed for every controller row

Every row of doc 08’s table was retrained with CMA-ES seed 1 and all rows were scored on the same 150 episodes.

My own fresh-seed re-check (90 episodes, seeds 9000+) gives fair τ=1 at 0.73 / 0.86 for the two seeds, real-trained 0.65 / 0.71, feed-forward 0.40 / 0.58, memoryless bound 0.53 — the same picture with the seed-1 outlier pulled in.

What survives two seeds. The fair τ=1 controller is above the memoryless bound, the privileged ceiling, the τ=0 controller and the z-only control on both seeds, with non-overlapping ranges; on long required moves it is above everything including the real-trained controller (0.66–0.72 vs 0.37–0.51) with the smallest spread of any row. The z-only control is at the bound on both seeds and last on displacement-while-blind on both. Those are the claims doc 08 rested on, and they hold.

What does not. Not one adjacent ordering in the table survives — every gap is smaller than the spread it separates. The feed-forward “floor” is not a floor: its second seed lands at 0.63, above the bound and above both τ=0 seeds. “Fair beats real-trained” holds only on long moves. And the τ=0 controller is indistinguishable from the z-only control overall.

3. Lessons

• An auxiliary target measures a capability; it does not create one. The counter head learned the half of its target that the recurrence gives for free and scored at chance on the half that needed prediction.
• One seed per row was never enough, and this table shows exactly how much was noise: seed spread of 0.1–0.25 against gaps of 0.02–0.08. The headline (memory beats the bound, and h is why) survived because it was measured against controls that differ by 0.2 or more; everything finer did not.
• “Floor” and “ceiling” are labels for a design, not guarantees. The feed-forward floor came out above the bound on one seed. With two seeds the honest floor is the z-only control, which sits on the bound both times.

Back to 09 — v3.1 results and lessons.