Experiments · E198b
With the reward's gate removed, is the brain's own gain signal still the worst?
Yes. It scored 136.9 against 153.2 for a fixed random gain, so the signal is better switched off.
In the log: queued (2026-09-19 22:08)
book/assets/diagrams/exp/E198b.svg).Results
EXPERIMENTS.md · line 12852E198b queued (2026-09-19 22:08). E198's verdict — live gain field worse than a fixed random
one — was measured on the gated reward, which E201b shows distorted both scale and ranking.
Two control arms (shuffle_frozen, random) re-run on the pure reward after E205; the live
arm is E201b's fly (136.87 ± 2.14 AUC, 284 ± 22.6 finds). Prediction, on record: the
ordering survives (live ≤ frozen ≈ random within one seed-sd) — the gate did not create the
gain field's failure, it only changed the scale; if live now leads by more than one sd,
E198's verdict is withdrawn outright. runs/e198b_chain.sh.
E198b — the gain-field controls on the PURE reward. live (E201b fly) 136.9 / 284;
frozen shuffle 143.2 / 293.5; fixed random 153.2 / 305.5. Prediction ("the ordering
survives; the gate did not create the failure") confirmed: the head-derived live gain
field is the worst of the three on the clean reward too, by 16 AUC_Q against random. E198's
verdict stands and is now unconditional. The modulation channel carries nothing. For
the book: switch it off (FORAGER_MODULATE=random is strictly better than the head's own).
The full record
This entry is written in 2 separate places in the log, shown here in log order.
EXPERIMENTS.md · lines 12852–12858
E198b queued (2026-09-19 22:08). E198's verdict — live gain field worse than a fixed random
one — was measured on the gated reward, which E201b shows distorted both scale and ranking.
Two control arms (shuffle_frozen, random) re-run on the pure reward after E205; the live
arm is E201b's fly (136.87 ± 2.14 AUC, 284 ± 22.6 finds). Prediction, on record: the
ordering survives (live ≤ frozen ≈ random within one seed-sd) — the gate did not create the
gain field's failure, it only changed the scale; if live now leads by more than one sd,
E198's verdict is withdrawn outright. runs/e198b_chain.sh.
EXPERIMENTS.md · lines 12960–12965
E198b — the gain-field controls on the PURE reward. live (E201b fly) 136.9 / 284;
frozen shuffle 143.2 / 293.5; fixed random 153.2 / 305.5. Prediction ("the ordering
survives; the gate did not create the failure") confirmed: the head-derived live gain
field is the worst of the three on the clean reward too, by 16 AUC_Q against random. E198's
verdict stands and is now unconditional. The modulation channel carries nothing. For
the book: switch it off (FORAGER_MODULATE=random is strictly better than the head's own).
