On the real reward, does the brain's own gain signal beat a fixed random one?

E198 — what the live head field carries, on the v5 reward (2026-09-18 23:19; queued)

E185/E189 showed it on the island reward: live gain field 39.9/40.4, frozen shuffle 41.1, fixed random 25.5 AUC_Q. Does the gap survive a real reward? Eight-element space (ce_8element, nothing refused), FORAGER_RUNG0=ece, MPS, 200 rounds × 4 × 2 seeds. Live arm = E195b's fly (every setting identical, same seeds; MPS runs reproduce bit-identically), so only the two controls run here: FORAGER_MODULATE=shuffle_frozen and random. Predictions, on record before either lands: A. live − fixed-random ≥ 15 AUC_Q (the E185 gap persists). B. frozen shuffle within one seed-sd of live (the field's heterogeneity is what matters, not its wiring-specific pattern — E189). If A fails on the real reward, the E185 gap was a property of the island reward's smoothness and the head's contribution on the real ladder is nil; that goes in the book as such. Chain runs/e198_chain.sh; logs runs/e198_shuffle_frozen.log, runs/e198_random.log.

E198, frozen-shuffle arm (2026-09-19 01:57). fly with FORAGER_MODULATE=shuffle_frozen: AUC_Q 91.90 ± 18.76, 188 ± 36 finds, best −101 — above the live arm (E195b fly: 73.54 ± 16.25, 158 ± 17). Prediction B (frozen within one seed-sd of live) holds as written, with the sign reversed: whatever the gain field contributes, its live updates are not it. Random arm running; prediction A (live − random ≥ 15 AUC_Q) decides whether a field is needed at all.

E198 result (2026-09-19 02:42) — N7 closed. Three gain fields, same head, eight-element space, v5 reward, 2 seeds:

Prediction A falsified with the sign reversed: live − random = −16.6 AUC_Q against a bar of ≥ +15 (E185's gap on the island reward was +14.4 the other way). Prediction B holds (frozen within one sd of live) but in the direction that matters least. The consequence pre-registered with the prediction applies: the E185 gap was a property of the island reward's smoothness, and the head's contribution on the real ladder is nil — worse than nil at two seeds: the live field is the worst of the three on every column. Combined with E195b (fly ties elitist on finds, loses on AUC_Q), the picture on a reward that ranks is: the whole-brain propagation supplies a proposal distribution that a random gain field drives as well as the head's own, and a mutate-the-best baseline drives better. What the head demonstrably still does: beat uniform and sparse by 10–20×, i.e. it is a competent proposal generator, not a learner of this reward. E195c's third seeds sharpen the counts; nothing queued can rescue A. For the book: this is the result.