With the reward's distorting gate removed, does the fly beat keep-the-best-and-mutate?

EXPERIMENTS.md · line 12819

E201b, elitist (2026-09-19 20:36), pure reward. AUC_Q 145.26 ± 10.95 (s.e.), 246 ± 3 finds, best −105, last-window 194 — against 88.90 / 169.5 / 147 under the gated reward (E195b). The icet gate had been removing about a third of the qualifiers (p capped at 0.5 whenever its single-decoration estimate returned 'unknown', plus its penalty). Every pre-fix count is therefore deflated by a history-dependent amount; only within-run comparisons survive from before 247ed9d. Fly arm running.

EXPERIMENTS.md · lines 12821–12850

E201b — the pure-reward baseline, and a reversal of last night's headline (2026-09-19 22:07)

fly − elitist: finds +38.0 against one seed-sd of 22.6 → the fly is ahead. AUC_Q −8.4 against one seed-sd of 15.5 → a tie. On the reward as it should always have been paid, the whole-brain fly finds more distinct qualifying alloys than keep-the-best-and-mutate and matches its rate. This reverses E195b's headline ("ties on finds, loses on AUC_Q") and with it the reading I reported — that the generator claim fails on a reward that ranks. It fails on the gated reward; on the pure one it does not. What survives from E195b: the fly's finds are shallower than elitist's (−66.8 vs −73.7, 7 meV), and its seed spread on finds is five times elitist's. What is withdrawn: "elitist ≥ fly on every column" and the E195b/E198 conclusion that the head's contribution on a real reward is nil — E198 must be re-run on the pure reward before its gain-field verdict can stand (queued as E198b). The two arms also differ from E195b in scale (246 and 284 vs 169 and 158): the icet gate was suppressing about a third of all qualifiers.

Bars restated for E202–E205 against this baseline, before E202's result lands. The late-stall premise is gone: the fly's windows rise here (154→196, 167→189), so E202's "last-window ≥ 139" is met before it starts and is void as written. New bars:

• E202 (local search): finds ≥ 307 (one seed-sd above 284) or mean depth ≤ −72 (elitist's level, from −66.8). Either is a pass; neither is a fail.
• E203 (imagine 4): mean depth ≤ −72, finds within one sd of 284. (ρ = 0.64 ≥ 0.2, so the arm is not void.)
• E204 (bennett): ρ ≥ 0.74 (up 0.10 from 0.64) and AUC_Q ≥ 134.7.
• E205 (all three, 3 seeds, own elitist): finds ≥ elitist's and AUC_Q within one seed-sd and depth within 5 meV of elitist's — the fly competitive on every column.