Once the search is paid by the new energy model, does it rank its finds?

E193 — the search on the non-flat reward (2026-09-18 16:29; operator: icet retired from the reward)

FORAGER_RUNG0=ece: the search is paid from v5's formation energy over 8 random 54-atom decorations in the DFT frame (drive = E_form − T·S_mix against the bcc pure elements; sigma = decoration spread ⊕ 10 meV); the MACE hull is a veto that fires only beyond 100 meV (its own measured error); everything else — kinetics, the icet ordering estimate, confirm — is unchanged. In-process predictor, 0.1–0.5 s per composition (E187 scan showed the old reward flat at ≈ −11 for every find). Checked on three known cells through the real build: Mo₀.₅Ta₀.₅ −76.1 (DFT −89.8), Mo₂₄Ti₁₀W₂₀ −81.1 (E187 pending), Hf₁₄Zr₁₃Ti₈Nb₁₁Ta₈ +77.9 → p = 0 (DFT +89.7). Head: E171b's, MPS, 200 × 4 × 2, tag e193. Predictions. 1. The finds are ranked: the spread of e_form_ece across a seed's finds is > 20 meV/atom (under the flat reward every find sat within ±1 of −11). 2. The finds move: ≥ 80% of finds have E_form(v5) ≤ −60 meV/atom and < 10% carry Hf+Zr+Ti > 0.4 (the flat reward promoted such cells; v5 pays them nothing). 3. AUC_Q and find counts are not comparable to E171b/E189 — the verdict changed — and are recorded, not scored. 4. The flywheel check that matters: the best find by v5 goes to DFT (E194) and lands within 15 meV/atom of its v5 value, as the last six cells did.

E193 interim (17:10) — the search on the v5 reward ran; its finds dump did not. MPS, 2 seeds: AUC_Q 24.1 / 30.5 (27.27 ± 3.21), finds 61 / 77, best −98 meV/atom (v5 frame) — prediction 3 (not comparable to the flat-reward arms) recorded as such. Predictions 1–2 (ranking spread and composition shift) need the finds by composition; the winners_by_seed field in the tagged results file is empty for both seeds — a defect in the dump path, being traced; if the finds cannot be recovered from this run it is re-run (17 min per seed).

E193 result (18:00) — the reward ranks; the finds move. Re-run (finds dumped): AUC_Q 24.1 / 30.5, finds 61 / 77 (85 / 130 raw winners), best −98 (v5 frame). v5 energies of the finds (recomputed post hoc — the run's screen values were not carried through verify(); fixed):

• Prediction 1 confirmed: spread of E_form across a seed's finds 62.7 / 60.0 meV/atom (range −96 to −34) against ±1 under the flat reward. The search can now tell good from better.
• Prediction 2, split: Hf+Zr+Ti > 0.4 among finds 7% / 2% (bar < 10% ✓); finds with E_form ≤ −60: 68% / 79% against my bar of 80% — missed on both seeds by a small margin. The bar was mine, not the ladder's: the qualifying line is a −40 meV drive at 90 K, and finds between −34 and −60 are legitimate under it. Recorded as a miss of the stated bar.
• Prediction 3: not comparable to flat-reward arms — recorded, not scored.
• The four best finds by v5 (runs/e193_top_finds.json): Mo₀.₆₂Ti₀.₃₄Nb₀.₀₄ (−96.2 ± 5.8), Mo₀.₆₄Nb₀.₃₃ (−96.1), Mo₀.₆₅Ta₀.₂₂Ti₀.₁₃ (−96.0), Mo₀.₆₈Nb₀.₃₃ (−95.6): Mo-rich binaries and ternaries on the Mo–Ti / Mo–Nb / Mo–Ta edges — new territory against the Mo–Ta basin the flat reward circled.