Experiments · E196
Partly. One matched within 2.7 meV/atom; the molybdenum–niobium one came out 28 meV/atom shallower than the energy model said.
In the log: prediction, written before pw.x (23:12)
book/assets/diagrams/exp/E196.svg).E196 — prediction, written before pw.x (23:12). E193's top four finds as 54-atom cells at the standard (60/720, 4×4×4, MV 0.02, local-TF 0.1), one decoration each (seed 0), vs refs_v3_fit.json. Find 1 = E194 (Mo33Ti19Nb2, on record −96.2). Finds 2 and 4 round to the same alloy (Mo35Nb18Ti1 / Mo36Nb18, L1 0.04 < 0.15), so find 4 stands for both. Cell-specific v5 predictions (the number DFT is held to, not the composition mean):
| cell | v5 (this decoration) | v5 random k=8 | E193 drive_cold |
|---|---|---|---|
| find3 Mo35Ta12Ti7 | −96.4 | −92.4 ± 7.0 | −76.1 |
| find4 Mo36Nb18 | −91.1 | −92.7 ± 7.2 | −90.2 |
Prediction: both DFT cells within 15 meV/atom of the cell-specific v5 value. A miss enters
v6 training (merge_dft_store.py); a miss > 30 meV on Mo36Nb18 says v5's Mo–Nb interior
is wrong, the one binary E193 keeps returning to. Queue: local, behind E191 (5 cells of 16
atoms) → E192 (8 × 54 atoms, ~3 h each) → E194 → E196; that is ~30 h of pw.x, so E196 does
not land tonight. Colab is not an option at the 45-min-per-job rule: a 54-atom standard
cell takes ~3 h on six cores and the L4 runtime has six. Builder:
scripts/validate/e196_cells.py; chain runs/e196_chain.sh; inputs runs/e196_dft/.
E196 result (2026-09-20 20:15) — the generator's finds at rung 4: two of three confirmed, and the miss is the one that was pre-registered.
| cell | DFT | v5 (this decoration) | diff | bar 15 |
|---|---|---|---|---|
| E194 Mo₃₃Ti₁₉Nb₂ | −88.1 | −96.2 | +8.1 | confirmed |
| find3 Mo₃₅Ta₁₂Ti₇ | −99.1 | −96.4 | −2.7 | confirmed |
| find4 Mo₃₆Nb₁₈ | −63.1 | −91.1 | +28.1 | falsified |
| Mo₃₅Ta₁₂Ti₇ at −99.1 meV/atom is the deepest first-principles formation energy the project has | ||||
| produced for an alloy nobody named before, and v5 called it to 3 meV. Mo₃₆Nb₁₈ is 28 meV | ||||
| shallower than v5 says — the pre-registered reading applies almost to the letter ("a miss |
30 on Mo₃₆Nb₁₈ says v5's Mo–Nb interior is wrong, the one binary E193 keeps returning to"). It is also where E201b's fly finds cluster (Mo-rich Mo–Nb), so the generator's favourite basin is one the model overstates by ~28 meV. Both rows stored; they are exactly the rows the rung-0 refit needs (this is where the flywheel does help — formation energies, not ordering).
This entry is written in 2 separate places in the log, shown here in log order.
E196 — prediction, written before pw.x (23:12). E193's top four finds as 54-atom cells at the standard (60/720, 4×4×4, MV 0.02, local-TF 0.1), one decoration each (seed 0), vs refs_v3_fit.json. Find 1 = E194 (Mo33Ti19Nb2, on record −96.2). Finds 2 and 4 round to the same alloy (Mo35Nb18Ti1 / Mo36Nb18, L1 0.04 < 0.15), so find 4 stands for both. Cell-specific v5 predictions (the number DFT is held to, not the composition mean):
| cell | v5 (this decoration) | v5 random k=8 | E193 drive_cold |
|---|---|---|---|
| find3 Mo35Ta12Ti7 | −96.4 | −92.4 ± 7.0 | −76.1 |
| find4 Mo36Nb18 | −91.1 | −92.7 ± 7.2 | −90.2 |
Prediction: both DFT cells within 15 meV/atom of the cell-specific v5 value. A miss enters
v6 training (merge_dft_store.py); a miss > 30 meV on Mo36Nb18 says v5's Mo–Nb interior
is wrong, the one binary E193 keeps returning to. Queue: local, behind E191 (5 cells of 16
atoms) → E192 (8 × 54 atoms, ~3 h each) → E194 → E196; that is ~30 h of pw.x, so E196 does
not land tonight. Colab is not an option at the 45-min-per-job rule: a 54-atom standard
cell takes ~3 h on six cores and the L4 runtime has six. Builder:
scripts/validate/e196_cells.py; chain runs/e196_chain.sh; inputs runs/e196_dft/.
E196 result (2026-09-20 20:15) — the generator's finds at rung 4: two of three confirmed, and the miss is the one that was pre-registered.
| cell | DFT | v5 (this decoration) | diff | bar 15 |
|---|---|---|---|---|
| E194 Mo₃₃Ti₁₉Nb₂ | −88.1 | −96.2 | +8.1 | confirmed |
| find3 Mo₃₅Ta₁₂Ti₇ | −99.1 | −96.4 | −2.7 | confirmed |
| find4 Mo₃₆Nb₁₈ | −63.1 | −91.1 | +28.1 | falsified |
| Mo₃₅Ta₁₂Ti₇ at −99.1 meV/atom is the deepest first-principles formation energy the project has | ||||
| produced for an alloy nobody named before, and v5 called it to 3 meV. Mo₃₆Nb₁₈ is 28 meV | ||||
| shallower than v5 says — the pre-registered reading applies almost to the letter ("a miss |
30 on Mo₃₆Nb₁₈ says v5's Mo–Nb interior is wrong, the one binary E193 keeps returning to"). It is also where E201b's fly finds cluster (Mo-rich Mo–Nb), so the generator's favourite basin is one the model overstates by ~28 meV. Both rows stored; they are exactly the rows the rung-0 refit needs (this is where the flywheel does help — formation energies, not ordering).
Built with PRISMWebsite and visualizations made using Claude