Experiments · E243

Can the ordering rung read a gradual ordering from how much of the change falls inside the window?

Withdrawn. A review found it would fail real passes at both window edges and misread frozen samples; it was replaced before any find was scored.

In the log: rung 1 reads the change inside the window, not only a peak (pre-registered 2026-09-24 02:13)

withdrawnDate 2026-09-24 02:13, as written in the logrung 2 · hull, MACE5 predictions · 1 result paragraphEXPERIMENTS.md lines 16437–16487, lines 16489–16493, lines 16495–16509

Pre-registration

  1. (1)
    The survivor Mo₅₁Ti₃₈W₃Ta₃ fails: f > 0.5. E240's own record already shows 23 meV/atom released between the random solution and 60 K, and nothing of that size above 1340 K on the diagnostic.
    no verdict written against it
  2. (2)
    At least 5 of the 11 fail or come out not established; at most 3 pass.
    no verdict written against it
  3. (3)
    Every find whose E240 rung-2 drive was positive (Mo₇₉Hf₁₉, Mo₇₀Hf₂₀Ti₁₀) is decided by rung 2 anyway, and is reported but not counted in
    no verdict written against it
  4. (4)
    1149 ± 114 K): f ≤ 0.35, most of its ordering above 1000 K. A known answer that fails voids the rule as written. Predictions for the finds (E240's 11 that passed rungs 0–2, re-run one at a time, about 50 min each on one core, overnight)
    no verdict written against it
  5. (60)
    is under 3 meV/atom there is nothing to order; f is noise and the reading is a pass. This passes both an alloy that is already ordered above 1000 K and stays so, and one that orders only below 90 K, as the requirement's own wording does (chapter 1). It fails the gradual in-window ordering the peak reader missed. Known answers first; the rule is not applied to a find until all of K1–K3 hold. K1 and K2 use the exact bcc nearest-neighbour Ising through the same code path (J = 19.12 meV, exact T_c ≈ 1410 K, as E215), and variants with J scaled: - K1. T_c ≈ 1410 K (ordered through the window): f < 0.25, pass. - K2a. T_c ≈ 500 K (inside): f > 0.5, fail. - K2b. T_c ≈ 50 K (below): f < 0.25, pass. - K3. Equiatomic Mo–Ta on e6 (E242
    no verdict written against it
The pre-registration, as written

E243 — rung 1 reads the change inside the window, not only a peak (pre-registered 2026-09-24 02:13)

Why. E242 found rung 1's blind spot. On the survivor the sampler ordered by 23 meV/atom on cooling (−83.7 → −106.5), gradually and with no peak on the 227 K grid. "No peak above 60 K" was scored as a pass. The requirement is that nothing happens between 90 and 1000 K; a peak is only one way something happens. At the time of writing, the survivor's fine-grid diagnostic (runs/e242_rung1_curve.log) had reached 1340 K and no point inside the window existed. The rule below is fixed before any.

The rule (new, alongside the existing peak reading, never replacing it):

  • Run. One cooling leg, same Hamiltonian (e6 ensemble mean), same 432-site cell, same 100 + 200 sweeps. Grid: 2600, 2000, 1400 K, then 60 K steps from 1340 to 60 K. E(T) is interpolated linearly to 1000 and 90 K.
  • The in-window share of the ordering. f = [E(1000 K) − E(90 K)] / [E(2600 K) − E(60 K)], the fraction of the energy the sampler releases on cooling that is released inside the window.
  • Pass if f < 0.25. Fail if f > 0.5. Not established between. Fail also on any dE/dT or fluctuation peak inside 90–1000 K on this grid.
  • Guard: if the total release E(2600) − E(60) is under 3 meV/atom there is nothing to order; f is noise and the reading is a pass.

This passes both an alloy that is already ordered above 1000 K and stays so, and one that orders only below 90 K, as the requirement's own wording does (chapter 1). It fails the gradual in-window ordering the peak reader missed.

Known answers first; the rule is not applied to a find until all of K1–K3 hold. K1 and K2 use the exact bcc nearest-neighbour Ising through the same code path (J = 19.12 meV, exact T_c ≈ 1410 K, as E215), and variants with J scaled:

  • K1. T_c ≈ 1410 K (ordered through the window): f < 0.25, pass.
  • K2a. T_c ≈ 500 K (inside): f > 0.5, fail.
  • K2b. T_c ≈ 50 K (below): f < 0.25, pass.
  • K3. Equiatomic Mo–Ta on e6 (E242 (4): 1149 ± 114 K): f ≤ 0.35, most of its ordering above 1000 K.

A known answer that fails voids the rule as written.

Predictions for the finds (E240's 11 that passed rungs 0–2, re-run one at a time, about 50 min each on one core, overnight):

  • (1) The survivor Mo₅₁Ti₃₈W₃Ta₃ fails: f > 0.5. E240's own record already shows 23 meV/atom released between the random solution and 60 K, and nothing of that size above 1340 K on the diagnostic.
  • (2) At least 5 of the 11 fail or come out not established; at most 3 pass.
  • (3) Every find whose E240 rung-2 drive was positive (Mo₇₉Hf₁₉, Mo₇₀Hf₂₀Ti₁₀) is decided by rung 2 anyway, and is reported but not counted in (2).

Decision. If the rule holds on K1–K3:

  • ECEVerifier rung 1 adopts it;
  • every E240/E241 verdict is re-scored, and E241's survivor choice is redone on the new verdicts;
  • the site's ladder page states the change.

If K1–K3 fail, the rule is withdrawn here and rung 1 stays as it is, with the blind spot stated on the ladder page. Code: forager/ladder/metropolis.py window_release(), tests on the exact Ising first. Sonnet reviews it adversarially before any find is scored.

E243 note (2026-09-24 02:15), grid. The pre-registered grid ends at 60 K. The expression first written for it, arange(1400, 59, −60), ends at 80 K, and 60 K is now added explicitly (tests/test_rung1_window.py). The survivor's E242 diagnostic already running (scripts/ordering/e242_rung1_curve.py) stops at 80 K. It stays an E242 diagnostic. The survivor's E243 score comes from its run on the 60 K grid with the other ten.

Results

EXPERIMENTS.md · line 16495

E243 — the rule as written is withdrawn before use (2026-09-24 02:35). An adversarial review (Sonnet) ran the rule at production scale on exact Ising models and found two defects before any find was scored:

  • False fails at both window edges. Short-range order builds over a range comparable to T_c itself, so the in-window share f leaks across the edges. An Ising with T_c = 90 K gives f = 0.57 ± 0.05 over five seeds, a fail; with T_c = 80 K it is never a pass; with T_c = 1050–1200 K it fails or is not established. K2b (T_c = 50 K) passed only through the 3 meV guard: its raw f was 0.55. K1–K3 could not see this, because a nearest-neighbour Ising ties magnitude to T_c.
  • The single cooling leg freezes. Acceptance falls to 0 below about 560 K for K1 and 200 K for K2a, and the colder points repeat one frozen configuration. Nothing in the rule checks that E(60 K) is equilibrium; for K1 and K2a the frozen state happened to be the ground state.

By this entry's own clause ("a known answer that fails voids the rule as written"), the near-edge variants fail, so the rule is withdrawn. The E243 chain was stopped before any known-answer or find curve was written. No find has an E243 score.

The full record

This entry is written in 3 separate places in the log, shown here in log order.

EXPERIMENTS.md · lines 16437–16487

E243 — rung 1 reads the change inside the window, not only a peak (pre-registered 2026-09-24 02:13)

Why. E242 found rung 1's blind spot. On the survivor the sampler ordered by 23 meV/atom on cooling (−83.7 → −106.5), gradually and with no peak on the 227 K grid. "No peak above 60 K" was scored as a pass. The requirement is that nothing happens between 90 and 1000 K; a peak is only one way something happens. At the time of writing, the survivor's fine-grid diagnostic (runs/e242_rung1_curve.log) had reached 1340 K and no point inside the window existed. The rule below is fixed before any.

The rule (new, alongside the existing peak reading, never replacing it):

  • Run. One cooling leg, same Hamiltonian (e6 ensemble mean), same 432-site cell, same 100 + 200 sweeps. Grid: 2600, 2000, 1400 K, then 60 K steps from 1340 to 60 K. E(T) is interpolated linearly to 1000 and 90 K.
  • The in-window share of the ordering. f = [E(1000 K) − E(90 K)] / [E(2600 K) − E(60 K)], the fraction of the energy the sampler releases on cooling that is released inside the window.
  • Pass if f < 0.25. Fail if f > 0.5. Not established between. Fail also on any dE/dT or fluctuation peak inside 90–1000 K on this grid.
  • Guard: if the total release E(2600) − E(60) is under 3 meV/atom there is nothing to order; f is noise and the reading is a pass.

This passes both an alloy that is already ordered above 1000 K and stays so, and one that orders only below 90 K, as the requirement's own wording does (chapter 1). It fails the gradual in-window ordering the peak reader missed.

Known answers first; the rule is not applied to a find until all of K1–K3 hold. K1 and K2 use the exact bcc nearest-neighbour Ising through the same code path (J = 19.12 meV, exact T_c ≈ 1410 K, as E215), and variants with J scaled:

  • K1. T_c ≈ 1410 K (ordered through the window): f < 0.25, pass.
  • K2a. T_c ≈ 500 K (inside): f > 0.5, fail.
  • K2b. T_c ≈ 50 K (below): f < 0.25, pass.
  • K3. Equiatomic Mo–Ta on e6 (E242 (4): 1149 ± 114 K): f ≤ 0.35, most of its ordering above 1000 K.

A known answer that fails voids the rule as written.

Predictions for the finds (E240's 11 that passed rungs 0–2, re-run one at a time, about 50 min each on one core, overnight):

  • (1) The survivor Mo₅₁Ti₃₈W₃Ta₃ fails: f > 0.5. E240's own record already shows 23 meV/atom released between the random solution and 60 K, and nothing of that size above 1340 K on the diagnostic.
  • (2) At least 5 of the 11 fail or come out not established; at most 3 pass.
  • (3) Every find whose E240 rung-2 drive was positive (Mo₇₉Hf₁₉, Mo₇₀Hf₂₀Ti₁₀) is decided by rung 2 anyway, and is reported but not counted in (2).

Decision. If the rule holds on K1–K3:

  • ECEVerifier rung 1 adopts it;
  • every E240/E241 verdict is re-scored, and E241's survivor choice is redone on the new verdicts;
  • the site's ladder page states the change.

If K1–K3 fail, the rule is withdrawn here and rung 1 stays as it is, with the blind spot stated on the ladder page. Code: forager/ladder/metropolis.py window_release(), tests on the exact Ising first. Sonnet reviews it adversarially before any find is scored.

EXPERIMENTS.md · lines 16489–16493

E243 note (2026-09-24 02:15), grid. The pre-registered grid ends at 60 K. The expression first written for it, arange(1400, 59, −60), ends at 80 K, and 60 K is now added explicitly (tests/test_rung1_window.py). The survivor's E242 diagnostic already running (scripts/ordering/e242_rung1_curve.py) stops at 80 K. It stays an E242 diagnostic. The survivor's E243 score comes from its run on the 60 K grid with the other ten.

EXPERIMENTS.md · lines 16495–16509

E243 — the rule as written is withdrawn before use (2026-09-24 02:35). An adversarial review (Sonnet) ran the rule at production scale on exact Ising models and found two defects before any find was scored:

  • False fails at both window edges. Short-range order builds over a range comparable to T_c itself, so the in-window share f leaks across the edges. An Ising with T_c = 90 K gives f = 0.57 ± 0.05 over five seeds, a fail; with T_c = 80 K it is never a pass; with T_c = 1050–1200 K it fails or is not established. K2b (T_c = 50 K) passed only through the 3 meV guard: its raw f was 0.55. K1–K3 could not see this, because a nearest-neighbour Ising ties magnitude to T_c.
  • The single cooling leg freezes. Acceptance falls to 0 below about 560 K for K1 and 200 K for K2a, and the colder points repeat one frozen configuration. Nothing in the rule checks that E(60 K) is equilibrium; for K1 and K2a the frozen state happened to be the ground state.

By this entry's own clause ("a known answer that fails voids the rule as written"), the near-edge variants fail, so the rule is withdrawn. The E243 chain was stopped before any known-answer or find curve was written. No find has an E243 score.

Related entries

Built with PRISMWebsite and visualizations made using Claude