Experiments · E247

If the requirement is one disordered phase throughout, rather than no transition inside the window, what changes?

Yes, the new rule works as intended. All seven known answers hold under the rewritten bars, and every candidate now fails: Mo₃₅Ta₃₃W₂₂Nb₅Ti₅, which orders at 1044 K, goes from 0.86 to 0. No candidate passes rungs 0–2.

In the log: the requirement is one disordered phase: rung 1's pass rule changes (pre-registered 2026-09-27 22:04)

confirmedDate 2026-09-27 22:04, as written in the logrung 1 · ordering3 predictions · 1 result paragraphEXPERIMENTS.md lines 17281–17325, lines 17405–17415

Pre-registration

  1. (1)
    All seven bars hold.
    no verdict written against it
  2. (2)
    All eight E245c candidates now score p_rung1 < 0.1. Mo₃₅Ta₃₃W₂₂Nb₅Ti₅ (1044 ± 43 K) goes from 0.86 to about 0; the seven not established stay at 0. The walk ends with no survivor.
    holds
  3. (3)
    No frozen-phase check runs (it only runs when p ≥ 0.1), so the re-read costs only the bootstraps. Also fixed in M1, recorded here. CI (.github/workflows/tests.yml) has failed on every run since 2026-09-13. The test machine lacked the science libraries, Python 3.11 cannot parse the code's PEP 701 f-strings, and tests that read data/ (not in git) failed instead of skipping. The workflow now tests on 3.12 and 3.13 with the libraries installed, and each data-dependent test names the file it needs and skips.
    holds
The pre-registration, as written

E247 — the requirement is one disordered phase: rung 1's pass rule changes (pre-registered 2026-09-27 22:04)

Why. External review (Jan, 2026-09-26) read the requirement as it is written: a single, disordered bcc phase from 90 to 1000 K. Rung 1 has scored "no transition inside the window": an alloy whose transition lies above 1000 K passed, because it has no phase change in the window. passes_window's docstring defends that ("an alloy that is already ordered at 90 K and stays ordered has no phase change either"); verifier.py says the opposite ("ordered throughout … a compound rather than a solid solution … a different material from the one asked for"). The operator adopted the reviewer's reading (2026-09-27). It is the first step of the review-filter plan (M1).

The rule, applied to every rung-1 reading from now on.

  • A significant transition passes only if it lies below 90 K: p_rung1 = P(every significant peak's corrected T_c draw < 90 K). A transition above 1000 K now fails.
  • A broad feature reaching 90 K or above: not established (p 0).
  • No significant peak, but the sampler froze at or above 90 K: the state there is ordered or segregated, so p 0.
  • No significant peak and the sampler never froze above 90 K: unchanged. "Still rising at the floor" scores P(T_c < 90 K); a quiet window scores 1.
  • passes_window (legacy paths) gets the same rule. The old rule stays available as rule="no_transition_in_window", so every earlier result can be reproduced.

Known-answer bars, rewritten before any re-read (stored E245c legs, seeds 6 and 7, re-read under the new rule, not re-sampled):

known answer exact new bar
K1 Ising 1410 K p ≤ 0.01 (fails: ordered through the window)
K4 Ising 1050 K p ≤ 0.01
K2a Ising 500 K p ≤ 0.01
K5 Ising 90 K 0.1 ≤ p ≤ 0.9: never a confident verdict at the edge
K2b Ising 50 K p ≥ 0.9
1D ring crossover p < 0.5
K3 Mo–Ta on e6 (~1100 K) p ≤ 0.01

Any bar that fails voids the change until it is understood.

Predictions. (1) All seven bars hold. (2) All eight E245c candidates now score p_rung1 < 0.1. Mo₃₅Ta₃₃W₂₂Nb₅Ti₅ (1044 ± 43 K) goes from 0.86 to about 0; the seven not established stay at 0. The walk ends with no survivor. (3) No frozen-phase check runs (it only runs when p ≥ 0.1), so the re-read costs only the bootstraps.

Also fixed in M1, recorded here. CI (.github/workflows/tests.yml) has failed on every run since 2026-09-13. The test machine lacked the science libraries, Python 3.11 cannot parse the code's PEP 701 f-strings, and tests that read data/ (not in git) failed instead of skipping. The workflow now tests on 3.12 and 3.13 with the libraries installed, and each data-dependent test names the file it needs and skips.

Results

EXPERIMENTS.md · line 17405

E247 RESULT (2026-09-27 22:51). Re-read from the stored E245c legs (scripts/ordering/e247_reread.py, runs/e247/).

  • (1) All seven bars hold. K1 (1425 K), K4 (1076 K), K2a (512 K) and K3 Mo–Ta (1120 K) score p 0.000: they fail. K5 at the edge scores 0.167, inside its 0.1–0.9 bar. K2b (still rising at the 60 K floor) scores 1.000. The ring stays not established (p 0).
  • (2) Holds. Every candidate scores p_rung1 < 0.1. Mo₃₅Ta₃₃W₂₂Nb₅Ti₅ goes from 0.860 to 0.000 (ordered at 1044 K, so ordered through the window); the seven not established stay at 0.
  • (3) Holds. No frozen-phase check was needed.

The walk ends with no candidate passing rungs 0–2 under the requirement as written. Also recorded: CI, rebuilt as above, passes in a clean Python 3.12 environment with no data/: 621 passed, 147 skipped (each naming its file), 0 failed.

The full record

This entry is written in 2 separate places in the log, shown here in log order.

EXPERIMENTS.md · lines 17281–17325

E247 — the requirement is one disordered phase: rung 1's pass rule changes (pre-registered 2026-09-27 22:04)

Why. External review (Jan, 2026-09-26) read the requirement as it is written: a single, disordered bcc phase from 90 to 1000 K. Rung 1 has scored "no transition inside the window": an alloy whose transition lies above 1000 K passed, because it has no phase change in the window. passes_window's docstring defends that ("an alloy that is already ordered at 90 K and stays ordered has no phase change either"); verifier.py says the opposite ("ordered throughout … a compound rather than a solid solution … a different material from the one asked for"). The operator adopted the reviewer's reading (2026-09-27). It is the first step of the review-filter plan (M1).

The rule, applied to every rung-1 reading from now on.

  • A significant transition passes only if it lies below 90 K: p_rung1 = P(every significant peak's corrected T_c draw < 90 K). A transition above 1000 K now fails.
  • A broad feature reaching 90 K or above: not established (p 0).
  • No significant peak, but the sampler froze at or above 90 K: the state there is ordered or segregated, so p 0.
  • No significant peak and the sampler never froze above 90 K: unchanged. "Still rising at the floor" scores P(T_c < 90 K); a quiet window scores 1.
  • passes_window (legacy paths) gets the same rule. The old rule stays available as rule="no_transition_in_window", so every earlier result can be reproduced.

Known-answer bars, rewritten before any re-read (stored E245c legs, seeds 6 and 7, re-read under the new rule, not re-sampled):

known answer exact new bar
K1 Ising 1410 K p ≤ 0.01 (fails: ordered through the window)
K4 Ising 1050 K p ≤ 0.01
K2a Ising 500 K p ≤ 0.01
K5 Ising 90 K 0.1 ≤ p ≤ 0.9: never a confident verdict at the edge
K2b Ising 50 K p ≥ 0.9
1D ring crossover p < 0.5
K3 Mo–Ta on e6 (~1100 K) p ≤ 0.01

Any bar that fails voids the change until it is understood.

Predictions. (1) All seven bars hold. (2) All eight E245c candidates now score p_rung1 < 0.1. Mo₃₅Ta₃₃W₂₂Nb₅Ti₅ (1044 ± 43 K) goes from 0.86 to about 0; the seven not established stay at 0. The walk ends with no survivor. (3) No frozen-phase check runs (it only runs when p ≥ 0.1), so the re-read costs only the bootstraps.

Also fixed in M1, recorded here. CI (.github/workflows/tests.yml) has failed on every run since 2026-09-13. The test machine lacked the science libraries, Python 3.11 cannot parse the code's PEP 701 f-strings, and tests that read data/ (not in git) failed instead of skipping. The workflow now tests on 3.12 and 3.13 with the libraries installed, and each data-dependent test names the file it needs and skips.

EXPERIMENTS.md · lines 17405–17415

E247 RESULT (2026-09-27 22:51). Re-read from the stored E245c legs (scripts/ordering/e247_reread.py, runs/e247/).

  • (1) All seven bars hold. K1 (1425 K), K4 (1076 K), K2a (512 K) and K3 Mo–Ta (1120 K) score p 0.000: they fail. K5 at the edge scores 0.167, inside its 0.1–0.9 bar. K2b (still rising at the 60 K floor) scores 1.000. The ring stays not established (p 0).
  • (2) Holds. Every candidate scores p_rung1 < 0.1. Mo₃₅Ta₃₃W₂₂Nb₅Ti₅ goes from 0.860 to 0.000 (ordered at 1044 K, so ordered through the window); the seven not established stay at 0.
  • (3) Holds. No frozen-phase check was needed.

The walk ends with no candidate passing rungs 0–2 under the requirement as written. Also recorded: CI, rebuilt as above, passes in a clean Python 3.12 environment with no data/: 621 passed, 147 skipped (each naming its file), 0 failed.

Related entries

Built with PRISMWebsite and visualizations made using Claude