[us-commits] [ehb54/ultrascan3] 7da8a3: somo/grpy: raise the shell-reduction error floor t...

[emre brookes]

  Branch: refs/heads/somo-dev
  Home:   https://github.com/ehb54/ultrascan3
  Commit: 7da8a315fa7ef14a7830927cdef90989f72fc445
      https://github.com/ehb54/ultrascan3/commit/7da8a315fa7ef14a7830927cdef90989f72fc445
  Author: Emre Brookes <emre.brookes at umontana.edu>
  Date:   2026-08-15 (Sat, 15 Aug 2026)

  Changed paths:
    M us_somo/develop/grpy/grpy_shell.hpp

  Log Message:
  -----------
  somo/grpy: raise the shell-reduction error floor to 0.75 of the raw gap

The estimate is max(safety * Richardson remainder, floor * raw inter-rung gap).
Validating the full 23-model grid with all five observables -- the previous
multi-observable grid ran only 12 of the 23 models, so eleven never had intrinsic
viscosity or rotational diffusion checked -- exposed two evaluations that passed
the estimator's stop test yet missed the requested tolerance (2GD1 intrinsic
viscosity at 1%: estimate 0.859%, true error 1.066%), plus eleven more that were
compliant but undercovered.

Tolerance compliance is a subset of estimator coverage: the ladder stops when the
estimate falls below the tolerance, so an estimate that covers the true error
makes stopping imply compliance. Verified on the grid -- no row missed its
tolerance while covered, and all 235 reduced rows converged.

Raising the floor rather than `safety` is deliberate. The two terms are combined
with max(), so while the floor binds the safety factor does nothing: with a
doubling ladder and the order at k_max = 3 the remainder is gap/7, and
safety * gap/7 stays under 0.5 * gap for any safety <= 3.5. Measured on the same
grid, safety = 2.0 fixed neither failure and drove the median case to retain every
bead, because it inflates the estimate for the well-behaved majority where the
floor is slack. The floor acts only on the high-observed-order rows that fail.

At 0.75 the grid is clean: 200/200 reduced evaluations inside their tolerance and
all 200 covered, worst case spending 55.3% of the allowed budget (was 106.6%), for
4.6 percentage points more beads retained on average.

floor_frac is exposed on ShellOptions rather than left hardcoded, alongside k_min,
k_max and safety. Measuring it required patching a private copy of this header,
and a second copy of an estimator is what makes validation results unattributable.

The file's own MEASURED note claimed 92/92 and 252/252; both were wrong (seven
observables counted, three research-only; and the narrow model grid). Corrected to
what this code actually returns -- the finest rung, not the extrapolation.

Co-Authored-By: Claude Opus 5 <noreply at anthropic.com>


  Commit: 69e372e6764e086a32759607f45e67c2831da51f
      https://github.com/ehb54/ultrascan3/commit/69e372e6764e086a32759607f45e67c2831da51f
  Author: emre brookes <ehb54 at users.noreply.github.com>
  Date:   2026-08-15 (Sat, 15 Aug 2026)

  Changed paths:
    M us_somo/develop/grpy/grpy_shell.hpp

  Log Message:
  -----------
  Merge pull request #515 from ehb54/ehb54-984-floor

somo/grpy: raise the shell-reduction error floor to 0.75 of the raw gap


Compare: https://github.com/ehb54/ultrascan3/compare/8423eafd59f9...69e372e6764e

To unsubscribe from these emails, change your notification settings at https://github.com/ehb54/ultrascan3/settings/notifications