[us-commits] [ehb54/ultrascan3] fdf85a: extrap_c0: report per-curve I(0)/c diagnostic (sur...

[emre brookes]

  Branch: refs/heads/ehb54-issue-920
  Home:   https://github.com/ehb54/ultrascan3
  Commit: fdf85a3561ea2a81545442e018d2402c0b912a51
      https://github.com/ehb54/ultrascan3/commit/fdf85a3561ea2a81545442e018d2402c0b912a51
  Author: ehb54 <brookes at uthscsa.edu>
  Date:   2026-07-21 (Tue, 21 Jul 2026)

  Changed paths:
    M us_somo/develop/src/us_hydrodyn_saxs_iqq_extrap_c0.cpp
    M us_somo/somo/doc/manual/somo/somo_saxs_extrap_c0.html

  Log Message:
  -----------
  extrap_c0: report per-curve I(0)/c diagnostic (surface evidence, not a verdict)

Closes out the reference-quality question. Two candidate automatic checks were built
and tested first, and both failed:

- a residual-derived sigma(q) as the regularization noise model: on simulated series
  swept from 4 to 32 concentrations it never beat the existing weighting, and did not
  improve with more curves, so the failure was not degrees of freedom.
- a high-q shape veto for reference quality: on simulated series with defects injected
  into the top curve it separated only a constant buffer-subtraction offset (8.7x), and
  missed mis-set concentration, aggregation and damage entirely (0% of seeds). Worse, the
  statistic tracks signal-to-noise, which rises with concentration, so in the CLEAN case
  the top curve scores best of the set (0.093 vs 0.145) -- a veto built on it is biased
  against ever firing.

Both fail for the same structural reason: the highest-concentration curve is at once the
highest-leverage point in the regression and the highest-SNR curve, so it looks good by
every measure available from inside the series. The information is not there.

So report the evidence instead of guessing. Each run now logs, per curve, the Guinier Rg
and I(0), the ratio I(0)/c, and its deviation from (a) the series median and (b) the trend
predicted by the OTHER curves. I(0)/c is constant in the dilute limit and falls smoothly
as the interparticle term grows, so a curve off the others indicates a mis-set
concentration or aggregation. The trend column is leave-one-out, so the curve cannot pull
the line onto itself -- the only form of this comparison that says anything about the
extreme concentration. Deviations beyond 15% are highlighted. Nothing is discarded or
re-selected on this basis; it feeds the reference selector, which the user drives.

On the 7-curve alpha-syn set this independently reproduces Mattia's assessment of the
0.694246 curve (+12.2% vs trend, i.e. the overestimated MW he reports). It does NOT
reproduce his call on 2.41176, which sits at -5.9% and is unremarkable by either measure.

Verified: all 9 regression baselines byte-identical (the diagnostic is log-only) and the
table is emitted on all 9 runs. Manual updated.

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



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