SUPERSEDED: the recommendation below is withdrawn. It hides poor narrowband
safety in an average score and its rho-based recovery erases remembered gain
when excitation stops. Retained as a record of the earlier study only.
Use paper_code_bridge_study results for the code-matching investigation.

Paper 2 simple indicator changes -- 2026-10-01

Recommendation: recovery='drive', guard=0.02. All other switches/default values unchanged.
Run: R = run_paper_benchmark(0.4, struct('recovery','drive','guard',0.02));

Selection and seed protocol
- Original seeds: 11,12,21,22, shared between fixed/active lambda cases.
- Tested 12 single changes and all 52 compatible pairs on these seeds.
- Original-seed winner: drive-ratio recovery + eta_q=0.12, J=0.02368.
- First holdout: add 1000,2000,3000 to each original seed (24 cases).
  Original-seed winner J=0.03773. Recovery alone J=0.02878.
- After this failed generalization, first holdout became development data.
  Eight already-tested alternative pairs were evaluated on these same seeds.
- Select using mean J over the original and three development seed sets.
  Exclude the existing soft confirmation gate from the recommendation because
  it introduces two constants (slope 2 and floor 0.1). It remains a comparator.
  Recovery + guard 0.02 wins among the evaluated eligible alternatives:
  development J=0.02987, including original seeds with equal weight per set.
- Freeze recommendation before the final untouched offsets 4000,5000,6000.
  Final seeds are {4011,4012,4021,4022}, {5011,5012,5021,5022},
  {6011,6012,6021,6022}. No tuning after seeing these results.

                         Original J   First holdout J   Final unseen J
Code as is                  0.02029        0.02231           0.02597
Paper as written            0.06101        0.04365           0.04965
Recovery + slow average     0.02368        0.03773           0.02761
Recovery + guard 0.02        0.03699        0.02750           0.02907

The slow-average pair again scores well on the final seeds, but its first
holdout was substantially worse. Retain the preselected guard pair. Across
all seven seed sets, its mean J is 0.02953 versus 0.03139 for the slow-average
pair. These are descriptive pooled results, not an additional untouched test.

Final unseen metrics, averaged over three seed sets:
Method                 Error wide  Error narrow  Unsafe %  Rough wide  Travel
Code as is               0.011971    0.032876       3.551     3.124      1.329
Paper as written         0.015396    0.056556      13.678    10.904      1.824
Recovery + guard 0.02     0.015058    0.028092       7.491    11.027      2.863

Remaining gap: unsafe time, roughness, and lambda travel. The recommendation
does not include smoothing. J does not penalize roughness or travel.
Final per-set J for the recommendation: 0.0298,0.0248,0.0325.

Exact equations are in paper_simple_equations.tex. Section 3.1.2 retains the
per-bin power fit. Only its recovery condition changes. Section 3.1.4 gains
the target offset h. Recovery adds no tuned parameter, guard adds only h.

Parameters for recommended run
New: h=0.02 (guard).
Unchanged: fs=2048 Hz, Delta t=0.125 s, Tw=0.5 s, band=2--80 Hz excluding
58--62 Hz, eta_s=eta_q=0.39, eta_A=4.54, sigma_u=0.0575/4.54,
l_floor=0.001, eta_g=0.12, r_g=0.73, g_th=1.75, kappa=0.625,
b_tar=0.4, lambda_min=0.2, t_ref=1.5 s (12 estimator reference updates),
G=logspace(log10(0.5),log10(6),25). lambda initially 1, log(g_s) initially 0.
All other modifications are off. The denominator safeguard is machine eps.

Definitions used for comparison
J=mean absolute margin error over eight cases + mean unsafe percentage over
cases 5--8 divided by 1000. Unsafe means true margin <0.38, evaluated t>=5 s.
Wide/narrow errors average cases [1,2,5,6]/[3,4,7,8]. Roughness is the mean
wideband 1000*median(abs(second difference of estimated margin)).
Use exactly the supplied runner's chordal loop margin and second-smallest
estimator-bin reduction. These results concern this benchmark and its noise
seeds, not a proof of safety or other plants/excitation regimes.

Artifacts and reproduction
- paper_simple_search.csv: all 105 evaluated method/stage rows.
- paper_simple_search.mat: one top-level result struct, options, seed-wise
  metrics and saved per-case tables. Original single/pair tables can be
  regenerated from the runner without changing the benchmark.
- run_paper_benchmark_seeds.m: copy of runner, changes RNG offset and save
  behavior only. Offset zero matches the original runner exactly.
- paper_simple_search.m: run stages single 1:14, pair 1:52, holdout 1:15,
  develop 1:8, final 1:16, in order. 'holdout' stores the first validation,
  subsequently reused as development data, not the final unseen evaluation.
- paper_simple_baseline.mat: saved as-written baseline before these edits.

Checks: as-written output unchanged exactly; seed-copy offset zero matches
that baseline exactly; Code Analyzer has zero notices for all changed/new
MATLAB source files. No changes to main checkout main/, manuscript or builder.
No physical workflows or commits.