AMD DROP-37 READER: EVIDENCE AND INTERPRETATION — 20 OCTOBER 2026

Target and terminology
The reader targets the independently numerically resolved population
density-mode count in simulations. FCM configurations are clusters of the
fitted partition at a specified resolution and data representation. They
are distinct from generating components, modes and modal basins. Correct
count does not validate observation assignments or substantive meaning.
One density mode is not absence of all structure.

Fixed reader
Five full-six-family forests, seeds 101/202/303/404/505; 194 raw and 231
processed predictors after removing 37 dimensional descriptors.
External v3 bundle SHA256:
3407a5a55c30a70e150afaa16a52ee4f2c8d3ed265941f3af8c4cb9d0c99035f
Forests trained on 1725 discovery populations. Own exact-K and one/several
logistic calibrators fitted on 1721 selection populations only. No
excluded-family calibrator is used. Candidate range 1:14; assessed output
support 1:8. Above-eight truths remain in evaluation denominators.

Common-unit validation
Independent bank: 1728 generated, 1719 numerically resolved, 9 unresolved.
Exact-count accuracy at original units: 69.75% for drop-37 versus 69.63%
for a discovery-trained matched absolute-feature control. Paired difference
0.12 percentage points, bootstrap 95% interval -0.81 to 1.05.
Common positive scaling of all coordinates by 0.25, 0.5, 2 and 4 changed no
drop-37 proposals in 1728 descriptor-level checks and 36 prespecified
end-to-end checks. This does not establish invariance to arbitrary
independent rescaling of variables or formal predictive equivalence.

Retrospective comparison on the historical final benchmark
1723 resolved populations, including four truths above eight; no new forests.
Drop-37 exact-count accuracy: 69.8201%, population bootstrap 95% interval
67.5566 to 72.0255. Historical complete reader: 75.1596%; paired difference
-5.3395 percentage points, interval -6.8485 to -3.8305.
The historical reader was fitted on discovery plus selection, whereas this
drop-37 reader uses discovery only. This is not a causal feature ablation.
Gap firstSEmax achieved 57.0517%, GMM-ICL 56.9356%, and GMM-BIC 55.9489%.
These are comparisons of complete procedures within the assessed domain.

Assignment recovery was evaluated independently of count accuracy. Using
1722 populations with fully assigned modal-basin references, mean selected
FCM-partition ARI was 0.7526 for drop-37 and 0.7630 for the historical reader.
Among the 1452 multimodal populations, respective means were 0.7218 and
0.7300. FCM supplied with the true modal count gave 0.7448; GMM-BIC gave
0.7964. Better count accuracy therefore did not imply better assignments.
The strict analysis excludes one partially assigned reference population;
it does not silently remove individual unassigned observations.

Prospective assessment with frozen reader and own calibrators
1728 generated, 1721 resolved, 7 unresolved; six resolved truths above 8.
Exact count: accuracy 68.5067%, mean reference confidence 70.9105%;
confidence excess 2.4039 percentage points, bootstrap 95% 0.6532 to 4.1538.
One/several: accuracy 92.9692%, mean reference confidence 93.2200%;
excess 0.2508 points, interval -0.7550 to 1.2861.
These are finite simulation-domain assessments, not individual guarantees.

Controlled distribution challenge with the same frozen reader
1824 populations, all modal truths resolved. Continuous factored densities,
aligned component centres and rotations; 19 scenarios across Gaussian,
Gumbel, logistic and quartic densities.
Exact count: accuracy 60.5263%, mean reference confidence 66.7108%;
excess 6.1845 points, bootstrap 95% 4.2333 to 8.0916.
One/several: accuracy 89.4737%, mean reference confidence 90.1499%;
excess 0.6762 points, interval -0.5684 to 1.9642.
Gaussian controls also showed exact-count overconfidence.
Among 565 reference values >=0.9, 50 exact-count errors occurred. A
post-result subgroup of 189 proposals of K=2 contained 46 errors; all five
forests agreed in those errors. This descriptive cutoff is not a K rule.

Completed training-intervention review
Two fixed interventions retained all six original families and the same
drop-37 predictors: adding stored challenge populations with equal bank mass,
and then controlling bank and modal-count masses to separate weight effects.
Each intervention fitted 100 forests: 19 complete held-out scenarios times
five seeds and five full fits for the prior bank. These are retrospective
comparisons on already examined banks, not independent confirmation.

Original / augmented / weight-controlled challenge exact-count accuracies:
60.53% / 67.43% / 67.05%. The weight-controlled gain over the frozen reader
was 6.52 percentage points (conditional scenario-bootstrap 95% 3.67 to 9.43).
Prior-bank exact accuracies were 68.51% / 68.33% / 68.39%.
The average challenge improvements are retained as positive findings.

In scenario 13, with three true modes in all 96 populations, false proposals
of one mode were 0 / 30 / 23. Thus controlling weights mitigated but did not
remove the substantial local regression. Both interventions repaired only
three of the 46 original high-reference K=2 errors. Scenario 14 exact-count
successes were 10 / 3 / 3 out of 96. The weight-controlled prior-bank
one/several accuracy was 92.45%, compared with the frozen reader's 92.97%.
No post-result noninferiority threshold was invented.

Both replacements are rejected for this release and archived as completed
experimental evidence. They are not an open adoption queue. No calibrators
were fitted for these rejected models, and no original-reader certainty was
transferred to them. The package retains the frozen reader's measured
performance and limitations; this reporting correction is not a claim that
all modal-count errors or arbitrary distribution shifts have been repaired.

Operational consequence
The package retains modal-count estimation. It reports proposed_k and
proposed_multiple_modes, forest agreement, and continuous conditional
simulation-reference values separately. Reliability for the distribution
being analysed is not established by N,D membership, high reference
confidence or forest unanimity. No automatic applicability threshold is
claimed. allow_extrapolation exposes reference values; it cannot validate
transport. The AMD peak is a numerical summary, not an automatic K rule.

Sources in the reproducibility workspace
rf_drop37_fresh_validation_20261014/VALIDATION_CLOSEOUT_20261016.json
drop37_certainty_prospective_20261018/VALIDATION_CLOSEOUT.json
drop37_applicability_20261019/EXPERIMENT_CLOSEOUT.json
paper_revision_20261020/FINAL_BENCHMARK_VERIFIED.json
paper_revision_20261020/assignment_results/ASSIGNMENT_VERIFIED.json
drop37_training_repair_20261019/INFORME_REVISION_APRENDIZAJE_20261020.txt
drop37_weight_control_20261020/INFORME_CIERRE_LECTOR_20261020.txt
TERMINOLOGIA_AMD.txt
Training-intervention results are not substituted into fixed-reader results.

Conceptual references
Hennig, C. (2015), What are the true clusters?
Pattern Recognition Letters 64, 53-62. doi:10.1016/j.patrec.2015.04.009
Chacon, J. E. (2015), A population background for nonparametric density-based
clustering. Statistical Science 30, 518-532. doi:10.1214/15-STS526
