NEWS

---

AMDconfigurations 0.4.0

• Separated modal-count proposals, forest agreement, conditional simulation reference values and applicability to the analysed distribution.
• reliability no longer reports simulation calibration as established reliability for a new distribution. Continuous own-calibrator values are in reference_reliability; existing feature, output and numeric guards remain.
• Added proposed_multiple_modes; has_structure remains a compatibility alias.
• Adopted empirical configurations / clusters of the fitted FCM partition at a specified resolution, distinct from components, modes and modal basins.
• Included subsequent prospective and distribution-challenge evidence, including exact-count overconfidence and confident unanimous errors.
• Closed the augmented-training and modal-weight-control review. Both tested replacements are excluded from this release because substantial local regressions remain. The frozen drop-37 forests and own calibrators are retained.
• Added retrospective historical-benchmark and modal-basin assignment results for the frozen reader, with their distinct training and evaluation scopes.

AMDconfigurations 0.3.0

• Adopted the full-six-family drop-37 reader. Its external v3 bundle excludes 37 dimensional descriptors, retaining 194 raw and 231 processed predictors, and removes the tested common-unit sensitivity.
• Added own exact-K and one/several correctness calibrators fitted on selection only. Their assessment on the previously examined independently sampled bank is retrospective; the unchanged forests retain their prospective validation.
• Reliability remains continuous per prediction, without an automatic cutoff. Numerical-grid, output, missingness and constant safeguards remain in force; grid membership does not certify real-data geometry.
• Require the versioned v3 drop-37 bundle and checksum. Historical v1/v2 bundles remain reproducible with matching historical releases. Excluded-family readers are archived diagnostics and are not deployed.
• Preserved the AMD definition, curve construction, assignments, feature-helper arithmetic, public historical snapshots and study classifications.

AMDconfigurations 0.2.1 (2026-10-12)

• Corrected the README model-manifest link for CRAN resubmission. The manifest remains available at the fixed v0.2.0 URL even though the external model directory is deliberately excluded from the CRAN source package.
• Updated package-version and installation references. The R implementation, bundled datasets, corrected-v2 forests and confidence calibrators are unchanged.
• Kept the original Ecography DOI as a plain identifier in the README; its publisher redirect returned HTTP 403 to automated URL checking. The cited article and DOI have not changed.

AMDconfigurations 0.2.0

Optional model-assisted interpretation and applicability

• The optional reader explicitly labels its target as the numerically resolved density-mode count used as the simulation reference, not the generating component count or a direct density measurement by AMD.

• Off-grid proposals themselves are labelled exploratory. Departures from omitted training-constant values suppress certainty, even when off-grid numerical values were explicitly requested. Predictions and models are unchanged.

• Resource diagnostics now distinguish exact partition distances from the frozen global-distance sample capped at 2000 rows. The prior blanket claim of exact distances was incorrect outside the assessed N grid.

• Bundle target, creation metadata and confidence-calibrator formulas are validated; unreachable legacy-loading branches were removed.

• Historical models, calibrators, feature-helper arithmetic, the AMD core and stored-partition diagnostics are unchanged. The same corrected-v2 bundle assembled for 0.2.0.9014 remains compatible.

• Added optional load_amd_predictor(path, sha256) and predict_amd() with external corrected-v2 forests. No model download or training occurs.

• Historical v1 bundles are rejected by the corrected extractor, not silently reinterpreted. Their frozen study artefacts remain unchanged outside this API.

• The corrected reader uses zero for a nonexistent left-pair margin. Its historical confidence calibrators were reassessed, without refitting, on the previously examined final bank; this is retrospective transfer evidence.

• Structure-decision reliability and exact-K reliability are reported separately. Numerical-grid membership does not certify the geometry of real data. Certainties are suppressed by default outside the assessed N/D grid, and for unsupported outputs or previously unseen descriptor-missingness patterns.

• Diagnostics have an explicit working-memory guard before fitting. Exact partition distances and the frozen global-distance sample are distinguished explicitly.

• The base curve, assignment and descriptive workflow still need no forests, ranger or digest. The published AMD definition is unchanged.

Stored-partition diagnostics

• Added diagnose_amd() for descriptive, read-only summaries of a saved curve, its memberships, hard/soft group sizes and adjacent partition transitions.
• The returned dictionary defines every diagnostic and its scale. Missing neighbours and summaries of empty hard groups remain NA, not imputed.
• ARI and NMI are invariant to group labels, including single occupied groups. Directional overlaps distinguish fragmentation from source mixing.
• No new fits, resampling, sample distance matrices, automatic K decisions or calibrated certainties are produced. The original data are not required.
• The published AMD formula, primary fit selection and assignment are unchanged. The optional supervised reader is a separate operation with external models.

AMD definition and curve construction

• AMD retains the published definition mean(apply(membership, 1, max)) - 1 / c. It is not normalised to [0,1].
• amd_score() is now an exported operation independent of the algorithm used to estimate memberships. It accepts the analytical c = 1 case, whose AMD is zero.
• Operational AMD curves still begin at c = 2.
• At each candidate c, the primary curve point now comes from the converged fuzzy c-means fit with the smallest objective. Random starts are fitting attempts, not statistical replicates.
• Every selected fit is stored. The final assignment no longer comes from an additional unrecorded fuzzy c-means run.
• c_peak is the numerical maximum of the primary curve. The historical c_opt name remains as an alias but no longer represents an automatic scientific choice.

Assignment and reproducibility

• The recommended workflow is assign_configurations(curve, c), with c explicitly chosen after inspecting the curve and relevant evidence. It extracts the exact fit that produced the selected curve point.
• Direct fitting with assign_configurations(data, c, ...) remains available as a provisional compatibility interface.
• Supplied seeds are now actually applied and recorded. Named seed matrices preserve common fits when a candidate range is extended.
• Objectives, iterations, convergence, selected starts, preprocessing, omitted rows and a process-level reproducibility manifest are retained.
• Fits that reach iter_max are retained as diagnostics but cannot supply the selected curve point or standalone assignment. Candidate ranges now fail early when their upper value is not smaller than the complete sample size.
• New amd_curve and amd_assignment objects carry schema_version = 2L.

Compatibility with saved 0.1.0 objects

• plot(), print() and summary() recognise saved public 0.1.0 amd_curve objects. They warn that the minimum-objective primary curve cannot be reconstructed and use only the stored historical summaries. A legacy c_opt marker is drawn only with type = "max", the series from which version 0.1.0 obtained it, and not on the stored mean.
• New curve objects retain the public 0.1.0 components c_opt, max, mean, raw and coordinates.
• Direct assignment objects retain c_opt, cluster, membership, centers and objective.
• The S3 plot methods use the active graphics device by default. The legacy compute_amd_curve(..., plot_curve = TRUE) interface retains its public default of opening a device; open_device has no effect when plotting is disabled.

Intentional incompatible changes

• estimate_sigma_equivalent() and all sigma-equivalent classes and methods have been removed from the canonical package.
• The numerical values of a curve can differ from version 0.1.0 because that release did not apply its advertised seeds, selected the largest AMD across starts and then produced coordinates with an additional unrecorded fit.
• The default graph now shows the primary minimum-objective curve rather than the mean across starts. type = "mean" and type = "max" remain available as diagnostics.

External model distribution

• Corrected-v2 forests are distributed separately from the package. The base curve, diagnostics and assignment workflow requires no model download, ranger or digest. Model loading requires an explicit trusted SHA-256.
• Historical v1 forests, calibrators and the local audit records remain unchanged. Corrected-reader confidence transfer is retrospective evidence, not a new prospective assessment or a claim of universal reliability.