Propose a modal resolution from an AMD analysis

Optional assistance for interpreting a complete AMD curve. It computes the frozen c = 2,...,14 AMD analysis (unless a compatible curve is supplied), derives curve, data, and candidate-partition descriptors, and applies five externally loaded random forests. Conditional reference values estimate correctness of the one-versus-several-modes decision and exact K under the reference simulation regime. Reliability for a new distribution remains unestablished; these values are not the probability that groups exist. The target is the numerically resolved number of density modes used as the simulation reference, not necessarily the number of generating components. AMD itself does not estimate a density.

Usage

predict_amd(
  data,
  predictor,
  curve = NULL,
  seed = 1L,
  allow_extrapolation = FALSE,
  verbose = TRUE,
  max_working_bytes = 512 * 1024^2
)

Arguments

data

Numeric matrix or data frame, with observations in rows.

predictor

An object returned by load_amd_predictor().

curve

Optional compatible amd_curve. Supplying it avoids repeating the principal AMD fits; it is checked against data and the frozen model contract.

seed

Positive integer controlling AMD starts, stability refits, and any large-data sampling used by descriptors. The caller's random-number state is preserved.

allow_extrapolation

Logical. Away from the validated numerical grid, retain conditional simulation-reference values in reference_reliability, labelled as unvalidated interpolation or extrapolation. This never establishes reliability for the analysed distribution and does not override feature or output-support guards.

verbose

Logical; report progress while computing a new curve.

max_working_bytes

Positive finite working-memory guard for the exact descriptors (default 512 MiB). The estimate is not a guarantee of total process memory and excludes the externally loaded forests. The frozen global-distance summaries use a seeded sample of 2000 rows when N > 2000; partition distances remain exact. This distinction is recorded in the resource diagnostics. No additional approximation bypasses the guard.

Value

An object of class "amd_prediction" containing the AMD curve, proposed numerically resolved density-mode count K, separate structure and exact-modal-count conditional reference estimates. reliability remains NA because applicability to the analysed distribution has not been established. reference_reliability contains the own-calibrator values subject to applicability guards. agreement contains forest vote fractions, not probabilities of correctness. has_structure is retained as a compatibility alias for proposed_multiple_modes. The object includes candidate scores, five-seed proposals, diagnostics, and applicability.

Details

This function does not replace inspection of the AMD curve, does not claim that its proposal is ground truth, and does not assign observations. When K > 1 is proposed, use assign_configurations() with the returned curve and a scientifically justified c.

The drop-37 forests omit 37 dimensional descriptors and use the corrected missing-left-pair convention. Their own correctness calibrators were fitted only on selection populations. A subsequent fixed-reader prospective bank gave 68.51 percent exact-count accuracy versus 70.91 percent mean reference confidence. A controlled distribution challenge gave 60.53 versus 66.71 percent. Thus the calibrated values cannot certify accuracy for a new distribution merely because N and D match the simulation grid.

Continuous exact-K and one/several values remain available separately in reference_reliability; no certainty threshold chooses or rejects K. By default these reference values are suppressed outside N = 150 or 600, D = 2, 6 or 15, assessed output support K=1:8, or assessed feature conditions. The mechanical candidate range remains 1:14. Departures from omitted training-constant predictors also suppress reference values. Neither forest unanimity nor a large reference value establishes distributional applicability. Count correctness does not validate FCM assignments or the scientific meaning of empirical configurations. See the installed READER_EVIDENCE.txt for evidence and limits.

Supplying a compatible curve avoids its 260 principal fits. The additional five computational restarts per c required by the reader are still run; these are not a bootstrap or other sample-level stability analysis. Curve checks establish row/dimension and objective consistency with the supplied data, not a cryptographic identity of a separately created data file.