Package 'hypermodR'

Title: Hyperspectral Modelling and Validation
Description: Interpretable and black-box hyperspectral modelling with explicit spectral, spatial and temporal data contracts, grouped validation, reproducible evaluation, optional isolated model runtimes, and an interactive research workbench. Model and instrument qualification is recorded separately from interface availability.
Authors: Raban Heller [aut, cre] (ORCID: <https://orcid.org/0000-0001-8006-9742>)
Maintainer: Raban Heller <[email protected]>
License: MIT + file LICENSE
Version: 0.0.0.9001
Built: 2026-11-04 22:06:33 UTC
Source: https://github.com/CTTIR/hypermodR

Help Index

Attach predictions to observation identities

Description

Attach predictions to observation identities

Usage

## S3 method for class 'hm_fit'
augment(x, newdata, ...)

Arguments

x

An hm_fit.
newdata

Explicit dataset accepted by hm_predict().
...

Arguments passed to prediction, including optional managed workers.

Value

Observation metadata with .fitted, and .resid for scalar numeric outcomes. Matrix outputs remain one explicit list column.

Inspect model and training scope

Description

Inspect model and training scope

Usage

## S3 method for class 'hm_fit'
glance(x, ...)

Arguments

x

An hm_fit.
...

Currently unused.

Value

One row recording model, task, training count, input track and the absence of inherited scientific qualification.

Queue uncertain training-pool cases for annotation review

Description

Queue uncertain training-pool cases for annotation review

Usage

hm_active_learning(
  predictions,
  pool,
  training_ids,
  heldout_ids,
  budget = 1L,
  seed = 1L,
  reference_revision,
  data,
  split
)

Arguments

predictions

Probabilistic hm_prediction aligned exactly to pool rows.
pool

Unlabelled data frame with observation_id and all task grouping columns.
training_ids

Character identities exactly matching the validated training partition.
heldout_ids

Character identities exactly matching the validated test partition.
budget

Positive integer number of pool cases to queue.
seed

Seed used only to break entropy ties.
reference_revision

Explicit immutable reference revision identifier.
data

Source data used to verify pool identities and grouping.
split

Validated grouped train/test split.

Details

Only training-access unlabelled rows may enter the pool. This creates a review queue; it does not acquire labels or modify annotations.

Value

List with an entropy-ranked pending-review queue, seed, budget, reference revision, run hash, annotation owner and labels_acquired=FALSE.

Admit reviewed resolved metadata as an exact install plan

Description

Admit reviewed resolved metadata as an exact install plan

Usage

hm_admit_backend(
  resolution,
  reviewed_terms,
  cache = tools::R_user_dir("hypermodR", "cache"),
  hardware_verified = FALSE,
  checkpoint = NULL,
  port = 8000L
)

Arguments

resolution

Result of hm_resolve_backend.
reviewed_terms

Full reviewed terms entries, preserving each package, version and artifact identity. Use the resolved entries after artifact notice review and set each reviewed flag to TRUE.
cache

External runtime cache.
hardware_verified

Whether exact driver/device ABI has been checked.
checkpoint

Optional exact local checkpoint artifact manifest for vLLM.
port

Unprivileged loopback server port.

Value

An hm_backend_plan bound to reviewed artifact terms and supplied device/checkpoint admission.

Bind an annotation project to its immutable export contract

Description

Bind an annotation project to its immutable export contract

Usage

hm_annotation_contract(project, label_map)

Arguments

project

annot_project.
label_map

Named positive integer class codes; zero is background.

Value

Serializable exact manifest binding and label map.

Open the local spectral modelling workbench

Description

Import, configuration and submission are separate explicit actions. Optional runtimes are never installed on launch. Each connection owns its state/jobs.

Usage

hm_app(launch = TRUE, host = "127.0.0.1", port = NULL, ...)

Arguments

launch

Run the application; FALSE returns a shiny.appobj.
host

Listening interface, default loopback.
port

Optional listening port.
...

Arguments passed to shiny::runApp.

Value

A Shiny application, invisibly when launched.

Create an isolated workbench state

Description

Create an isolated workbench state

Usage

hm_app_state()

Value

An independent session state environment.

Apply a frozen calibrator without learning from new outcomes

Description

Apply a frozen calibrator without learning from new outcomes

Usage

hm_apply_calibration(calibration, predictions)

Arguments

calibration

Frozen hm_calibration.
predictions

Probabilistic hm_prediction with the identical class map.

Details

Does not learn from new outcomes. Abstention marks low confidence without deleting rows.

Value

An updated hm_prediction with scaled probabilities, maximum-probability labels, logical abstain vector and calibration provenance.

Convert a qualified companion object

Description

Convert a qualified companion object

Usage

## S3 method for class 'cuvis_session'
hm_as_dataset(x, profile = NULL, ...)

## S3 method for class 'cuvis_measurement'
hm_as_dataset(x, profile = NULL, ...)

## S3 method for class 'annot_image'
hm_as_dataset(x, profile = NULL, observations = .hm_image_observations(), ...)

## S3 method for class 'annot_project'
hm_as_dataset(x, profile = NULL, ...)

hm_as_dataset(x, profile = NULL, ...)

## Default S3 method:
hm_as_dataset(x, profile = NULL, ...)

## S3 method for class 'hsi_cube'
hm_as_dataset(
  x,
  profile = NULL,
  observations = .hm_image_observations(),
  geometry = .hm_native_geometry(),
  ...
)

Arguments

x

Companion object.
profile

Explicit source profile where required.
...

Adapter-specific arguments.
observations

Explicit acquisition and biological identities.
geometry

Source spatial geometry.

Details

hyperspectR cube conversion preserves measured bands, masks and declared geometry. A cube exported by this package also retains selected-frame observations and band metadata; recorded content identity must still match on reimport. Explicit observations override restored identities. hyperannotatR conversion uses a bounded public tile and records manifest/revision identity. CUVIS session enumeration and SDK scaling remain blocked pending source-backed fixtures. Conversion does not infer calibration.

Value

HMDataset or HMSequence for an admitted class; unsupported classes and currently unqualified CUVIS sources raise typed errors.

Materialise one frame through hyperspectR

Description

Materialise one frame through hyperspectR

Usage

hm_as_hsi_cube(x, acquisition_id, frame_id = NULL, ...)

Arguments

x

Spatial HMDataset.
acquisition_id

One acquisition ID.
frame_id

One frame ID for sequences; NULL for an image.
...

Reserved.

Value

An hsi_cube with explicit native band sorting map in metadata.

Inspect local runtime capabilities without starting services

Description

Inspect local runtime capabilities without starting services

Usage

hm_backend_capabilities()

Value

List with OS, architecture, R version, Python executable path, and explicit unestablished GPU qualification.

Plan an isolated runtime installation

Description

Plan an isolated runtime installation

Usage

hm_backend_plan(
  model_id,
  device = "auto",
  profile = NULL,
  cache = tools::R_user_dir("hypermodR", "cache"),
  capabilities = hm_backend_capabilities()
)

Arguments

model_id

Candidate or custom model identity.
device

Requested device.
profile

Exact custom artifact manifest; no upstream profile is inferred.
cache

External user cache.
capabilities

Local capability record.

Value

An hm_backend_plan containing the validated profile, capabilities, cache, artifact-byte budget, actions and immutable digest.

Inspect a managed process

Description

Inspect a managed process

Usage

hm_backend_status(handle)

Arguments

handle

Owned process handle.

Value

List with PID, model, alive flag, exit status and plan digest.

Apply immutable training transformations

Description

Apply immutable training transformations

Usage

hm_bake(recipe, newdata)

Arguments

recipe

Fitted recipe returned by hm_prep().
newdata

Numeric observation-by-feature matrix with the original input column count and ordering.

Details

Reuses frozen training transformations without learning from new data.

Value

Numeric transformed matrix; rows retain input order and columns follow the fitted steps.

Admit a bounded Spectralverse benchmark protocol without accessing data

Description

Admit a bounded Spectralverse benchmark protocol without accessing data

Usage

hm_benchmark_protocol(protocol, comparison = NULL)

Arguments

protocol

List declaring version=1, program=Spectralverse, cohort, role, dataset_revision, reference_revision, license_admitted, person_grouping_verified, source_cases, roi_ids, input_track, context_budget, pretraining_overlap, prior_outcome_access, target_access and capabilities.
comparison

Optional second protocol requiring identical cohort/revisions/cases/ROIs/input/context and target access.

Details

Cohorts are HeiPor (primary_anatomy), xeno (development_species_perfusion) and ManU (human_admission). ManU requires verified person grouping. Each capability lane supplies model_id, task, platform, logical qualified and evidence. Admission is declarative and does not access data or execute a benchmark.

Value

List with validated protocol, content hash, target-role classification, capability-lane admission states, execution=not_run and scientific_qualification=FALSE.

Paired whole-unit bootstrap of fixed predictions

Description

Paired whole-unit bootstrap of fixed predictions

Usage

hm_bootstrap(
  evaluation,
  unit,
  draws = 200L,
  seed = 1L,
  level = 0.95,
  comparison = NULL
)

Arguments

evaluation

An hm_evaluation of fixed predictions.
unit

Reference column naming whole independent resampling units; connected biological/source components must not be split.
draws

Integer number of bootstrap replicates, at least two.
seed

Seed for reproducible unit sampling.
level

Interval coverage strictly between zero and one.
comparison

Optional exactly matched second evaluation for paired metric differences.

Details

Resamples complete units with replacement. Fitting and model-selection uncertainty are excluded. If any draw lacks valid metrics, intervals remain unavailable and status reports invalid draws.

Value

List with percentile interval, draw-by-metric resamples, row indices, sampled units, validity flags, invalid draw count, seed, level and status.

Create independent in-memory cached values

Description

Create independent in-memory cached values

Usage

hm_cache()

Value

A list with get and set functions; get returns an independent copy.

Compute a complete cache identity

Description

Compute a complete cache identity

Usage

hm_cache_key(data, state = list())

Arguments

data

Data, including all source values and band metadata.
state

Reader, calibration, model and preprocessing state.

Value

SHA-256 identity. Lazy sources include declared content hashes.

Freeze temperature scaling and abstention using an inner held-out partition

Description

Freeze temperature scaling and abstention using an inner held-out partition

Usage

hm_calibrate(
  predictions,
  reference,
  split,
  data,
  temperature_grid = c(0.5, 1, 2),
  abstention = 0,
  fit,
  outer_split
)

Arguments

predictions

Probabilistic predictions from fit on the inner held-out calibration partition.
reference

Exact source observation rows for the calibration partition, in prediction order.
split

Validated inner train/test split partitioning outer training.
data

Original source data used for fitting and split validation.
temperature_grid

Positive finite candidate temperatures.
abstention

Confidence threshold in ⁠[0,1]⁠ frozen into the calibrator.
fit

Training-only hm_fit that generated the supplied predictions.
outer_split

Validated outer split whose training rows are partitioned by split.

Details

Selects the first temperature minimizing calibration log loss. Requires full class support and verifies the supplied predictions against the fit; outer test outcomes cannot enter calibration.

Value

An hm_calibration containing selected temperature, abstention threshold, class map, fit/calibration identities, reference hash and grid scores.

Cancel only an owned modelling job

Description

Cancel only an owned modelling job

Usage

hm_cancel_job(job)

Arguments

job

Owned job handle.

Value

Confirmed job state, invisibly.

Inspect installation and separately supplied parity evidence

Description

Inspect installation and separately supplied parity evidence

Usage

hm_check_backend(
  model_id,
  tests = c("availability", "parity"),
  backend = NULL,
  parity = NULL
)

Arguments

model_id

Model identity.
tests

Requested checks.
backend

Installed backend.
parity

Named evidence including plan_digest and passed.

Details

Availability verifies the declared installation path, containment, artifact sizes/hashes and any Python environment inventory. Parity is caller-supplied evidence bound to the plan digest, reported as reported_pass rather than independently executed here. No scientific qualification follows.

Value

List with model identity, availability status, reported parity status and scientific_qualification=FALSE.

Compare matched predictions using paired whole-unit resampling

Description

Compare matched predictions using paired whole-unit resampling

Usage

hm_compare(
  evaluations,
  paired = TRUE,
  unit,
  draws = 200L,
  seed = 1L,
  family = NULL
)

Arguments

evaluations

List of exactly two matched hm_evaluation objects, ordered first minus second.
paired

Must be TRUE.
unit

Reference column identifying independent whole units.
draws

Integer number of paired bootstrap draws, at least two.
seed

Reproducible resampling seed.
family

Optional descriptive comparison-family declaration, retained without multiplicity adjustment.

Details

Requires identical cases, references, task, class map, estimand, hierarchy and metric names. This is descriptive inference without multiplicity-adjusted significance claims.

Value

An hm_evaluation list containing metric differences, paired bootstrap uncertainty, family and interpretation.

Map image coordinates through explicit geometry

Description

Map image coordinates through explicit geometry

Usage

hm_coordinates(geometry, coordinates, inverse = FALSE)

Arguments

geometry

Geometry with origin and homogeneous transform.
coordinates

Two-column x,y matrix of source indices (1-based).
inverse

Map world coordinates back to source indices.

Value

A two-column numeric matrix.

Construct a native spectral dataset

Description

Construct a native spectral dataset

Usage

hm_dataset(values, observations, bands, geometry = NULL, provenance = list())

Arguments

values

Numeric observation-by-band matrix or y,x,wavelength array.
observations

Data frame with unique character observation_id; images also require image_id and acquisition_id.
bands

Band metadata with numeric wavelength in nm. Unknown wavelengths require availability='unavailable'.
geometry

Explicit origin and invertible homogeneous transform for images.
provenance

Reproducible source and calibration metadata.

Value

An HMDataset.

Declare the dual_stream_mhsi upstream candidate

Description

Declare the dual_stream_mhsi upstream candidate

Usage

hm_dual_stream_mhsi(task, parameters = list())

Arguments

task

An hm_task.
parameters

Explicit mode, provenance and preprocessing declarations.

Value

An hm_model_spec; upstream execution remains admission-gated.

Declare independent reference endpoints

Description

Declare independent reference endpoints

Usage

hm_endpoint_contract(references)

Arguments

references

Data frame with observation_id, endpoint and value. Missing values remain missing.

Value

Validated declaration, without claiming source truth.

Evaluate predictions against an explicitly aligned reference

Description

Evaluate predictions against an explicitly aligned reference

Usage

hm_evaluate(
  predictions,
  reference,
  estimand = "pooled",
  resampling = NULL,
  hierarchy = c("subject_id", "acquisition_id")
)

## S3 method for class 'hm_evaluation'
print(x, ...)

## S3 method for class 'hm_evaluation'
summary(object, ...)

Arguments

predictions

An hm_prediction.
reference

Data frame with observation_id in exactly prediction order and task outcome/reference columns. Reconstruction targets are supplied as the numeric matrix attribute target. For predictions in fitted recipe coordinates, bake this target with the fitted recipe and set attribute target_space to predictions$provenance$output_space.
estimand

Either pooled or equal_class_subject_acquisition.
resampling

Optional named argument list passed to hm_bootstrap().
hierarchy

Two column names giving biological unit then acquisition for hierarchical categorical accuracy.
x

An object returned by hm_evaluate().
...

Additional arguments; forwarded to graphics for plotting, otherwise unused.
object

An object returned by hm_evaluate().

Details

The hierarchical estimand averages observations within class/acquisition, acquisitions within class/subject, subjects within class, then classes equally. Missing class support produces unavailable macro metrics. Anomaly scores are not calibrated risks; reconstruction spectral angles are radians and refer to the declared target coordinates. Named matrix reference/output columns must match exactly; no reordering is inferred. Survival uses comparable event-time pairs. Evaluation does not fit models or repair reference alignment.

Value

An hm_evaluation list containing named numeric metrics, task-specific per-class support/confusion, predictions, reference, estimand, hierarchy and optional uncertainty.

The input object invisibly.

A list of descriptive fields.

Generate an explicitly synthetic grouped spectral example

Description

Generate an explicitly synthetic grouped spectral example

Usage

hm_example(seed = 42L)

Arguments

seed

Nonnegative integer seed.

Value

An HMDataset with synthetic labels, subject groups and eight bands.

Export grouped annotations through hyperannotatR

Description

Export grouped annotations through hyperannotatR

Usage

hm_export_annotations(x, destination, groups, ...)

Arguments

x

Annotation project or session.
destination

Export directory.
groups

Explicit biological grouping table accepted by at_training_export.
...

Additional public export arguments.

Value

Companion export including integrity checks.

Export a reproducible workbench result

Description

Export a reproducible workbench result

Usage

hm_export_result(result, path, include_private = FALSE)

Arguments

result

Completed fit, prediction, evaluation and configuration bundle.
path

New export directory, never an existing run.
include_private

Include fitted objects and original prediction identities. The default exports configuration, aggregate metrics and a reproduction script.

Details

Reference reproduction requires an explicitly supplied admitted backend and artifact_root; it retrains in an owned bounded job. Private reference exports include checked learned files. The generated read_private_fit() helper resolves their location after moving the export; use hm_predict with an owned matching backend_handle to verify and predict. Installed runtime receipts and cache paths are not portable exports. The destination is reserved before writing. Its completion marker appears only after reproduction code and any private artifacts are finalized. Interrupted exports retain a failed destination with partial contents for inspection; the destination cannot be read as a completed run.

Value

Verified export directory, invisibly.

Extract admitted native upstream image features

Description

Extract admitted native upstream image features

Usage

hm_extract_native(
  model_id,
  cube,
  wavelengths,
  mask,
  admission,
  control = list()
)

Arguments

model_id

Named candidate identity (htc, hyperfree or hypersigma).
cube

Numeric full-frame HWC array; spectral channels remain measured.
wavelengths

Strictly increasing measured wavelengths in nm.
mask

Logical full-frame eligible-pixel mask.
admission

Explicit exact source, checkpoint and separate rights admission.
control

Runtime list with an owned backend_handle and optional timeout.

Details

Only htc, hyperfree and hypersigma extraction are admitted by this entry point. The worker must verify its adapter-specific exact source/checkpoint/input contract. Admission requires execute_source_acknowledged, code_terms_admitted and weight_terms_admitted flags. Extraction is not upstream downstream-task training or scientific qualification; other named architectures remain blocked.

Value

An hm_native_features list with worker-returned numeric features, identity metadata, qualified=FALSE and input digest. Feature shape is adapter-specific.

Fit a workflow using training indices only

Description

Fit a workflow using training indices only

Usage

hm_fit(workflow, data, split, control = list())

## S3 method for class 'hm_fit'
coef(object, ..., scale = c("fitted", "original"))

## S3 method for class 'hm_fit'
print(x, ...)

## S3 method for class 'hm_fit'
summary(object, ...)

Arguments

workflow

An hm_workflow.
data

An HMDataset of native spectra, or list with numeric values matrix, observations data frame, band_map and optional provenance.
split

One explicit train/test fold, normally from hm_split(); indices are source row positions.
control

Engine-control list: seed defaults to one; weights may align to all source rows or training rows. Optional tuning contains candidates, metric and direction. Reference workers require backend_handle and a new artifact_path.
object

An hm_fit.
...

Arguments forwarded to the fitted engine coefficient method where supported.
scale

Either fitted coordinates or original measured-band coordinates.
x

An hm_fit.

Details

Preprocessing and engine fitting use training rows only. Connected biological/source identities cannot cross train/test; inner folds must remain inside outer training. Spatial images/sequences require explicit supported routes and are never silently flattened. R reference spectral training currently admits reference:cnn1d, reference:spectral_vit, reference:masked_autoencoder, reference:vae, reference:distillation, reference:multitask and reference:domain_adversarial with the appropriate task; worker-level tensor families have a separate protocol. A fitted model is not independent scientific qualification.

Fitted coefficient availability depends on the engine. Original-band mapping is restricted to lm, glm, glmnet, pls, pcr and bayes_conjugate with globally affine scale, pca, select, derivative, resample or reflectance recipes. Nonlinear/missingness-dependent recipes, aliased coefficients and incompatible dimensions raise typed errors. Coefficients describe a fitted coordinate system and do not establish biological causality.

Value

An hm_fit containing workflow, fitted recipe, engine, training indices, validated split, original band map, input radiometric unit and hash-bound provenance.

Engine coefficients with scale and link attributes. Original-scale output is a named intercept/band vector (or engine list of vectors), with band_map attached.

Print methods return the input invisibly.

Summary methods return model identity and training-row count.

Import an explicitly profiled TIVITA measurement list

Description

Import an explicitly profiled TIVITA measurement list

Usage

hm_from_tivis(
  x,
  profile,
  observations = .hm_image_observations(),
  assets = data.frame(),
  registration_id = NULL,
  mask = NULL
)

Arguments

x

Unclassed tivis_measurement result.
profile

Explicit hm_tivita_profile declaration.
observations

Explicit source/group identity.
assets

Asset inventory with role and registration_id columns.
registration_id

Expected registration identity.
mask

Optional logical y,x validity mask; NA remains unknown.

Value

HMDataset; no device calibration is inferred.

Declare the htc upstream candidate

Description

Declare the htc upstream candidate

Usage

hm_htc(task, parameters = list())

Arguments

task

An hm_task.
parameters

Explicit mode, provenance and preprocessing declarations.

Value

An hm_model_spec; upstream execution remains admission-gated.

Declare the hyperfree upstream candidate

Description

Declare the hyperfree upstream candidate

Usage

hm_hyperfree(task, parameters = list())

Arguments

task

An hm_task.
parameters

Explicit mode, provenance and preprocessing declarations.

Value

An hm_model_spec; upstream execution remains admission-gated.

Declare the hypersigma upstream candidate

Description

Declare the hypersigma upstream candidate

Usage

hm_hypersigma(task, parameters = list())

Arguments

task

An hm_task.
parameters

Explicit mode, provenance and preprocessing declarations.

Value

An hm_model_spec; upstream execution remains admission-gated.

Declare the hypervision upstream candidate

Description

Declare the hypervision upstream candidate

Usage

hm_hypervision(task, parameters = list())

Arguments

task

An hm_task.
parameters

Explicit mode, provenance and preprocessing declarations.

Value

An hm_model_spec; upstream execution remains admission-gated.

Install declared files into an immutable user runtime directory

Description

Install declared files into an immutable user runtime directory

Usage

hm_install_backend(
  plan,
  accept_terms = FALSE,
  offline = FALSE,
  acknowledge_custom_code = FALSE
)

Arguments

plan

Resolved backend plan.
accept_terms

Explicit acceptance bound to this plan's terms and artifacts.
offline

Forbid network downloads.
acknowledge_custom_code

Explicit permission to run reviewed custom code.

Value

An hm_backend list with installation path, original plan and installed_unverified state.

Inspect admitted instrument capabilities

Description

Inspect admitted instrument capabilities

Usage

hm_instrument_profiles()

Value

A data frame. Synthetic profiles are not device qualification.

Poll a modelling job

Description

Poll a modelling job

Usage

hm_job_status(job)

Arguments

job

Owned job handle.

Value

State, elapsed time, result only after successful completion, and error.

Declare a model specification

Description

Declare a model specification

Usage

hm_model_spec(model_id, task, parameters = list())

Arguments

model_id

Single model identifier from hm_models() or an explicitly admitted custom identity.
task

An hm_task declaration.
parameters

Named list of engine-specific parameters; interpreted when fitting.

Details

This constructor declares intent and checks known core task compatibility. It neither installs a backend nor qualifies a model scientifically.

Value

An hm_model_spec list containing model_id, task and parameters.

List candidate model families

Description

List candidate model families

Usage

hm_models(task = NULL, backend = NULL, include_unavailable = TRUE)

Arguments

task

Optional task filter.
backend

Optional backend filter.
include_unavailable

Include unqualified candidates.

Details

Registry visibility denotes candidate declaration or setup availability, not installed execution or independent parity. Reference families have separate spectral/tensor routes; not every listed identity is accepted by hm_fit(). All returned scientific qualification flags remain FALSE.

Value

List of model records with identity, supported task declarations, candidate backend, availability/admission status and scientific qualification.

Declare supplied physical forward and inverse kernels

Description

Declare supplied physical forward and inverse kernels

Usage

hm_physics_kernel(forward, inverse, parameters = list(), source, units)

Arguments

forward

Function of concentrations and a parameter list, producing spectra.
inverse

Function of spectra and a parameter list, producing concentrations.
parameters

Serializable physical constants and assumptions.
source

Source identifier for the implemented physical model.
units

Named units for input and output quantities.

Value

A declared hm_physics_kernel. Supplying a function does not validate its physiological assumptions; independent standards remain necessary.

Explicitly pool all valid spatial pixels

Description

Explicitly pool all valid spatial pixels

Usage

hm_pool(x, mask = NULL, method = "mean", selection = list())

Arguments

x

Spectral dataset accepted by hm_read_block().
mask

Optional logical array matching the block validity dimensions; no missing entries.
method

Pooling rule; currently only mean is supported.
selection

Named block selection passed to hm_read_block().

Details

All eligible finite pixels in the selected support contribute equally; an empty eligible support is rejected.

Value

List with a one-row band matrix values, band_map, and provenance including eligible pixel count, equal weights and mask hash.

Index a portable spectral sequence without repairing chronology

Description

Index a portable spectral sequence without repairing chronology

Usage

hm_portable_sequence(
  frames,
  frame_index,
  observations,
  bands,
  geometry,
  provenance = list()
)

Arguments

frames

Numeric cubes or admitted RDS locators.
frame_index

Explicit frame table; time is source elapsed seconds, including missing or reversed values.
observations

Observation identities.
bands

Native band metadata.
geometry

Source grid.
provenance

Source evidence.

Value

HMSequence retaining original ordering and timing quality.

Predict with a fitted workflow

Description

Predict with a fitted workflow

Usage

hm_predict(object, newdata, ...)

## S3 method for class 'hm_fit'
hm_predict(object, newdata, ...)

## S3 method for class 'hm_fit'
predict(object, newdata, ...)

Arguments

object

An hm_fit.
newdata

Native spectral data in the same form as hm_fit data, with exactly matching measured band semantics.
...

Engine prediction controls; reference workers require control with the owned backend_handle.

Details

Applies the frozen training recipe. Measured wavelength/unit/fwhm/native/canonical band metadata present during training must match exactly. Prediction does not fit preprocessing or use new outcomes. The declared radiometric unit must also match; calibration identities may differ between acquisitions. Older serialized fits without this unit contract must be refitted. Reconstruction after preprocessing is in fitted recipe coordinates, recorded as provenance$output_space. Evaluate it against a target baked with that recipe and the matching target_space reference attribute, never against untransformed spectra.

Value

An hm_prediction in newdata observation order, with task-specific values and optional class probabilities.

Validate model predictions

Description

Validate model predictions

Usage

hm_prediction(
  values,
  task,
  observation_id,
  probabilities = NULL,
  class_map = NULL,
  provenance = list()
)

## S3 method for class 'hm_prediction'
print(x, ...)

## S3 method for class 'hm_prediction'
summary(object, ...)

## S3 method for class 'hm_prediction'
plot(x, ...)

Arguments

values

Prediction vector, or observation-by-output matrix for multilabel, unmixing and reconstruction.
task

An hm_task.
observation_id

Unique nonmissing character identifiers in prediction row order.
probabilities

Optional numeric observation-by-class matrix in ⁠[0,1]⁠, with columns exactly matching class_map; rows sum to one except for multilabel tasks.
class_map

Exact character class/label order, required for categorical and multilabel predictions.
provenance

List of source, fit and transformation provenance.
x

An object returned by hm_prediction().
...

Additional arguments; forwarded to graphics for plotting, otherwise unused.
object

An object returned by hm_prediction().

Details

Categorical labels must belong to the declared map. Unmixing outputs are nonnegative and sum to one unless task reference explicitly declares nonnegative-only abundances. This validation does not establish predictive accuracy or calibrated probabilities. Numeric scalar tasks (including survival risk) require finite numeric vectors; matrix tasks require finite numeric matrices. Named matrix outputs retain column identity.

Value

An hm_prediction list retaining values, task, identities, probabilities, class map and provenance.

The input object invisibly.

A list of descriptive fields.

The graphics plotting return value, invisibly.

Fit preprocessing on training spectra

Description

Fit preprocessing on training spectra

Usage

hm_prep(recipe, training, wavelength = NULL)

Arguments

recipe

Unfitted hm_recipe.
training

Numeric observation-by-feature training matrix.
wavelength

Optional measured wavelength vector in nm, aligned to input columns; required by wavelength-dependent steps.

Details

Call only on the training partition. Selection, derivatives, resampling and PCA can change output dimensions; resampling marks the track as adapted.

Value

A fitted hm_recipe containing learned step state, input/output wavelength grids, feature count and training hash.

Declare auxiliary language-model prompt provenance

Description

Declare auxiliary language-model prompt provenance

Usage

hm_prompt(text, origin, rendering = NULL, timestamps = NULL, images = NULL)

Arguments

text

Prompt text.
origin

Origin of instructions, independent of held-out outcomes.
rendering

Rendering manifest for image content.
timestamps

Frame timestamps retained in order.
images

Optional PNG/JPEG data URLs from an explicitly manifested rendering.

Value

An hm_prompt list retaining text, origin, rendering, timestamps, optional images, auxiliary_only flag and content digest.

Capture explicit reproducibility identities

Description

Capture explicit reproducibility identities

Usage

hm_provenance(objects = list(), files = character(), resources = NULL)

Arguments

objects

Named list of R objects whose serialized identities are hashed.
files

Character vector of existing file paths whose bytes are SHA-256 hashed.
resources

Optional resource declaration retained in the receipt.

Value

An hm_provenance receipt with object/file hashes, R version, platform and resources.

Read a bounded spatial block

Description

Read a bounded spatial block

Usage

hm_read_block(x, selection = list())

## S4 method for signature 'HMDataset'
hm_read_block(x, selection = list())

Arguments

x

Spatial HMDataset or HMSequence.
selection

Named acquisition_id, frame_id, rows, cols and bands selectors. Out-of-grid rows/cols explicitly request invalid NA padding.

Details

Rows and cols are unique integer source-grid indices; out-of-grid values request NA padding marked invalid. Bands are valid positive 1-based indices. Select exactly one acquisition; sequence frames retain requested order. Variable grids must be read one frame at a time. Validity requires every selected band finite and any declared spatial mask TRUE. The byte check bounds planned output allocation, not total resident memory; RDS sources have a separate bounded deserialization admission check.

Value

Named list with a numeric four-dimensional values array (y, x, wavelength, time), a logical three-dimensional valid array (y, x, time), observations, band_map, geometry and provenance. Default output budget is 256 MiB (option hypermodR.block_bytes).

Read and verify a completed run

Description

Read and verify a completed run

Usage

hm_read_run(path)

Arguments

path

Run directory created by hm_save_run().

Details

Requires matching complete receipt, object SHA-256 and COMPLETE marker before reading the object. Read only trusted RDS sources.

Value

The deserialized original R object.

Declare preprocessing steps

Description

Declare preprocessing steps

Usage

hm_recipe(steps = list())

Arguments

steps

List of step lists, each with a type and its required controls; see Details.

Details

Steps execute in list order. Learned quantities are estimated only by hm_prep() on its supplied training matrix. Supported types are impute (training column means), scale (training means and standard deviations), pca (optional integer components), select (explicit unique integer bands), normalize (row Euclidean norm), snv (row mean/standard deviation), msc (training mean reference), reflectance (explicit dark and white band vectors), derivative (adjacent finite differences per nm), resample (explicit wavelength grid within measured support), and continuum (positive upper concave envelope division). Missing values must be imputed before steps requiring finite data.

Value

An unfitted hm_recipe with native input track.

Describe confidence bins and coverage-risk without claiming OOD validation

Description

Describe confidence bins and coverage-risk without claiming OOD validation

Usage

hm_reliability(predictions, reference, breaks = seq(0, 1, length.out = 11))

Arguments

predictions

Categorical hm_prediction with probabilities.
reference

Exactly aligned reference data frame accepted by hm_evaluate().
breaks

Strictly increasing numeric confidence-bin boundaries from zero to one.

Details

Confidence is the maximum class probability. Empty bins or zero retained coverage have unavailable accuracy/risk. Confidence does not constitute a validated OOD detector.

Value

List of bin counts/mean confidence/accuracy, coverage-risk table, evaluation scores and an OOD interpretation.

Render a declared three-band auxiliary image

Description

Render a declared three-band auxiliary image

Usage

hm_render_hsi(cube, wavelengths, bands, limits, source_id, timestamp = NULL)

Arguments

cube

Finite height by width by channel numeric array.
wavelengths

Measured wavelength vector.
bands

Three distinct channel indices.
limits

Explicit common intensity range used for clipping and scaling.
source_id

Source provenance identity.
timestamp

Optional acquisition time.

Value

An hm_rendering list with a height-by-width-by-3 image scaled to ⁠[0,1]⁠ and a hash-bound rendering manifest.

Resolve official binary metadata into an exact reviewable runtime manifest

Description

This explicit network operation resolves metadata only. It does not download model weights or install packages. Published terms and inherited distribution notices must be reviewed before converting the result into an install plan.

Usage

hm_resolve_backend(
  backend = c("reference_cpu", "vllm_linux_cuda"),
  model_id,
  python,
  uv = unname(Sys.which("uv")),
  directory = tempfile("resolution-", tmpdir = tools::R_user_dir("hypermodR", "cache"))
)

Arguments

backend

reference_cpu. The former vllm_linux_cuda setup route is blocked pending a reviewed replacement for its retired vulnerable profile.
model_id

Exact model identity or reference family.
python

Explicit existing Python 3.11 or 3.12 executable.
uv

Explicit uv resolver executable (version 0.11.26).
directory

External persistent resolution directory.

Value

An hm_backend_resolution containing exact artifact profile, package metadata, external directory, hard gates and pending qualification.

Resolve an immutable public Hugging Face checkpoint without downloading weights

Description

Resolve an immutable public Hugging Face checkpoint without downloading weights

Usage

hm_resolve_checkpoint(
  repository,
  revision,
  directory = tempfile("checkpoint-", tmpdir = tools::R_user_dir("hypermodR", "cache"))
)

Arguments

repository

Public model repository owner/name.
revision

Exact 40-character commit SHA.
directory

New external metadata staging directory.

Value

List containing repository/revision, architecture, local checkpoint directory name, artifact hashes/sizes/URLs, unreviewed terms and pending qualification.

Declare bounded execution resources

Description

Declare bounded execution resources

Usage

hm_resources(
  cpus = 1L,
  threads = 1L,
  memory_mb = 1024,
  device = "cpu",
  timeout = 60,
  network = FALSE
)

Arguments

cpus

Positive integer logical-CPU count selected from inherited allowed affinity.
threads

Positive integer numerical thread budget, no larger than cpus.
memory_mb

Positive finite per-process virtual address-space limit in MiB for bounded child execution.
device

Single device identifier. General bounded children require cpu. A separately admitted CUDA backend may select an explicit cuda:N device, subject to inherited visibility and verified hardware ABI.
timeout

Positive finite time budget in seconds.
network

Logical network-use policy declaration; network access is not isolated by this launcher.

Details

Bounded child routes currently require Linux taskset, prlimit and timeout. They apply inherited-affinity CPU subsets, numerical thread environment settings before startup, per-process RLIMIT_AS and an external elapsed timeout. Memory limits concern virtual address space, not aggregate descendant RSS; no additional CPU quota or network isolation is supplied. CUDA selection is admitted only by the backend launcher; hm_with_resources remains CPU-only. GPU memory is not limited, and admitting a device does not establish live GPU or scientific qualification. This is not a security sandbox.

Value

An hm_resources list of the validated declarations.

Save an immutable run with an atomic completion receipt

Description

Save an immutable run with an atomic completion receipt

Usage

hm_save_run(object, path)

Arguments

object

R object to serialize and verify by readback.
path

New run-directory path whose parent already exists.

Details

Reserves a new directory, writes object.rds and a hash-bound receipt, and atomically finalizes COMPLETE. Existing paths are rejected. A failed write leaves a failed receipt for inspection; completion markers are removed.

Value

Normalized run-directory path invisibly.

Construct a spectral sequence

Description

Construct a spectral sequence

Usage

hm_sequence(
  frames,
  frame_index,
  observations,
  bands,
  geometry,
  provenance = list()
)

Arguments

frames

List of native y,x,wavelength arrays or declarative RDS sources.
frame_index

Frame metadata with character frame_id, sequence_id, acquisition_id and numeric source time in elapsed seconds (NA allowed). timing_quality marks ordered, unavailable, duplicate or reversed source timing; source order is retained.
observations

Observation metadata linked by frame_id.
bands

Band metadata in nm.
geometry

Explicit spatial geometry.
provenance

Source and timing declarations.

Value

An HMSequence.

Read native spectra without spatial reduction

Description

Read native spectra without spatial reduction

Usage

hm_spectra(x, selection = list())

## S4 method for signature 'HMDataset'
hm_spectra(x, selection = list())

Arguments

x

HMDataset containing native spectra.
selection

Named list with character observation_id and positive bands.

Details

Only native spectral matrices are admitted; images require explicit spatial reduction. Bands are unique positive 1-based indices in requested order.

Value

Named list with numeric observation-by-selected-band values matrix, selected observations in request order, band_map and source provenance.

Declare an extinction or response library

Description

Values must be supplied from measured or otherwise justified sources. This constructor records a declaration; it cannot verify the physical reference.

Usage

hm_spectral_library(values, wavelength, units, source, license)

Arguments

values

Numeric wavelength-by-component matrix.
wavelength

Strictly increasing wavelength centres in nm.
units

Character description of coefficient units.
source

Character source citation or persistent identifier.
license

Character declaration of reuse terms.

Value

An immutable-by-identity hm_spectral_library list.

Declare the spectralearth upstream candidate

Description

Declare the spectralearth upstream candidate

Usage

hm_spectralearth(task, parameters = list())

Arguments

task

An hm_task.
parameters

Explicit mode, provenance and preprocessing declarations.

Value

An hm_model_spec; upstream execution remains admission-gated.

Declare the spectralgpt upstream candidate

Description

Declare the spectralgpt upstream candidate

Usage

hm_spectralgpt(task, parameters = list())

Arguments

task

An hm_task.
parameters

Explicit mode, provenance and preprocessing declarations.

Value

An hm_model_spec; upstream execution remains admission-gated.

Plan grouped resampling

Description

Plan grouped resampling

Usage

hm_split(
  data,
  task,
  design = "nested_grouped",
  seed = 1L,
  folds = 5L,
  inner_folds = 3L,
  ...
)

Arguments

data

Observation data frame, HMDataset, or list containing observations.
task

An hm_task declaring grouping columns.
design

One of grouped, nested_grouped or temporal.
seed

Seed for reproducible grouped assignment.
folds

Integer outer-fold count, at least two.
inner_folds

Integer inner-fold count for nested_grouped, at least two.
...

Temporal controls: time is the numeric timestamp-column name (default timestamp), cutoff is required, and nonnegative look_back/embargo use the same units as time.

Details

Groups connected through any task grouping column or known biological/source identities stay together; overlapping source windows are connected. Grouped assignment balances component counts rather than stratifying outcomes and rejects outer categorical folds without full class support. Temporal training precedes cutoff minus embargo and look-back; groups exposed at/after cutoff are removed entirely from training. Source times are not repaired or reinterpreted. Temporal results retain the selected time-column identity and its observations; fitting rejects a stale clock or altered partition. Regenerate the temporal split after changing timestamps.

Value

An hm_split. Grouped designs contain folds (train/test row indices and optional inner folds), component membership, observation identities and hash. Temporal designs contain train/test/excluded indices, cutoff, look-back and embargo.

Stage a bound integer prediction for human review

Description

Stage a bound integer prediction for human review

Usage

hm_stage_annotations(predictions, project, label_map, contract, model)

Arguments

predictions

Integer y,x mask. NA is unknown; zero is background.
project

Destination annot_project.
label_map

Named class codes.
contract

hm_annotation_contract captured before prediction.
model

Model name, version and sha256 provenance.

Value

Companion staged patch with unknown mask receipt; never committed.

Start a process belonging to an installed runtime

Description

Start a process belonging to an installed runtime

Usage

hm_start_backend(
  backend,
  model = backend$plan$model_id,
  resources = hm_resources(),
  host = "127.0.0.1"
)

Arguments

backend

Installed backend receipt.
model

Expected model identity.
resources

An hm_resources CPU declaration, or explicit cuda:N for a CUDA plan with verified hardware ABI; threads must be one.
host

Bind address, restricted to loopback.

Details

Starts the declared verified command through the Linux bounded child launcher described in hm_resources with one numerical thread. CUDA plans require explicit cuda:N selection, verified hardware ABI and a selected device within inherited CUDA_VISIBLE_DEVICES when set. CUDA memory is not limited by this launcher; CPU memory is per-process virtual address space, not aggregate RSS. Network remains a policy declaration. Admission is not live GPU qualification. Handles are valid only in their owning R session.

Value

Mutable owned hm_backend_handle environment containing process, PID, owner PID, model identity and plan digest.

Stop only an owned runtime process

Description

Stop only an owned runtime process

Usage

hm_stop_backend(handle)

Arguments

handle

Owned process handle.

Value

Final backend-status list invisibly.

Bounded causal frame streaming

Description

Bounded causal frame streaming

Usage

hm_stream(predict_frame, max_frames = 256L)

Arguments

predict_frame

Function accepting a frame and its timestamp.
max_frames

Maximum frames before explicit reset.

Value

Mutable stream handle, with per-frame outputs and latency diagnostics.

Cancel a frame stream

Description

Cancel a frame stream

Usage

hm_stream_cancel(stream)

Arguments

stream

An hm_stream.

Value

The cancelled stream invisibly.

Push ordered frames into a stream

Description

Push ordered frames into a stream

Usage

hm_stream_push(stream, frames, timestamps, sequence_id)

Arguments

stream

An hm_stream.
frames

List of native frames; geometry may vary.
timestamps

Strictly increasing finite elapsed seconds.
sequence_id

Explicit recording identity.

Value

Outputs in source order, with timestamp and latency metadata.

Submit an isolated, cancellable modelling job

Description

Submit an isolated, cancellable modelling job

Usage

hm_submit_job(
  workflow,
  data,
  split,
  control = list(),
  resources = hm_resources()
)

Arguments

workflow

Validated modelling workflow.
data

Spectral dataset.
split

One explicit grouped train/test fold.
control

Fitting controls.
resources

Execution resource declaration.

Details

Input object size is checked before launching a bounded Linux CPU child. The launcher applies an inherited-affinity subset, per-process virtual address-space limit, numerical thread settings and external elapsed timeout; see hm_resources for limits and network policy. Poll hm_job_status() to collect completion or failure. Handles belong to the creating R session.

Value

An owned job handle; poll with hm_job_status().

Declare a modelling task

Description

Declare a modelling task

Usage

hm_task(outcome, type, unit, groups, reference = NULL)

Arguments

outcome

Single outcome-column name, or NULL for clustering, anomaly, reconstruction or unmixing.
type

One of classification, regression, ordinal, multilabel, segmentation, unmixing, reconstruction, clustering, anomaly, survival or temporal.
unit

Observation unit: spectrum, pixel, roi, image, sequence or subject.
groups

Nonempty character vector of biological/source grouping column names.
reference

Optional reference declaration; survival uses time/status column names, unmixing may declare abundance_constraint, and temporal tasks may declare horizon.

Details

Segmentation requires pixel/image units; temporal tasks require sequence/subject units. Declaring a task does not establish an executable backend for every task/model combination.

Value

An hm_task list recording outcome, type, unit, unique grouping columns and reference.

Declare a causal temporal reference

Description

Declare a causal temporal reference

Usage

hm_temporal_model(family, task, parameters = list())

Arguments

family

Framewise, timestamp_mean, temporal_cnn, lstm, gru, transformer or state_space.
task

An hm_task.
parameters

Explicit sequence and training declarations.

Value

An hm_model_spec with a separate package reference identity.

Declare an explicit TIVITA array profile

Description

Declare an explicit TIVITA array profile

Usage

hm_tivita_profile(
  wavelengths = NULL,
  device = "unknown",
  software = "unknown",
  evidence = "unavailable"
)

Arguments

wavelengths

Independently established wavelengths in nm; NULL preserves unknown.
device

Device declaration, not qualification.
software

Export software declaration.
evidence

Evidence identifier.

Value

Serializable profile list.

Select parameters using only grouped inner training folds

Description

Select parameters using only grouped inner training folds

Usage

hm_tune(
  workflow,
  data,
  split,
  candidates,
  control = list(),
  metric = NULL,
  direction = c("minimize", "maximize")
)

Arguments

workflow

An hm_workflow.
data

Source data accepted by hm_fit().
split

Validated outer split containing inner grouped folds; their test rows must cover outer training exactly once.
candidates

Nonempty list of complete model-parameter lists.
control

Fit-control list; resources must declare one CPU/thread. Optional max_trials must cover all candidates and seed controls reproducibility.
metric

Metric name returned by hm_evaluate(), or NULL for categorical log_loss and otherwise rmse.
direction

Whether smaller or larger mean inner-fold scores are preferred.

Details

Search is sequential inside one bounded child, so the declared elapsed timeout covers the entire search and final refit. The first tied candidate wins. Outer test outcomes are excluded from selection.

Value

An outer-training hm_fit refitted with the winning parameters and a tuning ledger containing fold scores, identities, seeds and protocol hashes.

Verify source identities against a provenance receipt

Description

Verify source identities against a provenance receipt

Usage

hm_verify_provenance(receipt, objects = list())

Arguments

receipt

An hm_provenance receipt.
objects

Named objects corresponding exactly to those originally recorded.

Details

Rehashes the recorded file paths and supplied objects. Does not compare current R/platform metadata or establish source truth.

Value

TRUE invisibly on matching hashes; otherwise raises an integrity error.

Verify a signed registry release

Description

Verify a signed registry release

Usage

hm_verify_registry(
  path,
  signature,
  key,
  trusted_keys,
  revoked_keys = character(),
  minimum_version = 1L,
  now = Sys.time()
)

Arguments

path

Registry JSON file.
signature

Detached SHA256 RSA signature file.
key

PEM public key file provisioned by the maintainer.
trusted_keys

Allowlisted SHA256 hashes of public key files.
revoked_keys

Revoked key hashes.
minimum_version

Minimum allowed integer release sequence.
now

Verification time.

Value

An hm_registry_verification list with verified flag, file digest, trusted key fingerprint and release sequence.

Query an explicitly configured OpenAI-compatible vLLM endpoint

Description

Query an explicitly configured OpenAI-compatible vLLM endpoint

Usage

hm_vllm_chat(
  endpoint,
  model,
  prompt,
  allow_remote = FALSE,
  disclosure = NULL,
  token = NULL,
  timeout = 30,
  max_bytes = 1024^2,
  generation = list(temperature = 0, max_tokens = 256L),
  .transport = .hm_vllm_transport
)

Arguments

endpoint

Explicit server base URL.
model

Exact served model identity.
prompt

A provenance-bearing prompt.
allow_remote

Explicit transmission consent for non-loopback endpoints.
disclosure

Description of the data transmitted to a remote endpoint.
token

Optional authentication secret, excluded from receipts and errors.
timeout

Maximum seconds.
max_bytes

Maximum request/response bytes.
generation

Deterministic generation settings.
.transport

Internal transport override for protocol testing.

Details

The endpoint must include the API base path expected before /chat/completions (for example /v1). Remote transmission requires HTTPS, allow_remote=TRUE and disclosure. Stock native HSI architectures are rejected. Responses must identify the exact served model, contain one choice and finish with stop. Text is auxiliary output with no calibrated probability interpretation.

Value

List with response text, exact model, prompt digest/origin, rendering/timestamps, generation settings, remote/disclosure fields and auxiliary-only, uncalibrated semantics.

Run an R expression in a bounded child process

Description

Run an R expression in a bounded child process

Usage

hm_with_resources(resources, code)

Arguments

resources

An hm_resources declaration.
code

Unevaluated expression evaluated in a bounded child with its referenced serializable caller bindings.

Details

Evaluation occurs in a fresh child, so side effects on ordinary parent R objects are not propagated. Use serializable inputs and results; live handles are unsuitable. Linux child limits are described in hm_resources. Nested evaluation inside an already bounded child must reuse the identical outer budget. Unsupported platforms/devices fail admission rather than executing without limits.

Value

The serialized value of code returned from the child.

Exchange bounded typed arrays with an owned worker

Description

Exchange bounded typed arrays with an owned worker

Usage

hm_worker_request(
  handle,
  operation = "handshake",
  array = NULL,
  channels = NULL,
  timeout = 30,
  max_bytes = 16 * 1024^2,
  metadata = list()
)

Arguments

handle

Managed process handle.
operation

One of handshake, predict, extract, fit or cancel.
array

Numeric array; dimensions and channel metadata are explicit.
channels

Expected measured channel identifiers.
timeout

Maximum seconds.
max_bytes

Maximum request or response bytes.
metadata

Additional JSON metadata.

Value

Decoded worker response list; any array payload is reconstructed as a finite numeric R array in declared dimensions. Cancellation returns backend status invisibly.

Assemble a modelling workflow

Description

Assemble a modelling workflow

Usage

hm_workflow(task, recipe, model)

Arguments

task

An hm_task.
recipe

An unfitted hm_recipe.
model

An hm_model_spec whose task is identical to task.

Details

Assembly validates compatible declarations; fitting and backend admission occur separately.

Value

An hm_workflow list containing the three declarations.

Validated spectral data container

Description

Validated spectral data container

Slots

sources

Native arrays or declarative RDS locators, never live handles.

observations

Identity and observation metadata.

bands

Native band mapping; wavelengths are in nm.

geometry

Explicit spatial transform; empty for native spectra.

provenance

Source, radiometry and calibration declarations.

data_kind

One of spectra, image or sequence.

assets

Distinct asset-role metadata.

Validated time-indexed spectral sequence

Description

Validated time-indexed spectral sequence

Slots

frame_index

Explicit frame identity and elapsed seconds; gaps are retained.

Summarize fitted coefficients in a standard table

Description

Summarize fitted coefficients in a standard table

Usage

## S3 method for class 'hm_fit'
tidy(x, conf.int = FALSE, conf.level = 0.95, ...)

Arguments

x

An hm_fit.
conf.int

Include conditional confidence or posterior intervals.
conf.level

Interval level strictly between zero and one.
...

Additional arguments; currently unused.

Details

Frequentist uncertainty is implemented for LM and GLM. Other coefficient engines retain unavailable uncertainty as NA. Intervals do not include model selection, preprocessing selection or biological transport uncertainty. Use coef(x, scale = "original") for an admitted affine map.

Value

A data frame with term, estimate, std.error, statistic, p.value, and optional conf.low and conf.high. Estimates use the fitted recipe coordinates. Bayesian rows are explicitly posterior summaries.