Command-line reference

Installing the package installs the commands below ([project.scripts] in pyproject.toml). Each section reproduces the command’s --help output. The guides show how the commands fit together: Quickstart and Runs and outputs for training, Distributed training for several GPUs, Parameter sweeps for sweeps, Analyzing saved checkpoints for checkpoint analysis and the collectors, and Compact checkpoints for compact checkpoints.

Training

dendritic-train

Entry point: dendritic_modeling.scripts.training.train_experiments:cli_main

usage: dendritic-train [-h] [--output_dir OUTPUT_DIR] [--run_name RUN_NAME]
                       [--data_dir DATA_DIR]
                       [--learning_strategy {dfa,ei_equilibrium,fa,freeze_branch_kl,freeze_branches,freeze_layers,homeostatic_control,local_ca,multi_stage,presentation_stream,recurrent,soma_dfa,standard,voltage_stabilization}]
                       [--error_broadcast_mode {scalar,per_soma,per_soma_shared,per_soma_tree,per_soma_shuffled,local_mismatch,low_rank,path_transport,pathway_vector}]
                       [--seed SEED] [--epochs EPOCHS] [--lr LR] [--topk_lr TOPK_LR]
                       [--blocklinear_lr BLOCKLINEAR_LR]
                       [--reactivation_lr REACTIVATION_LR] [--decoder_lr DECODER_LR]
                       [--train_samples_per_class TRAIN_SAMPLES_PER_CLASS]
                       [--val_samples_per_class VAL_SAMPLES_PER_CLASS]
                       [--validate-only]
                       config

Run dendritic modeling training and evaluation.

positional arguments:
  config                Path to the configuration YAML file.

options:
  -h, --help            show this help message and exit
  --output_dir OUTPUT_DIR
                        Override output directory specified in config.
  --run_name RUN_NAME   Override run name specified in config.
  --data_dir DATA_DIR   Override data.base_dir (dataset root) specified in config.
  --learning_strategy {dfa,ei_equilibrium,fa,freeze_branch_kl,freeze_branches,freeze_layers,homeostatic_control,local_ca,multi_stage,presentation_stream,recurrent,soma_dfa,standard,voltage_stabilization}
                        Override learning strategy specified in config.
  --error_broadcast_mode {scalar,per_soma,per_soma_shared,per_soma_tree,per_soma_shuffled,local_mismatch,low_rank,path_transport,pathway_vector}
                        Override the LocalCA error-routing mode specified in config.
  --seed SEED           Override experiment seed (for running multiple seeds per
                        config).
  --epochs EPOCHS       Override the number of main-training epochs.
  --lr LR               Override the main trainer's fallback learning rate.
  --topk_lr TOPK_LR     Override the main trainer's sparse-synapse learning rate.
  --blocklinear_lr BLOCKLINEAR_LR
                        Override the main trainer's dendritic-coupling learning rate.
  --reactivation_lr REACTIVATION_LR
                        Override the main trainer's reactivation learning rate.
  --decoder_lr DECODER_LR
                        Override the main trainer's decoder learning rate.
  --train_samples_per_class TRAIN_SAMPLES_PER_CLASS
                        Bound ImageNet training samples per class (for a short check
                        run).
  --val_samples_per_class VAL_SAMPLES_PER_CLASS
                        Bound ImageNet evaluation samples per class (for a short check
                        run).
  --validate-only       Load and validate the config; build the datasets, model,
                        analysis manager, calibrated gates, optimizer and trainer;
                        write construction diagnostics, then exit without training.

Compact checkpoints

dendritic-compact-checkpoint

Entry point: dendritic_modeling.scripts.deployment.compact_checkpoint:main

usage: dendritic-compact-checkpoint [-h] [--compression-config COMPRESSION_CONFIG]
                                    [--verify-full-model] [--device DEVICE]
                                    [--report REPORT]
                                    experiment_config source_checkpoint
                                    output_checkpoint

Export a trained model with losslessly frozen sparse topology.

positional arguments:
  experiment_config     Training YAML used to construct the original model.
  source_checkpoint     File holding the model state dictionary, such as
                        final_model.pt.
  output_checkpoint     Compact checkpoint to write; it must not exist.

options:
  -h, --help            show this help message and exit
  --compression-config COMPRESSION_CONFIG
                        Standalone deployment-compression YAML.
  --verify-full-model   Compare deterministic full-model outputs before export and
                        after a fresh strict compact-checkpoint load.
  --device DEVICE       Device used for conversion and optional full-model
                        verification.
  --report REPORT       Atomically write the complete JSON resource and verification
                        ledger to this new file; an existing file is refused before
                        anything is written.

Sweeps

dendritic-sweep

Entry point: dendritic_modeling.scripts.sweeps.sweep_manager:main

usage: dendritic-sweep [-h] (--config CONFIG | --analyze)
                       [--generate-only | --run | --dry-run] [--split-jobs SPLIT_JOBS]
                       [--output-dir OUTPUT_DIR]
                       [--exclude-nodes EXCLUDE_NODES [EXCLUDE_NODES ...]]
                       [--slurm-account SLURM_ACCOUNT]
                       [--slurm-partition SLURM_PARTITION] [--results-dir RESULTS_DIR]

Unified Sweep Manager for Dendritic Modeling

options:
  -h, --help            show this help message and exit
  --config CONFIG       Path to sweep configuration file (for generation)
  --analyze             Analyze a finished sweep
  --generate-only       Generate configs only, don't submit jobs
  --run                 Generate configs and submit jobs (the default)
  --dry-run             Show what would be submitted without actually submitting
  --split-jobs SPLIT_JOBS
                        Run the configs as this many sequential jobs instead of one
                        array job
  --output-dir OUTPUT_DIR
                        Directory the sweep directory is created in (generation), or
                        the analysis output directory (--analyze)
  --exclude-nodes EXCLUDE_NODES [EXCLUDE_NODES ...]
                        List of nodes to exclude from job scheduling (e.g., --exclude-
                        nodes node1 node2)
  --slurm-account SLURM_ACCOUNT
                        Override slurm_config.account in sweep config (e.g.,
                        my_account)
  --slurm-partition SLURM_PARTITION
                        Override slurm_config.partition with one approved compute
                        partition
  --results-dir RESULTS_DIR
                        Sweep directory to analyze (with --analyze)

Examples:
  # Generate and run any sweep (default behavior)
  dendritic-sweep --config configs/my_sweep.yaml

  # Generate configs only (no job submission)
  dendritic-sweep --config configs/my_sweep.yaml --generate-only

  # Analyze a finished sweep
  dendritic-sweep --analyze --results-dir output/unified_sweep_20240101120000

  # Ten sequential jobs instead of one array job
  dendritic-sweep --config configs/my_sweep.yaml --split-jobs 10

  # Dry run (show what would be submitted)
  dendritic-sweep --config configs/my_sweep.yaml --dry-run

  # Override SLURM account/partition at submission time
  dendritic-sweep --config configs/my_sweep.yaml --slurm-account my_account --slurm-partition gpu_partition

dendritic-analyze

Entry point: dendritic_modeling.scripts.sweeps.result_analyzer:main

usage: dendritic-analyze [-h] [--type {ei,stoch,local_learning,branch,general,noise}]
                         [--output OUTPUT_DIR] [--list-types] [--verbose]
                         [results_dir]

Result Analyzer for Sweep Analysis

positional arguments:
  results_dir           Sweep results directory

options:
  -h, --help            show this help message and exit
  --type {ei,stoch,local_learning,branch,general,noise}, -t {ei,stoch,local_learning,branch,general,noise}
                        Analyzer to use (default: detected from the generated configs)
  --output OUTPUT_DIR, -o OUTPUT_DIR
                        Directory for the tables and plots (default:
                        <results_dir>/plots)
  --list-types          List available sweep types
  --verbose, -v         Log debug messages and print the traceback of a failed
                        analysis

dendritic-collect-sweep-metrics

Entry point: dendritic_modeling.scripts.sweeps.collect_sweep_metrics:main

usage: dendritic-collect-sweep-metrics [-h] --manifest MANIFEST
                                       [--output-dir OUTPUT_DIR]

Collect declared completed sweeps into checkpoint-level metric tables. This command is
the non-inferential boundary between standard sweep outputs and config-driven
checkpoint summaries. It uses the package ``DataCollector``, requires an explicit
expected size for every sweep, attaches only declared design labels, and records
hashes for all collected configs and held-out performance files. It never aggregates
across independently trained seeds.

options:
  -h, --help            show this help message and exit
  --manifest MANIFEST   Collection manifest (schema_version:
                        sweep_metric_collection_v1)
  --output-dir OUTPUT_DIR
                        Directory to write into (default: the manifest's output_dir)

Analysis

dendritic-analyze-checkpoint

Entry point: dendritic_modeling.scripts.analysis.run_checkpoint_analysis:cli_main

usage: dendritic-analyze-checkpoint [-h] --config CONFIG --checkpoint CHECKPOINT
                                    --analysis-config ANALYSIS_CONFIG --output-dir
                                    OUTPUT_DIR [--filename FILENAME] [--device DEVICE]
                                    [--no-strict]

Run config-declared analyses on one frozen model checkpoint. This entry point
reconstructs the model and canonical dataset splits from the original training config,
loads a frozen state dictionary, and executes a standalone analysis YAML through
:class:`~dendritic_modeling.analysis.AnalysisManager`. It never trains or modifies the
checkpoint.

options:
  -h, --help            show this help message and exit
  --config CONFIG       Resolved training YAML
  --checkpoint CHECKPOINT
                        Frozen model checkpoint
  --analysis-config ANALYSIS_CONFIG
                        Standalone YAML containing the analysis profile
  --output-dir OUTPUT_DIR
                        Directory the analyzers and the provenance record write into
  --filename FILENAME   Result name of the analyzers' output files
  --device DEVICE       For example cuda, cuda:0, cpu
  --no-strict           Allow missing or unexpected checkpoint keys and record them

dendritic-submit-checkpoint-analysis

Entry point: dendritic_modeling.scripts.analysis.submit_checkpoint_analysis:main

usage: dendritic-submit-checkpoint-analysis [-h] --manifest MANIFEST [--pilot]
                                            [--rerun] [--generate-only]
                                            [--account ACCOUNT]
                                            [--partition PARTITION]
                                            [--max-concurrent MAX_CONCURRENT]

Prepare and submit YAML-declared frozen-checkpoint analysis arrays. The manifest is
figure-agnostic: every scientific job is explicit, all inputs are hash-bound, complete
inventories are immutable, and retries receive new execution ledgers. This command
schedules only the standard package checkpoint analyzer; it does not train models or
compute paper-local quantities.

options:
  -h, --help            show this help message and exit
  --manifest MANIFEST   Checkpoint analysis manifest (checkpoint_analysis_manifest_v1)
  --pilot               Run only the manifest's pilot_job_ids
  --rerun               Also run jobs whose provenance and completion artifacts exist
  --generate-only       Write the job files and the submission record; submit nothing
  --account ACCOUNT     Slurm account (replaces the manifest's scheduler.account)
  --partition PARTITION
                        Slurm partition (replaces the manifest's scheduler.partition)
  --max-concurrent MAX_CONCURRENT
                        Array tasks running at once (a positive integer; default: the
                        manifest's scheduler.max_concurrent_jobs, else all pending
                        jobs)

dendritic-collect-representations

Entry point: dendritic_modeling.scripts.analysis.collect_representation_accessibility:cli_main

usage: dendritic-collect-representations [-h] [--config CONFIG] [--jobs JOBS]
                                         [--output-dir OUTPUT_DIR]
                                         [--expected-count EXPECTED_COUNT]
                                         [--allow-incomplete]

Validate and collect frozen-checkpoint representation analyses. The collector consumes
the JSONL job declarations used to run ``dendritic-analyze-checkpoint``. It verifies
every declared result against its run provenance, keeps technical projection/probe
repeats within checkpoints, and writes stable CSV tables for downstream statistics and
plotting.

options:
  -h, --help            show this help message and exit
  --config CONFIG       Collection configuration file; replaces the other options
  --jobs JOBS           JSONL job declaration; repeat for pilot and production files
  --output-dir OUTPUT_DIR
                        Directory to write the collected tables into
  --expected-count EXPECTED_COUNT
                        Number of declared jobs the collection must contain
  --allow-incomplete    Collect completed declarations while recording missing outputs

dendritic-collect-path-interventions

Entry point: dendritic_modeling.scripts.analysis.collect_path_matched_intervention:cli_main

usage: dendritic-collect-path-interventions [-h] [--config CONFIG] [--jobs JOBS]
                                            [--output-dir OUTPUT_DIR]
                                            [--expected-count EXPECTED_COUNT]
                                            [--expected-support-hash EXPECTED_SUPPORT_HASH]
                                            [--allow-incomplete]

Validate and collect frozen-checkpoint path-matched interventions. The collector
consumes the JSONL declarations used by ``dendritic-analyze-checkpoint``. It binds
every result to the declared configuration, checkpoint, and analysis profile; verifies
the fixed-count support and record grid; keeps individual support draws in an audit
table; and averages technical support draws within each trained checkpoint for
downstream inference.

options:
  -h, --help            show this help message and exit
  --config CONFIG       Collection configuration file; replaces the other options
  --jobs JOBS           JSONL job declaration; repeat for pilot and production files
  --output-dir OUTPUT_DIR
                        Directory to write the collected tables into
  --expected-count EXPECTED_COUNT
                        Number of declared jobs the collection must contain
  --expected-support-hash EXPECTED_SUPPORT_HASH
                        SHA-256 that the selected index supports of every checkpoint
                        must have
  --allow-incomplete    Collect completed declarations while recording missing outputs

dendritic-collect-path-information

Entry point: dendritic_modeling.scripts.analysis.collect_path_matched_information:cli_main

usage: dendritic-collect-path-information [-h] [--config CONFIG] [--jobs JOBS]
                                          [--output-dir OUTPUT_DIR]
                                          [--expected-count EXPECTED_COUNT]
                                          [--expected-support-hash EXPECTED_SUPPORT_HASH]
                                          [--allow-incomplete]

Validate and collect frozen path-matched dendritic information analyses. The collector
binds every result to the checkpoint-analysis JSONL declaration and strict run
provenance. It preserves anatomical support draws in an audit table and averages those
technical draws within each trained checkpoint before downstream model-level
inference.

options:
  -h, --help            show this help message and exit
  --config CONFIG       Collection configuration file; replaces the other options
  --jobs JOBS           Job file (JSON Lines) of the analyzed checkpoints; repeat for
                        several files
  --output-dir OUTPUT_DIR
                        Directory to write the collected tables into
  --expected-count EXPECTED_COUNT
                        Number of declared jobs the collection must contain
  --expected-support-hash EXPECTED_SUPPORT_HASH
                        SHA-256 that the selected index supports of every checkpoint
                        must have
  --allow-incomplete    Collect completed declarations while recording missing outputs

dendritic-collect-branch-sensitivities

Entry point: dendritic_modeling.scripts.analysis.collect_branch_local_sensitivity:main

usage: dendritic-collect-branch-sensitivities [-h] --config CONFIG

Validate and collect frozen branch-local sensitivity analyses. The collector binds
every record to the declared checkpoint-analysis job and its strict-loading
provenance. It preserves integration mode, soma-relative depth, and branch component
as explicit columns, so summaries never depend on JSON list ordering.

options:
  -h, --help       show this help message and exit
  --config CONFIG  Collection configuration file

dendritic-collect-local-i-interventions

Entry point: dendritic_modeling.scripts.analysis.collect_layer_population_interventions:main

usage: dendritic-collect-local-i-interventions [-h] --jobs JOBS --output-dir
                                               OUTPUT_DIR --expected-count
                                               EXPECTED_COUNT

Validate and collect frozen local-I population intervention analyses.

options:
  -h, --help            show this help message and exit
  --jobs JOBS           Job file (JSON Lines) of the analyzed checkpoints; repeat for
                        several files
  --output-dir OUTPUT_DIR
                        Directory to write the collected tables into
  --expected-count EXPECTED_COUNT
                        Number of declared jobs the collection must contain

dendritic-collect-checkpoint-analysis

Entry point: dendritic_modeling.scripts.analysis.collect_checkpoint_analysis:cli_main

usage: dendritic-collect-checkpoint-analysis [-h] [--config CONFIG] [--jobs JOBS]
                                             [--output-dir OUTPUT_DIR]
                                             [--require-analyzer REQUIRE_ANALYZER]
                                             [--expected-count EXPECTED_COUNT]
                                             [--metric-pattern METRIC_PATTERN]
                                             [--omit-full-results]
                                             [--derive-information-null-excess]
                                             [--input-hash-validation {recorded_hash_chain,rehash}]

Collect generic registered-analyzer outputs from frozen checkpoints. This collector is
deliberately schema-agnostic. It validates the checkpoint-run provenance, preserves
each complete result dictionary in JSONL, and emits a long table of finite numeric
leaves. It performs no across-checkpoint inference and never treats list elements,
branches, somas, or technical draws as replicates. Analyzers with dedicated technical-
draw collectors should continue to use them.

options:
  -h, --help            show this help message and exit
  --config CONFIG       Collection configuration file; replaces the other options
  --jobs JOBS           Job file (JSON Lines) of the analyzed checkpoints; repeat for
                        several files
  --output-dir OUTPUT_DIR
                        Directory to write the collected tables into
  --require-analyzer REQUIRE_ANALYZER
                        Analyzer whose results to collect; repeat for several
                        (default: every enabled analyzer)
  --expected-count EXPECTED_COUNT
                        Number of declared jobs the collection must contain
  --metric-pattern METRIC_PATTERN
                        Shell-style pattern of the dotted metric paths to keep, such
                        as '*accuracy*'; repeat for several (default: every numeric
                        value)
  --omit-full-results   Leave the complete result dictionaries out of the JSON Lines
                        file
  --derive-information-null-excess
                        Add the observed-minus-null value of every information term
                        that has a *_null counterpart
  --input-hash-validation {recorded_hash_chain,rehash}
                        rehash: hash the input files again; recorded_hash_chain:
                        compare the job file's *_sha256 values with the provenance

dendritic-collect-perturbation-response

Entry point: dendritic_modeling.scripts.analysis.collect_perturbation_response:cli_main

usage: dendritic-collect-perturbation-response [-h] --config CONFIG

Collect tidy accuracy responses from frozen-checkpoint perturbation analyses. The
registered noise and multiplicative-gain analyzers deliberately save their native
result dictionaries. This collector validates those outputs against an immutable
checkpoint-analysis ledger and aligns every perturbed accuracy with the clean accuracy
from the same checkpoint. It emits one row per checkpoint and perturbation level; it
performs no across-checkpoint aggregation.

options:
  -h, --help       show this help message and exit
  --config CONFIG  Collection configuration file

dendritic-collect-gain-load-response

Entry point: dendritic_modeling.scripts.analysis.collect_gain_load_response:cli_main

usage: dendritic-collect-gain-load-response [-h] --config CONFIG

Collect frozen-checkpoint shared-gain/load factorial responses. The registered
analyzer emits one record per technical perturbation draw. This collector validates
immutable checkpoint provenance, verifies the complete factorial and common-random-
number contract, averages draws within checkpoint, and emits tidy draw- and
checkpoint-level tables. It performs no inference across independently trained
checkpoints.

options:
  -h, --help       show this help message and exit
  --config CONFIG  Collection configuration file

dendritic-collect-reactivation

Entry point: dendritic_modeling.scripts.analysis.collect_reactivation_dynamics:cli_main

usage: dendritic-collect-reactivation [-h] [--config CONFIG] [--jobs JOBS]
                                      [--output-dir OUTPUT_DIR]
                                      [--expected-count EXPECTED_COUNT]
                                      [--require-structured-coordinates]
                                      [--require-clean-analysis-source]

Collect reactivation operating-state records from frozen checkpoints. The generic
checkpoint collector deliberately preserves analyzer dictionaries without interpreting
module names. This dedicated collector adds the standard population-network
coordinates needed for analyses over network position, population identity, and soma-
relative dendritic depth. It performs no across-checkpoint aggregation; every emitted
module and metric row remains bound to one declared checkpoint.

options:
  -h, --help            show this help message and exit
  --config CONFIG       Collection configuration file; replaces the other options
  --jobs JOBS           Job file (JSON Lines) of the analyzed checkpoints; repeat for
                        several files
  --output-dir OUTPUT_DIR
                        Directory to write the collected tables into
  --expected-count EXPECTED_COUNT
                        Number of declared jobs the collection must contain
  --require-structured-coordinates
                        Require gate module names of population_network layers
  --require-clean-analysis-source
                        Require every run's analysis to have run from a clean Git
                        checkout, and all runs from the same commit (the commit that
                        trained the checkpoints is not checked)

dendritic-collect-synaptic-support

Entry point: dendritic_modeling.scripts.analysis.collect_synaptic_support:cli_main

usage: dendritic-collect-synaptic-support [-h] [--config CONFIG] [--jobs JOBS]
                                          [--output-dir OUTPUT_DIR]
                                          [--expected-count EXPECTED_COUNT]

Collect exact sparse synaptic-support records from frozen checkpoints. The registered
weight analyzer emits one scalar record per soma, dendritic depth, and pathway, plus
NPZ arrays containing the corresponding source profiles. Optional source-region
alignment records connect those profiles to a declared reference split without using
model activity. This collector validates all representations and preserves their
structural coordinates. It performs no across-checkpoint aggregation and never treats
somas, depths, pathways, or source coordinates as independent replicates.

options:
  -h, --help            show this help message and exit
  --config CONFIG       Collection configuration file; replaces the other options
  --jobs JOBS           Job file (JSON Lines) of the analyzed checkpoints; repeat for
                        several files
  --output-dir OUTPUT_DIR
                        Directory to write the collected tables into
  --expected-count EXPECTED_COUNT
                        Number of declared jobs the collection must contain

dendritic-collect-source-tuning

Entry point: dendritic_modeling.scripts.analysis.collect_source_tuning_support:cli_main

usage: dendritic-collect-source-tuning [-h] [--config CONFIG] [--jobs JOBS]
                                       [--output-dir OUTPUT_DIR]
                                       [--expected-count EXPECTED_COUNT]
                                       [--result-subdirectory RESULT_SUBDIRECTORY]
                                       [--result-filename RESULT_FILENAME]
                                       [--input-hash-validation {recorded_hash_chain,rehash}]

Collect split-safe source-tuning/support records from frozen checkpoints. The
``source_tuning_support`` analyzer emits scalar records by fitted checkpoint, network
layer, target soma, dendritic depth, pathway, and native source block. This collector
validates the split and coordinate contracts, flattens scalar alignment and preference
records into CSV, and preserves the checkpoint-analysis hash chain. It performs no
aggregation across fitted checkpoints or trained seeds.

options:
  -h, --help            show this help message and exit
  --config CONFIG       Collection configuration file; replaces the other options
  --jobs JOBS           Job file (JSON Lines) of the analyzed checkpoints; repeat for
                        several files
  --output-dir OUTPUT_DIR
                        Directory to write the collected tables into
  --expected-count EXPECTED_COUNT
                        Number of declared jobs the collection must contain
  --result-subdirectory RESULT_SUBDIRECTORY
                        Analyzer directory, inside each job's output directory, to
                        read
  --result-filename RESULT_FILENAME
                        Result name to read in that directory
  --input-hash-validation {recorded_hash_chain,rehash}
                        rehash: hash the input files again; recorded_hash_chain:
                        compare the job file's *_sha256 values with the provenance

dendritic-collect-input-region-interventions

Entry point: dendritic_modeling.scripts.analysis.collect_input_region_intervention:cli_main

usage: dendritic-collect-input-region-interventions [-h] [--config CONFIG]
                                                    [--jobs JOBS]
                                                    [--output-dir OUTPUT_DIR]
                                                    [--expected-count EXPECTED_COUNT]
                                                    [--input-hash-validation {recorded_hash_chain,rehash}]

Collect frozen-checkpoint input-region intervention summaries. This collector consumes
``input_region_intervention/final.json`` artifacts from standard checkpoint-analysis
jobs. It emits separate checkpoint-bound tables for classification, aggregate
responses, per-class responses, assigned-soma responses, and assigned-soma per-class
responses. It validates that results contain aggregate sufficient statistics only and
performs no aggregation over fitted checkpoints or trained seeds.

options:
  -h, --help            show this help message and exit
  --config CONFIG       Collection configuration file; replaces the other options
  --jobs JOBS           Job file (JSON Lines) of the analyzed checkpoints; repeat for
                        several files
  --output-dir OUTPUT_DIR
                        Directory to write the collected tables into
  --expected-count EXPECTED_COUNT
                        Number of declared jobs the collection must contain
  --input-hash-validation {recorded_hash_chain,rehash}
                        rehash: hash the input files again; recorded_hash_chain:
                        compare the job file's *_sha256 values with the provenance

dendritic-summarize-checkpoint-metrics

Entry point: dendritic_modeling.scripts.analysis.summarize_checkpoint_metrics:main

usage: dendritic-summarize-checkpoint-metrics [-h] --config CONFIG
                                              [--output-dir OUTPUT_DIR]

Create seed-level summaries from collected checkpoint metric tables. This command is
the inferential boundary between package collectors and render-only publication code.
It requires explicit grouping and experimental unit columns, rejects duplicate units,
and computes deterministic bootstrap intervals across independently trained
checkpoints. Optional paired contrasts reuse the package's standard paired-comparison
implementation.

options:
  -h, --help            show this help message and exit
  --config CONFIG       Summary configuration file
  --output-dir OUTPUT_DIR
                        Directory to write into (default: the configuration's
                        output_dir)

dendritic-select-validation-epochs

Entry point: dendritic_modeling.scripts.analysis.select_validation_epochs:cli_main

usage: dendritic-select-validation-epochs [-h] --config CONFIG
                                          [--output-dir OUTPUT_DIR]

Select training epochs with validation metrics and report held-out metrics. The
command consumes the long-form trajectory emitted by ``dendritic-collect-sweep-
metrics``. Epochs are selected independently within each trained checkpoint using
validation metrics only. Metrics from the selected epoch, including held-out test
metrics when declared, are then exported for seed-level summarization. Test metrics
are never eligible as selection criteria.

options:
  -h, --help            show this help message and exit
  --config CONFIG       Selection configuration file (validation_epoch_selection_v1)
  --output-dir OUTPUT_DIR
                        Directory to write into (default: the configuration's
                        output_dir)

dendritic-poisson-exposure-transfer

Entry point: dendritic_modeling.scripts.analysis.run_poisson_exposure_transfer:cli_main

usage: dendritic-poisson-exposure-transfer [-h] --config CONFIG --checkpoint
                                           CHECKPOINT --transfer-config
                                           TRANSFER_CONFIG --output-dir OUTPUT_DIR
                                           [--device DEVICE] [--no-strict]

Evaluate one frozen Poisson-trained checkpoint across exposure durations. The command
reconstructs the model from its resolved training YAML, loads the checkpoint strictly,
rebuilds the canonical test split at each declared exposure, and delegates all
scientific evaluation to the package ``PoissonExposureTransferAnalyzer``. It never
retrains the model.

options:
  -h, --help            show this help message and exit
  --config CONFIG       Resolved training configuration of the run (its config.json)
  --checkpoint CHECKPOINT
                        Model checkpoint to evaluate
  --transfer-config TRANSFER_CONFIG
                        Transfer configuration file (durations, draws, seeds)
  --output-dir OUTPUT_DIR
                        Directory to write the results into
  --device DEVICE       For example cuda, cuda:0, cpu (default: cuda when available)
  --no-strict           Allow missing or unexpected checkpoint keys and record them

dendritic-collect-poisson-transfer

Entry point: dendritic_modeling.scripts.analysis.collect_poisson_exposure_transfer:cli_main

usage: dendritic-collect-poisson-transfer [-h] --jobs JOBS --output-dir OUTPUT_DIR
                                          [--expected-count EXPECTED_COUNT]

Collect Poisson exposure-transfer outputs at the trained-checkpoint level.

options:
  -h, --help            show this help message and exit
  --jobs JOBS           Job file (JSON Lines) of the transfer runs; repeat for several
                        files
  --output-dir OUTPUT_DIR
                        Directory to write the collected tables into
  --expected-count EXPECTED_COUNT
                        Number of declared jobs the collection must contain

Configuration tools

dendritic-example

Entry point: dendritic_modeling.scripts.tools.export_example:main

usage: dendritic-example [-h] [--output OUTPUT]
                         [{adding_problem_gru,mnist_population,recurrent_decision_population,structured_tree_synthetic,synthetic_population}]

List or export a small installed DendriNet example.

positional arguments:
  {adding_problem_gru,mnist_population,recurrent_decision_population,structured_tree_synthetic,synthetic_population}
                        Example to export. Omit to list examples.

options:
  -h, --help            show this help message and exit
  --output OUTPUT       Write a new YAML file.

dendritic-upgrade-config

Entry point: dendritic_modeling.scripts.tools.upgrade_config:main

usage: dendritic-upgrade-config [-h] (--output-dir OUTPUT_DIR | --in-place)
                                [--report REPORT]
                                configs [configs ...]

Rewrite experiment or sweep YAML files written for earlier library versions.

positional arguments:
  configs               configuration files to upgrade

options:
  -h, --help            show this help message and exit
  --output-dir OUTPUT_DIR
                        write upgraded copies here, at their paths relative to the
                        inputs' common directory
  --in-place            overwrite the files (comments are lost)
  --report REPORT       write the change records as JSON