Choosing a QC Entry Point

Three functions can apply a trained model to radar data. They differ in what you hand them and what workflow they fit. All three honor config.task_mode (convolution or legacy) and the multi-pass cascade. All three produce identical QC results for the same model and data — the choice is about control flow and what you get back, not accuracy.

QC_scan(config::ModelConfig) — the default entry point

Config-and-directory driven. It reads config.input_path (a directory or a single CfRadial) and internally calls composite_prediction(config; QC_mode=true), writing QC'd fields (<var><QC_SUFFIX> for each config.VARS_TO_QC) back into the files. Use this for the normal "I trained a model, now clean this dataset" case. This is what workflow/06_qc.jl uses. Start here unless you have a specific reason not to.

composite_prediction(config; …) — the analysis / evaluation engine

Same cascade inference as QC_scan, but it returns (predictions, verification, indexers, pass_probs) and only writes QC'd fields if you pass QC_mode=true. Reach for this when you want the predictions and probabilities in memory — building metrics, plotting met-prob distributions, comparing thresholds, or any programmatic analysis where you don't (yet) want to mutate the CfRadials. It also backs run_evaluation. Pass skip_existing_met_probs=true to reuse previously written met_prob_pass_<i> fields instead of recomputing features (much faster for re-scoring).

composite_QC(config, files::Vector{String}[, models]) — the streaming / operational entry point

Driven by an explicit vector of file paths, not config.input_path. It runs each cascade pass file-by-file, writing met_prob_pass_<i> and mask_pass_<i+1> between passes, then writes QC'd fields. The 2-arg form loads the random forests internally from config.model_output_paths; the 3-arg form takes preloaded models (load once, QC many files). This is the right tool for realtime / aircraft-side pipelines and any driver that already has a file list and a ModelConfig reloaded via load_config("model_config_<EXPERIMENT_NAME>.jld2"). See REALTIME/Ronin_realtime.jl for a worked operational example.

Decision shortcut

• Cleaning a directory you've pointed config.input_path at → QC_scan(config).
• Need predictions/probabilities in memory, or computing metrics → composite_prediction(config; …) (or run_evaluation).
• Operational/streaming loop over an explicit file list, models loaded once → composite_QC(config, files, models).