Campaign accounting

With campaign_mode: true, each robustness course closes with a recorded outcome. The workflow revalidates the CLI receipt before recording the outcome in _campaign_ledger/records. Both local and container rules use this closure. A missing or invalid receipt stops the job.

A robustness extraction failure records readable source bindings and digest-bound source errors, including malformed contracts or RTSTRUCTs. The campaign closes that course as failed and proceeds to later courses. Aggregation retains the failed course in the denominator.

Robustness outcome	Campaign status	Measurement contribution
measured, without selection gaps	ok	Published measurement rows
measured, with selection_gap_targets	technically_missing	Published measurement rows
source_only_nonvolumetric	ok	Zero
not_applicable_no_target_structure	not_applicable	Zero
selection_gap_targets_unselected	technically_missing	Zero
not_applicable_no_planning_ct	not_applicable	Zero
failed_extraction	failed	Zero

radiomics_robustness.target_selection: patterns is the default and preserves case-insensitive whole-name fnmatch selection. targets adds every detected target to the configured patterns. Unknown modes are rejected.

The governed ROI class map and feature policy decide first. Mapped targets qualify. Mapped OARs, OAR subvolumes and planning helpers do not qualify, even when their names or interpreted types contain target labels. For example, CTVN_Prostate, GTVm - lopatka, GTVm_SUMA and GTVn4p are targets; Jelita w PTV, Jelita + PTV1, 2cm od PTV1, z_PTV1 and PTV1-poprzednio are not.

For unmapped names, the detector accepts a GTV, CTV, PTV or ITV token at the start or after a space, hyphen, underscore, dot or slash. A single p, n or m qualifier may precede the end, digits or a separator. GTVp, CTVn, PTVn1 and HR-CTV qualify; CTVx and GTVnode do not. The interpreted types GTV, CTV, PTV and ITV also qualify unmapped names. TREATED_VOLUME alone does not qualify.

A target must carry volumetric contour data or nonzero voxels. Template slots, POINT markers and zero masks remain in the source inventory but are not missed targets. Inventory-only helpers cannot enter the measurement base. In targets mode, apply_to_structures: [] requests detected targets alone.

Classification uses the source label. Canonical AutoRTS names use the TotalSegmentator class map; Custom Boolean structures use recorded configuration provenance. A same-named Custom ROI is copied support only when a standard source declares that name with contours. A contourless template slot cannot hide a contoured Custom target.

When selection matches nothing, the detector checks every evaluated source structure set. No detected target gives not_applicable_no_target_structure. A detected but unselected target gives selection_gap_targets_unselected. Both close the stage without aborting the workflow. A selection gap remains technically missing in the ledger, rollup, cohort admission and aggregation. It never counts as complete or not applicable; measured rows remain available.

No-match receipts record the mode, requested patterns and missed target names with their deciding rules. Default patterns mode with no exclusions preserves historical configuration hashes and measured columns. Non-default selection settings enter the effective configuration and measured parameter hashes, with extraction_configured_parameter_hash retaining the extractor hash. The separate target inventory records the evaluated source evidence on every CT path.

Rollup accepts robustness success only from a validated JSON receipt. Plain-text disabled and failed remain terminal states; other text is missing evidence. A newer validated receipt supplies the current status and detail. Metadata from a superseded ledger record, including its log and duration, is cleared. Rollup reports technically_missing_unit_count and the units with their reasons.

The aggregate workbook retains each admitted course in course_outcomes, with accounting_status, technically_missing, target_selection and nonmeasurement_reason. course_counts separates measured, technically missing, not-applicable and source-only courses. No-match courses contribute no raw measurement rows or ICC, CoV or QCD estimates. Evaluated source dispositions remain in source_dispositions.

Receipts carrying not_applicable_selection_unmatched are stale and require recomputation; startup moves only these receipts into _ROBUSTNESS_RECEIPT_QUARANTINE, retaining their bytes and an identity record. Migration is cached per resolved output directory and dependency content. Shared configuration mtimes and valid course receipts remain unchanged.

Campaign aggregation applies campaign_require_all_courses and campaign_min_completion_fraction (default 0.5). Upstream-failed courses stay in the manifest denominator. Tightening these gates reruns aggregation without rerunning per-course extraction. The Python API applies the same gates as the CLI.

Changes to the mode, patterns, source bytes, code or receipt bindings require a new run. Stage completion alone does not establish clinical or manuscript readiness.