# Source and historical review of wholesale sector claims

Run `wholesale_sector_journey_review.py` or native R `run_wholesale_sector_journey_review()`. Both read the packaged positions and financial-release CSV files, reconcile their complete balances, join only historically available releases, and bind the actual claim IDs and model input coordinates used by the seven sector histories. The optional `data`, `accounts_path`, `releases_path` and `artifact_paths` replace those inputs; `return_artifacts` exposes complete audit inputs. R additionally accepts named `engine_paths` for the input_lineage and wholesale_sector_governance companion examples. `verify_reference` enables reference checks; source/target integrity checks always run.

These are separate synthetic observations, not model-generated labels. Historical model selection, policy compilation/application and event execution come from the existing sector governance helper; this review adds policy decomposition and two actual independent executions of the frozen historical registry. The validator receives their full states, actual applied event IDs, the retained identical duplicate delivery and dated model/calibration/policy/source snapshots. Changing one replayed component probability by 0.01 must yield BLOCKED. Events are stably ordered by knowledge time and original sequence before both executions.

External rating observations are explicitly absent, so rating benchmarking returns NOT_APPLICABLE. The support-stage check uses the actual standalone contractual default event; transfer/intersection and unsupported-counterfactual outcomes remain unavailable. It does not estimate a causal support benefit. The single observed default has Brier loss (1-p)^2. The sovereign model artifact and report retain the complete pre_transfer_sovereign_payment_default target. A separately declared adapter binds the same case, PIT conditioning, physical one-year horizon and pre-transfer unpaid-claim observation to the validator's generic default event label. Its evidence records the original unrestricted cashflow 120 versus debt service 130, payment due and receipt, and dated observed default. The validator's final slot means this same standalone pre-transfer event, never a transfer-adjusted or supported sovereign PD. Other sectors already have the generic default target. The explicit synthetic observation contract is supplied in the data rather than inferred from a model outcome.

For public-service finance, frozen raw and calibrated predictions are compared on the same 20 mature observed claims (two defaults), with two censored cases excluded. The metric hand oracle is 0.1*(1-p)^2 + 0.9*p^2. Model PIT conditioning remains intact; the validation design's performing label describes the origin risk set, not a conversion to TTC. Historical predictions are external to the fitting cohort. There is no actual PD override: an empty override ledger and unchanged reviewed forecasts are audited explicitly.

All calibration and model parameters remain frozen. Full Python/native reports, including every artifact and input fingerprint, must agree under the ordinary comparison contract. Default tests select sovereign and public-service cases; CREDITRATINGSYSTEM_EXTENDED_TESTS=true exercises all seven histories. The source reconciliation still covers all seven positions in either mode.
