# Corporate and commitment historical source review

Run the Python example or native `run_wholesale_sme_journey_review()`. Optional `data`, `accounts_path`, `releases_path`, `corporate_data`, `commitment_data`, `observed_data` and named `artifact_paths` replace their corresponding packaged inputs. `verify_reference` controls hand-oracle checks, while input bindings and frozen inference checks remain mandatory. `return_artifacts` includes complete historical registries, requests, results and replay-validator inputs. Native `engine_paths` optionally names input_lineage, wholesale_corporate_lifecycle, wholesale_sector_governance and _wholesale_journey_tools.

The three separate populations are original corporate CO-001/claim-1 (365days), the original funded working-capital revolving facility (30days), and covenant-lite leveraged facility (731days including the leap day). Commitment identifiers in claim_ids refer to their original assessment facility IDs, not newly invented portfolio claims. Source balances are actual corporate80, active zero-drawn revolving0 and initial leveraged120; conditional future EAD and later actual default draw remain separate quantities. Historical source releases are read, joined, reconciled and checked against the exact financial input used by each pinned model.

The corporate resource freezes exactly the existing corporate-lifecycle fit, rather than introducing a new model. Original raw fits remain independently available in the existing lifecycle example. The same frozen artifact is used in both languages to preserve exact historical hashes. Its raw0.1 prediction is compared with the previously recorded0.1 reference forecasts: equality is an honest result, not evidence of improvement and not a calibration claim. Revolving and leveraged pipelines use their existing real calibration artifacts, including the revolving group argument, and independently verify every recorded cohort forecast against pinned inference.

Paired benchmark and override monitoring use the actual separate observed outcomes: 2/20 corporate, 2/20 revolving and6/20 leveraged defaults. The two additional commitment observations remain censored. The hand Brier formula is d/n*(1-p)^2+(1-d/n)*p^2. The validation label performing describes the origin risk set; original PHYSICAL/PIT model targets and horizons remain intact. No actual PD override was recorded, so the ledger is empty and reviewed forecasts are unchanged. Neither synthetic outcomes nor candidate predictions are passed off as causal evidence.

Actual existing histories select the effective model and retain frozen requests. Policies are compiled, applied and decomposed on their actual gate inputs. Two fresh historical executions are checked for full-state replay consistency; changed probability fails, identical duplicate deliveries are applied once. Corporate calibration snapshot explicitly says none applied. Default tests run the corporate history; CREDITRATINGSYSTEM_EXTENDED_TESTS=true adds both commitment histories. Full example execution always runs all three. Public-service SME coverage is in wholesale_sector_journey_review.
