# Dated investment and counterparty governance

`examples/structured_governance.py` and native R `run_structured_governance()` execute the governance stages for securitization issues, covered-bond issues, a receivables portfolio and a counterparty obligor. Each has an explicit subject type, product route, applicability bounds, dated frozen model, assignment, independent thirty-day contractual-payment-failure target and versioned investment or counterparty-limit policy. The source product recipe assessment is referenced by content hash. Its loss horizon and values are not silently reinterpreted as these separate thirty-day monitoring forecasts; counterparty PD remains at obligor level.

These complete synthetic analytical fixtures are intended for inspection and replacement with bank records. The frozen logit model gives probability 0.5 at the observed feature zero. It is a supplied reference model, not a fitted claim about these markets. The January 15 assessment carries forward that unchanged, still-supported January 1 feature under the same model. No outcome label becomes a feature.

Origination builds typed subjects, routes and checks applicability, selects the registered model, scores and composes an assessment. Policy specification 1.1 requires documented, dated nonapplicability of household affordability and verified legal evidence, while preserving eligibility and operational gates. With all explicit evidence the local recommendation is APPROVE; removing legal verification changes it to REFER. No recommendation executes a trade, limit or loan.

During performance and review, age and newly known payment-delay triggers control freshness. The event becomes visible only on its knowledge date and affects only its dependency-linked assessment; replaying the same event is a duplicate. A four-position out-of-time synthetic cohort has two contractual defaults, giving observed rate 0.5 and Brier score 0.25 for predictions 0.5. Negative outcomes require complete-through declarations and all performance metrics exclude immature horizons. A separately declared exchangeable feature-drift reference has KS distance 1 and the exact supplied permutation reference p-value 9/25. It is an analytical drift example, not evidence that the four positions establish a deployable drift threshold.

The sequential calibration routine consumes the actual zero cohort calibration error. Its explicitly degenerate zero-variation null and validation paths only demonstrate a zero-boundary arithmetic reference; they are not a market-specific stochastic calibration and must be replaced for operational surveillance. They do not trigger an automatic model change.

Exit adds a dated sale event. The unfinished forecast is censored; a forecast made after sale is outside the risk set and cannot become a nondefault observation. A separate main claim assessment is finalized and superseded on May 1 settlement, preserving both previous content hashes; the sold cohort control is a different entity. Registering a later model with a different intercept does not change exact replay of the original version-pinned request.

This stage provides origination, performing and review governance, plus event, risk-set exit and replay evidence. Survival exit curves, delinquency transitions, restructuring/workout economics and portfolio backtests are separate lifecycle components; this example does not claim all eight phases are complete.
