gaspatchio-model-scenarios

Build a ScenarioRun with a shocks dict, base_tables, and aggregations, then call plan.run(af, model_fn, audit=True). Keep model.py unchanged and put all scenario logic in run_scenarios.py, bridging shocked tables through the model_fn wrapper's assumptions_override.

Declare shocks such as MultiplicativeShock(factor=1.2, table="mortality_select") in the ScenarioRun plan. The plan stacks base tables with a scenario_id dimension and applies shocks per batch, so the model reads shocked values through a normal Table.lookup call.

Yes, build the shocks dict with a comprehension over both parameter grids, for example mortality × lapse multipliers, and hand it to ScenarioRun. Pivot the resulting aggregation DataFrame into a heatmap to visualize the 2D interaction.

Scenario-stacked tables require scenario_id as a lookup kwarg, but standalone runs use unstacked tables. Use the _maybe_scenario guard to pass scenario_id only when the table actually has that dimension, keeping the model debuggable under gspio run-single-policy.

plan.source_sha() gives a content hash over shocks, base tables, aggregations, and seed, plan.to_yaml() round-trips flat shock recipes, and plan.run(audit=True) writes a JSON audit sidecar. The report must pin the SHA and reference the sidecar path.

Use with_scenarios only for quick interactive exploration or pure interest-rate cross-joins. It provides no plan SHA, audit sidecar, or YAML round-trip, so promote the analysis to ScenarioRun once it settles.