Part of the baseline artifact pipeline: precompute the single-year US datasets (2026, 2027, 2025) and the 20 per-cohort national baselines per year as part of build/deploy, cache them content-addressed in GCS, bake them into the Modal image, and have the runtime load baselines instead of computing them.
Under the segmented-national default (#637), every plain US national macro request fans out to 20 fixed state-cohort children, and each child's baseline half is identical across all default-baseline requests — recomputed every time. Separately, the dataset prebuild layer's cost is paid through a manual prewarm ritual before pin-bump merges.
Phase 1 (this issue):
artifact_keys: canonical-JSON sha256 keys digesting the full input closure (versions, receipt content sha, certification fingerprint); deterministic Simulation ids; store layout. Keys carry country/region/scope so UK or state-level baselines need no schema change later.artifact_store: GCS client with write-once uploads, content-addressed manifests, deployed-environment markers.- Runtime reader: qualifying baselines get deterministic ids;
ensure() loads a baked artifact when present; loads are validated column-by-column against resolve_entity_variables and fall back to run() on any gap. No artifacts present -> today's behavior exactly.
precompute_app: ephemeral modal run app that plans against the store, computes only misses (wave 1 datasets, wave 2 cohort baselines via the executor's own request path), runs a determinism gate, and publishes the deploy manifest.
Later phases (separate issues): CI precompute job in the deploy pipeline; image fetch layer replacing prebuild_country_datasets (retiring the prewarm ritual and force_build); store/app GC; first measured pin-bump deploy.
Phase 1 gate before merge: full suite green; manual staging run of the precompute app populating the store (needs the artifact bucket provisioned + POLICYENGINE_ARTIFACT_BUCKET set), with a measurable no-op on immediate re-run.
Part of the baseline artifact pipeline: precompute the single-year US datasets (2026, 2027, 2025) and the 20 per-cohort national baselines per year as part of build/deploy, cache them content-addressed in GCS, bake them into the Modal image, and have the runtime load baselines instead of computing them.
Under the segmented-national default (#637), every plain US national macro request fans out to 20 fixed state-cohort children, and each child's baseline half is identical across all default-baseline requests — recomputed every time. Separately, the dataset prebuild layer's cost is paid through a manual prewarm ritual before pin-bump merges.
Phase 1 (this issue):
artifact_keys: canonical-JSON sha256 keys digesting the full input closure (versions, receipt content sha, certification fingerprint); deterministic Simulation ids; store layout. Keys carry country/region/scope so UK or state-level baselines need no schema change later.artifact_store: GCS client with write-once uploads, content-addressed manifests, deployed-environment markers.ensure()loads a baked artifact when present; loads are validated column-by-column againstresolve_entity_variablesand fall back torun()on any gap. No artifacts present -> today's behavior exactly.precompute_app: ephemeralmodal runapp that plans against the store, computes only misses (wave 1 datasets, wave 2 cohort baselines via the executor's own request path), runs a determinism gate, and publishes the deploy manifest.Later phases (separate issues): CI precompute job in the deploy pipeline; image fetch layer replacing
prebuild_country_datasets(retiring the prewarm ritual andforce_build); store/app GC; first measured pin-bump deploy.Phase 1 gate before merge: full suite green; manual staging run of the precompute app populating the store (needs the artifact bucket provisioned +
POLICYENGINE_ARTIFACT_BUCKETset), with a measurable no-op on immediate re-run.