Aligned to your shop. The physics stays put.
ARCNM resolves every physics input through a strict six-tier parameter cascade — first match wins, with no silent defaults — and aligns to a source's real costs at one clamped posterior, the environment-scoped cost offset. A handful of known costs is enough to finetune in a single pass, while the extractor, classifier, planner and physics stay industry-standard and untouched.
A six-tier parameter cascade
Every parameter resolves in order: environment overrides, inherited parent overrides, the environment-calibrated posterior, the platform-calibrated posterior, platform defaults, then hashed literature seeds. The first tier with a value wins, every resolution carries its provenance, and a miss raises ParameterUnresolved rather than a silent zero.
Alignment at one clamped posterior
Cost alignment happens only at the env-scoped cost-offset posterior, clamped to a safe band of 0.5 to 2.0. A multiplier outside that band signals real physics drift and is logged and suppressed, never quietly applied — so accuracy never means a distorted model.
One regression, two scopes
The same multilevel regression runs platform-wide, fitting cross-tenant priors from the whole corpus, and per environment, refining those priors into a tenant-private posterior for one plant, supply area or supplier — without leaking between tenants.
Calibrated uncertainty
Every source carries calibrated (conformal) uncertainty that tightens as its evidence accrues. Online drift detection flags when actuals start to wander, and oracle agreement with a wider-of rule keeps the interval honest instead of overconfident.