Problem
The current calibration setup can let high-cardinality or grouped targets dominate the L0 objective, while core population anchors such as household and person counts do not have explicit precedence or tolerance enforcement. That makes it hard to distinguish targets that must fail the build when missed from targets that should warn while we tune downstream behavior.
Proposed work
Add an epsilon-insensitive calibration target policy that resolves per-target loss weights, tolerances, enforcement levels, and provenance artifacts. The pipeline should pass those arrays into the L0 fit, hard-fail only required targets that exceed tolerance, warn on softer targets such as ACA targets, and publish target-policy artifacts through the stage outputs/manifest so each build can be interrogated.
Validation
Use focused unit tests for target-policy resolution, L0 argument passing, derived population anchors, artifact collection, and Stage 2 contract summaries. Also provide a small synthetic demo script for comparing the new loss mode against the prior grouped relative-loss behavior outside the default test suite.
Problem
The current calibration setup can let high-cardinality or grouped targets dominate the L0 objective, while core population anchors such as household and person counts do not have explicit precedence or tolerance enforcement. That makes it hard to distinguish targets that must fail the build when missed from targets that should warn while we tune downstream behavior.
Proposed work
Add an epsilon-insensitive calibration target policy that resolves per-target loss weights, tolerances, enforcement levels, and provenance artifacts. The pipeline should pass those arrays into the L0 fit, hard-fail only required targets that exceed tolerance, warn on softer targets such as ACA targets, and publish target-policy artifacts through the stage outputs/manifest so each build can be interrogated.
Validation
Use focused unit tests for target-policy resolution, L0 argument passing, derived population anchors, artifact collection, and Stage 2 contract summaries. Also provide a small synthetic demo script for comparing the new loss mode against the prior grouped relative-loss behavior outside the default test suite.