Report weighted target loss attribution in calibration fit

[anth-volk]

Summary

• consume authoritative per-target final-loss attribution from Microcosm diagnostics schema version 6
• reconstruct the audited historical weighting doctrines at read time for supported releases without backfilling published artifacts
• fail closed when reported or reconstructed attribution is incomplete or inconsistent
• replace the former Loss Sources calculation with weighted target error and add a separate target-weight view
• use weighted mean capped error for attribution-view coloring and update the overview metrics, tooltips, and presentation
• document the app's distinct loss and error concepts

Cross-repository contract

Producer implementation: PolicyEngine/microcosm#716 (merged; fixes PolicyEngine/microcosm#715).

The producer and consumer share the five row fields target_loss_weight, target_loss_weight_share, target_loss_scale, final_capped_scaled_error, and final_loss_contribution, plus the schema-version-6 target_loss_basis. Both sides use the same ordered-basis hash and a 1e-12 absolute or relative contribution tolerance.

Consumer fixtures mirror Microcosm's uniform, unequal/custom-scale, and structured-unavailable outputs. Cross-repository tests verify hash agreement, warning handling, reported attribution precedence, and contribution reproduction.

User impact

Calibration Fit can now show which targets carry the most calibration importance and which contribute most to weighted target error. Historical releases use deterministic, version-pinned fallbacks where exact reconstruction is possible; unsupported attribution remains unavailable without exposing internal provenance in the normal UI.

The External datasets page and its independent evaluation loss are unchanged.

Validation

• GitHub CI — passed
• bun test — 184 passed
• bun run lint — passed
• bun run build — passed
• git diff --check — passed
• local Calibration Fit presentation visually confirmed by the user

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
calibration-diagnostics	Ready	Preview	Aug 18, 2026 8:26pm