Conformal Coverage

This page is the public-facing surface for the VI-903 split-conformal calibration artifacts.

Important distinction:

• the receipt methodology keeps the strict temporal and uncertainty rules
• this page reports the offline held-out coverage evidence used to justify those intervals

Operational rule:

• the canonical artifact is generated offline by methodology/scripts/calibrate_conformal.py
• the published page should be copied from that artifact output rather than recalculating in-place
• for the current heterogeneous calibration evidence base, the default artifact is generated in stratified cell mode rather than a single global conformal fit

Artifact family:

• r2://vetted-methodology/reports/conformal/...

Current status for the latest real-data artifact:

• the current canonical artifact is generated in stratified cell mode
• the latest held-out empirical coverage is 0.9623
• this is good evidence that the interval calibration is behaving well on the current truth-bearing dataset
• however, that dataset is still dominated by shadow-domain truth rather than exact hosted provider truth

So this page should be read as:

• proof that the current methodology intervals are calibrated on real held-out data
• not proof that every hosted production cell already has exact hosted truth

Required published fields:

• methodology version
• calibration dataset date range
• nominal coverage
• held-out empirical coverage
• conformal grouping mode (cell, domain, or global)
• release status (pass, review, fail)
• per-cell coverage table

Until the hosted exact-cell gaps close, publish both:

• the conformal artifact values
• a short note describing whether the current release is driven by exact hosted truth, shadow-domain truth, or a mixture