Calibration Report

Every /cohort response returns a calibrated_return_pct band alongside the raw quantiles. These are the empirical coverage numbers behind those calibrated bands — validated on held-out anchors, not claimed. See the write-up at /blog/calibrated-retrieval.

Per-horizon offsets (uniform across cohort sizes)

horizon	n_val	offset [p10,p90]	raw cov	calibrated cov	offset [p25,p75]	raw cov	calibrated cov
5d	999	±1.72pp	70.2%	82.3%	±0.51pp	38.6%	48.9%
10d	999	±2.24pp	65.5%	78.8%	±0.91pp	35.8%	49.4%

Target coverage: 80% for [p10,p90], 50% for [p25,p75]. Raw bands systematically under-cover — the conformal offset widens each side so the empirical coverage hits nominal on held-out data.

Bucket-aware offsets (by cohort size)

horizon	bucket	n_val	offset [p10,p90]	raw cov	calibrated cov
5d	large	204	±1.10pp	66.2%	76.0%
5d	medium	790	±1.38pp	68.7%	79.2%
10d	large	126	±2.61pp	71.4%	84.9%
10d	medium	868	±2.75pp	67.3%	82.5%

Cohort size buckets: small (n<100), medium (100-199), large (200+). When a cohort falls in a bucket with too few calibration samples, the lookup falls back to the uniform per-horizon offset.

How to read this in an agent

Use calibrated_return_pct for sizing, stop placement, or any statement about uncertainty.
Use raw return_pct for ranking cohorts against each other — the raw medians are unbiased and comparable across horizons.
The calibration.coverage_80_validated field is the empirical 80%-band hit rate on the held-out calibration set. If your agent needs stricter guarantees, use that number to derate your confidence.
Offsets are refit after material changes to the embedding index or as calibration sample size grows. The current n_validation is the honest bound on how tight any coverage claim can be.