Previously the edition-level branch of analytics.getProjectRankings
(programId mode) pooled every juror's evaluations across every round
into a single z-normalization context. A juror's mean and stddev are
not stable across round types — quick intake screening produces a
very different grading profile than a deep evaluation round, and
mixing them yields a meaningless personal calibration.
The rollup now groups points by roundId, computes one balance context
per round, and aggregates per-project as the unweighted mean of the
per-round balanced averages. roundId mode is unchanged.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Adds a shared juror-balancing utility (z-score normalization per juror,
rescaled back onto the raw 1-10 scale) and wires it into:
- Admin reports page: Top-10 project table now shows "Raw Avg" and
"Balanced" columns side by side, and the summary stats row shows a
balanced-average tile. Sort defaults to balanced so harsh and lenient
graders no longer skew the ranking.
- Ranking dashboard: each project row shows a green/amber balanced-score
chip next to the raw average when the two differ by ≥0.05, making it
obvious when juror calibration moved a project's effective ranking.
Also adds AI Juror Calibration Advisory — a mutation that takes
anonymized per-juror stats, calls OpenAI, and produces a plain-language
explanation of the cohort's grading patterns plus per-juror severity
(normal / notable / outlier) with a one-sentence narrative. The advisory
describes the statistical balance that already runs; it does not
introduce a new weighting layer. Rendered as a panel in the Juror
Consistency tab when a specific round is selected.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
