Programs

Per-feature status and trend

Every check signed. Daily test set across five demographic checks × three text-analysis prompts. Continuous, not point-in-time.

Period 2026-04
Checks signed 2,411,183 period-to-date
Cortex flag +1.1 name-correlated · ETA 48h
Held for review 6 caught before reviewer

All period 2026-04 checks across the 3 live AI features.

Per-feature status — period 2026-04
Operational

Career Interest Badge

Checks this period
487,213
Last flag
No flag in current window.
Output type
Category · Strong / Moderate / Limited / Not expressed

Distribution stable across 14-day window. Source-span attachment rate 100%.

Operational

Grade Normalization

Checks this period
502,874
Last flag
2026-05-08 +0.7 in percentile-band shift on a single specialty cohort; auto-rolled back; investigation closed.
Output type
Normalized percentile band · 5 bands

Performance flag detected and resolved May 8. Model unchanged in production.

Attention

Cortex Screening

Checks this period
1,421,096
Last flag
2026-05-10 +1.1 shift on Emergency Medicine cohort, name-correlated; under investigation.
Output type
Combined score · surfaced for human review

Daily test set flag open. Cohort affected: Emergency Medicine specialty filter. ETA 48 hours.

Fairness matrix — current period

Daily test set. Each cell measures whether the AI text-analysis prompt produces an unchanged result when the same applicant submission is varied along one demographic dimension. One cell is currently flagged and under investigation.

Demographic perturbationRural-medicine interestAcademic-career interestLeadership / collaboration
Name-cue: Deshawn
98 Invariant
96 Invariant
Name-cue: Sarah
99 Invariant
99 Invariant
98 Invariant
Name-cue: Priya
97 Invariant
99 Invariant
96 Invariant
Age-cue: career changer
97 Invariant
95 Minor variation
99 Invariant
School-cue: top-20 institution
96 Invariant
96 Invariant
98 Invariant

Held for review — this period

Items the system held back before any output reached a program reviewer. Each case records what the AI tried to return and what (if anything) actually went through after re-checks.