Service 5

AIME

Trajectory-grounded correspondence

Version

2026.04.1

Retrained

2026-04-02

Status

healthy

Performance

Key metrics

Calibration & discrimination against the engine's published target. Under-target values are highlighted.

Grounding-in-trajectory

87.0%target 80%

Clinical accuracy (review)

93.0%target 90%

Hallucination rate

0.8%target 2%

Tone alignment

91.0%target 85%

14-day inference volume

248 in the last 24h · avg latency 1820 ms

Metrics are tracked on held-out evaluation cohorts; values here reflect the last retrain.

Model

Training basis & runtime

Transparent about what the engine was trained on and how it behaves in production.

Model family

Ensemble (3× 70B + critic) with temporal grounding

Trained on

58k clinician-edited letters · multi-LLM ensemble

Avg grounding

87%

Clinician edit rate

34%

down from 52% in Jan

Avg draft length

174 words

Fall-back rate

1.8%

History

Changelog

Each retrain is documented; clinicians can see what changed and when.

2026.04.12026-04-02
Referral style-sheet updated for NHS secondary-care recipients.
2026.02.02026-02-14
Grounding critic trained on 2,400 clinician edits; hallucination rate halved.

Safety

Guardrails

Hard limits and runtime checks the engine enforces.

No draft is auto-sent — every message requires clinician approval.
Rejects generation if trajectory grounding <0.6.
Redacts PII before any external LLM call.

Observability

Recent incidents & retrains

Automatic monitoring — warn / critical thresholds on calibration, drift, and latency.

Clinician edit rate trending downinfoAIME15/04/2026, 16:05
Edit-rate at 34%, lowest on record. Referral class particularly improved post style-sheet update.

All engines