Maternity Case Record Digitisation - Hallucination Rate: leaderboard

A 2026 benchmark of multimodal LLMs for digitising a handwritten South African maternity case record: each form image is extracted into a JSON template and compared field by field with reference JSON files. This board is the percentage of reference-empty fields incorrectly filled by the model, on the real registry forms; lower is better.

Metric: Hallucination rate (from 0-1, times 100): share of reference-empty fields incorrectly filled by the model (one minus null-field accuracy), on the real redacted Maternity Case Records (49 forms for Gemini and OpenAI models; 46 for Claude) of the UBOMI BUHLE Pregnancy Exposure Registry (South Africa), one form image per call with the optimised extraction prompt and JSON template, temperature 0, two attempts at valid JSON; lower is better. Source: arxiv.org. Saturation forecast: Estimated already saturated. 9 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1GPT-5.46#135
2Gemini 3.1 Pro (Preview)8#99
3Gemini 3.1 Flash Lite8#371
4Claude Opus 4.610#113
5Gemini 2.5 Flash11#463
6Claude Sonnet 4.612#173
7Gemini 2.5 Pro14#276
8GPT-5 Mini17#364
9Claude Haiku 4.523#583

No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=maternity-case-record-digitisation-hallucination-rate · How It Works · Data refreshed daily, snapshot 2026-11-21.