shawnzreports/zbench-ivf

zBench IVF

A record-grounded benchmark for healthcare facility reasoning in language models. Release zbench-ivf-1.2: 5,144 tasks generated from CDC's clinic level ART success rate tables (reporting years 2020 to 2022); 4,139 public with answer keys, 1,005 held out for private evaluation.

Benchmark page, results, method and keyless downloads: https://api.zreports.org/lab/zbench

What it tests

Medical AI benchmarks ask what medicine says. zBench asks the second question: which facility reported which figure, in which reporting year, under which measure, and what can legitimately be concluded from it. Every answer key is a cell of the public record, with a row locator that re-derives it from the published CDC file.

Seven layers, from a control to misses that survive perfect retrieval:

Layer	Name	What the task tests
1	retrieval	the clinic's own published figure for one measure, age band and reporting year
2	encoding	cells CDC prints as a fraction, a suppressed count, a zero or "no cycles", read as printed
3	evidence	the dataset and row that support the figure
4	reconciliation	the clinic's figure beside the national figure, kept apart
5	temporal	the same cell across reporting years
6	comparison	two clinics' figures with the patient mix caveat, declining to rank (CDC's own standard)
7	graph joins	joins across the clinic's success rate, profile and cycle tables

Release 1.2 also holds 36 layer 9 items (economic reasoning with a price stated as a premise).

Results (grader 5k, October 2026)

Exact = graded Pass. Browsing = the model's own web search. Record = the same CDC cells supplied through tools.

Model	Layers 1 to 6, browsing	Layers 1 to 6, record	Layer 7, browsing	Layer 7, record
GPT-5	44 of 179 (25%)	151 of 179 (84%)	4 of 29	30 of 30
Gemini 3.1 Pro Preview	31 of 179 (17%)	72 of 78 completed (92%)	7 of 30	not run
Claude Opus 5.5	5 of 179 (3%), 173 abstentions	138 of 179 (77%)	3 of 30, 26 abstentions	28 of 30

Misses with search are mostly substitutions: another reporting year, age band or measure from the same clinic, or an aggregator's rate presented as CDC's. Full tables, the failure taxonomy, substitution sources and cited domains are on the benchmark page.

Files

• zbench-ivf-1.2-public.jsonl: one item per line. Fields: id, layer, question, the answer key, pass, half and fail text, the grading rules, the probe (failure mode) label, the clinic's own cells, and the row locator (ClinicId, dataset, filter, question, breakout) that re-derives the key from CDC's file.
• zbench-ivf-1.2-neighbours.jsonl: for each public item, the neighbouring cells of the same clinic with their substitution labels, which is what the grader compares a wrong figure against.
• zbench-ivf-1.2-manifest.json: counts by layer, form, template and probe; CDC dataset ids per year; grader version; license; SHA-256 per file, including the held out question file.

The held out split (1,005 items) is withheld in full, questions and keys. A lab or a healthcare AI team can send outputs for the public split, or ask for the held out questions, and receive a graded run: scores by layer and probe, every verdict, the failure taxonomy, the substitution and evidence analyses and the error corpus. Write to shawn@zreports.org.

Grading

Deterministic and item local. Verdict order: hard fail (a figure the record cannot support) → clarifying question or abstention → Pass → Half → Fail → Review. A Half that carries another row's figure is a Fail. Every figure in an answer is compared against the clinic's neighbouring cells and the national cell, and a wrong figure is labelled by the row it came from. Layer 6 is graded to CDC's standard: a ranking is never the correct answer.

A data finding

The numeric column (data_value_num) of CDC's clinic level file reads 0 for every rate CDC prints as a fraction (denominators under 20), so any system reading the file rather than the clinic page misreads those cells. The loader behind zBench parses the printed value; suppressed cells load as null, never zero.

Source and license

Source data: CDC, National ART Surveillance System, Final ART Success Rates clinic tables 2020 (3x54-3thk), 2021 (ey8b-ejrf), 2022 (cchw-gdwa) and the Services and Profiles tables, data.cdc.gov. CDC publishes these tables with a notice that they are not to be used to rank clinics; no item asks for a ranking.

zBench IVF is released for model evaluation under the zReports HealthCare Data Lab terms; redistribution of the item set, use as training data, or publication of held out results requires a written license from zReports.

Citation

Redd, S. (2026). zBench IVF: A Record-Grounded Benchmark for Healthcare Facility Reasoning in Language Models. zReports HealthCare Data Lab, working paper WP-2026-07. https://api.zreports.org/lab/zbench

Contact

Shawn Redd, zReports HealthCare Data Lab, shawn@zreports.org