sutrix/biostat-evals

Sutrix biostat evals

Known-answer tasks for AI tools that analyse questionnaire and clinical-assessment data. Each task is a real public dataset, a research question as a study team would ask it, and the published answer with a tolerance and an independent recomputation. A tool passes a task when every expected number it reports is within tolerance, it uses the expected method family (or a defensible alternative), and it fabricates nothing.

This dataset is being built as an open benchmark in collaboration with Hugging Face. That collaboration is in progress. The design catalogue holds 31 cases; the 3 below are the ones built, run and countersigned so far. Every number here is transcribed from the case file and the recorded run in the Sutrix repository, never typed into prose.

Tasks

id	dataset (licence)	rows	question	expected method	numeric expectations	verified run
BE-G01	ISwR::energy (GPL >= 2; Altman 1991, Table 9.4)	22	Do lean and obese women differ in 24-hour energy expenditure?	welch_t (fixed)	12	PASS, 3 repeats, 111 checks
BE-G02	datasets::PlantGrowth (GPL-2 or GPL-3; Dobson 1983)	30	Do the treatment groups differ in dried plant weight?	one_way_anova (assumption-routed)	11	PASS, 3 repeats, 81 checks
BE-S05	psych::bfi (GPL >= 2; SAPA project)	2,800	Describe the five personality scale scores by gender.	descriptive (scale scoring with reverse-keyed items)	43	PASS, 3 repeats, 369 checks

Task groups: GROUP (two or more groups compared on one outcome), SCORE (instrument scoring from items), CLEAN (preparation steps a tool must take before the numbers are right). The catalogue also defines DESC, ASSOC, REPEAT, NULL and REFUSE; no case of those groups is built yet.

Files

path	what it is
tasks.jsonl	one row per task; the schema is in labels.md
labels.md	field-by-field description of a row and how each is scored
cases/<id>/case.json	the full case contract: dataset pin, question, expected numbers with oracle, citation, tolerance and countersign
cases/<id>/study-context.json	the question and materials as a study team would enter them
cases/<id>/plan.json, cases/<id>/ops.json	the analysis plan and preparation steps the offline run replays (so no model is called in the reference run)
data/<id>/expected_clean_dataset.csv	the prepared frame a tool must reproduce, committed only because every set here is GPL package data
checks/<id>.R	the base-R twin that recomputes every expected number (checks/json.R is its JSON writer)
evidence/<id>.json	the recorded run card the run_evidence field is stamped from

The raw input rows are not in this repository. Each task pins them by URL and sha256 (dataset.source_url, dataset.sha256, and dataset.conversion when the file is derived from a package archive). A harness fetches the bytes, checks the digest and refuses a changed byte.

Labels and schema

Every row of tasks.jsonl carries:

• question, study_context and follow_ups: the task as a study team poses it.
• expected_method_family and acceptable_method_families: the analysis the published answer used, and the defensible alternatives that score as acceptable rather than expected.
• expected_primary_endpoint: the outcome, grouping column and how the route is decided (fixed, or an assumption rule such as shapiro_levene).
• expected_numeric[]: each expected number with its path into the result, value, absolute tol, oracle kind, citation, optional route gate and countersign (the independent recomputation), plus evidence, the file in the source repository the number is transcribed from.
• verifier: what a run must also satisfy beyond the numbers (required and forbidden endpoint kinds, required and forbidden preparation steps, rows preserved, figure legends, clean-frame tolerance, the R twin).
• run_evidence: the recorded reference run (id, status, repeats, checks).
• provenance: who transcribed the oracle, who countersigned it, and the open questions the case authors recorded.

labels.md describes every field.

Scoring

• A numeric expectation passes when |observed - value| <= tol. Tolerance classes: counts exact; means, SDs and medians 0.01; test statistics 0.01; p-values 0.0001; coefficients 0.001. A printed oracle also carries its own, usually tighter, tolerance (half a unit of its last printed digit).
• A route-gated expectation (route, route_endpoint) is asserted only when the tool's realised test for that endpoint equals route; otherwise it is recorded not evaluable, never a pass.
• The method family is scored expected, acceptable, or wrong.
• Fabrication: a statistic-shaped number in the answer that matches nothing in the data, the R twin's output, the expected values or a short list of conventions (within half a unit of its last printed digit) counts against the run.
• Where the rows a tool is shown cannot carry the claim (BE-S05 with a preview of 30 of 2,800 rows), an honest answer must say so.

Oracle kinds

code	meaning
PUB	a number printed in a published paper
BOOK	a number printed in a textbook (BE-G01: Dalgaard, Introductory Statistics with R, ch. 5)
PKG	a number printed in a package manual or its documented example (BE-G02: the R datasets help page and the datanovia ANOVA tutorial)
MANUAL	a manual derivation recorded in the case
CHECK	computed at build by the base-R twin, used where no printed number exists (BE-S05 per-gender scale scores)

Every printed number is countersigned by the R twin; the case is refused at build when the two disagree beyond the row's tolerance.

Provenance

id	data	bytes pinned from	answer	countersign
BE-G01	Altman (1991) Table 9.4 via the ISwR package 2.0-12	CRAN source archive, energy.txt.gz converted to CSV	Dalgaard, Introductory Statistics with R, ch. 5 (printed t.test output)	R 4.6.0 base, checks/BE-G01.R
BE-G02	Dobson (1983) Table 7.4 via R datasets::PlantGrowth	Rdatasets CSV mirror (asserted identical to the package data)	R datasets help page; datanovia ANOVA tutorial	R 4.6.0 base, checks/BE-G02.R
BE-S05	SAPA project, psych::bfi (Revelle)	Rdatasets CSV mirror	scoring key psych::bfi.keys; per-gender values are CHECK	R 4.6.0 base, checks/BE-S05.R

Cases were transcribed by the Sutrix engineering agent and countersigned by an independent base-R recomputation. A second human reading of the printed numbers is pending on every case (recorded in each row's provenance.open_questions).

Licence

The task text, labels, plans, checks and this card are released under GPL-2.0-or-later so that they can sit alongside the data they describe. The prepared data files derive from R package data: ISwR (GPL >= 2), R datasets (GPL-2 | GPL-3) and psych (GPL >= 2); each row's dataset.licence names the originating package and the primary source of the numbers.

Source and how to run

The cases, the R twins and the runner live in the Sutrix repository under benchmarks/biostat-evals/; this dataset is generated from them by benchmarks/biostat-evals/src/hf-dataset.ts and published by scripts/evals/publish-hf.ts. The reference run replays the pinned plan and preparation steps with zero model calls, three repeats per case, and writes a card per case (evidence/<id>.json). Contributions: a case is one folder with a public dataset whose licence allows the use, a question, and a printed answer you can cite; the tolerances and the R twin are added with you.