Benchmark¶

318 benchmarks, generated from benchmarks/. Each one is a complete, runnable specification with its input data and the exact output an implementation must reproduce, and each page below shows that README, the source fixtures, the expected artifacts, and the YAML specification together.

Group	Count	Purpose
`schema-*`	44	Assess yamaa schema
`adam-*`	83	Assess SDTM to ADaM derivations
`bimo-*`	1	Assess BIMO clinical site dataset derivations
`sdtm-*`	53	Assess ODM to SDTM derivations
`negative-*`	137	Assess yamaa error handling

Rejection is half the contract. A portable specification language is defined as much by what it refuses as by what it computes, and a refusal is only portable if two implementations refuse the same thing in the same way, so every anti-pattern pins the exact error alongside the specification that provokes it.

The suite serves three audiences at once:

Users read it as a cookbook: "how do I express a baseline flag" has an answer to copy.
Developers run it as a conformance test: the expected files are byte-exact contracts.
AI agents learn the common derivations from one half and the anti-patterns that bound them from the other.

Start here¶

#	Benchmark	What it establishes
1	Basic Demographics	Reading collected items, handlers, and the declared keys as the grain
2	Lab Findings	Row construction: one template per collected test, `row_number` for the sequence
3	Lab BDS Build	A full Basic Data Structure build: parameters as row templates, then baseline, change and sequence as columns

Read a page top to bottom and the loop closes in five minutes: the Summary gives the intent, the specification sidebar gives the shape and the grain, and the expected artifact confirms the reading.