What was measured, and how

1 · The contract

Read on 2026-08-27 from demo-301/workflows/1_mappings/*.yaml — the study configuration that carries both the RBQM lens and the Safety charts, which is the realistic target for “bring your own study.” Every Raw_* domain in every spec: block, sentinel keys (names beginning _) dropped, merged across mappings.

og_init()'s default snapshot is smaller — 13 mapping workflows, 12 raw domains, 90 columns — because it omits ECG, inclusion/exclusion and the safety-chart mappings. Both numbers are correct; they describe different studies.

Two properties of the contract that shape any design:

• It is not CDISC. Raw_SITE alone asks for lower-case studyid, invid, site_status beside CamelCase InvestigatorFirstName, City, State, Country. Elsewhere it asks for subject_nsv, mdrsoc_nsv, toxgrg_nsv, mincreated_dts, agerep — operational EDC vocabulary, not a standard.
• Nothing marks a column optional. All 126 read as required, so the readiness check treats a decorative field and a metric's divisor identically. Every honest design has to introduce that distinction; this one does it by pricing rather than by a flag, because the cost of a column is a property of the workflows that run, not of the column.

2 · The delivery, and the matcher run

Seven files, 105 source columns, authored from the CDISC ADaM IG and SDTMIG variable lists plus EDC-vendor export conventions — written before the alias table, so the matcher gets no help from having been fitted to it. Both were written by the same author, which is the honest limitation of this measurement and is stated again at the bottom.

Tier 0 — no aliases at all

Exact case-insensitive name matching across the whole delivery settles 44 of 126. That is higher than the “a conformant CDISC package matches zero” framing suggests, and the reason matters: most of the 44 are studyid, subjid, country, sex, race, arm repeated across domains, plus the EDC vendor's City/State/Country, which are not CDISC at all.

Tiers 1 and 2 — the alias table

Three domains complete themselves entirely, and the two biggest holes are domains nobody sent rather than columns nobody matched. That shape is what makes “decline it, and here is what it costs” the right primitive: a third of the outstanding work is not mapping at all, it is a study telling the product what it does not have.

The six flagged weak are the interesting ones and they are why a matcher must be able to say “near”: Raw_SUBJ.enrollyn ← ITTFL, Raw_ENROLL.enroll_dt ← RANDDT, Raw_STUDCOMP.compyn ← COMPLFL, Raw_STUDCOMP.compreas ← DCSREAS, Raw_LB.battrnam ← PARCAT1. Each is defensible and each is a study-specific judgement — an intent-to-treat flag is not an enrolment flag, and randomisation is not enrolment. These are exactly the lines a reviewer needs to see in a diff, and exactly the ones a wizard would let a user click past.

3 · Pricing — what each gap actually costs

Traced through the metric workflows' own spec: blocks in demo-301/workflows/2_metrics/: 30 workflows, 136 spec entries, 47 distinct Mapped_* columns. The dependency is heavily concentrated.

Only 34 of the 126 required columns are named by any metric at all. The other 92 feed reporting, labels and the study masthead. That ratio is the load-bearing fact under this whole direction: the mapping job a user must actually get right is a quarter of the size the contract implies, and a diagnostic that orders findings by cost turns a 53-line chore into a 12-line one. Get the ordering wrong and the direction fails on volume — that is W4 on the design page.

The two executed failures

4 · The asset that already exists — and its first bug

safetyCharts ships nine metadata tables, 82 rows, each row carrying standard_adam and standard_sdtm defaults. It is MIT-licensed and @jwildfire is its author and maintainer. It contains two kinds of row, and the second kind is the more interesting one:

The 14 field rows are a value-level metadata layer that safetyGraphics has had for years and that gsm has nowhere — which is why hep_explorer.yaml ends up carrying four bare literal strings in lSettings.measure_values that a study must happen to match. The ADaM defaults in those rows even carry the units inside the string, "Alanine Aminotransferase (U/L)", which is exactly the unit problem showing through the vocabulary.

Bridging chart roles to gsm columns takes a hand-written table of 29 pairs. Doing that yields 52 alias rows covering 26 gsm columns for free.

5 · The landscape, read through one question

The thirteen-platform survey is not repeated here — the gap analysis holds it and the sibling artifact re-visited it for the screen-based direction. This slice asks one question of the same sources: when a platform's mapping is finished, is there a file, and does the study own it?

Documented, not verified: each row reflects what the platform's public documentation, package reference index or user guide describes. Nothing was tested, no demo requested, no vendor contacted, and vendor documentation flatters. Where a row is a judgement about the shape of the product rather than a documented feature, it is written as such.

Three readings of that table

• The open-source R half already does what this direction proposes — and does it by having no mapping surface at all. teal expects ADaM and stops there. That is convention over configuration in its purest form, and it works because ADaM is an industry standard with an entire profession behind it. gsm's raw layer has no such profession, so a design that copies teal's posture without supplying the generator and the diagnostic just reproduces tidyCDISC: reject the file, explain nothing.
• The commercial half maps the data for you, using people. At the top of the market nobody maps their own study; a services team does. That is the honest reason a wizard is not obviously the right answer for this product — the person who would use one is not the person who has the problem — and it is also the reason this direction concedes an audience (W1).
• Nobody produces a reviewable artifact. Not one surveyed platform leaves the study with a mapping a second person can review as a diff, that CI can replay, or that arrives in an inspection folder as evidence. Meanwhile the industry's own answer to “what is in this dataset” is define.xml — a machine-readable document, not a screen. open.gismo already snapshots the study's workflows into the study's own folder. This direction is the one that finishes that thought.

6 · Method, and what would change these numbers

• The contract was extracted with a script that reproduces .og_input_specs()'s rules — Raw_* only, sentinel keys dropped, merged across mappings — pointed at demo-301/workflows/1_mappings/ rather than an og_init() scaffold, so it describes a study that runs both lenses.
• The dependency graph was read from the metric workflows' own spec: blocks and mapped back from Mapped_* to Raw_* by domain name. Where a mapping renames a column between raw and mapped, that trace would be wrong; in demo-301's mappings the selects are identity for every column in the graph, so it is not.
• The environment was workr 1.0.0 / gsm.core 1.2.0 / gsm.mapping 1.1.3 / gsm.kri 1.5.0 / gsm.reporting 1.1.5, against open.gismo at dev v0.2.0.
• Confirmations run directly: the string source_col appears nowhere in open.gismo/R/ or inst/; og_run.R never calls og_validate(), while og_app.R does; no reader anywhere in the package touches haven, .xpt, sas7bdat, arrow, readxl or DBI — eleven read sites, all read.csv.