# Guided setup for open.gismo (2026-08-27) One of four competing directions explored in the open.gismo data-loading design session. This one commits to **a guided setup**: the user is walked through one decision at a time, in plain English, and the product does the inferring. It is optimised for someone who has never seen gsm and does not know what an analysis domain is. Requested by @jwildfire (2026-08-26): *"Maybe a ultradesign session for open.gismo? re-visit the research comparing other sites and do some mockups on how users will load their own data?"* Extends row **D1** of the [platform gap analysis](../platform-gap-analysis-2026-07-25/) — the data-mapping surface — from the guided-setup angle only. It deliberately does **not** repeat the competitor survey done by the sibling artefacts in this session. ## Pages | Page | What it holds | |---|---| | [`index.html`](index.html) | The design: six commitments, the staircase, the hard case walked end to end, what it costs, and five named weaknesses | | [`walkthrough.html`](walkthrough.html) | The mockup — eight clickable steps. Steps 2, 3 and 5 are live: change an answer and the join count, the preview and the ledger all move | | [`evidence.html`](evidence.html) | Method and verbatim output for all four measurements | ## The claim, in one line Six questions, all answerable by looking at your own rows, reach a real chart from a CRO delivery — and the design's shape is forced by the fact that the inference engine behind it was measured producing a **confidently wrong answer** on the study open.gismo ships. ## What was measured, and how Everything asserted about our own side was executed on 2026-08-27 against `open.gismo` @ `dev` 6d23c9b (v0.2.0), `demo-301` @ main, working copies of `safetyGraphics` and `safetyCharts`, with `gsm.mapping 1.1.3` / `gsm.kri 1.5.0` / `gsm.core 1.2.0` installed. - **M1 — the published name detector.** `safetyGraphics::detectStandard()` run unmodified against the fixture with `safetyCharts::meta_labs` and `meta_dm`: **3 of 9** lab roles (33%, and it names "SDTM" anyway), **5 of 7** on ADSL (71%), **0 of 4** hepatic analyte values. `hasColumn()` is exact membership — no fuzzy matching anywhere. - **M2 — a name-blind value profiler**, written for this session (`data/profile.R`), on the same lab file: **9 of 9** roles correct, including the four M1 missed. - **M3 — the same profiler on `demo-301/input/Raw_LB.csv`**: chose `visnum` as the lab result, `toxgrg_nsv` as the lower limit and `lb_dy` as the upper, scoring **1.000** — *above* the 0.849 it scored on the answer it got right. Re-ranking on a structural test (`data/profile2.R` — does some categorical column give exactly one reference range per level) puts the correct triple first on both lab files and demotes the impostor to second at purity 0.00. - **M3b — the join.** Every character column against every other: `SUBJID`↔`SUBJECT` shares 765 values, `USUBJID`↔`SUBJECT` shares **0**. `inner_join` on the latter yields **0 rows** from 57,200. A single constant prefix `AA-AA-000-0000` strips off `USUBJID` and recovers all 765. - **M3c — the emitted spec runs.** A spec written exactly as the interview would emit it, pushed through `gsm.mapping::ApplySpec()` and the study's own inner join: 57,200 rows over 765 participants, `lbstresn` numeric, `lb_dt` a Date. Six spec entries were dropped with **no message at any level** (`purrr::keep(~ .x$source %in% colnames(dfSource))`). - **M4 — the price of a partial map.** `gsm.kri::CalculateRiskScore()` on identical site data with twelve KRIs mapped versus six: Site Risk Score **8.3 → 16.7**, exactly double, with no warning. `srs0001` is the only workflow with no `spec:` at all; its denominator is `sum(dfMaxWeights$max_WeightMax)` over the analyses that *ran*. - **The staircase.** Every metric and module in `demo-301/workflows/` resolved through its `Mapped_*` dependencies to raw domains: `Raw_SUBJ` feeds 26 of 30 metrics and 10 of 12 modules; `Raw_SUBJ` + `Raw_LB` (29 columns, two files) reaches **6 of 12 report modules**. - **The ALT histogram** on walkthrough step 5 is the real distribution: 3,575 results over 765 participants, binned 0–100 U/L in steps of 5, reference range 7–41. The caption states the 737 above the upper limit and the 320 that fall beyond the right edge rather than the chart hiding them. ## The fixture, and what is invented about it All measurements use one constructed CRO delivery. **The values are real** — every row is demo-301 study data, unchanged. **The packaging is invented**, and that choice is mine: - `adsl.csv` (1,005 × 16) — `Raw_SUBJ.csv` renamed into ADaM, with `USUBJID` built as `studyid-subjid`. - `lab_central.csv` (57,200 × 11) — `Raw_LB.csv` renamed to a central-lab extract, with `SUBJECT` as a bare id, `LBORRES`/`LBORNRLO`/`LBORNRHI` as the *original* rather than standardised result and range, analytes respelled `ALT` / `SGOT` / `TBILI` / `Alk Phos`, and `toxgrg_nsv`, `lbblfl`, `lb_dy` removed. Two of those choices make the case harder and should be challenged: the original-versus-standardised result is what makes M1 score badly, and the three removed columns are what create the highest-friction moment. Both are real patterns; both are still mine. This is stated on `evidence.html` as well as here. The fixture itself is not committed — it is 57,200 rows of a study that already lives in this workspace. The generating recipe is in `evidence.html`; sample rows are committed as [`data/delivery.js`](data/delivery.js) so the walkthrough renders measured values rather than illustrations. The probe scripts are [`data/profile.R`](data/profile.R) and [`data/profile2.R`](data/profile2.R). ## What is invented in the mockup - The visual design of the interview. open.gismo has no setup route today. - `config/setup-log.md` does not exist; it is proposed as the artefact that keeps the user's plain English next to the machine-readable spec. - The step-5 chart is a static SVG from the measured distribution, not a live `safety.viz` render. - The four hepatic dropdown answers are pre-selected in the mock so the ledger can be shown in both states; the real interview would leave them empty. ## Known limitations - **The structural ranker fixes the failure I found.** It is not evidence that the ranker is sound in general, and the failure mode it guards is silent-wrong rather than loud-broken. - **The profiler is O(n³) in numeric columns.** Eleven is instant; an 80-column ADLB is over half a million combinations and would need pruning that reintroduces name heuristics. - **The dependency graph is demo-301's.** A study drawing on `gsm.mapping`'s wider domain set has more. - **Nothing was tested against a wide lab file** — one column per test rather than one row — where there is no categorical column for the structural test to group by. - **The competitor landscape was not re-surveyed here.** M1 is a direct run of one open-source implementation, not a survey. The session's landscape work lives in the sibling artefacts. ## Boundaries respected - **Nothing was built.** No change to `open.gismo`, `demo-301`, `gsm.mapping`, `gsm.kri`, `safetyGraphics` or `safetyCharts`. No issue, no branch, no PR. - **No writes outside `jwildfire`.** `gsm.mapping` and `gsm.kri` are Gilead-BioStats repositories and were read only. - **Read-only competitor research**, public sources only. - **Sibling artefacts untouched.** Three sessions ran this brief in parallel; `open-gismo-data-loading-2026-08-27/`, `og-data-loading-design-2026-08-27/` and `og-mapping-bench-2026-08-27/` were not read into this design's argument beyond confirming no overlap, and not modified. ## Design notes - Report chrome uses the obot report family tokens (executive overview 2026-07-21, goal atlas 2026-07-24, platform gap analysis 2026-07-25) with the same three-state light/dark contract. The product frame in `walkthrough.html` deliberately overrides them with open.gismo's own espresso theme, fixed light in both modes, because a mock of a product should look like the product. - Every state carries a glyph **and** a word. No colour-only status anywhere. - Nothing is fetched at view time — no CDN, no web fonts, no network calls, no build step. - Verified at a real 390px viewport: all three pages, no horizontal page scroll (`scrollWidth` 386), no element overflowing without its own scroll container, and the interview rail collapses from a sidebar to a horizontal step strip below 52rem. - The walkthrough's interactive paths were exercised in-browser, including the failure states: choosing the zero-overlap join key produces a refusal and a repair offer, and both a broken join and a wrong result column block the first-light step rather than drawing something plausible. --- Drafted by Claude Code using Opus 5 for review by @jwildfire.