cdisc-case-3

Build an interactive, self-contained HTML traceability explorer that links the full clinical-deliverable chain Objective -> Endpoint -> SDTM domain -> ADaM dataset -> TLF, with the Analysis Results Data (ARD) as the connective data layer. Assembles a node/edge graph from the pipeline artifacts (study-model.json, tlf-plan.json, analysis-spec.json, adam-spec.json) and per-table outputs (generated.md, ard.json, generate.R), then emits ONE standalone .html file (vanilla JS + SVG, zero external libraries/CDN/fonts) with an interactive node graph, click-to-highlight lineage, per-node detail panels, status badges, filters and search. Use this skill when the user wants to SEE how outputs trace back to objectives/endpoints and forward to their data — phrases like 'traceability', 'traceability matrix', 'lineage', 'objective-to-TLF', 'end-to-end trace', 'where does this table come from', 'ARD explorer', 'deliverables dashboard', 'traceability explorer', 'link objectives to tables', 'coverage of endpoints', 'which endpoi

Generate ADaM datasets from SDTM data using {admiral} R code, DRIVEN BY the ADaM spec (adam-spec.json) produced upstream by tlf-analysis-spec. The derived ADaM MUST satisfy the spec's datasets/parameters/variables/populations and its MANDATORY derivation rules: (1) day-based analysis-visit windowing that never drops unscheduled/early-termination visits, (2) LOCF records (DTYPE='LOCF', ANL01FL='Y') at the endpoint visit, (3) SITEGR1 pooling on RANDOMIZED (ITTFL='Y') counts. Enforces a data-quality gate on analysis-set N and checks conformance with metacore/metatools. Use this skill whenever the user has an adam-spec.json (or reviewed ADaM spec) plus SDTM datasets and wants ADaM datasets. Also trigger on 'ADaM derivation', 'ADaM generation', 'SDTM to ADaM', 'admiral code', 'derive analysis datasets', 'spec-driven ADaM'. Third step in the Protocol->SAP->TLF-plan->analysis/ADaM-spec->ADaM->TLF pipeline.

Turn a reviewed TLF plan into the analysis specs + ADaM variable spec that drive ADaM derivation and TLF generation. Use this skill when the user has a tlf-plan.json (from tlf-planner / tlf-plan-critic) and asks to 'author the analysis spec', 'build the ADaM spec', 'map the TLF plan to specs', 'create the analysis metadata', 'write the statistical method spec', 'produce ARS-aligned analysis specs', or 'run the bridge before ADaM / generation'. This is Step 2 of the TLF pipeline — the human-review gate between planning and compute: for every producible TLF it authors an ARS-aligned analysis-spec entry (analysisSet, groupingFactor, dataSubset, analysisVariables, methods with operations/model/lsmeans/contrasts, rounding), and aggregates the union of all TLFs' data needs into a reviewable adam-spec.json (datasets, parameters, variables, flags, the mandatory derivation rules, and the data-quality N-gate). It does NOT run R or derive ADaM — it produces specs only. Run before sdtm-to-adam and tlf-generator.

Generate the actual Tables, Listings, and Figures (TLFs) from an ARS-aligned analysis spec + ADaM data, ARD-first: compute every statistic into Analysis Results Data (ARD) with {cards}/{cardx}/emmeans, then render the ARD to the target display shape with {gtsummary} (tables) and {ggsurvfit} (Kaplan-Meier figures), applying SAS rounding at display time. Use this skill when the user has an analysis-spec.json (from tlf-analysis-spec) plus ADaM datasets and wants to produce the numbers and the displays. Also trigger on: 'generate TLFs/tables/figures', 'produce ARD', 'analysis results data', 'cards/cardx tables', 'gtsummary tables from ADaM', 'ANCOVA LS-means table', 'Kaplan-Meier figure', 'ADaM to tables', or 'compute the numbers for the tables'. This is Stage-3 of the Protocol->TLF pipeline and REPLACES/UPGRADES the older adam-to-tlg skill with an ARD-native numbers-then-format approach; prefer this skill over adam-to-tlg when an analysis spec exists.

Derive the exhaustive, traceable list of required TLFs (Tables, Listings, Figures) for a clinical study from its USDM metadata JSON. Use this skill when the user has a USDM / CDISC DDF study definition (or asks 'what TLFs/tables/listings/figures do we need', 'derive the TLF list', 'plan the reporting package from USDM', 'map objectives and endpoints to tables', or 'run the TLF planner'). This is the planning/inventory step that runs before ARD dataset creation and before {cards}/{cardx} TLF generation. It interprets objectives and endpoints, adds the ICH E3 regulatory tables no objective points to, fans out statistical analyses (imputation, subgroups, timepoints, sensitivity models), checks data feasibility, and produces a numbered, human-reviewable TLF index. Pair with the tlf-plan-critic skill, which independently audits the result for coverage and completeness.

Generalized self-learning loop for the CDISC Case 3 pipeline. Distil durable, per-skill lessons from the reviewer feedback captured during a run across ALL THREE review gates (plan / specs / TLFs). Reads /workspace/review_feedback.jsonl — one object per revise, each tagged with the skill it targets (tlf-planner, tlf-analysis-spec, tlf-generator) — plus the approved artifacts and each target skill's current references/lessons-learned.md; proposes an append-only markdown block per revised skill that would have produced an approvable result first pass. Writes { hasLessons, lessons: [ { skill, lessonAppendMarkdown } ], prTitle, prBody, summary } to /output/result.json. Append-only — never rewrites existing lessons. Triggers: 'learn from review feedback', 'propose skill lessons', 'codify reviewer comments into the skills', 'self-learning loop'. Used by the propose-skill-update agent step after the pipeline's reviews.

Independently audit a TLF plan (produced by the tlf-planner skill) for traceability and completeness before human review. Use this skill after tlf-planner, or whenever the user asks to 'audit/review/QC the TLF plan', 'check TLF coverage', 'verify every endpoint has a table', 'find missing or orphan TLFs', or 'run the coverage/traceability check'. It performs a two-way audit — forward (every objective and every resolved endpoint maps to at least one planned TLF) and backward (every TLF traces to an objective/endpoint or a regulatory rule) — plus conventional-completeness heuristics, then emits a coverage report with a verdict and, on gaps, a bounded re-run plan. It reads only the plan and the study model, never the planner's reasoning, so the check is genuinely independent.