QC Cockpit v0.1 · build 2026-08-18 · study CDISCPILOT01 · seeded from SDTM Mapper + ShellForge
5
SDTM domains
mapped & conformance-checked
21
TLF tables
21 R-validated
100%
Avg SDTM conformance
strong-model gate
0
Cells to review
LLM or conf < 0.50
End-to-end pipeline · Raw → SDTM → ADaM → TLF
Status board across the clinical-programming flow. Each stream is one source domain tracked from raw data through SDTM, ADaM and the final TLF output. Green = artifact present and checked; amber = downstream step pending.
SDTM domains — SDTM Mapper outputs
Per-domain coverage (share of required/expected variables mapped) and CDISC conformance from the strong-model auto-gate. Deterministic (temperature 0), audit-logged.
TLF tables — ShellForge annotations
Each table shell annotated to dataset.variable + where-clause + statistic. n_rule = exact/synonym rule match; n_llm = LLM-resolved (always reviewed); n_low = confidence < 0.50.
Structural validation: R as an executable validator of SAS
Both programs are generated from one annotation set; the R program is executed server-side. PASS = the R table built and rendered, confirming the row/column logic in the paired SAS and R programs is consistent. Cells sourced from the LLM or below 0.50 confidence are flagged REVIEW for a human.
Audit trail (append-only) — real run log
Every run records UTC timestamp, job id, stage, domain, model, temperature, coverage and the input file SHA-256 (first 8). 21 CFR Part 11-style provenance.
Reviewer sign-off
Sign-off log