Practical writing on data standards, validation, reporting, and the changing shape of regulatory submissions.
Models change, reprice, and get deprecated. If a validated workflow is welded to one provider, you inherit that instability. Own the workflow and the artifacts, and treat the model as a replaceable part.
Read articleGiving people an AI tool is deployment. Changing how the work actually gets done is transformation, and the gap between them is where most of the promised value is lost.
Read articleThe freeze point is the human-attested boundary where an AI draft becomes a controlled artifact. Before it, the model can stay flexible. After it, nothing it touched changes the regulated record.
Read articleThe real test is not whether an agent can produce something plausible. It is whether you can validate the result. Where you cannot check the output, autonomy has to stop.
Read articleA generative model will not return the same words twice, so classic reproduction fails. What you can reproduce is everything around the answer: the model, prompt, tools, data, and the exact environment.
Read articleIndependent QC assumes two independent minds. Run production and QC through the same foundation model on the same specification, and both can agree with each other and with the same mistake.
Read articlePrompt-level restrictions are guidance, not guarantees. Real isolation for a regulated agent is enforced by the platform: unmounted data, containerized runs, and a gateway that filters what leaves.
Read articleIf every run regenerates the same logic from a model, cost and variability compound. Use AI to write the first validated module, then freeze it, reuse it, and depend on the model a little less each study.
Read articleR4SUB turns the ready-to-file decision into a reproducible 0 to 100 Submission Confidence Index, from evidence to indicators to four pillars, with decision bands everyone agrees on once.
Read articleThe seven-column evidence schema that puts conformance findings and Define-XML metadata in one machine-readable table, and the two parsers in R4SUB that fill it.
Read articleGrade every analysis variable L0 to L3, then track L2-plus coverage as one number, so thin traceability shows up before an auditor finds it.
Read articleFailure Mode and Effects Analysis for a filing: a 1 to 125 Risk Priority Number, a tracked register with owners and mitigations, and a risk trend across dry runs.
Read articleThe role of ADSL, BDS vs OCCDS structures, derivation and traceability requirements, and how define.xml documents the mapping.
Read articleTemplated TLF generation, reproducible SAP-to-output pipelines, and where automation cuts CSR timelines — and where human review stays essential.
Read articleICH E6(R3) spreads the same decision across separate required sections, so a protocol states its key choices more than once. Review that checks for presence never sees them disagree.
Read articleAn adverse event table is a count of things that happened, as classified by a dictionary revised twice a year and applied by people making judgements. Both belong in the record.
Read articleHow absent observations are handled is part of what the analysis estimates. Deciding it at database lock, from whatever arrived, answers a question nobody asked.
Read articleICH E9(R1) asked sponsors to state precisely what treatment effect they are estimating. Many analysis plans still describe a method and leave the question implicit.
Read articleProduced at the end, it describes whatever was built. Written first, it becomes the specification the build is checked against.
Read articleEvery programming group says its outputs are traceable. Under audit, the chain usually holds in two places and thins in three predictable ones.
Read articleA modular, derivation based approach to ADaM that improves traceability and cuts the rework that comes with late data changes.
Read articleThe regulatory pilots that brought R based packages and interactive review applications to the FDA, and what they change for filing teams.
Read articleWhat it takes to run R and Python for regulated analytics, from package qualification through audit trails and reproducibility.
Read article