TLF quality control
that a reviewer can audit cell by cell.

Outputs arrive in a batch three weeks before lock. The QC programmer re-derives the first forty tables carefully and the last sixty quickly. Nobody chooses this; the calendar does.

Two outputs side by side, a programmer reading numbers across. A transposed digit in a footnote survives because nobody reads footnotes at eleven at night, and the compare utility was never configured for that layout.

The QC record says the output was checked and by whom. It does not say what was compared, with what tolerance, against what version of the data, or what was waived. The auditor asks exactly that.

The output shell and the analysis specification are parsed into a machine-readable definition: population, analysis variable, statistics, grouping, display precision, footnote rules. Ambiguities go back to the statistician as questions, not assumptions.

Each structured output is regenerated from the analysis datasets by independent code in your SAS or R environment, under your versions and macros, from the same data cut as production.

Production and re-derived outputs are compared cell by cell with explicit tolerances: numeric precision, rounding rules, ordering, labels, titles and footnotes. Figures and free-text listings are routed to a reviewer with the data behind them.

Mismatches are grouped by cause and linked to the cell, the derivation and the data. Your QC lead dispositions each one: production error, QC error, spec ambiguity or accepted difference. Nothing is closed without a named disposition.

Inputs, code versions, data cut, tool events, comparison results and dispositions are hashed and packaged per output and per run. The Evidence Room shows what this looks like on synthetic data.

Summary tables, shift tables, frequency and incidence tables, structured listings with defined sort order and columns. Compared deterministically, every cell, every run.

The parts that fail most often in manual QC and are read least: header text against the shell, N counts against the population flags, footnote references against the rules.

Kaplan-Meier curves, forest plots, narratives and unstructured listings are flagged for a person with the underlying data re-derived and attached, so review time goes to judgment rather than arithmetic.

Tolerances, precision rules, waiver categories and figure routing are documented and versioned before the pilot. Changes go through your change control.

The pipeline is tested on outputs with seeded discrepancies. Every seeded error must be found; every clean output must pass. The test evidence is part of the package.

Each comparison record references the shell version, the data cut, the code versions and the disposition. The packet reproduces on request.

Rules settle every case they can. A model is consulted only for the residual the rules cannot decide, and its output passes through the same rules before anyone sees it.

Processing runs in your environment, under your identity system. Nothing potentially unblinding reaches an outside model. The data flow diagram is available under NDA before kickoff.

Every output that matters is accepted by a person on your side whose name is on the record. Model output never writes a finding into a deliverable on its own.

Inputs, versions, tool events and decisions are logged as they happen, so the packet an auditor asks for exists before the audit does.

Output shells and the analysis specification, a QC sample set with known results, your comparison SOP and tolerances, access to the SAS or R environment and the analysis datasets.

Parser configuration and comparison rules in your estate, a comparison record per output, a discrepancy report per run, the evidence packet, validation evidence for the rule sets.

Your QC lead accepts each rule set before it is used on production. A named person clears every flagged figure and dispositions every mismatch.

Start with the study whose QC queue worries you most.

Tell us the phase, the environment and the lock date. We reply with a scoped proposal: pilot sample, rule sets, acceptance criteria.

Or start with the Validated Automation Review: two to four weeks, one workflow corridor, fixed scope. The plan is yours whether or not we build it.