Simulate a trial forward from a stated truth
Each area below is a frozen description of a disease, its visit schedule, its endpoint and its measurement panels. Pick one, state the true treatment effect, and the engine projects a single simulated patient set into raw → SDTM → ADaM → TLF → Define. Because the layers all derive from that one patient set they agree by construction, and because the true effect was an input, every run ships an answer key.
State what is true, then generate
Everything the engine emits is a pure function of these inputs. The same inputs give byte-identical data and the same 12-hex run id, so a run id is a durable reference to one exact dataset.
No area selected
Is the generator calibrated?
A synthetic-data tool nobody has calibrated is a random number generator with CDISC column names. Four gates are the proof. G-B re-simulates the same study over many seeds and asks how often the 95% CI covered the true effect — it should be about 19 in 20. G-C is its mirror: with no true effect, p < 0.05 should happen about 1 in 20, which catches the failure G-B cannot see — intervals that are far too wide.