Why we chose a deterministic geometry engine over an LLM — Zero Silent Error
A pipeline that extracts symbols, connectivity and time constants from nuclear control logic diagrams (CLD) without loss — the design principles behind a domain where silent failure is never acceptable.
In nuclear drawings, a 'silent failure' is an accident
A nuclear control logic diagram (CLD) is a document that demands licensing-grade traceability. An AI that misses a single symbol and moves on without saying anything — a silent error — is not acceptable in this domain.
That is why doAZ chose a deterministic geometry engine for the CLD pipeline, rather than a VLM.
Deterministic first, self-verified by simulation
A deterministic parser owns the entire geometry of the vector PDF. Feed in the same drawing and you always get the same result, and every extracted element carries its source coordinates.
On top of that we add one more layer — the extracted logic simulates itself for verification. By confirming that the extracted netlist actually behaves as the control logic it claims to be, we block extraction errors from flowing silently downstream.
P&ID and CLD, one stack
P&ID intelligence (extracting equipment, instruments, valves and piping objects along with control loops) and CLD intelligence are not separate technologies. KKS tags are the key that joins the two drawings — a single plant drawing intelligence stack that reads the drawings (L1) and verifies P&ID-to-CLD consistency (L2).
It is designed as a 100% open-source stack on the assumption of deployment into air-gapped networks, and data never leaves the site. Nuclear-grade reliability comes not from model performance but from honesty in the architecture.
We run a verification demo on one real set of your drawings.
Request a verification demo