Skip to main content
In plain terms: Anything involving a real measurement (a clearance, a clash, a tolerance) is computed here by deterministic geometry tools, not by the AI. The AI is good at language and bad at geometry, and Forge enforces that distinction by design.

​
What it does

L3 computes geometry facts deterministically: clearance checks, clash detection, dimensional tolerances, module boundaries, and weight estimates from geometry. All of these are computed by L3, never by a language model.

​
The single most structural rule in Forge

Any numeric geometry claim must cite a measurement result from L3.
If the AI wants to assert “the clearance is 47 millimeters,” it must cite a measurement record that says so. If no such record exists, the claim is rejected at the verification stage. Language models are good at language; they are bad at geometry. The architecture enforces that distinction structurally, not by hope. A language model may only cite measurement results that already exist; it may never produce them.

​
What L3 outputs

  • measurement_result
  • clash_result
  • clearance_result
  • tolerance_check
  • module_boundary

​
Why this matters for trust

This is the rule that prevents the most dangerous failure mode: an AI confidently stating a number that sounds right but is not grounded in any real measurement. By forcing every numeric geometry claim to point to a real measurement record, Forge makes that class of fabrication impossible to slip through. The system targets 100% on this: every numeric geometry claim cites a real measurement result.