ConstValidate one observed design reference as evidence.
The record is checked as a self-consistent observation, never as a design: a frame states how its own units are read, a primitive states what was on the sheet, a candidate states one proposed meaning, and an issue states what is still undecided. The validator's whole job is to keep those four layers from collapsing into each other, because the moment an observation is allowed to carry a conclusion, the design source of truth has been silently replaced.
So it refuses a candidate that cites a primitive nobody saw, an issue about nothing, an analysis that claims observed while producing no reading, and a non-observed analysis without a non-blank reason. Extra runtime properties on non-observed analysis objects are not inspected. It also refuses a settled scale that names no recorded candidate: an unknown scale is a fact about the drawing, and inventing one is how a plan image turns into metres nobody measured.
The same principle binds an analysis to the one frame it says it read: a candidate built from marks on another sheet was not produced by this reading, and filing it here would let a section's geometry inherit a plan's scale. Correlating two sheets is the authored building's job, through evidence that may cite any candidate of any document.
Observed points are in frame units and must lie within the frame's bounds (0 to width, 0 to height); each primitive family has its own point-count bounds, and only text and level-marker carry literal text. Frames are judged by validateDesignSourceFrame. Every id is unique and non-blank, the digest is a lowercase sha256: hex digest, and the media is a supported type. Findings are returned with $input paths and nothing throws; the asset bytes behind digest are not read here.
Observed design reference to validate.
Success, or every violation found.
Validate one observed design reference as evidence.
The record is checked as a self-consistent observation, never as a design: a frame states how its own units are read, a primitive states what was on the sheet, a candidate states one proposed meaning, and an issue states what is still undecided. The validator's whole job is to keep those four layers from collapsing into each other, because the moment an observation is allowed to carry a conclusion, the design source of truth has been silently replaced.
So it refuses a candidate that cites a primitive nobody saw, an issue about nothing, an analysis that claims
observedwhile producing no reading, and a non-observed analysis without a non-blank reason. Extra runtime properties on non-observed analysis objects are not inspected. It also refuses a settled scale that names no recorded candidate: an unknown scale is a fact about the drawing, and inventing one is how a plan image turns into metres nobody measured.The same principle binds an analysis to the one frame it says it read: a candidate built from marks on another sheet was not produced by this reading, and filing it here would let a section's geometry inherit a plan's scale. Correlating two sheets is the authored building's job, through evidence that may cite any candidate of any document.
Observed points are in frame units and must lie within the frame's
bounds(0 to width, 0 to height); each primitive family has its own point-count bounds, and onlytextandlevel-markercarry literal text. Frames are judged by validateDesignSourceFrame. Every id is unique and non-blank, the digest is a lowercasesha256:hex digest, and the media is a supported type. Findings are returned with$inputpaths and nothing throws; the asset bytes behinddigestare not read here.