Limitations
Measurements need supplied evidence.
Kelra is a measurement layer. It is not a visual reasoning model, screenshot-to-code generator, autonomous coding agent, or source editor.
- Kelra does not identify arbitrary named objects such as “the search icon” without supplied bounds, coordinates, runtime IDs, or supported evidence.
- Image-only registration estimates translation. It may abstain, and it does not infer rotation, scale, or affine motion without supplied correspondences.
- It does not provide physical units without calibration.
- Runtime-aware evidence currently targets instrumented SwiftUI apps and requires a synchronized snapshot from the caller.
- Pixel-to-runtime association is inferred from geometry and does not prove ownership or causality.
- Source influence candidates do not prove which source property caused a mismatch or what value to change.
- Kelra does not generate or modify application code.
- Claude Code, Cursor, Copilot, and other MCP clients have not been validated.