Evaluation
Track reference-resolution accuracy, OCR exact-match accuracy, OCR character error rate, chart value error, chart reconstruction similarity, object-color CIELAB distance, presentation theme consistency, uncached and cached latency by lens, external model call count, GLM input-token count, cache hit rate, provenance coverage, and low-confidence flagging.
The most important MVP behavior is that follow-up requests reuse stored visual assets and cached views rather than requiring another upload or recomputing unnecessary lenses.