The source separates three architectures: conventional OCR followed by LLM correction, VLM-native transcription, and agentic orchestration. It says an OCR-plus-LLM system may normalize an unusual but correctly read value even though the model does not see the original pixels. It also describes VLM outputs that omit rows from a dense table or repeat content while preserving a plausible table format. The source reports that ParseBench evaluates roughly 2,000 human-verified enterprise pages with more than 167,000 test rules, and says its leading methods reach about 90% content faithfulness.
The mechanism is that a decoder chooses likely tokens from both visual input and language context; weak visual evidence can be outweighed by the language prior. A token-level probability therefore reflects linguistic plausibility, not necessarily pixel-level support. The effect is most damaging in high-entropy fields—totals, serials, account numbers, and medical or product codes—where a plausible substitution can pass superficial validation.
Why now is that document systems are moving from OCR engines with glyph-level confidence outputs to VLM and LLM pipelines whose mistakes can look cleaner than the source. Teams that retain page-level string metrics may miss failures concentrated in the few fields that matter operationally.
Implications: production systems should retain a page-region reference for every consequential extracted field and test completeness, duplication, and cross-field consistency. Vendor evaluations should include the organization’s dense, unusual, and regulated documents rather than prose-heavy benchmark samples.
Evidence boundary: these claims and benchmark figures come from a vendor-authored LlamaIndex article, including discussion of its own ParseBench, so they establish a credible failure model but not an independent comparison of all OCR products.
Watch conditions: look for independent benchmarks that score visual grounding, field-level correctness, omissions, and fabricated values on real enterprise documents. Also watch whether document-extraction vendors expose auditable source-region citations and stable replay behavior for the same page.
This starts a distinct analytical record because no prior New Runtime record is supplied with a directly comparable examination of fluent, evidence-poor LLM OCR failure modes.