Internal Consistency Does Not Validate an LLM-Assisted Nutrition Pipeline: A Reference-Standard-Free Forensic...

Large language model (LLM) pipelines increasingly convert free-text meal descriptions into nutrient values, yet reference-standard validation is rarely feasible in production. We asked whether the controls such pipelines conventionally pass suffice and built a reference-standard-free battery any pipeline can run on its output.

We audited 13,073 meals logged by 513 adults through a WhatsApp conversational agent in rural Colombia (9 February 2026 to 6 August 2026). The pipeline is a two-stage cascade of commercial LLMs from different providers: a vision model renders photographs as text, while a text model maps these to nutrients. Six pre-specified tests were applied, including test-retest ICC(1,1) over recurring strings, stratified by modality because 48.3% of records carried a photograph.

The pipeline passed every conventional control (Atwater-reconstructed energy r = 0.9951; median relative error 0.0000) but failed reproducibility. Across 191 recurring strings (944 meals), energy ICC point estimates were 0.7061 (95% CI 0.5419-0.8217) text only and 0.5258 (0.3778-0.6513) with photographs; none reached 0.75. Modality is a design control, not an established effect, and the coefficient depends on which inputs recur, so it is not intrinsic. Under the strictest identity available-byte-identical text, no photograph-ICC was 0.6944. Intermediate-text inspection exposed a failure invisible at the output: 1.44% of photograph-accompanied records carried a vision-stage artefact string versus 0.01% of text-only records (ICC 0.0911).

Internal consistency does not validate an LLM-assisted nutrition pipeline. This design measures instability under recurring inputs without isolating its cause. Prospective test-retest reproducibility is a cheap, necessary acceptance criterion, and in a cascade, the intermediate representation must also be inspected.