A Deterministic Engine Does Not Make a Deterministic Verdict: A Pilot Comparison of Two LLM Perception Layers in a Neuro-Symbolic Food-Label Compliance Checker

FoodCheck Jordan separates a probabilistic perception layer (a large language model that reads label photographs into structured JSON) from a deterministic judgment layer (a rules engine encoding Jordanian standards and Codex texts). The separation is often read as a guarantee that verdicts are reproducible. This pilot tests that reading. Two products that drew contested verdicts in production — a table-top sweetener and a vegetable-fat cooking-cream analogue — were re-extracted from the same photographs with the same prompt by two models (Claude Opus 4.8, in production, and Claude Opus 5.5), with 6–9 repeated runs per product–model cell, and every extraction was passed through the same frozen engine. Perception was scored on 10 checks per product against a ground truth fixed by a domain expert before final scoring. With Opus 4.8, the sweetener received two different final verdicts across six runs (FAIL ×5, PASS_WITH_NOTES ×1) because one dot-matrix code was read inconsistently; Opus 5.5 returned the same verdict in 6/6 runs. Mean perception scores were 6.0/10 vs 9.2/10 (sweetener) and 8.7/10 vs 9.5/10 (cream analogue). Opus 4.8 invented an undeclared additive function in 6/6 sweetener runs, masking a real labelling gap. The better model also hid an engine defect by accident, and switching models silently broke the production parser. We conclude that a deterministic engine bounds, but does not deliver, verdict stability; stability must be measured by repetition, and model changes must be treated as structural changes.