Food literacy is increasingly measured to identify population needs, evaluate interventions, monitor change and inform public health policy, yet existing instruments present a fundamental measurement challenge: brief tools reduce respondent burden but may not comprehensively represent the construct, while comprehensive tools can be impractical for routine surveillance (1). The 100-item International Food Literacy Questionnaire (IFLQ-19) was developed according to the Vidgen and Gallegos food literacy framework, covering four domains and 11 theoretical components, from which 19 statistical components were derived during tool development (2). This comprehensive structure preserves theoretical breadth but makes the instrument too lengthy for routine use (3). To address this, a hierarchical "Russian doll" refinement approach is being used to develop four practical short-form tools nested within the original framework, each representing food literacy at a different level of specificity, from broad domain-level screening to detailed component-level assessment, so that users can select the tool that best fits their application, though these demands differ by level. This study aimed to test how refinement at different levels of this framework balances statistical robustness, theoretical comprehensiveness and practical feasibility. Secondary data from 503 Australian adults were analysed using a hierarchical, multi-criterion refinement approach integrating structural modelling and Rasch measurement. Two contrasting reduction strategies illustrate the trade-off between reducing respondent burden and maintaining measurement comprehensiveness. A statistically optimised 20-item version representing the four broad food literacy domains achieved excellent model fit (CFI = 0.988, RMSEA = 0.066) and high internal consistency (α = 0.914). However, item targeting showed limited discrimination among people with moderate-to-high food literacy, and reducing the questionnaire to broad domains sacrificed representation of more specific components that may be important for identifying intervention needs. Conversely, component-level Rasch refinement reduced the questionnaire from 100 to 74 items while preserving all 19 statistical components. Scores from the reduced and original versions remained strongly correlated across the four domains (Spearman's ρ = 0.974-0.989) but remained impractically long for surveillance settings. No single level of refinement satisfied statistical strength, theoretical comprehensiveness and practical feasibility together, demonstrating a genuine trade-off between competing priorities. These results show that a single reduction method does not serve every level of the framework equally well: the trade-off between discriminability, theoretical coverage and questionnaire length differs depending on the purpose of each intended tool. Current work therefore applies an Ant Colony Optimisation approach to identify item subsets balancing statistical performance and theoretical coverage, working towards a validated suite of four short-form IFLQ tools. Health practitioners need brief tools that avoid disrupting interventions during evaluation, and specific tools to guide investment decisions, needs heightened in priority-population interventions where long tools are impractical. Such a suite would support feasible monitoring of food literacy in national surveys and dietary assessment.