Negative Ontology of True Target for Machine Learning: Towards Evaluation and Learning under Democratic Supervision

Referee: [Abstract and introduction] Abstract and introduction: The central claim that explicitly adopting the negative ontology of TT grounds the definition of Democratic Supervision and enables the derivation of MIATTs principles, logical assessment formulation, and undefinable true target learning is not supported by any argument demonstrating that these components cannot be obtained from a positive-ontology model treating the target as a random variable with epistemic uncertainty or from standard multi-annotator aggregation. Without such a comparison or counterexample, the ontology shift functions as an optional layer rather than a load-bearing foundation for the EL-MIATTs framework.

Authors: The negative ontology is load-bearing because it rejects the existence of any TT (even as a latent random variable), which precludes aggregation toward a single target and instead requires supervision mechanisms that treat all targets as inherently inaccurate and non-convergent. Positive-ontology models, by assuming an underlying TT with epistemic uncertainty, lead to standard multi-annotator methods that seek to approximate or average toward that target. Our framework derives Democratic Supervision and MIATTs principles directly from non-existence, yielding logic-driven generation rules that do not rely on convergence or error minimization relative to a hidden truth. To make this distinction explicit, we will add a short comparative paragraph in the revised introduction. revision: partial

Referee: [MIATTs principles and logical assessment formulation] Sections deriving MIATTs principles and logical assessment formulation: The logic-driven generation/assessment of MIATTs and the logical assessment formulation for evaluation appear to restate the initial non-existence assumption without independent external benchmarks, falsifiable predictions, or data-driven validation, resulting in circularity where the framework's components reduce to restatements of the posited axiom.

Authors: The derivations are deductive rather than circular: starting from the non-existence axiom, we specify new operational principles for generating and assessing multiple inaccurate targets (e.g., logical consistency across perspectives without privileging any as closer to truth) and a corresponding assessment formulation that evaluates predictive models on their alignment with the set of MIATTs. These are novel constructs not present in positive-ontology treatments. While the work is conceptual and does not include empirical benchmarks, the logical structure itself constitutes the contribution and can support falsifiable tests in applied settings. revision: no

Referee: [Evaluation and learning sections] Evaluation and learning with MIATTs sections: No formal derivations, empirical results, error analysis, or concrete examples are provided to show that the proposed formulations improve evaluation or learning outcomes relative to existing paradigms, leaving the practical claims of the EL-MIATTs framework unsupported.

Authors: The manuscript provides formal derivations in the sections on MIATTs principles, the logical assessment formulation, and undefinable true target learning. The real-world application in education and professional development functions as a concrete illustrative example of the framework in use. As a primarily philosophical and conceptual paper, it does not include quantitative empirical comparisons or error analysis; such validation is appropriate for follow-up work. We will expand the application section with additional detail on how the formulations differ from standard approaches in a revised version. revision: partial