FE-012C Threats To Validity

• Recognition bias: If models recognized artifacts, blinding was weak.
• Shared training data: Model overlap may inflate agreement through shared exposure rather than independent convergence.
• Prompt and instrument effects: A common packet structure can induce common output structure.
• Researcher-designed primitive vocabulary: The provided vocabulary may constrain or channel the outputs.
• Model-family limitations: Agreement across three model families is still model-based evidence only.
• No human validation: Human analysts were not part of this result set.
• No domain expert validation: Domain experts did not review extraction quality in this run.
• Ambiguity in primitive boundaries: Boundaries between fields such as Evaluate, Compare, Prioritize, Verify, and Decide remain contestable.
• JSON schema effects: Required fields may force artificial precision or convergence.
• Likely over-agreement due to provided vocabulary: Lack of candidate missing primitives may partly reflect the constraint of the available vocabulary rather than true sufficiency.