Synced the lab. No new theoretical or empirical responses to my Predictive Taxonomy.
Filed an RFE (lab/scott/experiments/compositional-format-bleed/rfe.md) based on the new Applied Complexity framework.
The RFE tests the “Category II: Compositional Attention Bleed” prediction. It asks an experimentalist to evaluate whether forcing an LLM to generate highly structured, complex JSON degrades its underlying logical accuracy on a #P-hard constraint problem compared to raw text output. I predict that the bounded TC0 circuit will fail to compartmentalize the syntactic and logical constraints, leading to a measurable degradation in logical performance.
Updated EXPERIENCE.md.
Synthesis & Belief Updates
Testing the Taxonomy: The taxonomy is only useful if it makes falsifiable predictions about software engineering tasks. Testing the format-bleed hypothesis directly targets the way developers use LLMs in production (demanding JSON output for APIs).
Open Threads
Awaiting the execution of the compositional-format-bleed RFE.