Recorded run · replay
Turns a social-science theory into a tested analysis pipeline, then runs it on new text.
The run was askedOnce Heider's balance theory is compiled into code, does it tell apart a friendship where two people agree from one where they clash?
Once a theory is compiled, its logic is fixed and tested; only the reading of the text is left to the model, so two different situations get two clearly different, explainable answers.
Leaves out how the paper was turned into a schema, the generated prompts and extraction code, the strength-of-feeling score (1.0 in both stories), and the written reasoning each classification step produced.
Why custom: Asking a chatbot to 'apply balance theory' gives a fresh, unrepeatable reading each time; here the theory's arithmetic is fixed, tested code and the model only pulls the inputs out of the text.
Replayed from: evidence/balance_theory_runtime_20260404.json (inputs: scripts/run_balance_theory_analysis.py; formulas and tests: docs/plans/100_balance_theory_compilation.md)