Step 1 FINAL: two independent passes scoring 16 real LLM-agent systems against Beer's VSM (S1-S5,S3*,algedonic,recursion) from source code, ignoring the systems' own VSM labels. 128 cells x 2 passes = 256 rows. Inter-coder reliability: exact 63% (81/128), only 2/128 hard Present<->Absent flips, linear-weighted Cohen kappa 0.79 (substantial), unweighted kappa 0.44 (moderate). Disagreement concentrates on the ordinal… See the full description on the dataset page:
https://huggingface.co/datasets/latkes/cybernetic-agents-validation-step1-characterization-v2-2pass.