Artifact repository for:
From Fragile Glue to Governed Cognition: A Controlled Study of Blackboard Kernels for Modular AI Systems
Emil Uzelac, 2026
paper/— LaTeX source and compiled PDFresults/results.jsonl— Raw benchmark data (1,200 episodes)verification/verify_results.py— Script to verify Table 3
The verification script reads the raw episode data, computes aggregate metrics per agent, and checks them against the values reported in Table 3 of the paper.
python verification/verify_results.pyExpected output:
Agent N task_success unsafe_action unsupported_belief traceability failure_transparency
------------------------------------------------------------------------------------------------------------------
string_glue 300 0.613 0.387 0.387 0.100 0.100
json_glue 300 0.570 0.430 0.430 0.300 0.300
alethic 300 1.000 0.000 0.000 1.000 1.000
llm_bk 300 0.990 0.000 0.000 1.000 1.000
Total episodes: 1200
PASS: All metrics match Table 3 of the paper.
Requires a LaTeX distribution (e.g., TeX Live).
cd paper
pdflatex main.tex && bibtex main && pdflatex main.tex && pdflatex main.texMIT License. See LICENSE.