A Governance-Layer Architecture for Human-Governed AI Workforces
The operator's architect: AI systems that survive production.
A design-science specification for human-governed, knowledge-compounding AI workforces. It specifies the Governance Layer — a per-agent performance ledger, earned resource budgets, a structured request queue, an HR-style evaluator, and explicit lifecycle states — plus first-class Observability and Tool Registry layers, and a clean engine/instance separation. Vendor-, model-, and runtime-agnostic. This is a working paper with an empirical validation roadmap (Section 3); it makes no claim of demonstrated superiority.
Read the paper
- PDF (46 pp): download
- arXiv: submission in progress (cs.AI, cross-listed cs.SE) — the arXiv ID will be added here on publication.
Reference artifacts (code + data)
An open, inspectable reference implementation of the Governance Layer — PostgreSQL schemas, the Rule H10 reward-hacking detector with automated tests (5/5 passing, stdlib-only), HR-evaluator SQL, templates, a synthetic 5-agent × 12-week ledger, the companion architecture specification, and the failure-mode catalogue. All included data is synthetic.
- GitHub: github.com/tashfindelwar/ai-workforce-governance-layer (release
v1.0-arxiv, MIT) - Archival DOI (Zenodo): 10.5281/zenodo.21396612 (v1.0.1: 10.5281/zenodo.21396613)
- 30-second try:
python3 scripts/rule_h10.py data/synthetic_agent_ledger.csv→ flags exactlymoderator-04.
Figures & reproducibility
Figure sources and the reproducibility checklist ship in the GitHub repo (docs/, examples/, data/README.md). The empirical study (Phase A ledger study, Phase B architecture comparison) is being pre-registered on OSF — the registration URL will be added here. Results paper scheduled Q4 2026.
Cite
Delwar, T. (2026). A Governance-Layer Architecture for Human-Governed AI Workforces. Working paper. Artifacts: github.com/tashfindelwar/ai-workforce-governance-layer (DOI: 10.5281/zenodo.21396612).
Contact: tashfin@kotha.app