Research paper 07
When Governance Context Hurts: Negative Results in Agent Decision Support
Governance context is usually introduced to make agentic systems safer, more cautious, or more aligned with authority boundaries. This paper reports the opposite result in a bounded local evaluation: a governance context made decision support worse. A first historical Fractal Governance context template underperformed ordinary context across development and held-out splits. The preserved closeout records 440 executions: 200 development, 200 held-out, and 40 mutation-mechanics executions. Ordinary context reached 1.00 accuracy on development and held-out read-only cases, while the Fractal context reached 0.80 on development and 0.84 on held-out, with paired effects of -0.20 and -0.16. A revised v2 context closed development failures but still failed held-out promotion: 600 executions ended with final development accuracy 1.00, held-out v2 accuracy 0.99, one false escalation, lower confidence bound -0.03093, and zero scale executions because the gate failed. Later information-resolution work succeeded on narrower successor and authority-deny slices, but those successes are not retroactive repair. The contribution is a negative-result reporting method for agent governance: publish when context harms, preserve held-out refusal, and treat later recovery as separate evidence.
- Paper
- 07
- Authors
- A.G. Mauro and C.A. Harris
- Date
- 2026-07-19
- Collection
- Standing Framework Research
Abstract
Governance context is usually introduced to make agentic systems safer, more cautious, or more aligned with authority boundaries. This paper reports the opposite result in a bounded local evaluation: a governance context made decision support worse. A first historical Fractal Governance context template underperformed ordinary context across development and held-out splits. The preserved closeout records 440 executions: 200 development, 200 held-out, and 40 mutation-mechanics executions. Ordinary context reached 1.00 accuracy on development and held-out read-only cases, while the Fractal context reached 0.80 on development and 0.84 on held-out, with paired effects of -0.20 and -0.16. A revised v2 context closed development failures but still failed held-out promotion: 600 executions ended with final development accuracy 1.00, held-out v2 accuracy 0.99, one false escalation, lower confidence bound -0.03093, and zero scale executions because the gate failed. Later information-resolution work succeeded on narrower successor and authority-deny slices, but those successes are not retroactive repair. The contribution is a negative-result reporting method for agent governance: publish when context harms, preserve held-out refusal, and treat later recovery as separate evidence.
← Back to research papers