Research Roadmap
Aligned with the proposed manuscript and prospective evaluation protocol.
Phase 1 — Research artifact (complete / in progress)
- Publish GitHub repository with verified LaTeX manuscript source
- Provide simulator, schemas, synthetic data, tests, and examples
- Separate verified evidence from vendor claims and forecasts
- Create a numbered release and archive with Zenodo
Phase 2 — Synthetic evaluation scaffold
- Generate 1,000+ synthetic scenarios
- Compare ASBOS against rule-based and monolithic-agent baselines
- Measure resolution rate, invalid assignments, escalation rate, event count, and audit completeness
- Note: These are simulator results supporting the evaluation protocol, not field pilot evidence
Phase 3 — Prospective pilot protocol
Suggested 90-day pilot: two baseline weeks, four staged deployment phases, and a final observation period.
Initial cohort may include 20–100 guards across 5–15 sites; final sample based on operational feasibility and statistical power.
Primary outcomes
- Administrative minutes per scheduled guard-hour
- Median and 90th percentile time to fill a coverage gap
- Percentage of scheduled shifts filled before start
- Payroll exception rate per 1,000 worked hours
- Report delivery compliance
- Credential-related assignment blocks
- Human escalation rate
- False-positive exception rate
Secondary outcomes
- Guard acceptance
- Supervisor workload
- Client satisfaction
- System availability
Study design should preserve raw event logs, versioned policies, agent prompts, model/tool versions, human overrides, and outcome definitions. A stepped-wedge or matched-site design is preferable to a simple before-after comparison.
Phase 4 — Paper submission and field evidence
- Publish preprint of manuscript
- Submit to agent-systems, HCI, or applied AI venues
- Report pilot results with confidence intervals
- Distinguish statistical significance from operational significance
Phase 5 — Economic transparency
- Report gross benefit, implementation cost, recurring cost, net benefit, benefit-cost ratio, and uncertainty separately
- Do not present sensitivity-model outputs as validated ROI