Comparative experimental protocols
What the research will test
First modelling, then a limited pilot in a real organisation. Compare the protocol with a well configured ordinary workflow using the same tasks, data, resources and comparable AI support. Assess the contribution of meaningful confirmations and connected commitments: accepted outcomes, violations, recovery time after changes, errors and disputes, participant time and total costs. Benefits must justify additional work. Set success thresholds, stopping rules and evaluation procedures before trials; record negative and inconclusive results too.
How the protocol worksDraft protocols: distribution of authority and coordination mechanism are separate factors.
Comparable conditions
Within each authority structure, compare ObliNet with an explicitly defined baseline without ObliNet. Participants have comparable initial information, resources and AI support. Additional information acquisition and training count as costs. ObliNet receives no privileged access to simulator truth.
What to measure across all runs
Include every initiated case: successes, failures, refusals and unfinished activity at a fixed horizon. Measure shared-goal achievement, quality, resources used, elapsed time, work across all roles, computation, training, support and rework. Cost per successful outcome is a supplementary metric, not a substitute for complete accounting.
Mechanisms and causal contribution
Interpretation divergence, dependency visibility and revision latency are intermediate indicators. Disable condition checks, dependency display and revision separately. Assess their contribution against independent system-wide outcomes, not warning counts. Reduced disagreement alone does not establish benefit.
Before testing
Specify model and baseline versions, scenarios, authority rules, primary metric, quality, cost and autonomy thresholds, horizon and stopping rules, run-count rationale, random seeds, missing-data handling and analysis plan. Assess interaction between the two factors and uncertainty; publish null and adverse findings. Numerical parameters are not yet approved.
Adaptability after a change
After comparable external changes, measure time to an admissible revised plan, recovery of shared-goal achievement and additional costs. Faster revision is not an improvement if it harms quality or autonomy. Event-generation and adaptation-assessment rules are specified before execution.
| Distribution of authority | Without ObliNet | With ObliNet |
|---|---|---|
| Hierarchical | Baseline within a hierarchy | Same authority + ObliNet mechanisms |
| Decentralized | Baseline among autonomous participants | Same authority + ObliNet mechanisms |