Labs would exist so that policies can fail cheaply before expensive national implementation. Every project has a question, a method, a cost, an owner and a next gate. The portfolio below is a simulated first-year pipeline.
Q: Can aggregate procurement cycle-time be published without operational exposure?
Method: Data-classification review with security counsel.
Q: What is the causal effect of structured coaching on FLN outcomes?
Method: Pre-registered cluster RCT design.
Q: How accurate is satellite yield estimation vs. crop-cutting experiments in Indian smallholdings?
Method: Literature review + retrospective comparison in 3 states.
Q: What share of urban PM2.5 originates outside city boundaries?
Method: Chemical transport model calibrated on CPCB monitors; sensitivity analysis published.
Q: Can data reuse across departments be made auditable and consent-bound without a central database?
Method: Working prototype with federated verification; external security review.
Q: Does a 10-minute oral assessment reliably classify grade-level reading?
Method: Inter-rater reliability and test-retest across 200 districts.
Q: What share of the population is within 30 minutes of emergency-capable care, by district?
Method: OpenStreetMap + facility registry + road-speed calibration; 500 ground-truth routes per state.
Q: What does regulatory compliance cost a small firm as a share of revenue?
Method: Time-and-cost diaries in 600 firms across 5 sectors.
Q: Can a shorter instrument reproduce NSS catastrophic-expenditure estimates?
Method: Split-sample validation against full NSS module in 4 districts.
Outcomes, analysis plans, kill and scale criteria are published before a pilot begins.
External reviewers are named, disclose interests and their reports are published whole.
AI assists with literature discovery, comparison, consultation analysis and anomaly detection. Every AI output preserves sources, assumptions and uncertainty and passes human review. AI is never the final policy authority and is never used for sensitive political profiling.