From the Public Safety edition of September 18, 2026
arXiv: Triage Score — a counterfactual risk assessment instrument (Greiner/Imai et al.)
arXiv (stat.AP preprint) · Courts / criminal justice methodology · United States (methodological; pretrial RCT illustration)
- Publisher
- arXiv:2606.21050
- Original publication
- June 19, 2026 (submitted)
- Source retrieved
- 2026-09-19
What happened
Imai, Shin, Greiner, and Halen (submitted 2026-06-19) propose triage scores based on additive counterfactual utilities, with conventional risk scores as a special case that ignore outcomes under intervention. They argue effective decision-making requires comparing potential outcomes across alternative decisions and illustrate with their randomized controlled trial evaluating a pretrial risk score, finding triage scores can yield substantively distinct policy-evaluation and learning results. Newly relevant on 2026-09-18 because Norfolk's Harvard A2J plea pilot is described as implementing this triage-score approach with overlapping investigators.
Why it matters
Provides the methodological backbone SLED courts/prosecutors should demand when vendors or labs pitch 'better than risk scores' decision AI—clarify counterfactuals, utilities, and evaluation design.
Evidence and measured results
Opened arXiv abstract and HTML full text via curl; confirmed authors, submit date, triage vs risk definition, and pretrial RCT illustration.
Limitations and uncertainty
Preprint; illustration is pretrial PSA RCT, not the Norfolk plea field study; does not report Norfolk/Philadelphia results.
Put this evidence to work
Lighthouse Advisory interpretation, grounded in this source. Enriched 2026-09-19; this does not change the original publication date. Each role follows the same checklist so takeaways compare across sources.
Sales
Role takeaway
- Customer problem
- Conventional risk scores omit what happens if the agency intervenes, misaligning tools with plea and release decisions that change outcomes.
- Stakeholders
- Prosecutor leadership
- Pretrial services
- Court administration
- Academic methods partners
- Defense stakeholders
- Discovery questions
- Which decisions and counterfactual outcomes must the score support?
- Who sets and reviews utility weights?
- What experimental or quasi-experimental data exist?
- How will results be compared to existing risk tools?
- Value hypothesis
- Bounded value is clearer policy evaluation when tools estimate effects of actions—not automatic superiority over current practice.
- Potential engagement
- Methods workshop + utility elicitation sprint to decide whether a triage-score pilot is feasible with available data.
- Unsupported claims
- Norfolk plea pilot results
- Guaranteed fairness or accuracy gains
- That pretrial RCT findings transfer unchanged to pleas
Pre-sales engineering
Role takeaway
- Fit
- Statistical decision-support methodology for justice pilots; requires strong evaluation design rather than off-the-shelf risk API swap.
- Architecture and integration
- Model outcomes under alternative actions; configuration layer for utilities; reporting compares triage vs risk-score policies; human decision remains outside the model.
- Prerequisites
- Defined action set and outcomes
- Credible identification strategy or RCT
- Governance for utility parameters
- Constraints
- Preprint status
- Illustration uses pretrial PSA RCT, not plea bargaining field results
- Utility choices are normative
- Security
- Research data minimization; restrict who can alter utility parameters; version all scoring configs.
- Proof of value
- Replication of paper's policy-evaluation contrast on local holdout/experimental data
- Sensitivity analysis across utility weights
- Documented assumption checks
Delivery
Role takeaway
- Work
- Elicit utilities, prepare analysis pipelines, validate triage vs risk policies, and package advisory outputs for practitioners.
- Dependencies
- Agency data access
- Methods authors or equivalent statisticians
- Legal review of research use
- Ownership
- Research PI for methods integrity; agency lead for operational interpretation limits.
- Skills and adoption
- Train staff that triage outputs depend on stated utilities and counterfactual assumptions.
- Governance checkpoints
- Utility sign-off
- Pre-analysis plan
- External methods review before operational advice
- Acceptance criteria
- Written utility rationale
- Comparative evaluation report
- Advisory-only label on any pilot UI
- Risks
- Silent utility encoding of bias
- Over-transfer from pretrial to plea context
- Treating preprint as deployed product evidence
Implementation considerations
Lighthouse Advisory interpretation across the operating dimensions a public-sector buyer must settle before this evidence becomes a design. Each note answers the question under its heading for this specific source.
Architecture and integration
What must connect, and where does the AI sit in the workflow?
Model outcomes under alternative actions; expose utility parameters; keep recommendation UI separate from estimation; version configs.
Governance
Who approves, reviews and stays accountable for outcomes?
Utility weights are policy choices needing sign-off; preprint status requires independent validation before operational reliance.
Security and privacy
What data, permissions and controls need testing?
Experimental/administrative justice data used for fitting need research governance and access controls.
Accessibility and workforce
Who is affected, and what skills or accommodations follow?
Methods literacy for non-statistician prosecutors is required or agencies will misread triage outputs as simple risk labels.
Procurement
What should contracts, pricing and exit terms secure?
RFPs for decision AI should require counterfactual evaluation plans and utility documentation, not accuracy-on-risk alone.
Operating model
Which teams own the service once it runs?
Use triage methodology to structure pilots; do not equate paper publication with production readiness.
Publication history
- 2026-09-18Public Safety · Issue 124 resources
Stable resource ID: arxiv-triage-score-counterfactual-202606