Predictive-System Evaluation Workbench

A reusable twenty-four-dimension workbench for evaluating place-based and person-based predictive systems as complete sociotechnical interventions.

Read the working brief Choose a learning path

REAL-WORLD INTERPRETIVE Editorial synthesis or explanation derived from sources and explicitly open to correction.

Confidence: Moderate
Sources: 2
Independent corroboration: 0
Jurisdiction: International / varies by jurisdiction
Date range: Historical through August 2, 2026; program status and evidence vary
Review: WIP.76 source-bounded synthesis — independent legal, scientific, operational, accessibility, affected-community, and specialist review pending

Estimated reading: 30 minutes
Review date: 2026-08-02
Stable route: /psyops/pre-crime/predictive-system-evaluation-workbench

Source boundary · evidence boundary

“Pre-crime” is an analytical label for an anticipatory decision architecture, not one technology or proof that future conduct can be known. Record retrieval is not prediction; identity matching is not prediction of conduct; association is not guilt; descriptive mapping is not necessarily forecasting; a concern category is not a calibrated probability; a score, tier, list, map, or alert is not lawful authority for coercive action; formal human presence is not meaningful review without time, evidence, independence, authority, and an effective intervention path; forecast accuracy is not demonstrated crime reduction; program silence is not proof of termination; a renamed successor is not proof of functional discontinuity.

REAL-WORLD INTERPRETIVE

ANALYTICAL & SAFETY BOUNDARIES

Keep related capabilities, evidence, and authority states separate.

Capability, acquisition, pilot, deployment, operational use, autonomy, effectiveness, and legal authority are separate evidence claims.
Human review is meaningful only when the reviewer can inspect evidence, reject the recommendation, and change the outcome in time.
A source describes or analyzes a claim; it does not automatically establish that claim as the repository conclusion.

Analytical boundary: “Pre-crime” is an analytical label for an anticipatory decision architecture, not one technology or proof that future conduct can be known. Record retrieval is not prediction; identity matching is not prediction of conduct; association is not guilt; descriptive mapping is not necessarily forecasting; a concern category is not a calibrated probability; a score, tier, list, map, or alert is not lawful authority for coercive action; formal human presence is not meaningful review without time, evidence, independence, authority, and an effective intervention path; forecast accuracy is not demonstrated crime reduction; program silence is not proof of termination; a renamed successor is not proof of functional discontinuity.

Simulation safety boundary: Neutral, non-operational education. No real-person scoring, real targeting, weapons configuration, surveillance deployment, evasion, jamming instruction, or system-defeat guidance.

Source basis: Place-based, person-based, technology-stack, and comparative evaluation reports.

LEVEL 1

ORIENTATION

Why this matters

REAL-WORLD INTERPRETIVE

One-sentence brief

Accuracy claims are incomplete without a target, denominator, calibration, comparison, intervention, causal design, feedback-loop controls, and downstream legal analysis.

REAL-WORLD INTERPRETIVE

Three key points

Separate forecasting from patrol or intervention.
Distinguish police-recorded outcomes from independently observed conduct.
Measure overrides and abstention.
Require causal comparison.
Audit vendor conflict and opacity.
Track successor systems and corrections.

LEVEL 2

WORKING BRIEF

Evidence, context, and limits

REAL-WORLD INTERPRETIVE

Evaluate the complete intervention

The model, human workflow, operational action, feedback, and outcome measurement must be evaluated together.

Separate forecasting from patrol or intervention.
Distinguish police-recorded outcomes from independently observed conduct.
Measure overrides and abstention.

REAL-WORLD INTERPRETIVE

Rejected shortcuts

An arrest after intensified attention, a crime in a forecast cell, or an attack that did not occur cannot by itself validate prediction or prevention.

Require causal comparison.
Audit vendor conflict and opacity.
Track successor systems and corrections.

LEVEL 3

COMPLETE DOSSIER

Limitations, game links, and review context

DISPUTED / MULTIPLE ACCOUNTS

Known limitations and gaps

Public records are incomplete and date-bounded; silence is not proof that a capability or program ended.
Official or manufacturer descriptions are not automatically independent validation of performance, legality, deployment, or effect.
Operational status, thresholds, configurations, and outcomes can vary by place, mission, time, and user.
The page is educational and does not provide operational targeting, surveillance, evasion, attack, or system-defeat instructions.

REAL-WORLD INTERPRETIVE

Related PsychologicalWar.org analysis

No related page is required to understand this record.

REAL-WORLD INTERPRETIVE

Glossary cross-reference

Terms are defined for this site’s evidence method, not as universal legal or clinical definitions.

currentness evidence class human review redress successor continuity loitering munition autonomy kill web

REAL-WORLD INTERPRETIVE

Decision matrix

Predictive-system evaluation dimensions
Dimension	Question	Completion rule
Prediction target and time horizon	What event is predicted, over what period, and for what decision?	A complete, dated, source-linked answer is provided; unknowns remain explicit.
Unit of analysis	Is the unit a person, place, event, household, identity, network, record, or information location?	A complete, dated, source-linked answer is provided; unknowns remain explicit.
Base rate and denominator	How common is the target outcome in the relevant population and period?	A complete, dated, source-linked answer is provided; unknowns remain explicit.
Training and validation period	Are temporal leakage, policy changes, and distribution shift addressed?	A complete, dated, source-linked answer is provided; unknowns remain explicit.
Data provenance	Which data are police-generated, administrative, commercial, sensor-derived, or independently observed?	A complete, dated, source-linked answer is provided; unknowns remain explicit.
Outcome measurement	Does the outcome measure conduct, victimization, reporting, contact, arrest, charging, or conviction?	A complete, dated, source-linked answer is provided; unknowns remain explicit.
Feature construction and proxy risk	Which transformations create proxies for protected status, poverty, place, or prior attention?	A complete, dated, source-linked answer is provided; unknowns remain explicit.
Missing-data treatment	Who is absent, misrecorded, or represented by a default?	A complete, dated, source-linked answer is provided; unknowns remain explicit.
False positives	How many people or places are selected without the target outcome?	A complete, dated, source-linked answer is provided; unknowns remain explicit.
False negatives	How many target outcomes occur outside selection?	A complete, dated, source-linked answer is provided; unknowns remain explicit.
Calibration	Do displayed probabilities correspond to observed frequencies?	A complete, dated, source-linked answer is provided; unknowns remain explicit.
Subgroup performance	Are error, calibration, and burden reported across relevant groups?	A complete, dated, source-linked answer is provided; unknowns remain explicit.
Uncertainty and abstention	Can the system say unknown, defer, or request more evidence?	A complete, dated, source-linked answer is provided; unknowns remain explicit.
Baseline comparison	Does the system improve over transparent hotspot methods, simple rules, or professional analysis?	A complete, dated, source-linked answer is provided; unknowns remain explicit.
Intervention type	What action follows selection, and how intrusive or supportive is it?	A complete, dated, source-linked answer is provided; unknowns remain explicit.
Intervention fidelity	Did practitioners actually deliver the intended intervention?	A complete, dated, source-linked answer is provided; unknowns remain explicit.
Causal design	Does the evaluation separate analytical method choice, intervention, and outcomes with a credible comparison?	A complete, dated, source-linked answer is provided; unknowns remain explicit.
Displacement and diffusion	Did effects move elsewhere or benefit nearby areas?	A complete, dated, source-linked answer is provided; unknowns remain explicit.
Feedback-loop controls	Are police-generated outcomes prevented from self-validating future predictions?	A complete, dated, source-linked answer is provided; unknowns remain explicit.
Override and rejection rates	How often do humans reject, modify, or ignore outputs, and under what conditions?	A complete, dated, source-linked answer is provided; unknowns remain explicit.
Downstream legal authority	What independent lawful basis is required for each action?	A complete, dated, source-linked answer is provided; unknowns remain explicit.
Retention, correction, deletion, and appeal	Can data, derived features, designations, and downstream copies be corrected and removed?	A complete, dated, source-linked answer is provided; unknowns remain explicit.
Vendor conflict and proprietary opacity	Are commercial claims, conflicts, source code, thresholds, and audit access disclosed?	A complete, dated, source-linked answer is provided; unknowns remain explicit.
Independent replication	Can an independent team reproduce the evaluation and inspect negative results?	A complete, dated, source-linked answer is provided; unknowns remain explicit.

REAL-WORLD INTERPRETIVE