1. Black-box intake decision
1. Decisión de entrada de caja negra
The Module III MER is accepted as the sole technical evidence package. No metric is recalculated, no failed model is repaired, and no causal effect is invented.El MER del Módulo III se acepta como único paquete técnico. No se recalcula ningún índice, no se repara ningún modelo fallido y no se inventa ningún efecto causal.
| MER execution | EXECUTED WITH FLAGS | |
| Interpretation readiness | READY WITH GAPS | |
| Reason | Association, time, space, integration, diagnostics, and failure flags are available. Prediction has critical generalization flags; explanation is closed; causality is deferred. | Asociación, tiempo, espacio, integración, diagnósticos y banderas están disponibles. Predicción tiene banderas críticas; explicación está cerrada; causalidad diferida. |
2. Objective-function profile
2. Perfil de función objetivo
OBJECTIVE FUNCTION
Primary outcome: Accident_Severity
Direction: MINIMIZE
Desired movement:
Fatal → Serious → Slight
Secondary consequence objective:
Number_of_Casualties
Status: SEPARATE, not merged into severity
Decision horizon:
Operational / policy planning horizon — not estimable from the MER alone
Minimum meaningful change:
NOT DEFINED IN THE MER
Tolerance for false negatives:
LOW for Fatal and Serious cases
Tolerance for false positives:
UNDEFINED
Exposure-adjusted risk:
NOT AVAILABLE
The objective function is directionally clear, but the MER does not define a minimum meaningful reduction, exposure denominator, implementation horizon, or acceptable error trade-off. Therefore, action ranking can only be conditional and qualitative.La función objetivo es clara en dirección, pero el MER no define reducción mínima significativa, denominador de exposición, horizonte ni trade-off aceptable. El ranking solo puede ser condicional y cualitativo.
3. Integrated Evidence Profile (IEP)
3. Integrated Evidence Profile (IEP)
| Evidence family | Strength | Decision interpretation | Boundary | ||
|---|---|---|---|---|---|
| Association | MODERATE / CONDITIONAL | District shows the largest categorical association; Weather, Road Type, and Junction Detail show smaller measurable signals. | District muestra la mayor asociación; clima, tipo de vía y cruce muestran señales menores. | Small magnitudes, sparse expected cells, observational only. | Magnitudes pequeñas, celdas escasas, solo observacional. |
| Information | WEAK | All normalized MI values are small; variables carry limited standalone information under current coding. | Todos los NMI son pequeños; información individual limitada. | Estimator and sparse-category sensitivity. | Sensibilidad al estimador y categorías escasas. |
| Time | MODERATE / CONDITIONAL | Recorded severe-accident proportions vary by hour; daily aggregate severity shows autocorrelation. | La proporción severa varía por hora y la serie diaria muestra autocorrelación. | No traffic-exposure denominator; aggregated event patterns only. | Sin denominador de exposición; solo patrones agregados. |
| Space | MODERATE / CONDITIONAL | A small positive global clustering signal exists among recorded events. | Existe una pequeña señal global de agrupamiento entre eventos registrados. | Duplicated coordinate pairs and no exposure denominator. | Coordenadas duplicadas y sin exposición. |
| Cross-objective severity/casualties | WEAK | The relationship is measurable but very small; objectives should remain separate. | La relación es medible pero muy pequeña; objetivos separados. | No justification for combining outcomes. | Sin justificación para combinarlos. |
| Prediction | INSUFFICIENT / CRITICAL | The current models do not generalize acceptably out of time and perform poorly for Serious and Fatal classes. | Los modelos actuales no generalizan fuera del tiempo y fallan en Serious/Fatal. | Random CV materially overstates performance relative to chronological holdout. | CV aleatorio sobreestima frente al holdout cronológico. |
| Explanation | GATE CLOSED | No final explanation should be produced for an unapproved predictive model. | No debe explicarse un modelo predictivo no aprobado. | Await new predictive iteration. | |
| Causality | DEFERRED | No intervention effect may be claimed. | No puede afirmarse efecto de intervención. | No causal architecture. |
4. Evidence convergence and conflict
4. Convergencia y conflicto de evidencia
Convergent patternPatrón convergenteAssociation, information, and cross-objective results consistently indicate that most individual categorical signals are small.Asociación, información y objetivo cruzado indican consistentemente señales pequeñas.
Material conflictConflicto materialRandom stratified CV suggests moderate predictive structure, while chronological holdout shows weak real-time generalization and rare-class failure.CV aleatorio sugiere estructura moderada, pero holdout cronológico muestra generalización débil y fallo de clase rara.
Conditional spatial/temporal signalSeñal espacial/temporal condicionalPatterns are measurable among recorded accidents, but not interpretable as population risk without exposure.Patrones medibles entre accidentes registrados, no como riesgo poblacional sin exposición.
Data-quality convergenceConvergencia de calidadSparse categories, invalid conditional fields, duplicated coordinates, and source-label correction consistently constrain stronger conclusions.Categorías escasas, campos inválidos, coordenadas duplicadas y corrección de etiqueta limitan conclusiones.
5. Practical significance assessment
5. Evaluación de significancia práctica
| Finding | Statistical detectability | Practical usefulness | Decision status | |
|---|---|---|---|---|
| District association | Detectable | Potentially useful for targeted review, but not sufficient for intervention priority. | Potencialmente útil para revisión focalizada, no suficiente para priorizar intervención. | CONDITION |
| Hour-of-day differences | Detectable | Useful for descriptive scheduling or surveillance only; not risk estimation. | Útil para programación descriptiva o vigilancia, no para estimar riesgo. | CONDITION |
| Spatial clustering | Detectable | Useful for hotspot review among recorded events; not exposure-adjusted hotspot risk. | Útil para revisar hotspots de eventos, no riesgo ajustado por exposición. | CONDITION |
| Current prediction models | Technically executed | Not practically useful for deployment due to out-of-time and rare-class failure. | No útil para despliegue por fallo fuera del tiempo y clases raras. | RETURN |
| Severity–casualties relation | Detectable | Too weak to justify combining objectives. | Demasiado débil para combinar objetivos. | KEEP SEPARATE |
6. Variable Role Matrix
6. Matriz de Roles de Variables
| Variable / family | Role | Actionability | Module IV interpretation | |
|---|---|---|---|---|
| Local_Authority_(District) | Administrative / contextual marker | Not directly actionable | Use for stratified monitoring, data-quality review, and targeted investigation—not as a causal lever. | Usar para monitoreo estratificado e investigación, no como palanca causal. |
| Weather_Conditions | Observed environmental condition | Not controllable | Monitoring and operational adaptation candidate; not modifiable. | Candidato a monitoreo y adaptación operativa; no modificable. |
| Road_Type | Infrastructure/context marker | Potentially modifiable only through separate causal/infrastructure evidence | Action candidate hypothesis, not eligible action variable in this MER. | Hipótesis de acción, no variable elegible con este MER. |
| Junction_Detail | Infrastructure/context marker | Potentially modifiable | Requires location-specific design, exposure, and intervention evidence. | Requiere diseño local, exposición y evidencia de intervención. |
| Light_Conditions | Observed condition / possible infrastructure proxy | Partially modifiable | No meaningful association was established here; no action priority. | No se estableció asociación relevante; sin prioridad de acción. |
| Speed_limit | Policy variable | Modifiable | Not evaluable because 98.9% of records are at 30; descriptive only. | No evaluable porque 98.9% está en 30; solo descriptivo. |
| Junction_Control | Infrastructure/control variable | Potentially modifiable | Not eligible due to 19.4% invalid data and exclusion from core modeling. | No elegible por 19.4% inválido y exclusión del modelado. |
| Carriageway_Hazards | Observed hazard marker | Potentially modifiable | Not eligible because 99.2% “None” makes the evidence unstable. | No elegible porque 99.2% “None” vuelve inestable la evidencia. |
| Hour / time segment | Operational timing marker | Indirectly actionable | May guide surveillance allocation, subject to exposure-adjusted confirmation. | Puede guiar vigilancia, sujeto a confirmación ajustada por exposición. |
| Coordinates / spatial cluster | Location marker | Indirectly actionable | May guide site review, not direct intervention ranking. | Puede guiar revisión de sitios, no ranking directo de intervención. |
7. Action Variable Register
7. Registro de Variables de Acción
| Candidate | Modifiable | Temporal precedence | Causal support | Eligibility |
|---|---|---|---|---|
| Speed limit | YES | Likely before outcome | NONE IN THIS MER | OBSERVATIONAL ONLY |
| Junction control | YES / PARTIAL | Before outcome | NONE IN THIS MER + DATA FAILURE | NOT ELIGIBLE |
| Road design / type | PARTIAL | Before outcome | OBSERVATIONAL ONLY | CONDITIONAL HYPOTHESIS |
| Lighting infrastructure | YES / PARTIAL | Before outcome | NO SIGNAL IN CURRENT MER | NOT PRIORITIZED |
| Operational patrol / response scheduling by time | YES | Before some future events | NO EXPOSURE-ADJUSTED EFFECT | MONITORING-ONLY PILOT |
| Targeted site review by spatial cluster | YES | Before future action | RECORDED-EVENT CLUSTER ONLY | DIAGNOSTIC FIELD REVIEW |
No candidate passes the full causal gate for a claimed severity-reduction intervention. Therefore, the DER may recommend evidence collection, monitoring, and controlled pilot design—but not claim that changing any candidate will reduce accident severity.Ningún candidato supera la puerta causal completa. El DER puede recomendar recolección de evidencia, monitoreo y pilotos controlados, pero no afirmar reducción de severidad.
8. Scenario Portfolio
8. Portafolio de Escenarios
| Scenario | Description | Expected benefit | Main risk | Status | |||
|---|---|---|---|---|---|---|---|
| S0 — Current state / no action | Continue current data collection and analytical architecture unchanged. | Continuar datos y arquitectura sin cambios. | No implementation burden. | Sin carga de implementación. | Persistent rare-class failure, unstable prediction, unresolved exposure bias. | Persiste fallo de clases raras, predicción inestable y sesgo de exposición. | NOT PREFERRED |
| S1 — Predictive deployment now | Use current logistic or random-forest model operationally. | Usar el modelo actual operacionalmente. | Potential automated prioritization. | Priorización automatizada potencial. | High false-negative risk for Serious/Fatal; temporal generalization failure. | Alto riesgo de falsos negativos y fallo temporal. | REJECT / RETURN |
| S2 — Rebuild evidence architecture | Repair data defects, obtain exposure denominators, preserve chronological validation, redesign rare-class strategy, and re-execute Module III. | Reparar datos, obtener exposición, mantener validación cronológica, rediseñar clases raras y reejecutar Módulo III. | Highest information gain and best path toward defensible prediction. | Mayor ganancia de información y mejor ruta a predicción defendible. | Time and data burden. | Carga de tiempo y datos. | PREFERRED |
| S3 — Descriptive monitoring program | Use district, hour, and spatial patterns only for surveillance dashboards and field review, with explicit non-risk labels. | Usar distrito, hora y espacio solo para monitoreo y revisión de campo, con etiquetas no-riesgo. | Immediate bounded operational usefulness. | Utilidad operativa inmediata y acotada. | Users may overinterpret recorded-event patterns as risk. | Usuarios pueden confundir eventos con riesgo. | CONDITIONAL |
| S4 — Controlled intervention pilot design | Design a separate causal or quasi-experimental pilot for a selected modifiable road-safety action. | Diseñar un piloto causal o cuasiexperimental para una acción modificable. | Creates intervention evidence. | Crea evidencia de intervención. | Cannot be selected from the current MER alone; requires new design and domain expertise. | No puede seleccionarse con este MER; requiere nuevo diseño y expertos. | DEFER |
9. Risk–Benefit–Burden–Opportunity assessment
9. Evaluación riesgo–beneficio–carga–oportunidad
| Scenario | Benefit | Risk | Burden | Opportunity loss | |||
|---|---|---|---|---|---|---|---|
| S0 | Low | High risk of analytic stagnation. | Alto riesgo de estancamiento. | Low immediate | High loss from delayed data repair and model redesign. | Alta pérdida por retrasar reparación y rediseño. | |
| S1 | Superficial operational convenience. | Conveniencia superficial. | Critical due to rare-class and temporal failures. | Crítico por fallos de clases raras y tiempo. | Medium | High: deployment could consume trust and resources while missing severe cases. | Alta: puede consumir confianza y recursos mientras pierde casos severos. |
| S2 | High long-term evidence value. | Alto valor de evidencia a largo plazo. | Low safety risk; main risk is delay. | Bajo riesgo de seguridad; principal riesgo es demora. | High data/technical burden | Moderate if redesign is slow, but lower than deploying a failed model. | Moderada si el rediseño es lento, menor que desplegar modelo fallido. |
| S3 | Moderate immediate monitoring value. | Valor moderado inmediato. | Moderate overinterpretation risk. | Riesgo moderado de sobreinterpretación. | Low–medium | Low if clearly labeled and used alongside S2. | Baja si se etiqueta claramente y acompaña S2. |
| S4 | Potentially high causal learning. | Aprendizaje causal potencialmente alto. | Unknown until intervention is specified. | Desconocido hasta especificar intervención. | High design burden | High if causal evidence is indefinitely postponed. | Alta si la evidencia causal se pospone indefinidamente. |
10. Scenario ranking
10. Ranking de escenarios
| Rank | Scenario | Reason | Decision | |
|---|---|---|---|---|
| 1 | S2 — Rebuild evidence architecture | Best balance of safety, information gain, future decision quality, and avoidance of premature deployment. | Mejor balance de seguridad, ganancia de información y calidad futura. | SUPPORT |
| 2 | S3 — Descriptive monitoring program | Can create bounded immediate value if strictly labeled and not treated as risk or causality. | Puede crear valor inmediato acotado si se etiqueta estrictamente. | CONDITION |
| 3 | S4 — Controlled intervention pilot design | Important future path, but the current MER cannot select the intervention. | Ruta futura importante, pero el MER actual no selecciona intervención. | DEFER |
| 4 | S0 — No action | Avoids immediate burden but preserves known defects and opportunity loss. | Evita carga inmediata, pero conserva defectos y pérdida de oportunidad. | NOT PREFERRED |
| 5 | S1 — Predictive deployment now | Critical generalization and rare-class failures make operational use indefensible. | Fallos críticos hacen indefendible el uso operacional. | RETURN / REJECT |
11. Final Decision Engineering Report
11. Decision Engineering Report final
DECISION ENGINEERING REPORT — CASE 01
Case:
Road Accident System
Objective:
Minimize Accident_Severity while keeping Number_of_Casualties
as a separate consequence objective.
MER Readiness:
READY WITH GAPS
Integrated Evidence:
- Association: measurable but mostly small and conditional
- Information: weak
- Time: usable for descriptive monitoring only
- Space: usable for recorded-event review only
- Prediction: insufficient for deployment
- Explanation: gate closed
- Causality: deferred
Variable Roles:
- District, weather, hour, and coordinates are markers/context
- Speed limit, junction control, road design, and lighting are
potential action hypotheses only
- No variable passes the full causal action gate
Preferred Decision:
SUPPORT Scenario S2 — rebuild the evidence architecture
Conditional Parallel Decision:
CONDITION Scenario S3 — allow descriptive monitoring dashboards
only with explicit non-risk and non-causal labeling
Rejected Decision:
Current predictive deployment
Deferred Decision:
Selection of a road-safety intervention
Required Return:
Module I / II / III iteration with:
- source-label repair
- improved conditional-variable quality
- traffic-exposure denominators
- duplicate-coordinate review
- chronological validation as the primary standard
- rare-class strategy for Serious and Fatal outcomes
- re-evaluation of prediction before explanation
Final Module IV Status:
DER COMPLETE WITH CONDITIONS
12. Final Module IV decision
12. Decisión final del Módulo IV
The current predictive architecture is not supported for operational deployment. The evidence architecture should be rebuilt and revalidated. In parallel, the descriptive temporal, spatial, and district patterns may be used conditionally for surveillance and field-review purposes, provided they are not presented as exposure-adjusted risk, causal evidence, or intervention proof.
La arquitectura predictiva actual no se respalda para despliegue operacional. Debe reconstruirse y revalidarse la arquitectura de evidencia. En paralelo, los patrones descriptivos temporales, espaciales y por distrito pueden usarse condicionalmente para vigilancia y revisión de campo, siempre que no se presenten como riesgo ajustado por exposición, evidencia causal ni prueba de intervención.
Decision status: SUPPORT S2 · CONDITION S3 · RETURN S1 · DEFER S4.
Estado: SUPPORT S2 · CONDITION S3 · RETURN S1 · DEFER S4.
13. Monitoring, escalation, stop, and return triggers
13. Disparadores de monitoreo, escalamiento, parada y retorno
| Trigger type | Trigger | Response | |
|---|---|---|---|
| Monitoring | Track missingness, class balance, exposure coverage, coordinate quality, and out-of-time metrics. | Monitorear faltantes, balance, exposición, coordenadas y métricas fuera del tiempo. | Continue S2 redesign. |
| Escalation | Exposure-adjusted data and stable rare-class performance become available. | Se dispone de exposición y desempeño estable en clases raras. | Re-enter Module III and reopen explanation gate. |
| Stop | A dashboard is interpreted as causal risk or a failed model is used operationally. | Un dashboard se interpreta como riesgo causal o se usa modelo fallido. | Stop deployment and restore boundaries. |
| Return | Chronological holdout remains materially below baseline or Fatal/Serious recall remains unstable. | Holdout cronológico sigue bajo baseline o recall severo sigue inestable. | Return to data, ESS, MSS, and model redesign. |
14. Interpretation boundaries
14. Fronteras interpretativas
AllowedPermitidoRecorded accidents show small but measurable district, temporal, and spatial structure. The current models fail important out-of-time and rare-class criteria.Los accidentes registrados muestran estructura pequeña pero medible. Los modelos fallan criterios fuera del tiempo y de clases raras.
ProhibitedProhibidoClaiming that District, Weather, Road Type, time, or location causes severity; claiming that a current model is deployable; claiming that any variable is an intervention lever.Afirmar causalidad, despliegue o que alguna variable es palanca de intervención.
15. Case closing
15. Cierre del caso
The Module IV handbook passed its own test: it did not reward statistical significance, did not force a predictive success, did not confuse markers with action variables, and did not invent a road-safety intervention. It produced a bounded, transparent decision: rebuild the evidence architecture, permit conditional descriptive monitoring, reject current predictive deployment, and defer intervention selection until causal evidence exists.El manual del Módulo IV pasó su propia prueba: no premió significancia, no forzó éxito predictivo, no confundió marcadores con acción y no inventó una intervención. Produjo una decisión acotada y transparente.