boc/docs/design/DECISION_PIPELINE_v1.0.md

DECISION PIPELINE v1.0

Transformation Chain from Reality to Decision

Version	1.0
Status	DRAFT
Scope	All Landvex data flows, from collection to decision

---

The Pipeline

Reality
    ↓
Observation
    ↓
Evidence
    ↓
Finding
    ↓
Decision
    ↓
Action
    ↓
Business Impact

---

Step 1: Observation

Input	Photo, video, GPS, timestamp, sensor data
Transformation	AI detects objects, classifies, measures
Output	Observation (typed, located, timed)
Owner	Detection Engine

Example:

Input: Mobile photo of road + GPS coordinates
Transformation: AI identifies crack, measures 15cm width
Output: Observation {type: "crack", size: "15cm", location: [lat, lng]}

---

Step 2: Evidence

Input	Observations, history, GIS, weather, traffic
Transformation	Correlation, deduplication, context enrichment
Output	Evidence Bundle (linked observations with context)
Owner	Evidence Engine

Example:

Input: 3 crack observations + road age + traffic volume
Transformation: Correlate by location, check against historical data
Output: Evidence {observations: [obs1, obs2, obs3], trend: "increasing", confidence: 0.91}

---

Step 3: Finding

Input	Evidence Bundle
Transformation	Rules, thresholds, AI reasoning, pattern matching
Output	Finding (pattern, conclusion, severity)
Owner	Analysis Engine

Example:

Input: Evidence Bundle (3 cracks, trend increasing)
Transformation: Compare against degradation models, calculate severity score
Output: Finding {description: "Road degraded 15%", severity: "high", confidence: 0.87}

---

Step 4: Decision

Input	Finding + business rules + priorities + constraints
Transformation	Recommendation generation, priority scoring, action mapping
Output	Decision (recommended action, urgency, rationale)
Owner	Decision Engine

Example:

Input: Finding (road degraded 15%) + maintenance schedule + budget constraints
Transformation: Generate recommendation, calculate urgency, map to action
Output: Decision {action: "inspect", urgency: "14 days", rationale: "Safety risk"}

---

Step 5: Action

Input	Decision + user confirmation + resources
Transformation	Task creation, scheduling, assignment, tracking
Output	Action (scheduled, assigned, tracked)
Owner	Action Engine

Example:

Input: Decision (inspect in 14 days) + user approval + inspector availability
Transformation: Create work order, schedule inspection, assign team
Output: Action {task_id: "WO-2026-001", scheduled: "2026-07-16", assigned: "Team A"}

---

Step 6: Business Impact

Input	Action completion + before/after measurements + costs
Transformation	Impact calculation, ROI analysis, risk reduction quantification
Output	Business Impact (measurable effect)
Owner	Impact Engine

Example:

Input: Inspection completed + new measurements + actual costs
Transformation: Compare before/after, calculate risk reduction, quantify savings
Output: Business Impact {risk_reduced: "12% → 3%", cost_avoided: "$50,000", time_saved: "2 weeks"}

---

Step 7: Learning

Input	Business Impact + original Decision Object + actual outcomes
Transformation	Compare prediction vs reality, adjust models and rules
Output	Improved models, updated thresholds, better recommendations
Owner	Learning Engine

Example:

Input: Business Impact + original Decision {confidence: 0.87, action: "inspect"}
Transformation: Was recommendation followed? Did it produce desired effect? Was confidence correct?
Output: Learning {model_adjustment: "increase crack threshold by 5%", confidence_calibration: "0.87 → 0.92"}

Learning questions:

• Was the recommendation executed?
• Did it produce the desired effect?
• Was the confidence correct?
• Was the recommendation too aggressive or too cautious?
• Do rules or models need adjustment?

---

Control Intelligence

LandveX produces Control Intelligence, not AI analysis.

Control Intelligence consists of:

• Observations
• Evidence
• Findings
• Recommendations
• Business Impact
• Learning

Not: "AI analyzes the video" But: "LandveX produces a recommendation to inspect Road A12 within 14 days"

AI is implementation. Control Intelligence is the product.

---

Architecture Layers

┌─────────────────────────────────────┐
│  Presentation Layer                 │
│  Dashboard, API, Reports            │
├─────────────────────────────────────┤
│  Decision Layer                     │
│  Recommendations, Priorities        │
├─────────────────────────────────────┤
│  Intelligence Layer                 │
│  Findings, Analysis                 │
├─────────────────────────────────────┤
│  Knowledge Layer                    │
│  Observations, Evidence, History    │
├─────────────────────────────────────┤
│  Reality Layer                      │
│  Collection, Sensors, Mobile        │
└─────────────────────────────────────┘

Layer	Components	Responsibility
Reality	quiXzoom app, drones, sensors, cameras	Collect raw data
Knowledge	Detection Engine, Evidence Engine	Structure and enrich
Intelligence	Analysis Engine	Find patterns
Decision	Decision Engine	Generate recommendations
Presentation	Dashboard, API, Reports	Show decisions

Plus Learning Loop: Business Impact feeds back to Intelligence Layer to improve future recommendations.

---

Communication Principle

Not:

Video → AI → Score

But:

Observation → Analysis → Recommendation

AI is implementation. Recommendation is the product.

---

Target Customer Cases

Three customer types to validate the pipeline:

Customer Type	Decision	Why Important
Municipality	"Inspect or wait?"	Maintenance and prioritization
Property Owner	"Repair now or plan later?"	Cost vs risk
Contractor/Operations	"Which action first?"	Operational planning

Goal: Same Decision Pipeline works for three different customer types, not just three technical scenarios.

Validation Requirements

Before freezing Decision Pipeline v1.0:

1. Three real customer cases — municipalities, property owners, or contractors follow the full chain
2. Run each case through entire Decision Pipeline — from observation to business impact to learning
3. Document where pipeline breaks — not where you think it might break
4. Revise only after — model changes when data contradicts model, not before
5. Empirical validation — same Decision Object works for real data
6. End-to-end test — from observation to business impact

---

ändringshistoria

Version	Datum	Beskrivning
1.0	2026-07-02	Initial decision pipeline with 6 steps, 5 layers, validation requirements

---

STATUS

DRAFT — Awaiting empirical validation with real customer cases