- Reordered stories to reach First Verified Decision faster:
1. Mission Import (was 2) — proves we can receive real data
2. Dataset Explorer (was 3) — makes data visible
3. Annotation Workspace (was 4) — first human-in-the-loop
4. Decision Case (was 5) — first verified decision
5. Replay (was 6) — proves chain is reproducible
6. Session Management (was 1) — organizes when core works
- Added Golden Mission concept:
- Real mission that never changes, used as regression test
- Every new model runs against same mission
- See immediately if something got better or worse
- Added Review as first-class object:
- Observation → AI → Human Review → Approved/Rejected/Needs More Evidence
- Makes entire quality flow traceable
- Added Dashboard v1:
- Sessions, Missions, Decision Cases, Pending Reviews, Verified Decisions
- Big button: [Continue Reviewing]
- Work tool, not BI system
- Added vertical user journey (demo script):
- quiXzoom → photo → import → save → explore → AI observation →
correction → Decision Case → full chain viewer
- If this works, core is proven
- Added definition of 'First Verified Decision':
- Built on real observation data
- Reviewed by human
- Complete evidence chain
- Fully reproducible from raw data to recommendation
Rationale: Reach core proof faster, add organization later.
Golden Mission enables regression testing from day one.
Review object makes quality flow traceable.
- STOP Rule: No new pipeline until 20 real Decision Cases exist
- Field Readiness Gate: 5 questions before building any feature
- Sprint Goal: Every sprint must produce more verified Decision Cases
- Six stories:
1. Session Management — organize missions by location/date
2. Mission Import — upload video/images/GPS/EXIF, store immutably
3. Dataset Explorer — browse, filter, search, map view
4. Annotation Workspace — review/correct AI, version history
5. Decision Case — full chain: Observation→Evidence→Finding→Decision→Business Impact
6. Replay — step through mission chain (V1: simple playback)
- Not in first release (intentionally postponed):
GPU Queue, Hyperparameter Search, Distributed Training,
Benchmark, Canary Deployment, Auto Retraining,
Bias Dashboard, Drift Detection
- Definition of Done: Complete vertical slice from reality to verified decision
- Definition of Ready for Epic-002: 20 real Decision Cases + Field Readiness Gate
- Technical stack: React+TypeScript, Node.js+Express, PostgreSQL, S3/R2, Bull, Python AI service
- Quality gates: TypeScript strict ≥80% coverage, no secrets, OAuth 2.0, immutable audit log, GitOps
Rationale: Prove the system works end-to-end before scaling.
Focus on learning from reality, not building everything upfront.
- MVP Milestone: 'First Verified Decision'
- Developer films with quiXzoom, imports to Lab, corrects AI,
creates Decision Case, follows chain with full traceability
- When this works = first complete verifiable Control Intelligence pipeline
- Three development phases:
Phase 1 (Essential): Ingestion, Dataset Explorer, Annotation, Decision Case Viewer
Phase 2 (Scale): Replay, Benchmark, Evaluation
Phase 3 (Advanced): GPU Jobs, Hyperparameter Runs, Model Promotion, Canary
- Product Architecture: quiXzoom → Observations → Intelligence Lab →
Improved Models → LandveX → Better Decisions → Feedback → Intelligence Lab
- Two products: quiXzoom (observations), LandveX (decisions)
- Intelligence Lab = the factory that improves both
- New areas:
- Data Quality: Healthy/Blurred/Duplicate/Wrong GPS/Night/Rain/Occluded
+ Coverage (Roads, Buildings, Signs, Drainage, Vegetation)
- Decision Analytics: Acceptance Rate, Ignore Rate, Accuracy,
Insufficient Evidence, Data Collection Value
- Business value metrics, not traditional AI metrics
Rationale: Build MVP first, prove first real workflow, then scale.
Decision Cases are the heart. Data Quality explains model performance.
Decision Analytics measure business value.
- Internal development environment for Control Intelligence
- Core principle: 'Produces verified Control Intelligence, not AI models'
- Separate repo: landvex-intelligence-lab
- Navigation: Dashboard, Models, Datasets, Annotations, Training,
Evaluation, Decision Cases, Replay, Validation, Deploy, Settings
- Key features:
- Dashboard: AI status (models, datasets, jobs, cases)
- Mission Replay: click through entire chain
- Annotation: video + AI suggestion + manual correction
- Decision Cases: first-class objects, all playable
- Benchmark: compare YOLO, Grounding DINO, SAM, custom models
- Replay: find regressions between model versions
- Validation: field trials, scenario tests, decision tests
- Deploy: 'Promote Model' not 'Deploy' (dev → validation → pilot → prod)
- Experiments: link EP-1.0, DS-001, etc. to real data
- Target: New AI engineer understands in minutes:
'This is where we build, test, and verify LandveX Control Intelligence
before anything reaches production.'
Rationale: Single internal tool for all AI development. Centralizes
model training, annotation, validation, replay, decision chains,
regression tests, experiments, and model promotion.
- Added four pilot phases (product research, not marketing):
1. Collection — what can actually be detected?
2. Analysis — are decisions understandable?
3. Verification — was recommendation correct?
4. Reflection — what needs to change?
- Measurement principle:
- Not: Did AI find a crack?
- But: Did this become a decision a real person could act on?
- Document 'non-decisions' — equally valuable as clear decisions
- Film workflow, not just infrastructure:
- How you find area, choose mission, document
- What feels unclear, when you become uncertain
- When system saves time
- Observer mindset: document first, change model later
- Build model from real workflows, not assumptions
Rationale: First 20-50 real missions teach more than months of modeling.
- Created FIELD_TRIAL_LOG.md — observation protocol for real customer cases
- Log entry template with 10 fields
- Two example entries (accepted and rejected decisions)
- Questions the log answers: adoption, accuracy, calibration, rejection analysis
- Future KPIs: Decision Adoption Rate, Decision Accuracy, Time to Decision
- Updated MEMORY.md with Control Intelligence & Decision Model section
- Decision Pipeline v1.0 summary
- Decision Object v1.0 (STRUCTURE FROZEN)
- Status: VALIDATED FOR FIELD TRIALS
- Three target customer cases
- Future KPIs
- Key principle: No more modeling until first real customer case
- Added Decision Adoption Rate and Decision Accuracy as future product KPIs
- Not implemented yet — start collecting data now, calculate later
- Decision Adoption = accepted / total recommendations
- Decision Accuracy = correct / total decisions
Rationale: Stop modeling, start observing. First real customer case
will teach more than the last twenty documents combined.
- Status: STRUCTURE FROZEN (not LOCKED/FINAL)
- Frozen: field names, semantics, relationships
- Not frozen: implementation, algorithms, confidence calculation
- Three tests executed and PASSED:
1. Decision Invariance Test — all 8 scenarios use same 7-field structure
2. Evidence Variation Test — Decision Object identical regardless of evidence type
3. Explainability Invariance Test — chain Reality→Observation→Evidence→Finding→Decision
works for all actionable decisions
- Results summary:
- Manual Review: 6 PASS, 2 OBSERVATION, 0 FAIL
- Decision Invariance: PASS
- Evidence Variation: PASS
- Explainability Invariance: PASS
- Decision Quality Gate: PASS
- Verb Rule: PASS
- Overall: VALIDATED FOR FIELD TRIALS
- Internal validation complete
- Ready for real customer testing
- Not production truth yet
- Open questions documented (not added as fields):
- Observation A: Cost estimate for investment decisions
- Observation B: Explicit low confidence communication
- Next: Three real customer cases (municipality, property owner, contractor)
Rationale: Freeze structure before testing, run tests against locked model,
mark as 'Field Trial Ready' not 'Final'. Model changes when real data
contradicts it, not before.
- Added Step 7: Learning (feedback loop from Business Impact to Intelligence)
- Control Intelligence definition: LandveX produces Control Intelligence, not AI
- Three target customer cases defined:
1. Municipality — Inspect or wait? (Maintenance prioritization)
2. Property Owner — Repair now or plan later? (Cost vs risk)
3. Contractor/Operations — Which action first? (Operational planning)
- Validation requirements: Run each case through full pipeline, document breaks,
revise only after data contradicts model
- Communication principle: Observation → Analysis → Recommendation
(AI is implementation, recommendation is product)
Rationale: Stop modeling, start observing. Model changes when data contradicts
it, not before. Three real customer cases before freezing.
- Six layers (added Evidence between Observation and Finding):
1. Reality
2. Observation
3. Evidence (linked observations with context)
4. Finding
5. Decision
6. Business Impact
- Decision Object restructured:
1. Decision — what should user decide?
2. Why — why system recommends this
3. Evidence — what observations support this
4. Confidence — how certain (3 dimensions)
5. Consequence — what if nothing done
6. Action — next step
7. Business Impact — economic/operational meaning
- Confidence Model (3 dimensions):
- Observation Confidence: how certain is detection?
- Evidence Strength: how strongly supported?
- Recommendation Confidence: how certain is recommendation?
- Explainability Principle:
- Every Decision Card must be explorable
- User can click: Decision → Finding → Evidence → Observations → Reality
- Competitive advantage: traceability to source material
- Business Impact Model (4 dimensions):
- Risk, Cost, Time, Opportunity
- Three validation scenarios:
1. Road Crack — simple, common
2. Damaged Facade — complex, critical
3. Broken Road Sign — simple, regulatory
- Pass criteria: Same Decision Object works for all three
Rationale: Decision Intelligence, not BI. Evidence-backed decisions
are the core product. Explainability is competitive advantage.
Validation against real scenarios before freezing.
- DASHBOARD_AUDIT_PROTOCOL.md (v1.0, LOCKED):
- Five criteria: 3-Second Rule, Journey Principle, Reality→Decision,
Dashboard Principle, Decision Density
- Blink Test: 3-second exposure, consistent answers = pass
- Information-to-Decision Ratio: measure objects needed per decision
- Audit form with pass/fail for each criterion
- Decision matrix: all pass = proceed, some fail = adjust Dashboard,
most fail = review Foundation
Rationale: Dashboard is the stress test for Foundation v1.0. If Dashboard
passes without Foundation changes, the base is robust enough to scale.
If Dashboard fails, adjust Dashboard first — not Foundation.
- DASHBOARD_PRINCIPLE.md (v1.0, LOCKED):
- Four questions: What, Where, How serious, What now
- Information hierarchy: Score -> Observations -> Map -> History -> Details
- Decision Density: Dashboard = 1-3 decisions max
- Not BI: helps user make decision, not consume statistics
- Verification checklist
- EXPERIENCE_AUDIT_QA.md updated:
- Screen 4: Landvex Dashboard (Phase: Manage)
- Dashboard-specific questions: 3-second understanding, today's decision,
80% ignorable, primary CTA
- Dashboard Principle check: What/Where/How serious/What now
- Decision Density target: 1-3 decisions
- One-Sentence Test: understand area performance and what to act on
- Ready for Landvex Dashboard screen (stresstest for Foundation v1.0)
- JOURNEY_PRINCIPLE.md (v1.0, LOCKED):
- Five phases: Discover, Capture, Process, Outcome, Manage
- Maps to both quiXzoom and Landvex
- Screen mapping table
- Balance check reveals Capture/Process/Outcome are empty
- Verification: every screen must belong to a phase
- EXPERIENCE_AUDIT_QA.md updated:
- Screen 3 renamed: Public Map -> Mission Marketplace
- Added Phase column: Discover
- One-Sentence Test updated: which missions not how much
- Added Journey Principle reference
- Screen 4: Dashboard (Landvex) - awaiting screen
- Screen 5: Results/Decision View (Landvex) - awaiting screen
- Next: Landvex Dashboard audit (stresstest for Foundation v1.0)
Rationale: Journey Principle prevents screens from floating without
context. Mission Marketplace better reflects the page's purpose.
Dashboard is the most complex screen - best stresstest for principles.
- LANDVEX_DESIGN_CONSTITUTION.md:
- Added §0a: Capability Parity Principle
- All user functions must work on smartphone
- Desktop may enhance, never require
- Two-question rule for every new feature
- Desktop as enhancer (not exclusive)
- Examples: Payments, History, KYC
- Verification checklist for every new feature
- RFC_DEFINITION_OF_DONE.md:
- Added criterion #3: Capability Parity
- Can be used on smartphone; desktop is enhancement
- Total criteria: 14 (was 13)
- Exit criteria: must pass all 14
Rationale: quiXzoom targets field workers. Phone is primary or only
computer. Phone must always be a full client. Desktop is a more
comfortable workspace for some tasks, not a prerequisite. This is a
constitutional principle, not implementation.
- Updated audit framework with fixed 6-question matrix:
1. 3-Second Rule: understand purpose immediately
2. Three-Step Rule: Input → Processing → Outcome visible
3. Primary Decision: know what to do
4. Information Hierarchy: most important is largest
5. Cognitive Load: what can be removed
6. Call to Action: clear next action
- Added One-Sentence Test:
- Can user describe screen in one sentence?
- Examples for Onboarding, Map View, Dashboard, Results
- Two+ sentences = screen tries to do too much
- Added placeholders for next screens:
- Screen 3: Dashboard (Landvex)
- Screen 4: Map View (Landvex/quiXzoom)
- Screen 5: Results/Decision View (Landvex)
- Priority order: Dashboard → Map View → Results → Search → Settings
- Updated quiXzoom onboarding audit with new matrix format
Rationale: Fixed matrix ensures consistent evaluation across all screens.
One-Sentence Test reveals overloaded screens. Placeholders ready for
next screens. Priority reflects user journey impact.
- EXPERIENCE_AUDIT_QA.md (v1.0, DRAFT):
- Audit framework: 6 questions per screen
- Screen 1: quiXzoom Onboarding — 3s ✅, R→D ⚠️, Decision ✅
- Issue: 'kr' used instead of USD/EUR (Valutaregel violation)
- Issue: Launch date creates hesitation
- Screen 2: Landvex Process — 3s ✅, R→D ✅, Decision ✅
- Issue: Technical terms in step descriptions
- Action items: currency update, date removal, verb simplification
- Next screens: Dashboard, Map View, Search, Mission Details, Results
- COMPONENT_MATURITY.md (v1.0, LOCKED):
- Three levels: Foundation → Validated → Canonical
- Foundation: RFC approved, passes 13 DoD criteria
- Validated: used in 3 real screens, passes Experience Audit
- Canonical: used in 5 screens, 2 sprints stable, other components inherit
- Promotion process with verification requirements
- Current status: all 4 components at Foundation level
Rationale: Observe first, standardize second. Components become Canonical
only after proving value in real screens. Experience Audit identifies gaps
before building components. Maturity model prevents premature standardization.
- FOUNDATIONS-v1.0.md updated:
- Added Landvex Core Principle and Design Constitution to frozen foundations
- Added Core Principle Rules to invariants (2.4)
- Reordered component priority: Modal, Search, Map Panel, Navigation, Table, Data Grid
- Reflects business value: search and maps are central to Landvex
- EXPERIENCE_AUDIT.md (v1.0, LOCKED):
- Four questions for every screen:
1. Can purpose be understood in 3 seconds?
2. Is 'Reality → Decision' visible?
3. Is it obvious what the user should do?
4. Does everything contribute to the decision?
- Three-second rule levels with verification method
- Audit process: select screens → apply questions → prioritize fixes → verify
- Example audit format
- Core Principle Check for RFCs: 4 questions before approval
- COMPONENT_TEMPLATE.md updated:
- Added Core Principle section to template
- Requires: decision support, cognitive load reduction, 3-second understanding
- RFC_DEFINITION_OF_DONE.md updated:
- Added Core Principle as criterion #2 (was 12 criteria, now 13)
- Exit criteria: must pass all 13
Rationale: Freeze foundation with Core Principle as highest authority.
Experience audit ensures all screens support decision-making.
Core Principle Check prevents components that don't serve the user.
Business-value prioritization: search and maps before tables.
- LANDVEX_CORE_PRINCIPLE.md (v1.0, LOCKED):
- Single promise: From reality → decision
- Five steps: Question → Collect → Analyze → Calculate → Decision
- Business verbs, not technical terms
- Three-second rule: 3s/10s/60s understanding levels
- Gauge: Video → AI → Score
- Applies to all communication, UX, documentation, AI
- LANDVEX_DESIGN_CONSTITUTION.md:
- Added §0: From Reality → Decision as highest principle
- Five steps in Swedish
- Three-second rule
- Gauge explanation
- Positioned before Introduction as foundational principle
Rationale: Landvex sells one thing — from reality to decision.
Every screen, illustration, and presentation must follow this flow.
Business verbs make it human. Three-second rule makes it scannable.
Gauge makes it instantly understandable.
Four-point validation:
1. COMPLETENESS ✅ PASS
- All 6 Button states covered
- All 7 Input states covered
- All 10 Select states covered
- All 7 Card states covered
2. ORTHOGONALITY 🟡 PASS WITH NOTES
- 3 minor overlaps identified (surface.raised vs overlay, text.secondary vs muted, action.disabled vs text.muted)
- All resolved by distinct use cases
- 1 ambiguity noted (brand vs action), documented in contracts
3. MINIMALITY 🟡 PASS WITH NOTES
- 9 unused tokens identified
- All justified: needed for upcoming components (Modal, Link, Alert) or dark mode
- No tokens should be removed
4. STABILITY ✅ PASS
- Modal: 0 new tokens needed
- Navigation: 0 new tokens needed
- Table: 0 new tokens needed
- Map Panel: 0 new tokens needed
VERDICT: APPROVE with minor notes. Semantic Color System is stable enough to freeze.
Action items:
1. Freeze Semantic Color System v1.0
2. Update Input, Select, Card to semantic aliases (raise to ~80%)
3. Create Foundation Freeze milestone
4. Proceed with Modal, Navigation, Table, Map Panel
- RFC-002 Input:
- Token Dependencies updated to semantic aliases with primitives
- All tokens now map through color.surface, color.border, color.text, color.status
- No direct color references (e.g., semantic-red-500 → color.status.danger)
- Helper/error text uses semantic tokens
- RFC-003 Select:
- Token Dependencies updated to semantic aliases with primitives
- Dropdown, option, and error states use semantic tokens
- Selected option uses color.action.primary at 10% opacity
- No direct color references
- RFC-004 Card:
- Token Dependencies updated to semantic aliases with primitives
- Surface, border, text hierarchy use semantic tokens
- Shadow tokens mapped to elevation system
- No direct color references
- COMPLIANCE_MATRIX.md:
- Input: 33% → 83%
- Select: 33% → 83%
- Card: 33% → 83%
- All now have Governance ✅, Tokens ✅, Motion ✅, Accessibility ✅, AI ✅
Rationale: Semantic model first. Components never reference primitives.
Token contracts (purpose, allowed, forbidden) ensure correct usage.
Dark mode implementable via token swap without component changes.
- AI_DESIGN_PRINCIPLES.md:
- Restructured into 5 parts: A. Decision Process, B. Component Selection,
C. Token Usage, D. Design Review, E. Escalation
- All rules use RFC 2119 normative words (MUST/MUST NOT/SHOULD/MAY)
- Clearer escalation path: when to create DDR, when to flag uncertainty
- COMPLIANCE_MATRIX.md:
- Tracks readiness for all 29 components (Tier 1/2/3)
- 6 dimensions: Governance, Tokens, Motion, Accessibility, AI, QA
- Foundation documents tracked separately
- Button at 82% (reference for others)
- docs/platform/README.md:
- Reserved directory for cross-cutting governance
- Future: AI Governance, Product Governance, Review Process,
Quality Gates, Release Governance, Decision Frameworks
- Separates platform governance from design system
- LANDVEX_DESIGN_SPECIFICATION.md:
- Button renamed to RFC-001: Canonical Button Component
- Structured with: Cognitive Purpose, When to Use/NOT Use,
User Expectation, Interaction Contract, Failure Behaviour,
Accessibility Contract, AI Contract
- Sets standard for all future RFC component specs
Rationale: AI rules before components (AI writes components).
Compliance matrix makes progress visible. Platform governance
separates cross-cutting concerns from design system. RFC format
for components ensures production-ready specifications.
- Motion System v1:
- Renamed duration tokens from time-based (fast/slow) to intention-based
(instant, interaction, transition, navigation, layout, context-change,
attention, background)
- Updated all component motion tables to use intention-based tokens
- Rationale: AI reasons about purpose, not milliseconds
- Button gold standard (chapter 8.3.1):
- Added RFC-level sections: Cognitive Purpose, When to use/NOT use,
User Expectation, Interaction Contract, Failure Behaviour,
Accessibility Contract, AI Contract
- Sets the bar for all future component specifications
- AI_DESIGN_PRINCIPLES.md (v1.0, LOCKED):
- Core principle: AI must be a design system citizen
- Before-action rules: read hierarchy, check existing, apply New Component Rule
- Token rules: no hardcoded values, semantic first, lifecycle, deterministic
- Component rules: inheritance first, use template, document anti-patterns
- Motion rules: intention-based duration, reduced motion, no decorative animation
- Accessibility rules: WCAG 2.1 AA, no ARIA without purpose
- AI Contract for Button: when to choose Button vs Link vs Icon Button
- RELEASE_DEFINITION.md:
- Defines Landvex Design System v1.0 release gates
- Foundation, Tokens, Foundations, Components (Tier 1/2/3),
Validation, Production
- Current progress: ~55%
- Next milestones: Grid+Elevation, Input/Select/Card, Brand Palette
Rationale: AI rules before components (AI will write components).
Intention-based motion tokens for better AI reasoning. Button as RFC
sets the standard. Release definition creates clear target.
- TOKEN_PHILOSOPHY.md:
- Added Token Lifecycle (Draft → Experimental → Stable → Deprecated → Removed)
- Added Design Review Gate (6 mandatory checks before new token)
- Added deterministic principle: two AI agents must arrive at same token choice
- Restructured into 8 sections with clear hierarchy
- LANDVEX_DESIGN_SPECIFICATION.md:
- Restructured into 4 parts: GOVERNANCE → TOKENS → COMPONENTS → VALIDATION
- Added part headers and updated table of contents
- Aligned with Token Philosophy lifecycle and review gate
Rationale: Governance before values. Prevents uncontrolled token growth
and ensures AI-assisted design development is deterministic and traceable.
Bernt