docs: create roadmap (4 phases)

Phases:
1. GNN Verifier Foundation: GNN-01, GNN-02
2. Hybrid Verifier Integration: GNN-03, ENS-01, ENS-02
3. Graph-Aware Proposer: SELF-01, SELF-02
4. Evaluation Framework: EVAL-01, EVAL-02, EVAL-03

All 10 v1 requirements mapped to phases.

Author: vmehtacode

.planning/REQUIREMENTS.md

@@ -71,22 +71,22 @@ Which phases cover which requirements. Updated during roadmap creation.
 
 | Requirement | Phase | Status |
 |-------------|-------|--------|
-| GNN-01 | TBD | Pending |
-| GNN-02 | TBD | Pending |
-| GNN-03 | TBD | Pending |
-| SELF-01 | TBD | Pending |
-| SELF-02 | TBD | Pending |
-| ENS-01 | TBD | Pending |
-| ENS-02 | TBD | Pending |
-| EVAL-01 | TBD | Pending |
-| EVAL-02 | TBD | Pending |
-| EVAL-03 | TBD | Pending |
+| GNN-01 | Phase 1 | Pending |
+| GNN-02 | Phase 1 | Pending |
+| GNN-03 | Phase 2 | Pending |
+| SELF-01 | Phase 3 | Pending |
+| SELF-02 | Phase 3 | Pending |
+| ENS-01 | Phase 2 | Pending |
+| ENS-02 | Phase 2 | Pending |
+| EVAL-01 | Phase 4 | Pending |
+| EVAL-02 | Phase 4 | Pending |
+| EVAL-03 | Phase 4 | Pending |
 
 **Coverage:**
 - v1 requirements: 10 total
-- Mapped to phases: 0
-- Unmapped: 10
+- Mapped to phases: 10
+- Unmapped: 0
 
 ---
 *Requirements defined: 2026-01-27*
-*Last updated: 2026-01-27 after initial definition*
+*Last updated: 2026-01-27 after roadmap creation*

.planning/ROADMAP.md

@@ -0,0 +1,87 @@
+# Roadmap: Grid Guardian
+
+**Created:** 2026-01-27
+**Core Value:** Detect anomalies in energy distribution networks without labeled data, using physics constraints and self-play learned patterns on graph-structured grid data
+**Config:** depth=standard, parallelization=false
+
+## Milestone: v1.0 — GNN-Based Anomaly Detection with Self-Play Validation
+
+### Phase 1: GNN Verifier Foundation
+
+**Goal:** Build a topology-aware GNN verifier that understands grid structure and can score anomalies based on spatial relationships between nodes.
+
+**Requirements:** GNN-01, GNN-02
+
+**Success Criteria:**
+1. Graph construction pipeline transforms SSEN metadata into PyTorch Geometric Data batches with nodes (households, feeders, substations) and edges (physical topology)
+2. GNN model (GAT/GraphSAGE + temporal layer) processes graph-structured input and outputs per-node anomaly scores
+3. Model achieves >85% accuracy on held-out synthetic anomalies with <30ms inference latency for batch_size=32
+4. Oversmoothing is prevented (node embeddings remain distinguishable across grid hierarchy levels)
+
+**Depends on:** None
+
+---
+
+### Phase 2: Hybrid Verifier Integration
+
+**Goal:** Replace the MLP Verifier with a three-layer ensemble that combines physics constraints, GNN patterns, and cascade logic into the self-play training loop.
+
+**Requirements:** GNN-03, ENS-01, ENS-02
+
+**Success Criteria:**
+1. Physics constraint layer (hard bounds from SSEN: voltage, capacity, ramp rates) validates inputs before GNN processing
+2. Cascade early-exit logic skips GNN computation for clear physics violations, reducing average inference time
+3. HybridVerifierAgent integrates into existing SelfPlayTrainer, producing verification rewards for training
+4. Hybrid verifier improves detection accuracy over baseline rule-based verifier on held-out test scenarios
+
+**Depends on:** Phase 1
+
+---
+
+### Phase 3: Graph-Aware Proposer
+
+**Goal:** Enhance the Proposer to generate topology-respecting anomaly scenarios where disturbances propagate through connected nodes (e.g., COLD_SNAP cascading through neighbors).
+
+**Requirements:** SELF-01, SELF-02
+
+**Success Criteria:**
+1. GraphAwareProposer generates scenarios that respect grid topology (anomalies occur on connected subgraphs, not random nodes)
+2. Cascade scenarios (COLD_SNAP, OUTAGE) propagate through physically connected neighbors with configurable decay
+3. Self-play training maintains scenario diversity (each of 5 types appears in at least 15% of batches, no mode collapse)
+4. Generated scenarios are visually distinguishable and pass domain plausibility review
+
+**Depends on:** Phase 2
+
+---
+
+### Phase 4: Evaluation Framework
+
+**Goal:** Validate the GNN-based approach through rigorous multi-level evaluation comparing against baselines and measuring physics compliance.
+
+**Requirements:** EVAL-01, EVAL-02, EVAL-03
+
+**Success Criteria:**
+1. Precision, recall, and F1 scores computed against physics-violation ground truth and synthetic scenarios
+2. GNN Verifier performance compared against IsolationForest, Autoencoder, and DecompositionAnomalyDetector on identical test splits
+3. Physics compliance rate measured (percentage of detected anomalies that violate known physical constraints)
+4. Ablation study quantifies contribution of each component (GNN-only vs. physics-only vs. hybrid ensemble)
+5. Results documented with statistical significance testing and visualization for FYP report
+
+**Depends on:** Phase 3
+
+---
+
+## Progress
+
+| Phase | Name | Status | Requirements | Coverage |
+|-------|------|--------|--------------|----------|
+| 1 | GNN Verifier Foundation | Not started | GNN-01, GNN-02 | 2/10 |
+| 2 | Hybrid Verifier Integration | Not started | GNN-03, ENS-01, ENS-02 | 3/10 |
+| 3 | Graph-Aware Proposer | Not started | SELF-01, SELF-02 | 2/10 |
+| 4 | Evaluation Framework | Not started | EVAL-01, EVAL-02, EVAL-03 | 3/10 |
+
+**Total Coverage:** 10/10 requirements mapped
+
+---
+*Roadmap created: 2026-01-27*
+*Last updated: 2026-01-27*

.planning/STATE.md

@@ -0,0 +1,58 @@
+# Grid Guardian — Project State
+
+## Project Reference
+
+See: .planning/PROJECT.md (updated 2026-01-26)
+
+**Core value:** Detect anomalies in energy distribution networks without labeled data, using physics constraints and self-play learned patterns on graph-structured grid data
+**Current focus:** Phase 1 - GNN Verifier Foundation
+
+## Current Position
+
+- **Phase:** 1 of 4
+- **Plan:** 0 of 0 (not yet planned)
+- **Status:** Not started
+- **Progress:** [..........] 0%
+
+## Performance Metrics
+
+| Metric | Value | Notes |
+|--------|-------|-------|
+| Plans completed | 0 | - |
+| Requirements done | 0/10 | - |
+| Phases done | 0/4 | - |
+
+## Key Decisions
+
+(None yet)
+
+## Blockers
+
+(None)
+
+## Accumulated Context
+
+### Technical Decisions
+(None yet)
+
+### Open Questions
+- SSEN metadata schema for graph construction (needs exploration in Phase 1)
+- Optimal GNN hyperparameters for 100K node graphs (will tune in Phase 1)
+
+### Deferred Items
+- Uncertainty quantification (UQ-01, UQ-02) - v2
+- Elexon BMRS live integration (DATA-01, DATA-02) - v2
+- FYP report and presentation (ACAD-01, ACAD-02) - v2
+
+## Recent Activity
+
+- 2026-01-27: Project initialized, roadmap created with 4 phases
+- 2026-01-27: Research synthesis completed (MEDIUM-HIGH confidence)
+
+## Session Continuity
+
+**Last session:** 2026-01-27
+**Next action:** Plan Phase 1 via `/gsd:plan-phase 1`
+
+---
+*Last updated: 2026-01-27*