sync: update docs/ROADMAP.md

docs/ROADMAP.md

@@ -1,7 +1,7 @@
 # 🚀 NeuralAI Development Roadmap
 
-**Version Target: 6.0 (The Workstation Pivot)**
-**Last Updated: June 6, 2026**
+**Version Target: 6.0 (The Workstation Pivot) + 7.2 Resilient Runtime**
+**Last Updated: July 15, 2026**
 
 ---
 
@@ -67,21 +67,30 @@
 
 ## 📊 System Status
 
-- **Main Service:** **READY** (Unified `file neural_core_service.py`)
+- **Main Service:** **READY & RESILIENT** (`webui_service.py` v7.2.0 — auto-restart, memory watchdog, ZO-native inference backend)
 - **Voice Service:** **READY** (ElevenLabs v2 Migrated)
-- **Model:** SmolLM2-360M-Instruct + DPO v14.0 (In Progress)
+- **Model:** SmolLM2-360M-Instruct + DPO v15.0 (Deployed)
 - **Context:** System-wide (Expanding)
 
-### Latest DPO Run (v13.0)
+### Latest DPO Run (v15.0)
 
-- **Training samples:** 363
-- **Validation samples:** 41
+- **Training samples:** 597 (expanded from 302)
 - **Epochs:** 3
-- **Best checkpoint:** `checkpoint-69`
-- **Best validation loss:** `0.7232`
-- **Eval progression:** `1.6713` → `0.8969` → `0.7232`
-- **Alignment improvement:** `56.73%` reduction from first eval to best checkpoint
-- **Completed:** `2026-05-17 08:41 UTC`
+- **Steps:** 450
+- **Final training loss:** `0.305`
+- **Reward margin:** `~0.5` → `~3.5`
+- **Hardware:** Apple Silicon MPS (MacBook Air M4)
+- **Duration:** `730.5s` (~12 min)
+- **Completed:** `2026-07-11 20:00 UTC`
+- **Adapter:** live on HF `Subject-Emu-5259/NeuralAI`
+
+### Deployment & Inference (updated July 15, 2026)
+
+- **Host:** ZO Computer (Free plan, 4 GB RAM) at `neuralai-deandrewharris.zocomputer.io`
+- **Inference backend:** `LLM_BACKEND=zo` → ZO native `/zo/ask` using the user's own BYOK model (`byok:0d3567f7-f521-42b0-8adf-65c9b036cf89`). Uses **0 MB local RAM**.
+- **Why not local:** Loading PyTorch + SmolLM2-360M on a 4 GB box used ~6.2 GB → OOM-kill loop that paused the service. The ZO backend removes that dependency entirely (see `docs/INCIDENT-2026-07-14-NEURALAI-PAUSES.md`).
+- **Local PyTorch backend:** still available via `LLM_BACKEND=local` on GPU/Colab-class machines (≥8 GB RAM) for offline/private inference.
+- **Resilience:** supervisor auto-restart, `/api/health` keepalive, and a memory watchdog that runs GC before any OOM. Service verified stable after the July 15 fix (free RAM returned to ~3 GB).
 
 ---