KARL V6 — Cognitive Twin Architecture

# KARL V6 — Cognitive Twin Architecture ## Autonomous Session Driver with Persistent Context **Version**: 6.0.0-design **Date**: 2026-04-01 **Status**: Architecture (pre-implementation) **Supersedes**: V5 `twin_session_driver.py` V5 failed after turn 15 for one reason: the 4B model is a **reaction machine**, not a **planning machine**. It sees 80 lines of terminal output, generates a plausible next prompt, and forgets everything. It has no model of where it is in a project, no memory of what it already tried, and no self-monitor to notice it's looping. It's a parrot reading a scroll — fluent in the present, amnesiac about the past. V6 does not fine-tune to fix this. Instead, it **externalizes everything the model cannot hold internally** — project state, turn history, repetition detection, task graph — and **injects that context into every single prompt** as structured scaffolding. The model's job shrinks from "figure out the whole session" to "given a complete briefing, pick the next one sentence." That is a problem a 4B model can solve. # PHASE 1: PRIME ## Core Insight — Why Small Models Fail at Multi-Turn Session Driving