Context-compensation scaffold for LLM evaluation prompts — disclose, gate on evidence, hedge on thin

Bootstrap a fresh Claude Code session with the conventions, grounding triggers, and tool-map already loaded — per-project, marker-anchored, idempotent. Stop re-deriving the same rules at minute 30.

LPCI PROVED: Stateless LLM maintains conversational coherence across 20 turns via refreshing language scaffold. No conversation history. The scaffold IS the memory. Hermes Labs 2026.