KGM-Based Skill Dispatch

Knowledge bundle capturing the design decisions, evidence, and implementation of using KGM (Knowledge Graph Memory) as a skill dispatch mechanism for weaker LLM models.

Concepts

Concept	File	Summary
Problem & Evidence	problem-evidence.md	Granite 8B fails to scan # Skills prose; behavioral score progression 0/3 → 1/3 → KGM solution
Architecture	architecture.md	Separate JSONL indexes, AgentSkill entity format, symlink workaround, reindex pipeline
AGENTS.md Evolution	agents-md-evolution.md	Template token reduction 2100 → 1362, section reordering, Discovery Tiers elimination, Rule 8 rewrites
Testing Infrastructure	testing-infrastructure.md	template-check.sh, template-behavioral.sh, template-eval.sh, merge-score-entry.sh, score tracking
Design Principles	design-principles.md	Lower cognitive bar, prose → scripts, zero-arg wrappers, ledger integrity