What problem does it solve? It prevents impulsive or unverified adoption of external open-source resources (skills, hooks, rules, MCP servers, methodologies) into a Claude Code project by enforcing a structured evaluation protocol with baselines, benchmarks, and recorded decisions. ## Core Features & Use Cases - 5-Phase Protocol: Phase 0 Baseline (sovereignty questions + latent-capability audit), Phase 1-2 Differential Benchmarking (A/B/C/D quadrant + 8-item context match), Phase 3 Essence Extraction (MVE distillation with adaptation-depth scoring), Phase 4 Grafting (timeboxed git worktree pilot with kill switch), Phase 5 Validation (7-day observation + ADR). - Mechanized Enforcement: 8 hooks block forbidden patterns such as skipping baselines, concurrent pilots, permanent pilots, cooldown bypass, and scope creep, writing structured records to baseline_snapshots and external_evaluations SQLite tables. - Use Case: When a user says "I saw this GitHub tool, should we integrate it?", the skill runs a latent-capability audit, classifies the trigger, quantifies a baseline, and either rejects the impulse or guides a timeboxed pilot with a binary integrate-or-rollback decision. ## Quick Start Ask the AI to evaluate whether a specific external tool or workflow should be integrated into the project using the tianji-pavilion protocol.
