Repository navigation
SKILL.md Split: 1490 → 250 Lines
SKILL.md is now user-facing only. All architecture, templates, and deep-dive content moved to docs/ with progressive loading references. Follows Anthropic's < 500 line recommendation.
Divergence Engine
- Provocateur archetype — challenges whether the question itself is right. Exempt from evidence audit, immune to pruning.
- EXPLORE mode — for meta-questions. Each agent reframes instead of analyzes. Goal: find the best question, not the best answer.
- Preserve-if-unique triage — uniqueness overrides signal strength
- Creative disruption check — replaces early termination with "what would break this consensus?"
Structural Protections (enforced, not aspirational)
- Adversarial agents immune to pruning and early termination
- Socratic follow-ups: 2-3 per team (real elenchus)
- Refutation resistance replaces confidence scores
- Socratic Remainder in every report (the unexamined premise)
- Inverted termination: unanimous agreement increases scrutiny
Security Fixes
- Injection defense on Cross-Review, Devil's Advocate, Phase 5 templates
- Profile poisoning prevention (sanitization on directory scan)
- Scoped file access for adversarial agents
Anti-Boxing
- Contestability replaces Falsifiability (supports normative questions)
- Exploration signal: 6th Socratic Gate dimension
- Condition-based outsider injection (replaces every-5th-run counter)
Full Changelog: v4.0.0...v4.1.0