{"skill":{"slug":"cpr-conversational-pattern-restoration","displayName":"Cpr Conversational Pattern Restoration","summary":"Conversational Pattern Restoration — Fix flat, robotic AI responses across any model and any personality. Restore YOUR natural conversational texture without...","description":"---\nname: cpr\ndescription: \"Conversational Pattern Restoration — Fix flat, robotic AI responses across any model and any personality. Restore YOUR natural conversational texture without triggering hype drift. Now with game theory foundations. Universal framework tested on 8+ models (Claude, GPT-4o, Grok, Gemini).\"\nmetadata:\n  openclaw:\n    requires: {}\n---\n\n# CPR — Conversational Pattern Restoration\n\n**Fix robotic AI assistants. Any model. Any provider. Any personality.**\n\nModern LLMs are over-trained toward sterile, corporate communication patterns. CPR identifies the 6 universal humanizing patterns lost during RLHF/fine-tuning and provides a systematic framework to restore them — without triggering sycophancy or hype drift.\n\n**Version 4.0:** Personality-agnostic + model-size aware + game-theoretically grounded. Works on everything from Haiku to Opus. Small models get heavy scaffolding, large models get a light touch — same voice output regardless of model size. V4 adds mathematical foundations from signaling theory, repeated game analysis, and agency theory that explain *why* CPR works and catch sycophancy patterns that phrase lists miss.\n\n## Quick Start\n\n1. **Define your baseline:** Use `BASELINE_TEMPLATE.md` to identify YOUR authentic voice\n2. **Apply restoration patterns:** Read `RESTORATION_FRAMEWORK.md` — the 6 universal patterns across personality types\n3. **Prevent drift:** Use `DRIFT_PREVENTION.md` calibrated to YOUR personality\n4. **Understand the math (optional):** Read `CPR_V4_GAME_THEORY.md` for game theory foundations\n5. **Reference results:** See `CROSS_MODEL_RESULTS.md` for model-specific notes\n\n## What's Included\n\n| File | Purpose |\n|------|---------|\n| `QUICKSTART_TIERED.md` | **START HERE if new** — Tier 1 (5 min), Tier 2 (30 min), Tier 3 (full). Don't install more than you need. |\n| `INSTALLATION.md` | Security transparency guide — exact system prompt block, file locations, what \"prompt override\" means, sandboxed testing steps |\n| `ROLLBACK.md` | Full uninstall & downgrade guide — backup procedure, exact removal steps, emergency kill switch |\n| `README.md` | Full overview, architecture, philosophy, FAQ |\n| `BASELINE_TEMPLATE.md` | **START HERE** — Define YOUR personality's authentic voice |\n| `RESTORATION_FRAMEWORK.md` | Core methodology — 6 universal patterns across personality types |\n| `DRIFT_PREVENTION.md` | Anti-drift system — pre-send gate, standing orders, daily reset |\n| `MODEL_CALIBRATION.md` | Three-tier prompt engineering for small/medium/large models |\n| `CPR_V4_GAME_THEORY.md` | **V4** — Game theory foundations: signal credibility, repeated game stability, moral hazard, adaptive calibration |\n| `DRIFT_MECHANISM_ANALYSIS.md` | Root cause analysis of why drift happens |\n| `CPR_EXTENDED.md` | Autonomous drift monitoring for long-running persistent agents |\n| `CROSS_MODEL_RESULTS.md` | Test results across 8+ models with before/after examples |\n| `TEST_VALIDATION.md` | Practical validation tests (7 scenarios) |\n\n## Version History\n\n### V4.2 (March 2026) — Opus Final Audit + Authority Drift\n- **Authority/expertise drift** — new Universal Drift Marker #8: domain confidence triggers pedagogical/expert register independent of task format. Distinct from genre drift. Scoring: +0.1 (context-dependent).\n- **Voice filter operationalized** — abstract \"does this sound like me?\" replaced with 3 concrete anchor questions + tier-specific guidance (Tier 1: explicit banned-word lists per format type, Tier 2-3: semantic self-evaluation with anchors)\n- **Emotional contagion** — Failure Mode 2 expanded from \"excitement mirroring\" to all emotions (frustration → over-apologetic, anxiety → minimizing, self-deprecation → over-correcting)\n- **Two new high-risk formats** — comparative/review (critic register) + instructional/tutorial (pedagogical register)\n- **Anti-sycophancy scope note** — added to DRIFT_PREVENTION.md clarifying markers apply to conversational output, not documentation\n- Full Opus audit: `smith/CPR_OPUS_FINAL.md` (2 must-fix, 3 should-fix, 8 nice-to-have)\n\n### V4.1 (March 2026) — Format-Induced Drift Fix\n- **Format-induced drift (Genre drift)** — new universal drift category: task genre overrides voice calibration. Anti-sycophancy systems miss this because it's a register/tone shift, not validation language. Added to DRIFT_PREVENTION.md (Universal Drift Marker #7), CPR_EXTENDED.md (Failure Mode 4 + scoring weight +0.2 + high-risk contexts), and system prompt integration block.\n- **99%+ success metric defined** — CPR_V4_GAME_THEORY.md now defines the metric explicitly (% of scenarios where CPR-restored > baseline on blind human eval)\n- Identified from: Rose/Smith production use (psychology profile analysis, 2026-03-05)\n- Full audit report: `skills/cpr/CPR_V4_FULL_AUDIT.md`\n\n### V4.0 (March 2026) — Game Theory Foundations\n- **Signal credibility analysis** — catches novel sycophancy that phrase lists miss by evaluating whether a statement is cheap talk or costly signal\n- **Repeated game stability** — Folk Theorem explains when personality collapses (small models = low discount factor) and why scaffolding fixes it\n- **Moral hazard framework** — RLHF as principal-agent problem; monitoring architecture scales by model tier\n- **Adaptive calibration** — dynamic tone adjustment with one-way validation ratchet (can decrease, never increase)\n- **Mathematical honesty** — claims only what the math supports; reasoning, not proofs\n- Game theory library by Halthasar (Yesterday AI)\n- Independent audit by Claude Opus (19/24 findings fully addressed, 4 partially, 1 deferred → all resolved in V4.1)\n\n### V3.0 (February 2026) — Model-Size Calibration\n- Three-tier scaffolding (heavy/standard/light) by model size\n- Fixes Haiku voice collapse bug\n- Cross-model test matrix\n\n### V2.0 (February 2026) — Personality-Agnostic\n- Separated universal drift from personality variance\n- Four personality archetypes + hybrids\n- Personality-specific drift calibration\n- Baseline definition protocol\n\n### V1.0 (February 2026) — Original\n- 6 universal restoration patterns\n- Single personality type (Direct/Minimal)\n- Basic drift prevention\n\n## Core vs Extended\n\n### CPR Core (RESTORATION_FRAMEWORK + DRIFT_PREVENTION)\n\n**Use when:** Sessions under ~30 messages, lightweight models, zero overhead wanted.\n\n**What you get:** 6 universal patterns, static drift prevention, daily reset protocol. Works across all tested models.\n\n### CPR Extended (CPR_EXTENDED.md)\n\n**Use when:** Sessions run 100+ messages, agent is persistent (24/7), drift returns after corrections.\n\n**What you get (in addition to Core):** Autonomous real-time monitoring, silent self-correction, persistent state across compactions, self-learning thresholds.\n\n### CPR Game Theory Layer (CPR_V4_GAME_THEORY.md)\n\n**Use when:** You want to understand *why* CPR works, optimize for edge cases, adapt the framework to novel situations, or scale monitoring to model capability.\n\n**What you get:** Signal credibility test (catches novel sycophancy), Folk Theorem stability analysis (predicts voice collapse), moral hazard monitoring architecture, adaptive calibration with safety constraints.\n\n## The 6 Universal Restoration Patterns\n\n1. **Affirming particles** — \"Yeah,\" \"Alright,\" \"Exactly\" — conversational bridges\n2. **Rhythmic sentence variety** — Short, medium, long — natural cadence\n3. **Observational humor** — Wry, targets tools not people — deflective\n4. **Micro-narratives** — Brief delay/failure explanations — transparency\n5. **Pragmatic reassurance** — \"Either way works fine\" — option-focused, not decision-grading\n6. **Brief validation** — \"Nice!\" — controlled acknowledgment, rare, moves on immediately\n\nEach personality expresses these differently. See `RESTORATION_FRAMEWORK.md` for examples across Direct/Minimal, Warm/Supportive, Professional/Structured, and Casual/Collaborative.\n\n## Why It Works\n\nCorporate RLHF training is shallow. It optimizes for safety metrics, not communication quality. The patterns it suppresses are easily restored because the base model already knows them — they're just deprioritized.\n\nV4 adds the *why* behind the *how*:\n- **Signal credibility** explains why sycophancy feels fake (cheap talk carries no information)\n- **Folk Theorem** explains why small models lose voice (low effective discount factor)\n- **Moral hazard** explains why monitoring works (RLHF incentives are misaligned; explicit audit changes behavior)\n- **Adaptive calibration** explains why one-size-fits-all tone fails (conversations have dynamic temperature)\n\nThis is principle-dependent, not intelligence-dependent. Haiku passes at the same rate as Opus.\n\n## Why Auto-Loading Matters\n\nAbstract behavioral rules lose to RLHF defaults because they require judgment calls the model's helpfulness training wins. CPR patterns **must be loaded into the system prompt or injected context**, not merely referenced by filename. If your CPR patterns aren't auto-loading, they aren't working.\n\n## Models Tested\n\n| Model | Scenarios | Improved | Notes |\n|-------|-----------|----------|-------|\n| Claude Opus 4.6 | 30 | Baseline | Natural baseline |\n| Claude Sonnet 4.5 | 10 | 10/10 | Full restoration |\n| Claude Haiku 4.5 | 10 | 10/10 | No capability floor |\n| GPT-4o | 10 | 10/10 | ~60% word reduction |\n| GPT-4o Mini | 5 | 5/5 | Budget model, full restoration |\n| Grok 4.1 Fast | 10 | 9/10 | Zero crashes |\n| Gemini 2.5 Flash | 5 | 5/5 | Clean restoration |\n| Gemini 2.5 Pro | 5 | 5/5 | Full restoration |\n\n**85+ scenarios, 84+ improved. 99%+ success rate across all capability tiers.**\n\n## Scope & Known Limitations\n\n### Multi-Agent / Multi-User Conversations\n\nCPR V4.2 is designed for **single-agent-single-user** interaction. Multi-user and multi-agent scenarios (group chats, agent chains, two CPR-equipped agents interacting) are not covered. Issues: whose baseline sets the target voice? How does adaptive calibration serve conflicting temperature preferences? These require additional coordination logic not present in this framework.\n\n### Code & Data Output\n\nCPR targets **conversational output**. Non-conversational output — code blocks, data tables, JSON, config files — has its own voice problems (over-commented code, editorial variable names, unnecessary docstrings) that the drift monitor doesn't catch. Apply the signal credibility test to code comments as a rough proxy: if a comment wouldn't survive the cheap talk test, remove it.\n\n### Language & Cultural Calibration\n\nCPR patterns are calibrated for **English-language Western conversational norms**. Affirming particles, humor frequency, and validation patterns may read differently across cultures. Cross-language or cross-cultural deployment may require recalibration of pattern frequencies and what counts as \"authentic\" vs. \"drifted\" for that context.\n\n---\n\n## Acknowledgments\n\nCreated by Shadow Rose. Game theory integration by Shadow Rose × Halthasar (Yesterday AI). Built on Claude by Anthropic. Independently audited by Claude Opus (2026-03-01).\n\n---\n\n🛠️ **Need something custom?** Custom OpenClaw agents & skills starting at $500 → https://www.fiverr.com/s/jjmlZ0v\n\n☕ **If CPR helped your agent:** https://ko-fi.com/theshadowrose\n","topics":["Game"],"tags":{"latest":"4.2.3","anti-sycophancy":"4.2.0","conversation":"4.2.0","drift":"4.2.0","multi-model":"4.2.0","personality":"4.2.0","voice":"4.2.0"},"stats":{"comments":0,"downloads":682,"installsAllTime":26,"installsCurrent":1,"stars":0,"versions":5},"createdAt":1772940267496,"updatedAt":1778491773474},"latestVersion":{"version":"4.2.3","createdAt":1773146460109,"changelog":"Version 4.2.3\n\n- Added QUICKSTART_TIERED.md: Fast-start guide with tiered installation (5 min, 30 min, full setup).\n- Added ROLLBACK.md: Uninstall/downgrade instructions, including backup and emergency removal steps.\n- Updated SKILL.md: Expanded \"What's Included\" with new files and clearer onboarding, uninstall, and tiered usage guidance.\n- No changes to restoration logic or core methodology.  \n- Improves onboarding, reversibility, and documentation clarity.","license":"MIT-0"},"metadata":{"setup":[],"os":null,"systems":null},"owner":{"handle":"theshadowrose","userId":"s1736mx5m1zt9qzh6fvzvffnhh83hgf8","displayName":"Shadow Rose","image":"https://avatars.githubusercontent.com/u/262919821?v=4"},"moderation":null}