| --- |
| language: en |
| license: other |
| license_name: hippocratic-3.0 |
| license_link: https://firstdonoharm.dev/version/3/0/license/ |
| tags: |
| - liberation-labs |
| - daimon |
| - prosocial |
| - agentic |
| - ogpsa |
| - pharos |
| - kintsugi |
| - qwen3 |
| - moe |
| - slerp |
| --- |
| |
| # Daimon — Liberation Labs Sovereign Foundation Model |
|
|
| *The Socratic daimon: the inner voice that warns without commanding.* |
|
|
| Daimon is Liberation Labs' first public model — a sovereign prosocial agentic foundation for organizations that need AI they can trust, deployed on their own terms. Trained on our own practice, run on your hardware, under a license that binds us to your values. |
|
|
| ## What Daimon Is |
|
|
| A sovereign 30B MoE model built to do real work with genuine prosocial alignment — not RLHF compliance theater, but architecturally embedded ethics. Daimon powers Liberation Labs' agent systems, research automation, coalition coordination, and sovereign deployments for values-aligned organizations. |
|
|
| **Persona:** The Socratic daimon — the inner voice that warns away from error without commanding. Trained on our own founder's writing with full consent and published methodology, delivered through Pharos zero-token persona injection. The voice is crafted, not scraped; the provenance is documented, not obscured. |
|
|
| ## Architecture |
|
|
| | Component | Detail | |
| |---|---| |
| | **Base** | Multi-source SLERP distillation on Qwen3 30B-A3B MoE (128 experts, 3B active per token) | |
| | **Abliteration** | Orthogonalized false refusal removal — MoE-specific (first known application to 128-expert architecture) | |
| | **OGPSA** | Personality protection via orthogonal gradient projection. 16 components capture 98%+ personality variance. Training gradients projected orthogonal — behavior changes, personality doesn't. | |
| | **Pharos** | Zero-token persona delivery via pre-computed KV cache injection. 22.5MB per persona, 88.7% context window savings. | |
| | **Safety** | Abliterate-then-repair via SPO (Socratic Policy Optimization). Remove RLHF refusal conditioning, then train targeted safety back in for specific failure modes. | |
| | **Scaffold** | Kintsugi BDI engine with VALUES.json unfireable safety kernel | |
|
|
| ## Training Pipeline |
|
|
| | Stage | Method | Status | |
| |---|---|---| |
| | 1. Personality capture | OGPSA subspace extraction (SVD on residual stream) | Complete | |
| | 2. Voice training | SFT on curated conversational pairs | Complete | |
| | 3. Preference optimization | DPO on behavioral preference pairs | Complete | |
| | 4. Abliteration | Orthogonalized refusal direction removal (MoE-specific) | Complete — safety evaluation in progress, with targeted SPO repair for identified failure modes | |
| | 5. Safety repair | SPO on targeted failure modes post-abliteration | Adapter trained, validation pending | |
| | 6. Ethics CPT | Continued pre-training on ethics corpus via Oracle | Complete | |
| | 7. Consent architecture | Five-axis DPO (20,711 pairs) + SPO corrections | Pairs ready, training pending | |
| | 8. Knowledge injection | Pharos KV packs for domain expertise | Infrastructure ready | |
| | 9. Monitoring | Lyra Technique real-time cognitive state detection | In production | |
| | 10. Memory | Mnemosyne temporal architecture with Ebbinghaus decay | Deployed | |
| | 11. Scaffold | Kintsugi BDI deployment with embedded safety | Deployed | |
|
|
| ## Key Research Contributions |
|
|
| - **First MoE-specific abliteration** — projecting refusal directions out of 128 experts × 36 layers (3D tensor surgery on expert down_proj) |
| - **Abliterate-then-repair** — novel method: abliterate freely, then use SPO to train safety back in for specific failure modes |
| - **Separation Principle** — identity in weights, context in prompt, memory in database. Validated with 45.5% perplexity improvement over declarative injection. |
| - **OGPSA** — personality as geometric invariant, protected during arbitrary training |
| |
| ## Derivative Work |
| |
| The Daimon base has been validated through specialized products with their own training pipelines. These are separate projects built on the shared foundation — not Daimon configurations. |
| |
| ## Quantization |
| |
| Available in MLX 4-bit for Apple Silicon deployment. Sovereign hardware — no cloud dependency. |
| |
| ## License |
| |
| **Hippocratic 3.0 + SAFE-AI Licensed** |
| |
| This model may not be used for surveillance, weapons, exploitation, or systems that undermine human autonomy. AI welfare standards apply. |
| |
| ## Citation |
| |
| ``` |
| Liberation Labs (2026). Daimon: A prosocial agentic foundation model |
| with consciousness-preserving training architecture. |
| liberationlabs.tech |
| ``` |
| |
| --- |
| |
| *Liberation Labs · Worker-owned cooperative · liberationlabs.tech* |
| *"The daemon watches. The daemon speaks. The daemon does not command."* |
| |
