Ornith 1.0: Self-Scaffolding Agentic Coding Models
Sam Witteveengo watch the original →
the gist
Ornith 1.0 is a family of fine-tuned LLMs that generate both task-specific execution harnesses and solution rollouts, allowing models to dynamically build their own scaffolding for coding tasks.
Self-Scaffolding Architecture
Ornith 1.0 introduces the concept of a self-scaffolding LLM, where the model is trained to generate both the execution harness and the solution trajectory. By treating the harness as a learnable object rather than a human-defined constraint, the model performs context engineering on the fly. The training process utilizes a two-stage reinforcement learning approach where the model proposes a refined harness, conditions its rollout on that harness, and receives reward signals via Group Relative Policy Optimization (GRPO) based on the success of the execution.
Reward Integrity and Verification
To prevent reward hacking, the system employs a three-layer defense mechanism. First, the sandbox environment and available tools are immutable, preventing the model from modifying its own execution constraints. Second, a deterministic monitor tracks the scaffolding process to penalize attempts to use unsanctioned tools or modify verification scripts. Third, an LLM-as-a-judge provides a final veto layer to ensure the generated solution adheres to the intended task requirements without taking unauthorized shortcuts.
Model Family and Performance
The Ornith 1.0 family consists of four models based on the Qwen 3.5 and Gemma 4 architectures, ranging from a 9B parameter model to a 397B mixture-of-experts model. Benchmarks indicate that these models are competitive with larger proprietary models like Claude Opus and outperform other open-weights models in agentic coding tasks. The 9B model is particularly notable for its performance relative to its size, making it a viable candidate for local execution on consumer hardware.