Ammonix/AmmonixRCM-Writer-9B

AmmonixRCM-Writer-9B: Specialized Claim Paperwork Writer

AmmonixRCM-Writer-9B is a 9 billion parameter language model developed by Ammonix, serving as the frozen local language model for the Ammonix RCM Agent. Its sole function is to write claim paperwork after each decision made by the agent, without influencing the decision-making process itself. This model is the exact artifact evaluated in the RCM paper (https://doi.org/10.5281/zenodo.22871212).

Key Capabilities & Performance

This model is a LoRA (r=32) fine-tuned on 1,339 self-generated synthetic pairs, rewarded by a simulated payer's verdict. It demonstrates superior performance in reducing claim rejections and financial losses in a controlled synthetic environment:

• Significantly lower rejection rates: In evaluations with 200 demonstration claims and 525 decision points, this model resulted in only 16 rejected claims, compared to 48 for Qwen 27B, 44 for Ornith 1.5 9B (base), and 37 for Claude Opus 5.
• Reduced financial impact: It minimized money lost on rejected claims to $7,709, substantially less than the $27,039 (Qwen 27B), $23,933 (Ornith 1.5 9B), and $21,157 (Claude Opus 5).
• Synthetic data training: Trained exclusively on fully synthetic data from the Cardessa synthetic claims world, ensuring no real patient, provider, or payer data was used.

Limitations and Intended Use

• Synthetic environment only: Trained and evaluated solely on synthetic data; it has no exposure to real claims, payer rules, or denial codes. Its writing style is tailored to one simulated payer.
• Writer, not a decision-maker: Functions strictly as a writer within the agent's harness; its outputs are unchecked outside this system.
• Not for production use: Not intended for production billing, clinical use, or submission to real payers. It is a research demonstration on a fully synthetic claims world.

This model is ideal for research, educational, and evaluation purposes within synthetic medical claims processing environments.