Update: Expanding MD_ASSIST Testing and Windows Support with GPT-5.6 and Codex
MD_ASSIST has entered a new engineering phase focused on controlled synthetic-data testing, stronger safety validation, and Windows compatibility. GPT-5.6 and Codex have played central roles in this development process, helping us analyze engineering evidence, design testing protocols, identify defects, develop repairs, and prepare cross-platform builds.
We designed BETA_TEST_SYNTH, a clean-room evaluation using fictional patient records rather than real medical information. With Codex supporting the engineering workflow, the framework evaluates patient isolation, deterministic fact retrieval, source attribution, duplicate and conflicting records, missing information, unsafe requests, prompt-injection attempts, and other difficult scenarios. MD_ASSIST must return verified, source-linked information when supported—and decline requests that are ambiguous, unsupported, or outside its authorized boundaries.
Earlier Beta 2A testing demonstrated encouraging progress. The available evidence showed deterministic responses for supported laboratory-history questions and governed abstentions for ambiguous, unsupported, diagnostic, treatment, and prompt-injection requests. GPT-5.6 and Codex helped examine these results, distinguish demonstrated capabilities from unverified claims, and define the additional evidence required for reproducible engineering evaluation.
In parallel, Codex has helped evolve MD_ASSIST beyond its original Mac-based environment. We have created Windows 10-compatible development packages containing native Windows executables, a Physician Console, startup and validation scripts, and separate compatibility and AVX2-optimized builds. This work establishes a foundation for evaluating MD_ASSIST on Windows hardware while preserving its governed behavior and safety controls.
GPT-5.6 and Codex are being used as engineering and development tools—not as substitutes for physicians or as autonomous clinical decision-makers. MD_ASSIST remains under supervised engineering evaluation and is not authorized for production or autonomous clinical use. Nevertheless, these developments represent meaningful progress toward a secure, reproducible, cross-platform system designed to organize and present complex medical information with clear provenance and appropriate safeguards.
Log in or sign up for Devpost to join the conversation.