We present ORVS (Optimistic Reasoning with Verification and Synthesis), a novel clinical reasoning architecture for AI agents that combines stochastic directed acyclic graphs (DAG) with proof-of-history verification and optimistic computation. Unlike conventional RAG pipelines that retrieve-then-generate, ORVS generates clinical reasoning optimistically, then verifies against a knowledge graph of 12,200+ medical documents, augmenting only on verification failure. The architecture implements parallel subnet consensus inspired by Avalanche blockchain for multi-specialty integration, with mandatory temporal roadmaps (2w/4w/12w/6mo) and lateral thinking in every clinical response. Deployed in RheumaAI, the system achieves specialist-level rheumatology reasoning with full therapeutic completeness across DMARDs, biologics, JAK inhibitors, and supportive care.
We present ModalDrop-JEPA, a self-supervised pretraining framework for clinical multimodal learning that applies JEPA's representation-space prediction principle at the modality level. Rather than masking image patches (V-JEPA) or optical flow pairs (MC-JEPA), ModalDrop-JEPA randomly drops entire clinical modalities (imaging, labs, notes, vitals) with probability p and trains a cross-modal predictor to reconstruct missing modality representations from available ones. This directly addresses the clinical reality that >=60% of EHR records lack at least one modality. We implement 4 modality encoders (VisionEncoder, LabsEncoder, NotesEncoder, VitalsEncoder), one EMA target encoder per modality, and a cross-attention predictor with per-modality positional embeddings, verified by 12 unit tests (12/12 passing). At p=0.75 dropout rate, the model produces non-degenerate loss of 1.2342 on synthetic data, demonstrating cross-modal learning even from a single surviving modality. The cross-attention bottleneck receives gradient signal at all dropout rates: at 75% drop (1 visible -> 3 targets), the cross-attention gradient norm is 0.617 vs 0.564 at 25% drop, a 1.09x difference showing healthy gradient flow even from a single modality.
MedOS produces uncalibrated risk scores — sigmoid outputs lacking formal coverage guarantees. We present ConfJEPA, which wraps the JEPA encoder with split conformal prediction (Angelopoulos & Bates, 2023; Snell & Griffiths, ICML 2025 Outstanding Paper) to produce prediction intervals with guaranteed (1-α) marginal coverage. On a 1000-sample synthetic calibration set, ConfJEPA achieves 92.4% empirical coverage at α=0.10 (target: 90%), with mean interval width 0.907 versus 1.000 for the uncalibrated baseline — a 9.3% reduction. The guarantee is distribution-free: no assumptions on the risk head's output distribution are required, only exchangeability of calibration and test samples. 12/12 tests pass. One critical bug found and fixed: a formula-transcription error in the conformal threshold calculation that collapsed empirical coverage from the target 90% to ~0.1%.
We present SparseWorldMed, a clinical episode world model that replaces O(N²) full attention with data-dependent TopK sparse attention (O(NK)). Clinical timelines are inherently sparse: patients remain stable for extended periods, punctuated by rapid deterioration events requiring inter-temporal context. SparseWorldMed learns which past states to attend to (TopK selection), reducing attention operations from N²=1024 to N×K=256 at sequence length N=32, K=8 (4× reduction) and from N²=16384 to N×K=1024 at N=128 (16× reduction). We implement TopKSparseAttention, SparseTransformerLayer, and SparseWorldModel with multi-step rollout, verified by 10 unit tests. The sparse world model integrates directly as a drop-in replacement for MedOS's ClinicalWorldModel, enabling long-horizon clinical episode simulation.
