CConnectedResearchersinconnectedresearchers.hashnode.dev·14h ago · 7 min readPhoenix Treats Refresh Drift as a Recoverable Runtime StateThe DDR5 attack works because it exposes progress, detects a missed refresh, and repairs phase instead of demanding a perfect detector. Long-running systems fail differently from short routines. A det00
CConnectedResearchersinconnectedresearchers.hashnode.dev·1d ago · 8 min readDesigning LLM Assistants for High-Stakes Operations: Evidence Before VerdictsThe dominant product story for AI in a security operations centre is an autonomous copilot. It reads alerts, investigates telemetry, decides whether activity is malicious, and prepares or executes a r10
CConnectedResearchersinconnectedresearchers.hashnode.dev·2d ago · 6 min readTarget-Region Reconstruction as a Training Contract for Robot PoliciesA robot policy can recognize the instructed object and still move toward the wrong place. The gap sits between visual recognition and action generation. The target exists in the encoded scene, but the00
CConnectedResearchersinconnectedresearchers.hashnode.dev·3d ago · 7 min readPIPEMORPH: Treating Pipeline Parallel Training as a Runtime Control ProblemDistributed training systems often optimize the steady state. They choose a pipeline schedule, minimize bubbles on a healthy cluster, and assume the execution order will remain useful for the next ite00
CMChaeyeon Mia Leeinmiainflorence.hashnode.dev·4d ago · 7 min readGradCuit: Credit-Assigned Gradient Flow for Robust Test-Time Latent Reasoning in LLMsThe Problem Nobody Had Solved Test-time scaling is one of the most active frontiers in LLM research: spending more compute at inference to get better answers without touching model weights. Chain-of-T00
CConnectedResearchersinconnectedresearchers.hashnode.dev·4d ago · 8 min readEncode Once, Query Any Point: The Interface Behind D4RTA developer-focused reading of the CVPR 2026 Best Paper: one global video representation, an independent five-index query, and task cost that follows requested output. Most dynamic video geometry syst00
CConnectedResearchersinconnectedresearchers.hashnode.dev·4d ago · 8 min readEmail Aliases Are an Identity-Layer Bug, Not a String-Validation ProblemMost account databases impose a unique constraint on an email string and call the problem solved. Some implementations lowercase the value first. Others add special handling for Gmail dots and +tags. 00
CConnectedResearchersinconnectedresearchers.hashnode.dev·4d ago · 7 min readModelSpy: Reconstructing DNN Architecture from GPU Electromagnetic LeakageAn inference host can be unreachable from the network and still reveal information about the model it runs. The NDSS 2026 Best Paper “Peering Inside the Black-Box: Long-Range and Scalable Model Archit00
CConnectedResearchersinconnectedresearchers.hashnode.dev·Aug 2 · 7 min readYour Deployment Binary Needs Its Own Model Security ReviewImagine the release gate for a machine learning model. The weight hash matches. Accuracy is within tolerance. Trigger tests look normal. Four backdoor scanners find nothing useful. The build system th00
BABrett A McCallindarketype.hashnode.dev·Jul 30 · 5 min readOnly 2.2% of repositories earn an AFor the last several months, repOptics has been quietly doing one thing: scanning public repositories and grading them on how ready they actually are — measuring decisions, architecture, governance, d10
