research Cross-Model SAE Feature Transfer Can sparse autoencoder features learned in one LLM be matched and causally transferred to a different LLM? Tokenizer-robust signature matching with 4-control steering validation. Subliminal Unlearning Controlled transfer and mechanism audits — does an unlearning method change the resource surface for recovering held-out facts from behaviorally floor-equivalent models? Lazarus — Cross-Model Transfer Attacks on LLM Unlearning Systematic test of whether unlearning is robust to cross-model transfer attacks across 4 architectures, 3 unlearning methods, and 3 attack vectors. "Universal Forgetting is Fragile." Ghosts in the Model Machine unlearning and interpretability study showing that forgotten behavior can be restored through activation patching. ImmunoPath Immunotherapy decision-support pipeline for predicting immune biomarkers from H&E slides using MedGemma and related medical foundation models. Multi-Domain GRPO Reasoning Google Tunix Hackathon project teaching Gemma2 to reason transparently with multi-domain GRPO training. spon-extension Extending the SPON paper (arXiv:2512.12744) with layer-wise allocation and mechanistic interpretability — where to place SPON biases and what they encode. GRAFT — Cross-Model Residual Transplant Zero-shot cross-model residual-stream transplant of DeepSeek-R1-Distill-Llama-70B activations into Llama-3.3-70B on AMD MI300X. Affective Concealment — Where Models Learn to Mask Measuring and localizing affective concealment in post-trained LLMs. Do instruction-tuned models suppress emotional affect that base models show, and where in the network does that happen? development ANVIL Compliance-grade RAG over ASME standards with auditable provenance, typed retrieval, and bounded agentic tool use. evidiff — Evidence-Locked Code Migration An agent that migrates dependencies and proves every changed line — source to claim to rule to usage to patch to passing test to independent reviewer. MIB Doc Challenge — PDF Adjudication Ledger Provenance-first evidence ledger for adversarial PDF adjudication. Air-gapped, deterministic, with dual-OCR and a role-aware trust ladder. MythForge Agentic mythology engine that expands a seed concept into an interconnected world using Gemini, LlamaIndex, Neo4j, Mem0, and Gradio. AMNESIAC — Reverse-Turing Interrogation The Turing test, inverted — now you have to prove you're human. An AI interrogator reads your pulse via rPPG and micro-expressions from a 478-point face mesh, all on-device.