White papers
Full-length research documents and detailed technical write-ups. Read like a paper, not a blog post.
-
The Black Hole Information Paradox: Physics, Formulas, and an Interactive Simulator
A technical reference covering the black hole information paradox from Hawking radiation through the Page curve, island formula, and ER=EPR — with documentation for a Feistel-cipher-based simulator of the full process.
PhysicsToolsquantum information Read paper -
GRIMOIRE: Static Binary Triage via PatchSpec Two-Model Inference
Technical reference for GRIMOIRE — a static binary analysis tool that combines entropy scanning, heuristics, and a two-model Ollama inference pipeline (PatchSpec) to produce verdicts, analyst reports, and YARA rules without executing the binary.
AIResearchmalware analysis Read paper -
minidet: A Queue-Based Multi-Signal Network Threat Detection Sidecar
Architecture and design rationale for minidet — a Suricata sidecar that runs four signal layers (LIMEN, BPB, GRIMOIRE, PCAPR) at different costs and latencies, correlates verdicts across destination IPs, and emits enriched EVE JSON.
AIResearchnetwork security Read paper -
PCAPR: Protocol Reverse Engineering and Security Analysis from Raw Packet Captures
Technical reference for PCAPR v0.6.0 — a tool that infers protocol structure, detects security anomalies, and generates dissectors, detection rules, and fuzzing artifacts from pcap files without prior protocol knowledge.
AIResearchnetwork security Read paper -
PetriDish: A Mechanistic Interpretability Workbench for Transformer Language Models
Architecture and experiment results for PetriDish, a 22-endpoint FastAPI workbench for real-time activation interception, causal patching, head ablation, and steer vector injection in any HuggingFace causal LM.
AIResearchmechanistic interpretability Read paper -
Cross-Model Inversion in Whitespace Robustness: The Pythia Anomaly
Pythia-1.4B produces per-sentence whitespace robustness rankings that are anti-correlated with Llama-3.2-3B and Mistral-7B (r = −0.485 to −0.512). The inversion strengthens at 2.8B before reversing at 6.9B. Tokenizer and frequency controls rule out the simplest explanations.
AIResearchmechanistic interpretability Read paper -
Context Entanglement in Whitespace Robustness: The Q6/Q7 Dissociation
Causal patching (exact hidden state transplant) recovers the whitespace merge effect at layer 3. Steer vectors (averaged direction injection) at the same layer are null. The dissociation establishes that the effect is context-entangled in residual space — not extractable as a universal direction.
AIResearchmechanistic interpretability Read paper -
Sharpening Circuits in BPE-Tokenized Language Models: Early QK Routing and Late OV Projection
Whitespace perturbations increase model confidence in ~34% of merge cases. We characterize the two-stage circuit responsible: early BOS-anchoring heads (layers 0–3) are load-bearing via QK routing; late OV projections (layers 15–27) carry the sharpening signal independently of attention patterns.
AIResearchmechanistic interpretability Read paper -
The Human Baseline Problem: Why AI Code Detectors Fail Across Domains
Stylometric classifiers achieve 0.92–0.99 F1 within a single corpus but collapse to near or below random chance cross-corpus. We identify the mechanism, confirm it across three independent corpora, and provide a continuous diagnostic scorer.
AIMLcode detection Read paper