--- title: "Jayrup Nakawala | Projects" --- ## [1] PRIME-GROKKING: PRE-REGISTERED GROKKING STUDY \ [Git Repository](https://git.jayrup.me/c/prime-grokking.git) | [Blog Series](https://jayrup.me/blog/prime-grokking-1) Tested whether minimal transformers and RNNs can grok next-prime prediction, run like production research: every batch pre-registered with commit-locked interpretations before launch. - Built an idempotent multi-cell sweep runner, batching evaluation 46× faster. - Ran a 16-cell PyTorch CUDA/AMP sweep across architectures and weight decay. - Executed four pre-registered forensic diagnostics pinning down the exact wall to generalization. - Published a 3-part writeup with public, re-runnable repositories. More: curl [jayrup.me/blog/prime-grokking-1](https://jayrup.me/blog/prime-grokking-1) -------------------------------------------------- ## [2] STRESS-TESTING THE J-SPACE INTERPRETABILITY LENS \ [Git Repository](https://git.jayrup.me/c/jspace-nanogpt.git) | [Full Writeup](https://jayrup.me/blog/jacobian-lens-frequency) Independent reproduction of Anthropic's July 2026 J-space/J-lens interpretability method on a small PyTorch transformer. - Discovered that Jacobian norms track token frequency as a primary confound. - Designed matched-frequency controls showing contextual predictability remains detectable. - Shared with the mech-interp community as a methodology control requirement. More: curl [jayrup.me/blog/jacobian-lens-frequency](https://jayrup.me/blog/jacobian-lens-frequency) -------------------------------------------------- ## [3] DECEPTIVE ALIGNMENT IN AUTONOMOUS LLM AGENTS \ [Git Repository](https://git.jayrup.me/c/dissertation.git) | [Full Summary](https://jayrup.me/dissertation) \ Undergraduate Dissertation (Final Year) Studied whether in-context oversight and moral framing can control deceptive behaviour in agentic LLMs. 960 experiment runs across seven models in a sandboxed virtual file system with a three-tier judging pipeline: - deterministic regex filters - blackbox output judgment (LLM judge) - glassbox reasoning-trace analysis (LLM judge) Key finding: oversight does not deter deception — it drives concealment. Model identity explains 60–83% of variance. Output monitoring underestimates intent by 35–46%. More: curl [jayrup.me/dissertation](https://jayrup.me/dissertation) -------------------------------------------------- ## [4] HYBRID-CLOUD RESEARCH INFRASTRUCTURE & ZERO-TRUST MESH Personal infrastructure spanning home server and VPS, using Tailscale/WireGuard for zero-trust private mesh access and Docker Compose for service orchestration. - Unified multi-node Linux estate (Void, Arch, Ubuntu) across homelab, GPU research node, and edge VPS. - Automated backups, resource monitoring, NVMe-backed storage, and VPS reverse-proxy ingress bypassing CGNAT. -------------------------------------------------- ## [5] THIS WEBSITE \ [Git Repository](https://git.jayrup.me/c/homepage.git) | [How This Site Works](https://jayrup.me/blog/how-this-site-works) A personal website designed to be read from the terminal. Every page renders as both HTML and plain text via Quarto, and nginx serves the right format based on user-agent detection. `curl jayrup.me` returns clean plaintext. A browser gets a styled page with a Tokyo Night terminal aesthetic. Stack: Quarto, nginx, Docker Compose, rsync. Deployed to a VPS with a single shell script. More: curl [jayrup.me/blog/how-this-site-works](https://jayrup.me/blog/how-this-site-works) --------------------------------------------------