1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
|
---
title: "Jayrup Nakawala | Projects"
---
<nav>
[Home](https://jayrup.me) | [Blog](https://jayrup.me/blog) | [Projects](https://jayrup.me/projects) | [CV](https://jayrup.me/cv) | [Dissertation](https://jayrup.me/dissertation) | [Git](https://git.jayrup.me)
</nav>
## [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)
--------------------------------------------------
|
