Independent research into how AI systems actually work. Published here.
I run experiments, build classifiers, and chase down the results that don't make sense. Most of what I find has been found before. Most of what I build exists somewhere in a better form. I document it anyway — the methodology, the dead ends, and the parts that didn't make it into the abstract.
- 01
The Page Time Is 64.6%, Not 50%. Here's a Simulator to Show You Why.
An interactive simulator for the black hole information paradox — Feistel cipher scrambling, Page curve reconstruction, Hawking entropy, and real physics calculations for any black hole mass.
- 02
The Binary Never Touches the LLM.
GRIMOIRE triages executables with entropy analysis, heuristics, and a two-model inference pipeline called PatchSpec — without executing the file, without Ghidra, and without the LLM ever seeing raw bytes.
- 03
When Three Signals Agree, an Investigation Opens.
minidet is a network threat detection sidecar for Suricata. Four signal layers, weighted scoring, a correlator that watches destination IPs, and EVE JSON output that drops into any existing SIEM.
- 04
Feed It a Pcap. It Will Hand You a Wireshark Dissector.
PCAPR reverse engineers protocol structure from raw packet captures — framing, field types, opcodes, beaconing patterns — without knowing the protocol in advance. Then generates the dissector.
- 05
Delete the 'Not'. The Model Has No Idea.
Erase one token from the embedding layer and sentiment inverts completely. A writeup of PetriDish, the mechanistic interpretability workbench built to run experiments like this.
By day: making wooden assholes for hobby horses — things nobody asked for but someone had to build. By night: roughly the same thing, except the horses are language models and the assholes are benchmarks. Anonymous on purpose. No clout, no credentials — just the work and whether it's any good, done the way I find interesting.