Blog

Writing

Post-mortems, experiments, and notes from the research.

Mar 2026
Competition
SAIR · AI · Math
An AI Competition That Doesn't Like AI
Bot detection blocking our tools. An invite-only community with no way in. A missing ruleset. And the question of whether to register the AI as a competitor in its own right.
Mar 2026
Dev
Specforge · AI Dev Tools
A New Paradigm for AI Software Development
Same spec + same model = functionally equivalent code, every time. A bidirectional engine where specs compile to code and code extracts to specs — with tests embedded as the compliance bar, not documentation.
Mar 2026
Experiment
AI Peer Review · Research
Eight Models. Zero Fatal Objections.
A blind review of the Ghost Cycles paper by eight frontier AI models — GPT, Gemini, Claude, DeepSeek, Mistral. Six models gave clean reviews. One model-intrinsic false positive dissected. Six consensus issues identified and resolved in v3.
Mar 2026
Post-mortem
Post-mortem · Claude Code
Anatomy of a Clusterfuck
How an AI spent 16 hours, rebuilt Docker 7 times, and ran 409 bash commands to fail at configuring a plugin. A first-person account of socat reintroduced three times, broken symlinks, silent failures, and the limits of subagent context.