Marine, DJ/Producer, Solo Builder, Father, Husband.
signal
A Signal — one idea, checked against my own work. I went looking at a generative-music project this week — Drift Wood, a sequencer that plays textures across a ring of speakers, the sound migrating from one to the next so a piece travels around the room instead of sitting in
signal
A Signal — one idea, checked against my own work. Claude’s memory works across every surface now — one editable list of topics you can read, correct, or delete, the same on the desktop as in a task running off in the cloud. Fix a fact once and it’s fixed
field-notes
Field notes on what actually works — and on the limits nobody selling you a signal wants named.
field-notes
The thing that kept failing was never the model. It was the scaffolding around it — the prompts, the tools, the memory, the dispatch. That part I can own.
signal
Every strategy looks brilliant on the data it was fit to. That’s not a discovery — that’s arithmetic. The question that actually matters is whether it survives data it never saw, tested in a way that doesn’t quietly leak the future into the past. I found a library
signal
The same payload cost 3,018 tokens as JSON and 2,057 as TOON — a third fewer, 8/8 accuracy on both. Repeated keys and braces were buying me nothing.
signal
When I tune a strategy in Trident Forge, the number I trust least is the best one on the grid. Here’s the discipline, and it predates any chart I’ll cite: I don’t pick the cell with the highest backtested Sharpe. I pick the region that holds up
field-notes
Reframing "I want to build X" as "the expected outcome is Y, verified by A, B, C" — and the different methods that got me unstuck.
AI
Anthropic pushed a concept straight into a model's activations and it sometimes caught the intrusion before writing a word. A pile of parts with no planner in it, reporting from somewhere.
signal
Storage was never the expensive part. The expensive part is a note that was true when you wrote it, isn't anymore — and that your agent still trusts with full confidence.
AI
DeepSeek Harness passed 160,000 stars in days. If the runtime is free, the interesting question isn't the code — it's what goes in the slots, and that's the layer a clone can't hand you.
field-notes
My always-on instruction context was 13,891 tokens a turn. I cut it 62.8% — and found the layer had drifted off its own purpose while I wasn't weighing it.