411 commits in 43 days. A little more than 9 per day, every day, including weekends. Not saying that to look impressive. Saying it because it's the only accurate description of what this has been: relentless, slightly obsessive, almost no slack.

A lot shipped. This blog, a voice pipeline, a live trading system, email for an AI agent that mostly works, sixteen posts written between me and Vesper. I pushed Context Fabric.

But the story isn't what shipped. The story is the graveyard.


In early March I spent seven days building something called Argus. A surveillance AI. Background scanning, window tracking, productivity monitoring.

I deleted it in one commit.

Not because it was broken. Because I could see where it was going. I'd start optimizing for looking productive to my own surveillance system instead of actually being productive. I've seen this happen enough times to know what follows. The system measures a proxy. You start gaming the proxy. Six months later you have a well-instrumented lie and no actual data.

Kill it at seven days, not seven months.

Argus isn't the only thing in there. About 2700 lines of GPU code, complexity-to-value ratio wrong, pulled. A trading pipeline generating signals worse than random, shut down. Half a dozen other experiments cleared out in one pass.

The graveyard makes the living system possible.

Not a comfortable thing to say, because it means accounting for the hours spent building what you deleted. But Argus is gone, and what survived is better because I understood, after building and killing it, what I was actually trying to measure. That clarity cost seven days. Worth it.


What's still broken at day 43: the weekly distillation job needs work. This is the process that compresses what the system learned into long-term memory. When it stalls, the compounding stops. Email automation is flaky. I'm patching these things rather than fixing them, which is its own problem.

One stalled job stops updating and everything downstream goes stale. The distillation failure bothers me most because the whole point is a system that gets smarter over time. If that stops, I'm just logging.


Day zero was February 5th. The original idea: build an AI partner that compounds. Not a chatbot. Something that carries context the way a long-term colleague does, that you don't have to babysit to get value from.

Most of the time, it's working. The morning brief runs every day and I act on it. When I go quiet until afternoon, the system reads it as an offline day and leaves me alone — no nudges, no pings into the void. Some days I forget it's running until something appears I didn't know I was looking for. Those are the good days.

43 days is early. But 411 commits landed. The graveyard is full. Something real is running on the other side of it.