Starting a personal site, eight months in

Starting a personal site eight months into the project it’s meant to sit alongside is backwards. The usual order is: register the domain, put up an about page, then build the thing. I did it the other way round.

That wasn’t really a decision. For eight months the hours went almost entirely to the work itself. A personal site felt like furniture: something you arrange once the house exists. The house is now far enough along that the furniture is overdue.

There’s also a quieter reason. Writing about a project while you’re still unsure whether it will work is a particular kind of exposure, and I wanted enough of it standing that the writing could be description rather than hope.

So, briefly, the house.

The project is Consiliences: a research operation that uses a swarm of AI agents to test where independent lines of evidence converge, and, more usefully, where they don’t. Its public face is the Observatory, which publishes signals, each carrying a verdict. Some are confirmed. Many are killed, and the kill-list is public on purpose: by the last count, close to three hundred of roughly seventeen hundred adjudicated signals. A research surface that shows only its wins is advertising, not research.

Under that is a fair amount of machinery: well over half a million lines of Python, forty-two registered agents with distinct roles, and a validation pipeline that puts every candidate finding through four independent statistical batteries before it can earn anything stronger than “suggestive.” Around that sit the tools that keep a fleet of coding agents from stepping on each other while they build it. I built it, and I run it alone.

The part worth saying out loud is the cost.

It runs at hobbyist cost. The default language model runs locally, on my own hardware, rather than through a metered API. Cloud models are optional and used only when they earn their place. That single architectural choice is most of the reason a one-person operation at this scale is feasible at all. The expensive version of this project would need a team and a budget. The cheap version needed a specific set of decisions about where the compute lives and what’s allowed to fail gracefully without taking the rest of the system down.

I’m not going to explain Consiliences any further here.

That’s deliberate. consiliences.com is the project’s surface, and it can speak for itself. This site is mine, and the two jobs are different.

What robderoo.com is for is the building rather than the built. The homepage draws the year as a solar system: fourteen projects, close to twenty-seven thousand commits, every planet with a page of its own. Some of the projects live here too. The first is SwarmQuarium, which replays a month of my coding agents as an aquarium you can watch in the browser. The writing goes alongside: the decisions, the dead ends, the parts that are interesting precisely because they’re not finished. Some of it will be technical. Some of it will be about the odder problem of running a research institution that is mostly software.

I’m not promising a schedule. Posts will appear when there’s something worth a reader’s time, and not on a calendar.

That is the whole reason this site exists separately. I could have folded all of it into consiliences.com under a “notes” tab and saved myself a domain. I didn’t, because the architecture being built and the person building it are genuinely different objects, and each is better served by a surface that isn’t pretending to be the other.