unit-of-trust.
A scrollytelling essay arguing that agent decision legibility should be a first-class layer of the agent stack, not an afterthought.
Heads up: this is still a draft. I'm actively iterating on the writing and the scroll interactions, so the live version will keep changing.
What it is
A scrollytelling essay (live demo: unit-of-trust.vercel.app) that ties together three of my other projects into one argument: agent decision legibility should be a first-class layer of the agent stack, not something you bolt on after the fact. The through-line is pretty simple. With agents, the thing you actually need to trust isn't the output, it's the decision that produced it. So that's the unit the essay keeps coming back to.
The three projects it pulls from:
- recommend-arena — a recommender where the explanation ended up mattering more to me than the rank. Getting a good ordering wasn't really the hard part; being able to see why something got recommended was.
- observe-ui — collapses an agent's raw event log into decisions you can actually interrogate, instead of scrolling a 200-line trace trying to reconstruct what happened.
- wine-gpt — a small GPT-2 I built from scratch. The lesson I took from it was "deliberate out loud and defer": make the model show its reasoning as it goes, and hold off on committing to an answer until it has to.
Why I made it
I'd written these three things separately, and at some point I realized they were all kind of circling the same idea without me saying it out loud. I kept building tools to see why a system did something, not just what it did. This is me trying to write that down in one place, with the scroll interactions doing some of the explaining that prose can't.
It's a draft, so honestly some of the sections still feel a bit unfinished to me. I'd rather put it up and keep editing than sit on it.