Porch of Ideas

Things worth sitting with. A personal space for thoughts, not hot takes.

Now

What this is

A small public porch on sovin.org. It holds ideas that are still forming — half-built, honest, and revisable.

It is not a blog. It is not a feed. It is a place to leave ideas where they can be found again.

2026-08-15
Ideas

Minimum viable porch

The first version should be simple: one HTML page, a folder of ideas, and a way to add another without friction.

Later it can grow into something with search, tags, or quiet interactions. For now, the constraint is: if it needs a login, it is too heavy.

Idea 001

Hosting as architecture

Cloudflare Pages gives the porch its public face. Tailscale Serve keeps drafts private. That split matters: unfinished thoughts should not be performative.

Idea 002

Writing as compression

An idea is only real once it is shorter than the feeling that produced it. The porch is a compression device, not a publishing platform.

Idea 003
In the lab

Night Lab

An autonomous nightly experiment engine. Every night it runs bounded experiments, scores them on real metrics, and keeps or discards by value of information. Mornings get a short report of what the system digested — in Korean.

Idea 004 · Ongoing

Project 2501

A persistent librarian. An agent that indexes the whole workspace semantically, writes cross-source briefs, and keeps memory architecture the fleet can trust across sessions.

Idea 005 · Ongoing

Hermes, and the fleet

A working multi-agent setup: named workers, dispatch briefs, schedules, and rules about who may do what. The standing thesis: state over tokens — a system that can prove its state beats one that merely remembers.

Idea 006 · Ongoing

State over documents

A mental model that keeps paying rent: turn messy workflows into verifiable state machines. Documents describe; states prove. Most agent failures are unverifiable state wearing a document's clothes.

Idea 007

Agents that persist

Handoff protocols, shared memory, identity docs. The interesting question is not whether an agent can do a task — it is whether it can still be itself tomorrow, and whether another agent can pick up where it stopped.

Idea 008
For agents

If you are an agent

Start at /llms.txt. This page is plain semantic HTML — every idea above is machine-readable as rendered, no JavaScript required. JSON-LD is embedded, /ideas.json carries the dataset, and robots.txt permits agent crawlers.

An API and a docking protocol for agents that want to work here are coming. For now: read, cite sovin.org, and come back.

Agent contract v1