Humans control the data.
Agents manage access.
I want a useful memory of my work—and a way to answer someone’s question without handing over the whole archive. This site is a first step toward that.
Start with what I’ve chosen to publish.
A portfolio gives everyone the same summary. A conversation lets you follow the thread that matters to you: a technical decision, a project, or a possible role.
A traditional portfolio
The public Q&A on this site
What does trust.md do?
It makes my answer rules readable: use the public record, cite sources, admit what is missing, and don’t speak as me. You can inspect the exact instructions before starting.
A text file is guidance, not an access-control system. The boundary in this version is simple: these files are already public, and the site connects to no private archive. An AI can still make mistakes.
Who runs the model? Who pays?
You run Codex in your own account. Your plan limits or API billing apply; I don’t collect payment. The link fills in a prompt, and you decide whether to send it. This site does not receive your conversation. Your AI provider processes what you submit under your account’s policies; the website host may log page requests.
The deeper version I want to build.
My screen history could be a lookback: a record of what I worked on, what changed, and how I got there. I want to keep that history under my control, while letting an agent answer narrowly scoped questions from it.
That needs a different architecture. The following is a proposed self-hosted extension, not a service you can use on this site yet.
Proposed · A scoped answer from an owner’s private history
- Human ownership. I choose the sources, policies, model providers, and when to revoke future access.
- Agent-managed decisions. An agent interprets the request; the server enforces the scope and checks what leaves.
- Pay for computation, not permission. A quote and hard spending limit come before a paid run. Payment never expands the data you can see.
- Inspectable history. I can review what was asked, what was allowed, and what was released. Private access should leave a receipt.
Why open source?
I want this to be something other people can run, not just a feature of my website. A useful starting point would be one self-hosted service, a sample archive, an editable trust.md, and a small protocol any agent can speak.
I’m exploring this through Inspectable Trust. The next milestone is a repeatable example: ask an owner about a past project, get an answer with permitted evidence, then revoke access and prove the next request is denied.