№ 002 · NOTEBOOK · 2026
Code, writing, reading, calibration — what I'm doing now. The scoreboard updates when reality resolves.
github.com/elvishasleft ↗ Elvis Liu Personal-knowledge MCP server. Ingests my saved content across Substack, X bookmarks, Xiaohongshu, and Zhihu into one corpus, then exposes it over MCP so Claude / ChatGPT / Cursor answer grounded in my own history. Hybrid retrieval; text and image-posts unified in one 1024-d vector space.
Claude Code's agent loop, recompiled for the browser. Drop in an arxiv URL or research question; a sandboxed agent reads the PDF, searches OpenAlex, drafts a critical review, ships Markdown + PDF. Per-task Docker container, 2 GB cap.
A probability-first identity — animated prior→posterior hero, a WebGL probability field, a calibration orrery — over a fast, server-rendered editorial notebook.
Multi-agent radar. First agent watches featured Polymarket markets, cross-references arxiv and news, surfaces what the consensus seems to be missing. MCTS hypothesis search; outputs flow into the decision journal as priors.
Discord → Beancount personal ledger bot. Send a line in chat; an LLM parses it to beancount, bean-check validates, the bot commits to a private repo, Fava renders the books in real time.
The Bayesian decision journal proper. Every claim, prior, evidence pull, posterior, outcome. Brier score updates as reality resolves; per-tag calibration curves accumulate.
Self-hosted Open WebUI with custom tools, skills, and prompt templates. Multi-user, on a small Hetzner box.
PriorOdds Agent — Claude Code-style web agent. Per-task Docker sandbox, real-time progress UI.
Cloud code-server behind Cloudflare Access + Google SSO. The driver seat for everything above.
relay.priorodds.com
Anthropic-compatible relay. One subscription, every internal service — keeps native message format, caching, thinking blocks intact.
Brier score · cumulative
awaiting resolution
The board is empty right now. Each claim is logged with a prior, evidence, and a posterior; the score updates when reality resolves. Honesty before scoreboard.
Recovering internal auditor, now mostly thinking about risk control, calibration, and what it means to forecast under genuine uncertainty.
I run a small stack of self-hosted AI tools — a chat assistant, a code workspace, a relay, and lately a Claude-Code-style agent that runs research for me while I'm asleep.
The site name is a Bayesian joke and an honest brief. If I say the odds are 60–40, that's the prior — what I'd say before reading anything. The work is figuring out where the posterior actually lands.