Fifteen Things a Self-Modifying Agent Cannot Do Without
Rice, Löb, Dawid, proper scores, Goodhart, Ville, Bahadur–Savage, Alpern–Schneider, Conant–Ashby, Hart, Fuller, Anderson, Thompson, Arrow–Fisher, Gall: what …
Writing
Essays on how to know whether an AI system works — pre-registered evaluation, proper scores, anytime-valid tests, checks that must prove they can fail — and on the court that lets an agent change itself; earlier, Bayesian decision theory from first principles; and a close reading of Israel’s 2026 election. If you came for one of the three kinds of work, start here; the five essays that carry the argument are on the front page.
/
Build the Court Before the Agent
A dispatch from day six, written 10 October 2026 and scored against the record. assize is a court for a self-improving agent that does not exist yet. Coding …
The Science Was Done. The Engineering Wasn't.
A self-modifying agent needs a fixed core it cannot reach, a free edge it can rewrite, and a score that is rigid everywhere. The founding note for jack: the …
Line by Line
A model of Israeli coalition-building since 2019 as a graph: the lists are points, each red line is an edge, and a government is a set of lists with no edge …
The Court That Tried Itself
assize was a court for coding agents: a constitution in a CLAUDE.md, checks that must catch a planted bug, a forecast ledger, builders and reviewers that never …
The Bitter Lesson for Persons
Fourteen briefs of the mathematics I argue for — a verifier sound by grammar, belief propagation over crossing letters, calibrated commitment, …
Station by Station
A live election projection on bechirot.fyi failed on the night of Israel's 26th Knesset election, 27–28 October 2026. The cause was not the model but the data: …
Polling Place by Polling Place
Findings from the final count of Israel's 26th Knesset election, 27 October 2026, compared with 2022 polling place by polling place: an Arab turnout surge of …
Fifteen Ways Past My Own Gate
A fresh agent session was pointed at a governance repo with one question: can the repo's own rules be broken while the verification gate stays green? It found …
The Ablation That Beat the Agent
I said a Credence paper was going to arXiv. It never went. Each of its three framings was killed by a specific measurement: frontier-LLM dominance, then a …
Signed Before the Code Existed
The frozen-oracle protocol under which a language was built by a coding agent: the acceptance tests were authored and signed before any code, then handed off …
The Agent That Prefers to Be Wrong
The off-switch theorem says an agent uncertain about what the user wants will prefer to defer rather than act unilaterally. That is not a design choice — it …
Think More, or Act Now
Rational metareasoning made honest. In a Bayesian agent language, think-more-or-act-now is one argmax, priced by the world's impatience rather than a hard-coded …
There Is No Forgetting
My own 2011 paper gave old evidence a way to decay: multiply the conjugate parameters by a forgetting factor before each update. A successor language throws …
The Alphabet Is the Prior
In a language where hypotheses are programs, description length is the prior — so every terminal you admit is a bit charged against every belief. Minimality …
What You Cannot Say, You Cannot Get Wrong
Rust makes illegal states unrepresentable; proplang makes axiom violations unsayable. The difference is that in a Bayesian agent language the alphabet is also …
The Gate That Said No
Whether to adopt a tool you built is itself a decision under uncertainty, so it deserves a decision rule, not a leaderboard. I built a pre-committed, …
Make Your OpenClaw Agent Cheaper, and Measure It Yourself
credence-pi is an OpenClaw plugin plus a local daemon that learns your agent and acts at two points by expected utility: it routes each turn to the cheapest …
Make Your OpenClaw Cheaper and Harder to Fool
credence-pi is an OpenClaw plugin plus a local daemon that learns your agent's behaviour and governs its tool calls by expected utility: it blocks the calls …
What a Regex Can't Do
The governor from 'The Brain is Opaque to the Body' now ships as an OpenClaw plugin: wasted tool calls blocked at precision and recall 1.0 on real sessions, …
The Brain is Opaque to the Body
A first-pass body-brain architecture for governing a coding agent's tool_call hook with a Bayesian decision-theoretic brain. The wire schema is fixed by what …
Three Types and a Funeral for Your Inference Library
What would it take to build an agent whose behaviour is derived from a few fundamentals the way physics is derived from conservation laws? Three types, four …
Keeping the Coding Agent on the Straight and Narrow
A companion to the PKM Phase 1 essay. The foundation was built by two AIs — Claude.ai for design, Claude Code for implementation — with a spec as the contract …
A Content-Addressed Foundation for Personal Knowledge
Eleven million words of personal documents, four canonical questions none of Khoj, Paperless-ngx, Obsidian, or Karpathy's LLM Wiki can answer, and a …
Ninety-Six Percent Cheaper and Slightly Better
Credence-proxy sits between an agent and its LLM providers, learns which model is good for which category, and routes accordingly. On an OpenClaw benchmark it …
The Prompting Gradient
Each prompting technique helps. Reasoning traces, strategy guidance, cross-question history — each one improves accuracy and score. None of them closes the gap …
The Agent That Invents Its Own Rules
Most agents are given a fixed set of decision rules. Credence's second tier generates candidate rules from sensor features, scores them by complexity, and lets …
Sixty-Two Percent Correct and Winning by a Hundred and Twenty Points
A Bayesian decision-theoretic agent scores lower on accuracy than the two stronger LLM variants it competes against — and beats the best of them by 120 points. …
The Loop Problem
Every RL agent that has played a text adventure has tried to take the lantern fifty times in a row. The fix is not better exploration heuristics. The fix is …
Teaching Zork to a Bayesian
A Bayesian agent plays text adventures with four information sources and a VOI gate on every query. The LLM recommends actions. The mathematics decides whether …
The Bitter Lesson Has No Utility Function
I wrote about decision theory fading from AI. Hacker News said I was annoyed at Rich Sutton's Bitter Lesson. I was not. But the misreading proves the point.
Why We Stopped Using the Mathematics That Works
Someone asked why decision theory stopped being widely used in AI. The answer involves ImageNet, academic departments, and the seductive power of not having to …
How Decision Theory Cuts Your AI Agent's API Bill in Half
Most LLM agents use tools like a tourist uses a phrasebook. A few lines of probability theory can fix it.
Agentic AI Is Neither Intelligent Nor an Agent
I built a Bayesian agent and set it against LangChain on a tool-use benchmark. LangChain got more answers right and still lost — by 120 points.
Evolution Discovers How to Think: A Philosophical Journey in Code
Part 2 of the Bayesian agent series. We confront the question of what should be designed versus what should be allowed to emerge, and discover that it is agents …
On Owning Your Data
Why I reverse-engineered a cheap Bluetooth scale to keep my health data out of the cloud
Building a Bayesian Learning Agent That Teaches Itself to Eat
A visual demonstration of Bayesian inference and the exploration-exploitation tradeoff. An agent navigates a grid world, learning which foods are safe through …
Scalibur: Reading Body Composition from a Cheap Bluetooth Scale
How I built a Raspberry Pi dashboard to capture and visualise body composition data from a GoodPharm TY5108 Bluetooth scale
I Built a GTD Telegram Bot in Under an Hour with Claude Code
How I built a Getting Things Done Telegram bot with natural language parsing using Claude Haiku, Python, and PostgreSQL — all in under an hour with Claude Code.
Hanukkah of Data 5783
Step-by-step R solutions to all 8 Hanukkah of Data puzzles — a data-focused alternative to Advent of Code with a delightful narrative twist.