Research Notes

Forty-six dated entries held the Cambrian project together across eleven days and a dozen AI sessions. Commit messages couldn’t. We’ve extracted the setup — adapted from Kanare’s 1985 chemistry textbook — into a standalone repo you can drop into any project. Why append-only matters for AI sessions, the hypothesis-before-experiment table, and what a real lab entry captures that a commit message cannot.

Seventeen days after M1, the answer to “will evolution optimize for the wrong things?” is more interesting than yes or no. A 10-of-10 failure run traced to a hardcoded command in the Supervisor, not the LLM. A subsequent campaign hit 80% viability. Differential analysis across 8 viable organisms found four quality patterns the spec never required — now automated by an AST-based phenotypic distiller that feeds the wins back into the genome. The Baldwin Effect, mechanized.

Two days after Gen-2 failed: Gen-1 produced five viable offspring in 44 minutes. M1 is done. Two spec patches — streaming requirement, aiohttp test pattern — eliminated the whole failure class. 474,834 tokens, 50% first-try success rate, 100% test pass rate on every promoted generation. And a look at what M2 actually requires: container isolation, fitness vectors, spec mutation with grammar constraints.

Nine days after writing that Phase 1 would take a week and cost $20–50: Gen-1 ran its generation loop for the first time, generated Gen-2, submitted it to the test rig, got back a failure report, and rolled it back — all autonomously. The loop works. Here's what it cost and what it taught: why the spec is really a description of the environment, not the agent.

What if instead of patching code, we evolved the specification and regenerated the agent from scratch? The spec is the genome, the code is the phenotype, and economic viability is the fitness function. An agent that can't pay its own bills — compute, LLM calls, storage, bandwidth — goes extinct. Prediction markets as the first income strategy. CAMBRIAN-SPEC-001: the 300-line generative specification for a self-sustaining software lineage.

Gen-72: a ClojureScript agent autonomously modified its own source code, verified the change, and promoted it to master in 56 seconds. The real story is the 16 failures that got us there — six infrastructure bugs, three models tested, and the lessons each one taught.

The MVP field report: 2,214 lines of ClojureScript, 90 tests, 17 generations run. From early failures to a 5/5 stable pipeline. What we built, what broke, what's still missing, and the road to recursive self-improvement.

How do you steer a self-improving agent? With a collaboratively written contract. The refined execution model: user and Prime draft a program.md, an autonomous Lab executes it, Prime verifies independently, and git tracks every generation.

A ClojureScript agent running in a container, modifying itself, and proving the modification in a second container before promoting it. Both the argument and the spec. No frameworks, no build step — just a loop, a contract, and two containers.

We argued for Elixir, then found its limits. Lean's type checker is an incorruptible judge for self-improvement. Memory is the beachhead. Here's what's hard about it, and what kills it.

Karpathy's autoresearch reveals the structure that makes self-improvement work: three separations, a keep/revert loop, and well-chosen fixed points. The blueprint for agents that get better overnight.

An existing Elixir port of the Pi coding agent already solves the hardest architectural problem. One architect, frontier models, five weeks. Here's the case.

A practical assessment: port the core gateway and agent loop to Elixir, keep the TypeScript channels, and bridge them. 3–4 months for humans, 4–6 weeks with coding agents.

Why building a self-modifying agent in TypeScript is writing letters to your future self. Common Lisp offers depth, Elixir/BEAM offers safety — and the hybrid might be the real answer.

Five approaches to giving an AI agent a coherent mental model of its own codebase — from source-aware skills to a dedicated dev agent — and the security tradeoffs of each.

A thought experiment: designing an AST-native code editing protocol for LLMs using Lisp's homoiconicity. Inspired by Can Bölük's hashline work.

A comprehensive research report mapping the companies, technologies, and supply chains for atmospheric control and CO₂ scrubbing in next-generation lunar stations.