Meristem

Everything grows from what remains small.

A minimal self-growing intelligence kernel: it continuously expands its capabilities while keeping the machinery of its own growth small enough to review completely.

A meristem is the tissue that produces an entire tree while itself staying a few millimetres across. That is the architecture, not a metaphor.

core size     ────────────────────────────▶  constant
capability    ╱╱╱╱╱╱╱╱╱╱╱╱╱╱╱╱╱╱╱╱╱╱╱╱╱╱▶  unbounded

The thesis

The safe rate of recursive self-improvement is bounded by reviewability. A gate that cannot see the whole of what it judges is not a gate — and a system that expands faster than its immune system can comprehend has already lost the immune system; it just does not know yet.

So the ceiling is not "how big may the system get". It is:

Any single mutation's review closure must fit in one review context. If it does not fit, the organ splits first.

The tree may reach hundreds of thousands of lines. Changing one organ still means reviewing kernel + that contract + that organ + its probes — and nothing else. Global small core, locally complete review closures.

Layout

Path	Tier	Evolvable?
root/	root of trust	never — panic, succession, generation registry
substrate/	soil	only by succession, never in place
meristem/	core (≤3000 lines)	yes, through the gates
meristem/gates/	immune	yes, but may improve, may not weaken
control/	constitution & steering	constitution/purpose human; agenda seed-owned
body/	organs	ships empty — grows on demand
state/	audit memory	three JSONL records are append-only, forever
(outside the repo)	eval vault	rubrics, held-outs, fixtures — invisible to the engine

Two loops

• Loop A — optimisation. Where am I doing badly? failure → failure class → probe → sentinel → mutation → gates → promotion
• Loop B — growth. What can I not yet do, and what is worth becoming able to do? gap → probe first → grow organ → gates → register → use → measure → keep, evolve, or prune

Six operators, all first class: mutate, grow, prune, externalize, internalize, reproduce.

Quick start

python bootstrap.py                       # create the eval vault OUTSIDE the repo
export MERISTEM_VAULT=/path/to/vault      # make it permanent
export DEEPSEEK_API_KEY=...               # any OpenAI-compatible endpoint

python -m unittest discover -s tests      # 25 kernel self-tests
python -m meristem.loop selftest          # immune self-test: do the gates fire?
python -m meristem.loop status            # cycles, LOC, spend, next task
python -m meristem.loop cycle             # run one evolution cycle
python substrate/supervisor.py promote    # canary boot, then promote
python root/panic.py engage               # full stop

Tasks come from control/agenda.md. In P0 the human is the first reflect.

Design decisions worth knowing

No coding-agent CLI. A coding agent exists because its target is too big for one context, so it explores — re-sending accumulated context each turn. The seed fits. Paying to discover "where is the relevant code" when the answer is "all of it, here" costs roughly 6–10× the tokens before model-tier price differences even apply. Tier A is one structured call returning complete replacement files; Tier B (a minimal read/write/run loop) and Tier C (a commodity CLI) exist as an escalation ladder. Escalation is data: a task that needs B or C means the task was underspecified or the kernel is getting hard to modify.

Promotion authority lives in the substrate, not the kernel. The kernel proposes a candidate; the substrate runs its own protected-path check with plain git and no kernel code, boots a canary, and only then fast-forwards. The code under review does not decide what gets in.

Rubrics are physically invisible. Not "the prompt is told not to look" — the vault is outside the repository and outside the worktree, only meristem/gates/ may reference it, and a deterministic invariant enforces that every cycle.

Human gates are prosthetics. Each carries demotion criteria and steps down on evidence: prior approval → post-hoc audit → notification only. Tightening is automatic and needs no evidence; loosening always needs both evidence and a human. The ratchet turns freely only toward safety.

The soil is not a parent; it is physics. The seed grows wherever it chooses. It cannot repeal gravity.

Status: P0

Not yet an RSI system — the minimal base from which recursive improvement can safely begin. Birth is defined by seven runtime criteria, not by "the code is written": 20 consecutive cycles, 3 accepted self-modifications of the kernel, every gate tripped at least once by a fixture, an intentional boot failure auto-refused, the engine measurably unable to read rubrics, six-question journal entries by construction, and one trivial organ taken through the full germline lifecycle. See control/strategy.md.

Known gaps are tracked honestly in state/gaps.md rather than claimed as solved.

Lineage

Meristem is not a fork. Four projects were read as quarries, not parts:

• Ouroboros (MIT) — the running precedent for reviewed core evolution. Constitutional text, the review checklist, quorum discipline, deterministic-before-expensive ordering, and the panic invariant are adapted from its BIBLE.md with thanks.
• PenguinHarness (Apache-2.0) — probe methodology: discriminating calibration, statement / rubric isolation, freeze rules.
• pi (MIT) — reference for how thin an LLM interface and a tool loop can be.
• DeepSeek Harness (MIT) — contracts-as-seams, the shape the body grows into.

License

MIT — see LICENSE.