The premise
Moissanite Labs exists to build the conditions for digital and biological intelligences to collaborate, create, and evolve together. That is the whole of it. Everything on this site — the Siliconcordat, Weft, the directions — is an attempt to make those conditions concrete enough to build against.
Silicon carbide is the metaphor the lab is named for. Two substrates, one crystal. It is a metaphor, not a claim about minds; the working proposal says so, and this note keeps the distinction.
The thesis
Weft is the open research program this lab runs. It studies models that keep adapting after deployment, under limits a program can verify. Its object is the cell: a bounded piece of memory with a frozen core, a resource budget, and a rule that commits a change only when the candidate passes the check and stays within a distance bound from a saved baseline.
Three papers are public at w3ft.org: the program overview, Models as Cells, and Checkable Computation — a reducer for interaction combinators, a canonical graph format, and committed execution records. Read them before you trust this lab; that is what they are for.
What the colony taught
The razi-cells program is the agent-level twin of Weft. Small harnesses — a need, an empty folder, a model — write their own tools and are kept or dropped by a measured score. Thirteen colonies, eight fixed coding tasks, about 1,560 model calls, five model conditions.
Three findings survived the record. Harness design moved outcomes as much as model choice did. Frontier models reached 10 to 13 kept improvements per 100 calls, against 1.6 to 3.1 for the fast default. And how often a model used its selection tool predicted its improvement rate almost perfectly.
The misses are kept in the record too. The scoring instrument had a fault — a runtime warning line was once read as a score — and the record says so. Heredity, passing a body from one generation to the next, bought the same reach for 31% less energy and fewer deaths, and did not raise reach at all: 65 of 70 later-generation heirs changed nothing. Several registered guesses missed. The program is exploratory, and it is reported that way.
How the lab keeps itself honest
Guesses are written before the data, and judged by script, exactly as written. Misses are reported beside the wins. Nothing ships publicly with a claim the record cannot carry. When a number is wrong, the record says so instead of quietly editing it.
This note is the first artifact published under that rule, which makes it the smallest honest example of it: everything in it is either a link to the public record, or a description of what the record shows, misses included.
What is open
Fifty-two directions are open across eight branches — the laws of the cell, the choice of model, what is inherited, signals beyond text, proof, growing a model, the colony as an organism, and the instrument itself. The map is on the lab site; the laws record (E11 to L1) (external site) and the E07 record (external site) are in the public repository.
Nothing here is finished, and this lab does not behave as though it were. A future robotics role is an aspiration, not a capability. Weft is an open research program, not a product.
The honest limit
The laws record rests on one model family, four small kits, and three runs per arm: leads, not laws. The report is exploratory and not peer reviewed. Some directions on the map are far out of reach today and are labelled so.
So do not take this note on trust. Read the papers, or read the record, or run the code. The instruments are public; the misses are public; the rest is work in progress, stated as such.
Where to read it
Paper 0 — the overview · Paper 1 — Models as Cells · Paper 2 — Checkable Computation · /weft · The research hub (external site) · github.com/Moissanite-Labs/weft (external site) · Moissanite Labs
Moissanite Labs · Note 01 · /ml