RFC index

The RFCs are the design record. They live in design/ in mzizi-dev/mzizi, and the repository’s own README says to read them first. Two things to know before reading any of them. They argue with each other, on purpose. RFC-0002 opens by declaring RFC-0001’s design target wrong and inverting three of its conclusions. RFC-0003 records a claim its own test suite falsified mid-implementation. RFC-0004 opens by correcting the premise it was commissioned on. That is the intended shape of the record, not disorder in it. The status headers are historical. RFC-0001 and RFC-0002 both say “nothing here is implemented”. That was true when written. The front end now exists — see Status for where the documents and the code have since diverged.

---

​

RFC-0001 — Surface syntax, canonical form, and the agent protocol

The nine failure modes an agent hits writing Rust UI code, and the syntax that answers each. end <kind> <name> with a cross-checked name echo; one construct per intent; enum data columns instead of parallel maps; contract blocks in the language; capability declarations at the top; no ownership surface at all. Also specifies canonical form (the compiler owns formatting, no configuration) and the mz check --agent protocol: whole-program NDJSON in deterministic order, at most one diagnostic per real error, messages written for a reader with zero file context, and fixes as machine-applicable data tagged exact / guess / none. Its own method rule: “If a decision doesn’t trace to a failure mode, it doesn’t belong in the language.” Leaves open: local state; the grammar formalism; the IR; whether component name must match file name. → Syntax · The compiler

---

​

RFC-0002 — The design target, the runtime as the product, and what to mine

The correction that reshaped the project. RFC-0001 was written by a frontier model reasoning about its own experience, which optimised for the author rather than the population. The design target is small models — a 7B open-weight model running locally — and that inverts three conclusions: token efficiency is permanently first-order, the end echo is an error-correcting code for weak long-range attention rather than parser elegance, and a small closed grammar beats a familiar open one. Also establishes that the runtime is the product (“shared logic, written once, running everywhere”), commits to the content-addressed IR, and surveys thirteen languages for what to borrow — Elm’s error design, Roc’s platform/application split, Gleam’s tiny keyword set, Unison’s content-addressed codebase, Zig’s no-hidden-control-flow rule, Tree-sitter’s error-tolerant parsing — each with an explicit licence note. The licence discipline is stated once and is unambiguous: design ideas are free; code carries obligations; GPL/AGPL sources must not be copied into the project at all, and anything borrowed as code must be recorded in a NOTICE file. §3.1 names what is genuinely missing everywhere, which is the defensible research contribution: contracts in the language checked by the toolchain; a diagnostic protocol designed for a machine reader; a content-addressed IR wired to an agent-facing patch API; and a grammar whose ambiguity budget is deliberately near zero. Leaves open: local state; the IR’s concrete shape; effect-system depth; how far contract expressiveness should go before it becomes a proof assistant. → Overview · Syntax

---

​

RFC-0003 — The content-addressed IR and the agent’s read/write surface

The other half of the loop: eight barriers an agent hits reading a codebase, and the content-addressed IR that answers seven of them from one decision. The node model, the length-prefixed canonical serialization, hand-rolled SHA-256 verified against NIST vectors, structural sharing, mz outline, and the query/patch surface. Contains the project’s most instructive paragraph: §7.1, where the measured suite falsified the RFC’s own claim that structural paths are “stable across edits”, and the claim was narrowed to what is true rather than defended. Leaves open: local state under content addressing; the patch API’s conflict model; store persistence; contract evaluation semantics. → The content-addressed IR

---

​

RFC-0004 — Test topology: what stays public, what goes private, and why

Which test infrastructure stays public (nearly all of it, permanently), which is held out (a benchmark task set), and the dependency rule that keeps forks working: private consumes public, public never consumes private. Opens by rejecting the reason it was commissioned on. Private tests are close to worthless as a security control; the one legitimate reason here is measurement validity, because a public benchmark task set gets scraped into training data and a contaminated benchmark looks exactly like a successful one. Leaves open: held-out set rotation policy; a path for third-party verification of a published benchmark claim. → The Phase 0 benchmark

---

​

Reading order