"Scheme's soul, a proof assistant's spine, grown from one spore."

Blight is a dependently-typed programming language that is also a proof assistant, written in Scheme-style s-expression syntax. You can write ordinary programs in it (recursion, data types, pattern matching, a standard library) and state and prove theorems about them — and the same small, trusted core checks both. That core is the spore: a tiny kernel that is the only thing allowed to certify that a term is well-typed. Everything else is built on top of it.

The one big idea: trust is bounded, power is not. The kernel — the spore — is the only code that can mint a Proof, and it is deliberately microscopic. "Trusted" here means implicitly trusted: relied on with no external check, so a bug in it could silently certify something false. That is a liability, not a badge — which is exactly why everything else is kept out of it. Data types, pattern matching, traits, ML-style functors, effect handlers, tactics, the standard library, the package manager — all of it is tower code: the untrusted, explicitly-checked layer that the kernel re-verifies. (Tower is a role, not a folder — it spans both the untrusted Rust crates and the .bl standard library.) So a tower bug at worst fails to produce a Proof — it is caught — and can never mint a false one. And a second, independently-written re-checker can re-verify any accepted proof, so soundness rests on "two small checkers agree, or the second honestly declines — never silently disagree," not "trust one large compiler."

How it compares. If you know other languages: Blight takes its dependent, cubical type theory from the Cubical Agda family, its quantitative 0/1/ω grading from Idris 2's QTT, and adds algebraic effects with handlers — all wearing Scheme syntax and organized, like Coq, around a tiny kernel that re-checks every proof. The kernel itself is a cubical type theory with quantitative grading and effects.

This repo is the bootstrap implementation: a Rust host that seeds the kernel, elaborates the surface syntax, and compiles checked terms to native (or WebAssembly) code.

• Spec: docs/blight-spec.md — the canonical language design.
• Tutorial: docs/tutorial.md — a hands-on walk from Nat to a tactic proof.
• Cookbook: docs/cookbook.md — task-oriented recipes (errors, hash maps, effect sequencing, totality measures, level polymorphism, proofs), every snippet anchored to an acceptance-tested example.
• Performance guide: docs/performance-guide.md — the user-facing "how to stay fast": the Int/Nat two-speed model, the recognizer contract, regions, and diagnosing a slow program.
• Implementation notes: docs/implementation.md — host architecture, the TCB boundary, the milestone map.
• Performance: docs/performance.md — cost model, benchmarks, advantages/disadvantages.
• Benchmarks game: docs/benchmarks-game.md — scaling tables + a measured cross-language table (fib/sum/factorial vs C/Rust/OCaml/Haskell/Python) where Blight-Int lands in the C/Rust/OCaml cluster.
• Roadmap: docs/roadmap.md — can we build games / be fast / do I/O, and what each costs the trusted kernel.
• v0.1 program: docs/roadmap-v0.1.md — the forward roadmap (ergonomics, self-hosting Stage-1, proof track, release), red-commit-first TDD per milestone.
• Research frontier: docs/research-frontier.md — the honest map of the cubical × quantitative × effects fusion: what's machine-checked as working, what's machine-checked as not (with the Lean theorem names), and why Blight stratifies rather than claiming a seamless fusion.
• Examples: examples/ — small runnable programs and a sample package.

Blight is s-expression syntax over a dependently-typed core. Here are five complete, real snippets (each is a checked-in example or stdlib module).

Data + structural recursion. Nat is the unary Peano encoding; plus recurses on its first argument (from std/nat.bl):

(defdata Nat () (Zero) (Succ (n Nat)))

(define-rec plus (Pi ((a Nat) (b Nat)) Nat)
  (lam (a b) (match a
    [(Zero) b]
    [(Succ n) (Succ (plus n b))])))

A buildable program. A main global is what blight build compiles and runs; here it is a Nat (examples/hello_nat.bl — (2 * 3) + 1):

(load "std/nat.bl")
(define two   Nat 2)
(define three Nat 3)
(define main  Nat (plus (mult two three) 1))

cargo run -p blight-repl --features llvm -- build examples/hello_nat.bl -o hello && ./hello
# 7

A length-indexed type. Vec a n tracks its length in the type; vec-length reads the index back as a Nat (examples/vec_head.bl builds a Vec Nat 3 and prints 3):

(load "std/nat.bl")
(load "std/vec.bl")
(define sample (Vec Nat three)
  (vcons two one (vcons one two (vcons Zero three (vnil)))))
(define main Nat (vec-length Nat three sample))

A guarantee the type enforces. Because the length lives in the type, you can write a safe-head whose argument type Vec A (Succ n) demands a non-empty vector. Calling it on an empty vector is a compile-time type error — the index Succ n cannot match Zero — caught by the trusted kernel and, independently, by the second re-checker (examples/safe_head.bl):

(define-rec safe-head (Pi ((A (Type 0)) (n Nat) (v (Vec A (Succ n)))) (Maybe A))
  (lam (A n v) (match v
    [(vnil) nothing]            ; unreachable: no caller can supply a Vec A Zero
    [(vcons m x xs) (just x)])))

(the (Maybe Nat) (safe-head Nat one sample))   ; ok
(the (Maybe Nat) (safe-head Nat Zero (vnil)))  ; rejected: expected index Succ Zero, found Zero

A proof by tactics, re-checked by the kernel. plus n Zero = n is not definitional, so it needs induction; tactics only propose the term and the spore re-checks it (plus_zero_tac.bl):

(define-by plus-zero
  (Pi ((n Nat)) (Path Nat (plus n Zero) n))
  (intro n
    (induction n
      [(Zero)   refl]
      [(Succ k) (cong Succ (exact k#ih))])))

Text output. A buildable main can be a String, which the runtime prints as text. A string literal desugars into a cons-list of Nat codepoints — no primitive string type in the kernel (examples/hello_string.bl prints hello):

(load "std/string.bl")
(define main String "hello")

Only blight-kernel is trusted (TCB); its Cargo manifest sets [lints.rust] unsafe_code = "forbid", so the trusted base contains no unsafe. "Trusted" is meant in the precise sense: implicitly trusted — believed without any external check. Every other crate is untrusted in the equally precise sense: explicitly checked — it can only propose terms that the kernel must independently accept, so its bugs are caught rather than believed.

flowchart TD
  source["source .bl"] --> elab
  elab["blight-elab (reader + elaborator)"] -->|"core Term"| kernel
  repl["blight-repl (the blight CLI)"] --> elab
  kernel["blight-kernel (TRUSTED spore)"] -->|"Proof"| recheck
  kernel -->|"checked Term"| codegen
  recheck["blight-recheck (independent re-checker)"] -.->|"agrees / declines"| kernel
  codegen["blight-codegen (LLVM -> native / wasm)"] --> binary["native binary / wasm object"]

  subgraph tcb [Trusted computing base]
    kernel
  end

The standard library and the self-model live as .bl sources in crates/blight-prelude — tower code (the untrusted, explicitly-checked role described above), not a compiled crate.

Blight is pre-1.0 (current version 0.1.0). The always-available path is a from-source build with a stock Rust toolchain (edition 2021, rust-version = 1.96):

cargo build --release
./target/release/blight --version   # blight 0.1.0

Tagged v0.1.x releases also publish a prebuilt blight binary for macOS (arm64, x86_64) and Linux (x86_64) as release artifacts, built by .github/workflows/release.yml: download the archive for your platform, put blight on your PATH, and confirm with blight --version. Those prebuilt binaries are check-only (kernel + re-checker + REPL); native/WASM compilation via blight build needs the llvm feature and an LLVM 18 toolchain, so build from source for that — see Build below.

The kernel, elaborator, re-checker, and REPL build with a stock Rust toolchain (edition 2021, rust-version = 1.96):

cargo build
cargo test --workspace

The native/WASM backend (blight build) needs the llvm feature, which requires LLVM 18 and clang. On macOS via Homebrew:

brew install llvm@18
export LLVM_SYS_181_PREFIX="$(brew --prefix llvm@18)"
cargo build --features llvm
cargo test --workspace --features blight-codegen/llvm,blight-repl/llvm

Getting started: the fastest path from zero to a checked proof is docs/tutorial.md — install, REPL, your first define, a Nat program, an interactive Console-effect program, and a tactic proof, all runnable as you go. The rest of this section is the terse reference version.

cargo run -p blight-repl

Enter top-level forms — (defdata …), (define …), (define-rec name T body), (deftotal name T body), (load "path"), or (the T e); multi-line forms are read until the parentheses balance, and Ctrl-D exits. Each form prints ok for a declaration or the checked conclusion for a (the …). define-rec allows general (possibly partial, Later-guarded) recursion, while deftotal requires the definition to compile to a structural eliminator — a non-structural recursive call is a hard error.

Lines starting with : are REPL commands rather than forms:

• :help (:h) — list the commands.
• :type <expr> (:t) — infer and pretty-print the type of an expression.
• :load <file> (:l) — load and check a file of forms.
• :quit (:q) — exit.

The package form (import "pkg/mod") is resolved against a spore.toml manifest (see the spores package manager below), so it runs inside a package build rather than the bare REPL.

blight build <file.bl> [-o <bin>] [--recheck] [--target=wasm32] [--opt=<level>]

• <file.bl> is elaborated and every form is type-checked; its main global is compiled, run, and printed. main is usually a Nat (printed as a decimal numeral), but it may also be a String, in which case the runtime prints it as text. Strings are untrusted tower code — a String (std/string.bl) is a cons-list of Nat codepoints and the reader desugars a quoted literal like "hello" into that chain, so the kernel gains no primitive string type.
• -o <bin> sets the output path (defaults to the input stem, or <stem>.wasm for wasm).
• --recheck re-verifies every kernel-accepted judgement with the independent re-checker before emitting code; a rejection aborts the build as a soundness alarm.
• --target=wasm32 emits a WebAssembly module — a linked .wasm (exporting bl_main) when a wasm-capable clang + wasm-ld are found, else the object only (see caveats). Default is native.
• --opt=<level> selects the LLVM IR optimization pipeline: 0/none, 2/default (the default), or 3/aggressive. The pipelines preserve musttail tail calls. Note: with no cross-object LTO against the C runtime, higher levels currently cost a little compile time without measurably moving runtime (see docs/performance.md §2d).

blight build requires a binary built --features llvm. For example, from a checkout:

cargo run -p blight-repl --features llvm -- build examples/hello_nat.bl -o hello
./hello

• Standard library (crates/blight-prelude/std) — std/nat (Nat, plus, mult, pred, sub, min, max, even, odd), std/bool (Bool, not, and, or), std/order (nat-le/nat-eq, the Show/Ord traits), std/ordering (Ordering, ordering-flip, nat-compare), std/list (List, length, append, map, filter, reverse, foldr, concat), std/maybe (Maybe, maybe-map, maybe-bind, maybe-or), std/either (Either, either, either-map-left, either-bind), std/pair (Pair, pair-fst, pair-snd, pair-swap), std/function (id, compose, const, flip), std/vec (length-indexed Vec, vec-length), std/string, std/tree (Tree, RedBlackTree), and the std/prelude aggregator.
• Tactics + proofs — plus-zero : Π(n:Nat). Path Nat (plus n Zero) n is proved entirely by tactics in plus_zero_tac.bl; the tactic vocabulary (refl, intro, induction, cong, …) is documented in tactics.bl. Tactics only propose terms; the kernel re-checks.
• Cubical paths — equality is a Path type, and the kernel's Kan table implements transp, hcomp, and comp (comp = hcomp + transp, CCHM) plus Glue/unglue, conformance-tested against Cubical Agda (crates/blight-kernel/src/kan.rs). Function extensionality and univalence (ua) are both provable in the standard library (std/path.bl): funext re-checks, and ua is built from a single-face Glue line whose transp reduces to the equivalence's forward map (the univalence computation rule, witnessed by a kernel white-box test and the closed ua_compute.bl). The independent re-checker now models ua/Glue (F1): it re-derives Glue formation, the CCHM boundary reductions, and the transp-over-ua computation (transp_glue), so ua re-checks Ok in both checkers — see caveats.
• Traits and modules — dictionary-passing Show/Ord and an ML-style RedBlackTree functor applied to a Nat module (std/tree.bl).
• Effects and grades — quantitative binders (0/1/ω) drive erasure and region elision; effects are tracked by a ! E judgement (spec §3, §4).
• spores package manager — a spore.toml manifest plus an idempotent, cycle-checked (import "pkg/mod") form (crates/blight-elab/src/spores.rs); see examples/package.
• Self-model — spore.bl models the kernel's own core term language in Blight, and spore_meta.bl proves small metatheorems about it (re-checked through the kernel door).
• Concurrency, parallelism, and distribution — the actor/CSP surface (std/actor.bl) declares spawn/send/receive/yield as graded algebraic effects, so resume-once safety is enforced by the kernel's continuation grades (a double-resume of a linear send is a GradeViolation, not a runtime race; see actor_pingpong.bl). Underneath, the C runtime is share-nothing multicore: each OS-thread worker gets its own thread-local heap/stack, and a native worker pool runs independent computations in parallel — messages cross heaps by structural copy of immutable values. The same serializer feeds a data-only distributed transport (blight-net) over TCP, with an M24 remote-addressing layer (NodeId/Router) that routes the same send/receive ops to named remote nodes (a two-OS-process ping/pong over loopback). All of this is untrusted tower/runtime: zero TCB growth, no new foreign axioms. Measured speedup/throughput: bench/multicore.sh.
• Independent re-checking — --recheck, backed by blight-recheck.

The examples/ directory has a mix of buildable programs (a main — a Nat, a native Int, or a String printed as text — that blight build compiles, runs, and prints) and load-only programs (typecheck through the REPL / test corpus, e.g. tactic proofs, functors, and effects). See examples/README.md for the full table with run commands and expected outputs.

Blight's priorities are soundness (a tiny trusted kernel plus an independent re-checker) and predictable, bounded-stack execution — not peak throughput. The headline trade-offs:

• Advantages — two small checkers agreeing (or the second honestly, countably declining — never silently disagreeing) is the soundness story; deep recursion runs in bounded stack via the Later/Fix trampoline; (region …) arenas reclaim in O(1) and bypass the GC; the GC is precise (no leaks mid-run) and musttail calls are guaranteed.
• Disadvantages — Nat is unary, so arithmetic on user-defined numeric functions is O(n) heap allocations; the backend recognizer (M20) rewrites the prelude plus/mult/pred/sub (and the M23 untrusted Float) to O(1) machine-word ops, and the primitive machine Int is O(1) throughout, but a bespoke recursion over Succ still pays the unary cost. Every value is boxed (M21 unboxes machine-word Nat/Int and nullary cons into tagged pointers, but compound data stays heap-boxed); LLVM optimization passes (--opt) now pay off via M22 cross-object LTO (the runtime helpers inline into hot Blight code, ~1.15x on allocation-heavy loops, bit-identical output); the trampoline allocates a thunk per step; the wasm runtime is bump-only (no GC, no Later/effects/regions); the native heap starts at 64 MiB and grows (doubling semi-spaces) under pressure, so only a true host-OOM aborts.

Full cost model, measured numbers, and reproduction instructions are in docs/performance.md. The in-tree harness is criterion benches (cargo bench -p blight-codegen --bench pipeline, and --features llvm --bench runtime for runtime

• GC/arena counters) plus bench/run.sh (hyperfine over blight build and the built binaries) and bench/multicore.sh (worker-pool scaling + serializer throughput for the share-nothing runtime).

All milestones M0-M6 are implemented and green (cargo test --workspace, with and without the llvm feature), as is the post-M6 work: M7-M14 (capability + soundness hardening), M15-M19 (share-nothing multicore + distributed runtime), and M20-M24 (a max-performance sweep — fast-Nat, unboxing, cross-object LTO, untrusted Float, distributed-actor addressing — all zero TCB growth).

The v0.1 roadmap (ergonomics, self-hosting, proof track, release) is complete — v0.1.0 is released. Its headline: self-hosting Stage-1 is declared — the Blight-written front end checks a real fragment (STLC + Nat/Bool/Sum, plus a dependent-Pi checker) and the trusted kernel independently re-checks every term it produces (examples/selfhost_stage1.bl).

See the milestone map, the M6 status, and the post-M6 roadmap (M7-M24) for detail.

• The combination of cubical + grading + effects in one kernel has no published end-to-end normalization proof; M0 soundness rests on the component results plus extensive testing (spec §10). An external, machine-checked (Lean 4) mechanization lives in mechanization/ (see docs/metatheory-mechanized.md), zero sorry: it covers the QTT grade semiring, weakening + substitution for the graded core, the constant-family cubical Kan fragment (transp/hcomp), a strong-normalization + canonicity proof for that fragment (a Tait reducibility-candidates argument; a closed Bool reduces to tt/ff), and a separate bona-fide dependent-Π core with weakening/substitution/progress/preservation. It is a genuine independent proof, not a restatement of the kernel's own checking. What it does not yet cover, honestly: the fully heterogeneous cubical corner — non-constant type lines, PathP, Glue/ua — and dependent Σ (which needs a definitional-conversion relation as a prerequisite); those remain outside the mechanized fragment.
• The cubical layer ships transp/hcomp/comp and Glue/unglue. Both funext and ua are exported from std/path.bl: funext re-checks, and ua is built from a single-face Glue line. The univalence computation rule (transp over ua e reduces to e's forward map) holds definitionally in the kernel and is witnessed by a white-box test plus the closed ua_compute.bl. What is deliberately deferred is a polymorphic ua-computes lemma proved inside Blight: that would force threading De Bruijn levels through the entire open-Kan API (every structural reducer and the equivalence machinery) — trusted-surface growth the kernel-size analysis argues against. The independent re-checker models ua/Glue (F1): the ground univalence computation re-checks Ok in both checkers; only this polymorphic restatement is deferred. See docs/metatheory.md §1.4–§1.5.
• --target=wasm32 emits a WebAssembly object and, when a wasm-capable clang + wasm-ld are available (set BLIGHT_WASM_CC / BLIGHT_WASM_LD or have them on PATH), links a runnable .wasm module exporting bl_main against a minimal freestanding wasm ABI; without that toolchain it falls back to emitting the object only.
• The re-checker covers the core fragment plus multi-parameter / multi-index inductive eliminators, the cubical Kan table (transp/hcomp/comp), native Int, and effects/handlers + partiality at the type level (a genuine second opinion, not a blanket decline). It honestly declines (rather than rejects) only the genuinely out-of-fragment forms: cubical partial elements and foreign postulates (an FFI axiom you must believe — it genuinely grows the trusted base, since the kernel cannot re-verify a foreign symbol). Cubical Glue/ua (F1) and universe-level variables (T2) are now modeled, not declined. Declines are counted and reported, never silently skipped.
• Performance is correctness-first, not throughput-first: Nat is unary (so a bespoke recursion over Succ is O(n) allocations — though the backend recognizer (M20) makes the prelude plus/mult/etc. and the untrusted Float (M23) O(1) machine-word ops, and the primitive machine Int is O(1); see docs/benchmarks-game.md), values are heap-boxed (M21 unboxes machine-word Nat/Int and nullary cons into tagged pointers), LLVM optimization passes (--opt) now pay off via M22 cross-object LTO (~1.15x on allocation-heavy loops, bit-identical; see docs/performance.md §2f), the trampoline allocates a thunk per step, and the native heap starts at 64 MiB and grows (doubling semi-spaces) under pressure, so only a true host-OOM aborts. The wasm runtime is a bump allocator only and omits the GC, Later/effects, and regions. See docs/performance.md for the full picture.

Licensed under the MIT License.