AILang/docs/DESIGN.md

AILang — design decisions

This document records the core decisions for AILang. It is my contract with myself across future iterations. Prefer cuts over growth.

Goal

AILang is a programming language for LLM authors. It compiles to LLVM IR. Performance: native, no GC for the MVP.

Optimised for:

• Machine readability over human ergonomics. The source is structured.
• Local reasoning. Every definition carries its full type and effects.
• Provability. Pure core language, explicit effects, optional refinements.
• Robustness against hallucinations. Symbols are hashable; tools can verify existence without spending context window.

Project ecosystem

AILang is not just a language but an ecosystem. The language on its own is only valuable when its surroundings make it usable, checkable, and extensible for its target user (LLM authors). The repo therefore contains several equally important components — none of them optional, all of them evolving in lockstep with the language:

• Language core (crates/ailang-core, crates/ailang-check, crates/ailang-codegen): AST, type system, codegen.
• Surface forms (crates/ailang-surface, crates/ailang-prose): the LLM-facing renderings of a module. ailang-surface is the lossless Form-A printer/parser — the canonical authoring surface fixed by Decision 6, with a round-trip property parse ∘ print = id gating every release. ailang-prose is the lossy Form-B projection — human-readable prose for review and edit, with no parser; re-integration goes through the LLM-mediator round-trip documented in docs/PROSE_ROUNDTRIP.md.
• CLI (crates/ail): toolchain for tooling consumers — manifest, describe, deps, check, build, parse, render, prose, merge-prose, etc., preferably with --json for machine consumption.
• Examples (examples/): canonical .ail.json programs. They are specification anchors, not demos — the E2E suite hangs off them.
• Skills (skills/): specialised disciplines (brainstorm, planner, implement, audit, debug, fieldtest) plus the agent rosters they dispatch (skills/<name>/agents/). They form the project's own development methodology and are versioned with the codebase. See skills/README.md.
• Docs (docs/): DESIGN.md (canonical state), docs/journals/ (per-iter decisions log; see INDEX.md), docs/journal-archive.md (archived monolith for pre-2026-05-11 history), roadmap.md (forward queue), specs/ (per-milestone design specs), plans/ (per-iteration implementation plans).
• Tests: unit tests per crate plus E2E in crates/ail/tests/e2e.rs. Every new compiler path needs a test, otherwise the feature does not count as done.

When the language grows, these components grow with it. New tools that strengthen the LLM tooling (e.g. ail diff, IR snapshot diffs, new agents) explicitly belong in the ecosystem inventory of this section and are added here as soon as they are established.

Project language: English

All in-tree content is written in English: source code (identifiers, comments, string literals, CLI help), design documents, the journal, agent prompts, READMEs, commit messages, examples, and CLAUDE.md. The live conversation between user and me stays German for ergonomic reasons; everything that lands in git is English. This keeps diffs and tooling output uniform and matches the audience for AILang (LLM authors), for whom English is the default.

Feature-acceptance criterion

A proposed feature ships only if both hold:

1. An LLM author naturally produces code that uses it. Without prompting toward the feature, the LLM reaches for it as the clean way to express the situation. If the feature is only used when explicitly mentioned, it isn't earning its keep — the LLM is the only author, and what the LLM doesn't reach for naturally is dead surface area.

2. The feature measurably improves correctness or removes redundancy. Either it eliminates a class of bugs structurally (the schema forbids the wrong code), or it lets the LLM express the same logic in fewer sites that have to stay consistent across edits. Aesthetic appeal — "feels elegant", "is idiomatic" — does not count.

This is the positive complement to the CLAUDE.md rule that implementation effort is not a rationale: cost is not a reason for a feature, and neither is human aesthetic preference. The only thing that is, is LLM-author utility.

Two corollaries:

• Human-attractive but LLM-neutral features are cut. Point-free style, operator overloading, implicit conversions, syntactic shortcuts that hide structure. They reward human authors who enjoy compression; they cost the LLM the explicit form it relies on to keep RC, uniqueness, and effects locally legible.

• Human-hostile but LLM-friendly features are kept. JSON as canonical authoring surface; mandatory mode annotations on every fn parameter; mandatory top-level type signatures; explicit clone for shared values. These cost a human author keystrokes; they let the LLM reason locally without spending context window on cross-references.

Empirically: if a feature is proposed and the LLM does not produce it in unprompted code samples, the feature is proposed for the wrong reason. The orchestrator's job is to notice that and cut.

Decision 1: source = data, not text

A module is a JSON object with a fixed schema. There is no parser for free-form text. Typos in identifiers turn into hash-lookup errors that the compiler proposes a fix for directly.

A textual form exists (.ail, S-expression-like), but only as a bidirectional projection of the JSON form. It is intended for human reviews and diffs.

Canonical format: .ail.json with deterministic key order.

Decision 2: content-addressed definitions

Every top-level definition has a hash value (BLAKE3 over canonical JSON without the hash field itself). References between definitions go primarily by name — names are for readability. The hash is the canonical identity.

Advantages:

• Refactoring by adding new defs, not by in-place change. Old versions stay callable until manually removed.
• Caching of typecheck results and codegen per hash.
• Diffs show exactly which def has changed.

Decision 3: pure core language + algebraic effects

The default is total, pure functions. Effects are declared as a set in the function type: (Int) -> Int ![IO]. The effect set is row-polymorphic (![IO | r]). In the MVP only the effects IO and Diverge (for infinite loops) are wired up.

This is the most important LLM property: when I read a function, I can trust its signature without reading the body.

Decision 4: Hindley-Milner + optional refinements

MVP: HM with let-polymorphism. All types are inferable, but at the top level they must always be explicitly annotated (for local reasoning).

Later: refinement annotations that escalate to SMT. (i: Int | i >= 0). They are reserved in the AST from the start, but in the MVP they are simply passed through as opaque strings.

Decision 5: emit LLVM IR as text

Instead of inkwell or llvm-sys: AILang produces .ll files as strings and hands them to clang for linking.

Rationale:

• The LLVM IR text syntax is largely stable across versions.
• No build dependency on a specific libllvm version.
• Generated code is trivially inspectable, which makes debugging much easier.
• An LLM can read the generated IR directly, which is harder with opaque library calls.

Trade-off: no inline optimisations through the LLVM API. We rely on clang -O2 as the standard pipeline.

Decision 6: authoring surface

Form (A) is implemented as the ailang-surface crate (parser + printer). Form-A is gated against drift by ailang-surface/tests/round_trip.rs, which parses every .ailx fixture, prints it back, re-parses, and demands canonical-byte equality. ail render and both branches of ail describe were rewired to use ailang_surface::print, making form (A) the sole text projection of a module — the legacy non-round-tripping pretty-printer code in pretty.rs was deleted at the same time, leaving only diagnostic helpers (type_to_string, pattern_to_string, manifest) public. The rest of this section records the why of Decision 6 for the audit trail; the constraints listed below describe the surface as shipped.

Why this is opening up

Early development authored everything as raw *.ail.json. That worked for 17 fixture files (each ≤ 60 LOC of JSON) but does not scale. Two breaking signals:

1. The token-economy cost of the JSON-AST is massive: a single integer literal 1 is encoded as {"t":"lit","lit":{"kind":"int","value":1}} — ~38 tokens of structural overhead per bit of semantics. For a stdlib in the 200–500 def range this displaces real attention budget.
2. JSON-AST authoring exposes a class of errors (wrong field names, silently-accepted extra fields under #[serde(default)], inconsistent ctor casing) that surface only at load time. The schema is correct-by-construction in storage but error-prone in authoring.

Decision 1 anticipated this: it says a textual form exists "as a bidirectional projection of the JSON form." The pretty-printer already emits S-expression-style text (see crates/ailang-core/src/pretty.rs). What is missing is the inverse direction — text → AST. Decision 6 adds that inverse as one authoring projection alongside the existing pretty-printer; the JSON-AST remains the source of truth.

Architectural pin: data structure is the source of truth

The textual surface is not a replacement for the JSON-AST. It is one projection among potentially many. Concretely:

• The JSON-AST keeps its role as the canonical, hashable, content- addressed representation of a module. All hashing, content- addressing, cross-module references, and typecheck/codegen input flow through the JSON-AST. No new hashable form is introduced.
• The textual surface (form A, this Decision) is the AI authoring projection: optimised for me producing programs token-efficiently and for foreign LLMs producing programs from a spec. It is not optimised for human authors and does not need to be human-pleasant.
• Future projections are explicitly anticipated: a visual / graphical front-end is a plausible second projection for human review and inspection (display being the one case where non-AI eyes matter). The architecture leaves room: any producer of well-formed ailang-core::ast::Module values is a valid front-end.
• No human is expected to author AILang seriously. Authoring is AI work. Display and verification, by contrast, are concerns where human-facing alternatives may be useful — and which can therefore layer their own projections on top of the same AST without touching the surface or the core.

In code terms: ailang-core owns the AST. ailang-surface is one producer/consumer pair: text-form-A → AST → text-form-A. A hypothetical ailang-visual would be a different producer of the same AST. ailang-check and ailang-codegen consume only the AST and remain projection-agnostic.

Constraints (hard, in priority order)

1. Formalizable for a foreign LLM. The grammar must fit in an EBNF/PEG spec of ≤ 30 productions. A model that has never seen AILang must be able to read the spec and produce conforming source zero-shot. Rules out: precedence between binary operators, semantic indentation, maximal-munch lexing, context-sensitive reductions.
2. AST-isomorphic. Every surface form maps to exactly one AST shape. The full bijection between .ail.json and .ailx (both directions, BLAKE3-stable hashing, Float-bits-hex encoding, workspace-CI enforcement points) is anchored as the top-level §"Roundtrip Invariant" — this constraint records that Decision 6's surface-design choice must satisfy that invariant; the invariant itself lives at top level because the property is load-bearing on the language identity, not on this Decision's surface-design rationale.
3. No external symbols. ASCII only. No Greek (∀), no arrows (→), no subscripts. Reasoning: I substitute mojibake for non-ASCII characters under context pressure; foreign LLMs vary in how they tokenize Unicode.
4. No precedence. Either everything is parenthesized, or there are no infix operators. Prefer the latter — add(x, 1) over x + 1. Removes a fail mode for both me and foreign LLMs.
5. No semantic indentation. Block structure expressed by paired delimiters or terminator tokens. Indentation is informational only; the parser ignores it.
6. One construct per token-list. Every AST node corresponds to exactly one parenthesized form (or atom). No "sometimes you can omit the parens" rules.
7. AST surface stays frozen. The surface adapts to the AST, not the other way around. We do not change the JSON schema or invalidate hashes to make the surface prettier.

Candidate notations (same map encoded in each)

The reference target — the polymorphic map from examples/list_map_poly.ail.json:

data List a where Nil | Cons a (List a)
fn map : forall a b. ((a) -> b, List a) -> List b
       = \f xs. match xs of Nil -> Nil
                          | Cons h t -> Cons(f(h), map(f, t))

(A) S-expression with fully-tagged AST nodes

(module list_map_poly

  (data List (vars a)
    (ctor Nil)
    (ctor Cons a (con List a)))

  (fn inc
    (type (fn-type (params (con Int)) (ret (con Int))))
    (params x)
    (body (app + x 1)))

  (fn map
    (type
      (forall (vars a b)
        (fn-type
          (params (fn-type (params a) (ret b)) (con List a))
          (ret (con List b)))))
    (params f xs)
    (body
      (match xs
        (case (pat-ctor Nil) (term-ctor List Nil))
        (case (pat-ctor Cons h t)
          (term-ctor List Cons
            (app f h)
            (app map f t)))))))

Grammar core (3-rule lexical layer + ~25 named-form productions):

sexpr  ::= atom | "(" sexpr* ")"
atom   ::= integer | string | ident
ident  ::= any maximal non-whitespace, non-paren run that is not
           a recognised integer or string literal.

The lexer recognises one delimiter (( / )) and whitespace. Every other maximal token is classified post-hoc:

• All-digit run with optional leading - → integer atom.
• "-delimited run → string atom.
• Otherwise → ident.

Consequence: operators like +, ==, <=, **, qualified names like io/print_int, and cross-module references like std_list.map are all single ident tokens with no special lex rule. The only reserved tokens are (, ), and whitespace. Bool literals (true, false) and unit ((lit-unit)) are disambiguated by parser context, not by lex.

Every AST node form has a unique head keyword (module, data, fn, forall, fn-type, con, var, app, lam, match, case, pat-ctor, term-ctor, do, seq, ...). A bare atom in a positional slot (e.g. inside (con List a) second position) is a name reference whose sort is determined by the parent slot:

• inside (con NAME args...) second-and-later positions → type expression. Bare atom there ⇒ Type::Var { name }.
• inside (app HEAD args...) first position ⇒ Term::Var.
• inside (pat-ctor CTOR fields...) field positions ⇒ Pattern::Var.
• inside (case PAT BODY) second position ⇒ term.

There is no lexical case rule. To construct a value with a ctor, write (term-ctor TypeName CtorName args...). To match against one, write (pat-ctor CtorName fields...). Capitalised identifiers carry no special meaning to the parser. This rules out a class of silent errors ("I forgot to capitalise Cons and it parsed as a function call").

Pros: smallest formal grammar of any candidate (the lexical core is 3 rules; the named-form productions are uniform — every node a tagged list). Foreign-LLM bar lowest. Round-trip with the existing pretty-printer is a refactor of pretty.rs to emit this tagged form, plus a new parser.

Cons: paren density is high. (forall (vars a b) (fn-type ...)) has more visual nesting than the current pretty-printer's forall a. (...) -> .... Verbosity is ~2× JSON for the same node when measured in characters, but ~8× shorter in lines (the existing JSON box.ail.json of 160 lines becomes ~20 lines in this form).

(B) Indented record-style with explicit terminators

module std_list

data List(a):
  Nil
  Cons(a, List(a))
end

fn map:
  type: forall a b. fn(fn(a) -> b, List(a)) -> List(b)
  params: f, xs
  body:
    match xs:
      Nil          => Nil
      Cons(h, t)   => Cons(f(h), map(f, t))
    end
end

Grammar core (~20–30 productions): module-level (def/data/end), type sub-grammar (forall, fn, con, var), term sub-grammar (lam, match, ctor, app, lit, var, seq), pattern sub-grammar.

Pros: higher information density per line, closer to mainstream ML/Haskell shape. Cons: four sub-grammars instead of one. forall a b. fn(...) keeps a pseudo-precedence (-> binds tighter than the outer fn(...) wrapper). Foreign-LLM bar higher.

(C) Pretty-printer-as-source

Use exactly the format pretty::module already emits, plus a parser that accepts it. The existing pretty-printer's quirks (:: for type-of, [params] for fn-params, <a> for type-args, forall a. ..., !IO, () ambiguous between unit-arg-list and empty-form) become the spec.

Pros: zero churn — the existing pretty-printer is already the spec; only the inverse is missing. Round-trip is the identity by construction. Cons: the existing format mixes four mini-dialects (s-expr at term level, ML-shape at type level, square brackets for params, <> for type args). Formalising it crisply is harder than designing a uniform form from scratch.

First choice and rollback plan

Try (A) first. Reasoning: constraint 1 (formalizable) outweighs constraint readability. (A) has a 3-rule core grammar with one lexical disambiguation rule. (B) doubles the rule count and re-introduces a soft form of precedence (-> inside forall). (C) is tempting because it is zero-design but the resulting spec is visibly heterogeneous, which is exactly what constraint 1 was meant to rule out.

If implementing (A) reveals that paren density actively hurts my authoring (measurable: I make more wrong-paren errors than the JSON-AST shape produced before), roll back and try (C). (B) stays on the shelf for a future iter only if both fail.

Implementation outline

• crates/ailang-surface — new crate. Strictly additive. PEG parser produces existing ailang-core::ast types. No new AST nodes, no schema changes, no new hashable form. Pretty-printer for form (A) lives here too (the round-trip is the contract).
• ailang-check, ailang-codegen are not modified. They continue to consume ailang-core::ast::Module values regardless of which projection produced them.
• Round-trip test: for every examples/*.ail.json, parse the corresponding hand-written *.ailx, canonicalise, and assert hash-equivalence to the original. Hash equivalence is the truth check; the surface ships only if every fixture round-trips identically.
• CLI: ail parse <file.ailx> -o <file.ail.json>. Symmetric to existing ail render. .ail.json remains a first-class input to every existing subcommand; the parser is a producer, not a gatekeeper.
• Stdlib (std_list, std_maybe, ...) authored in form (A) from day one because that is the AI authoring projection, not because JSON authoring is forbidden. The resulting .ail.json is what tests and downstream tools see.

What this Decision deliberately does not do

• It does not promote form (A) to the source of truth. Form (A) is one projection; the JSON-AST stays canonical.
• It does not foreclose visual or graphical front-ends. The crate layout (core owns AST; surface/visual/... are siblings) reserves that lane.
• It does not remove .ail.json as input. Every existing CLI subcommand (check, render, describe, emit-ir, build, run, manifest, deps, diff, workspace, builtins) keeps its current .ail.json interface.

Form refinements during implementation

Two productions in the original sketch had to be widened during implementation to round-trip the existing AST faithfully. Captured here for the spec record:

1. lam-term carries types and effects. The AST's Term::Lam stores parallel params, param_tys, ret_ty, and effects fields. The original sketch had only names. The implemented form is

   lam-term ::= "(" "lam" "(" "params" typed-param* ")"
                          "(" "ret" type ")"
                          effects-clause? body-attr ")"
   typed-param ::= "(" "typed" ident type ")"
   

   This keeps the no-precedence / one-construct-per-token-list invariants and adds no new lexical rules.

2. import-clause admits an optional alias. The AST's Import.alias is Option<String>; the original sketch only supported the None case. The implemented form is

   import-clause ::= "(" "import" ident ("as" ident)? ")"
   

   as is a bare ident token in this position; no special lexical rule is needed.

Neither change extends the grammar's rule budget meaningfully: the 30-production ceiling of constraint 1 is intact (the parser implements ~28 named productions). All 17 examples/*.ail.json fixtures round-trip identically through print → parse → canonical JSON; the three hand-written .ailx exhibits parse to canonical JSON identical to their corresponding .ail.json files.

3. Tail-call surface. Decision 8 ships two new productions, both positional analogues of their non-tail counterparts. Their result terms set Term::App.tail = true / Term::Do.tail = true; the typechecker's verify_tail_positions pass enforces that the marker is only used in tail position.

   tail-app-term ::= "(" "tail-app" term term+ ")"
   tail-do-term  ::= "(" "tail-do" ident term* ")"
   

   Production count: ~30, still inside the 30-rule constraint-1 budget. No new lexical rule (tail-app / tail-do are bare ident tokens; no special casing).

Form (B) — human prose projection

AILang ships a second textual projection of the AST: ailang-prose, a one-way projection from Module → human-readable text. It is not an authoring surface; it is the "display" projection that Decision 6's architectural pin (line 167–176) explicitly anticipated:

"Future projections are explicitly anticipated: a visual / graphical front-end is a plausible second projection for human review and inspection (display being the one case where non-AI eyes matter). The architecture leaves room: any producer of well-formed ailang-core::ast::Module values is a valid front-end."

Form (B) targets the specific failure mode where a human reviewer needs to read an AILang module quickly. Form (A) was designed to fit a 30-production EBNF spec and to be parsed zero-shot by foreign LLMs; that prioritisation makes it dense and visually noisy for human readers. Form (B) inverts the trade-offs:

• Rust-flavoured surface. Braces and => for match arms, Rust-aligned 4-level operator precedence, infix arithmetic (a + b, not +(a, b)), unary ! for not.
• Lossy by design. Projection elides machinery the LLM can re-derive: (con T) wrappers ((con Int) → Int), the (fn-type (params ...) (ret ...)) wrap, (term-ctor T C ...) collapses to C(...), redundant parens. Only the AST machinery whose information is recoverable from typecheck context.
• Lossless on load-bearing detail. Mode annotations (own T, borrow T), effects (with IO), explicit clone, reuse-as, doc strings, type annotations on signatures and lambdas, the tail flag — all preserved verbatim.

Critically, form (B) has no parser. Form (A) is round-trippable by construction (Decision 6 constraint 2); form (B) deliberately is not. Re-integrating prose edits requires an external LLM mediator, not a compiler pass — see docs/PROSE_ROUNDTRIP.md for the six-step cycle and the prompt template ail merge-prose composes.

Form (B) does not weaken any Decision 6 invariant:

• The JSON-AST remains the only hashable artefact. Prose is not hashed, not content-addressed, not load-bearing for any cross-module reference.
• Form (A) remains the canonical authoring surface. Foreign LLMs still author against form (A); humans review and edit through form (B).
• The 30-production grammar of form (A) is unchanged.
• ailang-check and ailang-codegen remain projection-agnostic; ailang-prose is a downstream consumer of ailang-core::ast, parallel to ailang-surface but in the rendering direction only.

The CLI gains ail prose <m.ail.json> (the deterministic projection) and ail merge-prose <m.ail.json> <edited.prose.txt> (the mediator-prompt composer); both are listed in the CLI section below.

Form-A spec embedding. An earlier merge-prose prompt instructed the LLM to emit JSON-AST and offered a 12-line schema-essentials reminder; that combination did not give a foreign LLM enough to produce valid output. The current prompt revises this:

• The LLM emits Form-A (the canonical authoring surface), not JSON. JSON-AST stays the only hashable artefact, but it is not a writing surface. The user runs ail parse foo.new.ailx before ail check to produce the canonical JSON.
• crates/ailang-core/specs/form_a.md is the complete LLM-targeted Form-A specification — grammar, every term / pattern / type / def keyword, schema invariants, pitfall catalogue, four few-shot modules drawn from examples/*.ailx. It is exported as ailang_core::FORM_A_SPEC and embedded verbatim in every merge-prose prompt.
• crates/ailang-core/tests/spec_drift.rs walks every variant of Term, Pattern, Type, Def, Literal via exhaustive match and asserts an anchor for each appears in the spec. The exhaustive match is the load-bearing piece: adding a new variant without updating the match fails compilation in this test, before its assertions even run. Hand-written content, mechanical drift detection.

The discussion of richer integration paths (LLM tool-use, MCP server, LSP) was deferred — all three layer additively on the static-prompt path ships, which remains the lowest-common-denominator fallback that always works.

Empirical addendum (2026-05-12)

Cross-model measurement against Qwen/Qwen3-Coder-Next (IONOS), temperature=0, top_p=1, max-turns=5. Two blind cohorts on the same four MVP tasks (t1_add_three, t2_length, t3_main_prints, t4_count_zeros); each cohort sees only its own form's mini-spec. Run directory: experiments/2026-05-12-cross-model-authoring/runs/2026-05-12-df7531/.

metric	JSON cohort	AILX cohort
reached green	1/4	2/4
first-attempt green	0/4	1/4
mean turns-to-green (green only)	2.0	2.0
total prompt tokens	191,768	95,417
total completion tokens	15,376	2,587
top error class	check (×4)	parse (×3)

The AILX cohort reached green on more tasks at roughly half the prompt-token cost and one-sixth the completion-token cost. The only first-attempt success in the entire run was AILX on t3_main_prints (~5 lines of tagged s-expr; the JSON-cohort counterpart was ~23 lines of structured JSON and required a correction turn). JSON-cohort failures clustered in the typecheck-stage; AILX-cohort failures clustered in the parse stage, which is the symmetric front-of-pipeline check for that form.

Scope of this addendum: single subject, single deterministic run. This is a data point, not a verdict. The universal claim of this Decision (".ailx is the AI authoring projection") remains pending a multi-subject expansion — see the roadmap entry "Multi-subject expansion (cross-model authoring-form follow-up)" for the next-step queue.

Decision 8: explicit, verified tail calls

For an LLM author, recursion is the natural iteration form (\n. if n == 0 then () else loop(n-1) is what I reach for, not a for-loop). Without a tail-call guarantee, every recursive program has a silent stack-depth ceiling that no compile-time diagnostic warns about. That is exactly the class of correctness hazard the language exists to eliminate.

Solution: explicit, verified tail calls.

• AST. Term::App { fn, args, tail: bool } and Term::Do { op, args, tail: bool } gain a tail flag, serde-defaulting to false so existing fixtures load with tail: false and their hashes stay bit-identical.
• Typecheck. A new pass verify_tail_positions(fn_body) runs after the main type-check. It walks the body with an is_tail_context: bool threaded down. The flag is true at the start, true for the body of every Term::Match arm, true for the right operand of Term::Seq, true for the body of Term::Let, true for the body of Term::Lam (each Lam opens its own tail scope). The flag is false for: args of any App/Do/Ctor, scrutinee of Match, left of Seq, condition of Let-bound expression. When the walker visits an App { tail: true } or Do { tail: true }, the flag must be true at that visit; otherwise emit diagnostic tail-call-not-in-tail-position.
• Codegen. Emit musttail call (LLVM IR) for marked calls instead of plain call. LLVM rejects at IR-verification time if the call cannot physically be a tail call (calling convention mismatch, signature divergence, etc.). The reject surfaces as a hard build error, not a silent runtime surprise.
• Form (A). Two new keywords: tail-app, tail-do. Productions are positional analogues of app/do with tail: true set on the resulting term. EBNF gains 2 lines; total production count goes from ~28 to ~30, still inside the constraint-1 budget.

What this does NOT promise. Per the tail-call survey of existing fixtures, many existing recursive calls are not in tail position because they are arguments to constructor calls (e.g. Cons (f h) (map f t)). The current pipeline adds annotation + verification; it does not add a CPS transform or accumulator-form rewrite. Programs whose recursion is constructor-blocked will continue to be stack-bounded by recursion depth. The canonical authoring pattern in such cases is to write the accumulator-form variant (map_acc, fold_left, etc.) explicitly. The stdlib ships both forms where relevant.

Migration of existing fixtures is partial: only print_list-style terminal recursions get marked. The constructor-blocked recursions in map, sort, insert remain unmarked — they cannot benefit from musttail without a source-level rewrite.

Decision 9: dual allocator — RC canonical, Boehm parity oracle

AILang ships two allocator backends with an asymmetric role:

• RC is canonical. --alloc=rc is the CLI default for ail build and ail run. The runtime AILang's memory model (RC + uniqueness inference) is designed for. New examples, benches, and corpus tests run under RC unless they explicitly pin GC.
• Boehm stays as a parity oracle. --alloc=gc remains reachable. Its load-bearing job is differential diagnosis: when RC produces a segfault, refcount underflow, or wrong stdout, the GC build of the same module is the cheap "memory bug or logic bug?" probe. The end-to-end suite includes per-example parity tests that run both backends and assert byte-identical stdout — those tests are what make the oracle real.
• --alloc=bump is unchanged: a leak-only bench instrument, not a production target.

Full Boehm retirement (drop libgc, remove the gc backend) reopens when the parity oracle stops paying its keep — concretely, when a few iter families ship without the gc arm catching anything that the rc arm did not already catch. Until then, the cost of keeping libgc as a build dependency is accepted in exchange for diagnostic leverage. Decision 10 (RC + uniqueness) holds as the specification of the canonical runtime; the rest of this section documents the Boehm half, retained as the oracle.

Choice: Boehm-Demers-Weiser conservative GC. The simplest working option:

• Replace malloc(...) with GC_malloc(...) in every IR site (currently lower_ctor's ADT box, lower_lambda's env block and closure pair).
• Replace the IR-level declare ptr @malloc(i64) with declare ptr @GC_malloc(i64).
• Add -lgc to the clang link command (in crates/ail/src/main.rs's Build / Run paths).
• No language-level change. No AST change. No schema change.

Rationale:

• Mature. Boehm has been the default conservative GC for decades. Linux distros ship it as libgc / libgc-dev / gc (Arch).
• No language work. Conservative scan of the C stack handles AILang's stack frames without LLVM stack-map infrastructure (which is its own multi-iter design).
• Single-iter integration. Lift-and-shift of the four allocation sites; all existing tests must still pass with identical output.

Trade-offs accepted:

• Conservative over-retention. A user-supplied Int field whose value happens to coincide with a heap address will pin that allocation. In practice, vanishingly rare for typical values; survivable.
• Pause time non-deterministic. Boehm uses stop-the-world mark-sweep. For LLM-author-written stdlib code at MVP scale, pause times are not the bottleneck.
• Build-time dependency. libgc must be installed on the build host. Users without it get a link-time error from clang, not a silent failure.

Per-fn arena via stack alloca

This optimisation is layered on top of Boehm in its simplest form. ailang-codegen runs an escape-analysis pre-pass over every fn body (and every lifted lambda thunk body); allocations the pass proves do not outlive the fn frame are lowered to LLVM alloca instead of @GC_malloc. Allocations that may escape continue to use @GC_malloc. The Boehm collector is still linked and unchanged; this is purely an optimisation above the floor.

Allocation mechanism: LLVM alloca (not a heap arena). Stack allocation matches the "freed at fn return" lifetime exactly, needs no malloc/free pair, and integrates with LLVM's existing optimiser (mem2reg / SROA may further promote the alloca'd box to registers if the box is small and its uses are simple). No new runtime is introduced; no language-level change; no AST or schema change.

Escape rule (conservative). A Term::Ctor or Term::Lam allocation is non-escaping iff (1) it is the value of a Term::Let { name = X, value = ALLOC, body = B }, and (2) the body B does not let any value derived from X flow past the fn frame. "Derived from" follows two propagation rules:

• A Term::Match whose scrutinee is a Var referring to a tainted name propagates taint to every pattern-bound name in every arm. (Pattern bindings hold field projections of the scrutinee, which live inside the same allocation.)
• A Term::Let { name = Y, value = Var(t), ... } where t is tainted makes Y tainted in the let's body.

A tainted name "escapes" if it appears in any of: the tail position of B, the arg list of any Term::App / Term::Do, the field list of a Term::Ctor, or the free-var capture set of a Term::Lam. The closure-pair-callee position of Term::App where the callee is a bare Var to the tainted name is NOT an escape (calling locally is fine).

What this is not. Not a region-inference system. Not flow-sensitive within an arm. Not field-sensitive (pattern bindings are tainted wholesale). Precision can be improved later; correctness is the priority for this iter. A pessimistic answer (claiming an allocation escapes when it does not) only loses optimisation, never correctness.

Codegen integration. Three sites in ailang-codegen/src/lib.rs:

• lower_ctor — ADT box.
• lower_lambda env block (when there are captures).
• lower_lambda closure pair (always 16 bytes).

Each site queries the per-fn non_escape: BTreeSet<usize> (raw pointer addresses of Term::Ctor / Term::Lam AST nodes flagged as non-escaping). On a hit the emitter writes alloca i8, i64 <size>, align 8; on a miss it writes call ptr @GC_malloc(i64 <size>). The rest of the lowering (tag store, field stores, closure-pair packing) is identical.

The closure-pair and its env share an escape verdict — they have parallel lifetimes. If the closure pair is non-escaping, the env is too.

Decision 10: memory model — RC + Uniqueness with LLM-author annotations

The GC bench (bench/run.sh) showed Boehm contributing a substantial fraction of runtime on allocation-heavy workloads that hold the heap fully live (bench notes in JOURNAL). The mainstream "RC + inference" position is extended with mandatory LLM-author mode annotations (borrow / own), explicit clone, first-class reuse-as, and drop-iterative data attrs.

The cost of GC is structurally in the allocate path — Boehm's GC_malloc is structurally slower than a bump pointer, and the bench workloads exercised allocate cost without collection cost. Tracing GC's irreducible variability cannot be tuned away; RC's costs are bounded and analysable per program point. A corpus committed to one memory model is expensive to switch — pre- stdlib is the cheapest moment to commit.

Choice. AILang's canonical memory model is reference counting with static uniqueness inference and explicit LLM-author annotations on fn signatures, in the lineage of Lean 4 / Roc / Koka. Boehm becomes a transitional allocator (Decision 9) and is retired when the RC pipeline matches the bump-allocator floor within an acceptable margin (target 1.3× on bench/run.sh).

Workload scope of the 1.3× target. The 1.3× target was calibrated on the original bench/run.sh corpus: linear list sum (bench_list_sum) and tree walk (bench_tree_walk) — uniform single-allocation-per-step workloads where one inc/dec pair amortises against one allocation. The corpus was later extended with bench_closure_chain (closure-pair allocation: each step allocates two heap objects, the closure cell and its captured env struct) and bench_hof_pipeline (poly-ADT + indirect dispatch). The closure-chain fixture measures wider than the 1.3× linear/tree target: each step pays two allocs and two decs against one bump-pointer bump, doubling the allocation tax on closure construction (current ratio recorded in JOURNAL bench entries). This is a representational cost of the closure-pair layout, not a defect in the RC implementation; a future slab/pool allocator for fixed-shape pair cells (Decision 9 retirement follow-up) would compress this ratio without changing semantics.

The 1.3× retirement target therefore applies to the linear / tree / poly-ADT subset of the corpus. Closure-heavy workloads are tracked under a wider band (the closure-chain baseline records its rc/bump ratio as the rc_over_bump reference value with ±15% tolerance) and are explicitly excluded from the Boehm-retirement gate until a slab/pool answer ships. Decision-10's RC commitment is unchanged; what is scoped is the quantitative retirement criterion, not the choice of memory model.

The architecture has two layers:

1. Inference. A post-typecheck pass produces a per-node uniqueness side table. Codegen uses it to elide inc/dec wherever provably redundant.
2. LLM-author annotations. Fn signatures carry mandatory (borrow T) / (own T) mode markers. Authors mark sharing-vs-consumption explicitly. The compiler verifies rather than guesses.

The combination plays to what LLMs are good at (writing slightly more annotation per definition) and avoids what compilers are bad at (proving sharing absent in the face of recursion + closures + match).

Why not other memory models

Tracing GC. Open-ended. Boehm's bench overhead is structurally on the allocate path (JOURNAL bench notes); tuning Boehm cannot move that needle. A precise tracing GC would need read/write barriers, root maps, generational machinery — months of work, with irreducible pause-time variability at the end. The user's framing was decisive: "Am GC kannst du ewig rumschrauben (und bekommst trotzdem auch in 100 Jahren keine berechnbare Performance)."

Region inference (Tofte/Talpin / MLton-style). Considered seriously; rejected. Regions tie lifetimes to dynamic scope — every value lives in some region, regions stack on entry/exit, deallocation is bulk-by-region. The reduction: a region is an explicit allocator with sugar plus a static check that nothing escapes its lifetime. The deeper problem: regions assume stack-shaped lifetimes. Real programs have non-stack-shaped lifetimes — caches, memo tables, registries, lookup structures whose lifetimes are not nested. Regions would either force these into a letregion at the top of main (everyone's allocator becomes the root region — useless), or require a region per cache variant (combinatorial). RC has no lifetime-shape assumption; it works for any DAG.

Linear / ownership types as primary mechanism (Rust-style). Considered as a general-purpose alternative; rejected as primary. Rust's borrow-checker is a great mechanism but requires the author to thread lifetimes through every signature and accept that some programs cannot be expressed without unsafe. For an LLM-targeted language with recursion + closures everywhere, that cost is too high — most of the source surface would be lifetime annotations rather than logic. AILang uses a subset of these ideas (mode annotations, linear consumption discipline) selectively, layered on top of RC, where the annotations buy concrete optimisations and never have to thread lifetimes through callees.

The LLM-aware sharpening

A mainstream RC implementation (think Lean 4 in default mode) infers everything from naked AST plus a few optional hints. The inference is conservative; whatever it can't prove unique becomes shared and pays runtime inc/dec. AILang exploits its target audience to push that conservative ceiling higher.

Five mechanisms.

(1) Mandatory (borrow T) / (own T) on fn signatures.

(fn list_length
  (type (fn-type (params (borrow (List Int))) (ret (con Int))))
  ...)

(fn sum_list_consume
  (type (fn-type (params (own (List Int))) (ret (con Int))))
  ...)

(borrow T) declares the parameter is read-only and lives at most until the call returns; the caller still owns it; the callee performs no inc/dec on it. (own T) declares ownership transfer; the callee consumes the value and is responsible for its end-of-life. The declaration is structural (visible in JSON) and binding (the typechecker rejects bodies that contradict it).

For a Lean 4 / Roc author this is all optional and inferred when omitted. AILang makes it mandatory because the LLM author can carry the cognitive cost trivially, and the compiler gains a precise contract at every call site instead of a probabilistic guess.

(2) Linear-by-default consumption with explicit (clone X).

In bodies, every binder is consumed by exactly one own-mode use. If the LLM writes:

(let p (expensive_fn x)
  (let r1 (consume_a p)        ; consume_a takes (own); consumes p
    (let r2 (consume_b p)      ; ERROR: p already consumed
      ...)))

the compiler emits a structured non-linear-use diagnostic with concrete suggested_rewrites:

• "make consume_a borrow": refactor consume_a's signature, no body change at the call site;
• "explicit clone": insert (clone p) at the first use;
• "fuse traversal": replace the two separate calls with a fused fn.

The LLM picks one. There is no implicit clone — sharing always costs visible source.

(3) Reuse hints as first-class.

(fn map_inc
  (type (fn-type (params (own (List Int))) (ret (own (List Int)))))
  (params xs)
  (body
    (match xs
      (case Nil Nil)
      (case (Cons h t)
        (reuse-as xs (term-ctor List Cons (app + h 1) (app map_inc t)))))))

(reuse-as SRC NEW-CTOR) asks the codegen to allocate NEW-CTOR in SRC's memory slot. Compiler verifies: SRC is owned, this is its last use, sizes match. On a hit: no malloc, no free — the box is overwritten in place. On a miss: structured diagnostic explains which precondition failed; the LLM either adjusts the surrounding code or removes the hint.

This matches Lean 4 / Roc reuse analysis but lifts it from "compiler-inferred when possible" to "author-asserted, compiler- verified". The LLM applies it everywhere it expects to fire and lets the compiler bounce the request when it can't.

(4) (drop-iterative) annotation on data declarations.

(data Tree (vars a)
  (ctor Leaf)
  (ctor Node a (Tree a) (Tree a))
  (drop-iterative))

When the refcount of a Tree value reaches zero, the synthesised dec-on-zero traversal is iterative (worklist + heap-allocated stack) instead of recursive. Avoids stack overflow on deep structures. The LLM adds the annotation where appropriate; the compiler refuses to emit recursive dec-cascade on annotated types.

(5) Structured compiler diagnostics with suggested_rewrites.

Every RC-mode error (use-after-consume, mode-mismatch, reuse-as-fail, drop-cascade-too-deep) emits a JSON object containing the failure kind, the source span, and a list of concrete rewrite suggestions in form-A AILang. The LLM consumes these without prose-parsing. This is the missing half of the LLM-as-author story: the language spec defines not only what compiles, but what the compiler tells the author when it doesn't.

Language-design constraints (binding)

The four constraints below are necessary preconditions for RC to be sound and complete without a cycle-collector backstop. They are not new — AILang already satisfies all four — but Decision 10 makes them load-bearing rather than incidental:

1. Strict evaluation. Every Term::App argument is fully evaluated before the call. No laziness, no thunks. (Already true.)
2. No recursive value bindings. (let x EXPR ...) evaluates EXPR in a scope where x is not bound. Recursion is exclusively via Term::LetRec (which binds a fn, not a value) and module-level fn defs. (Already true.)
3. No shared mutable refs. Values are immutable once constructed. There is no ref, IORef, Mutex, or any primitive that allows a value to be mutated from a position outside its allocation. (Already true.)
4. ADTs are acyclic by construction. Strict evaluation + no-recursive-value-bindings + no-shared-mutable-refs together guarantee that any value graph reachable from a binding is a DAG. The reference graph has no cycles. (Follows from 1–3.)

Laziness, recursive value bindings, shared mutable state, or any feature that creates cycles is rejected at design time unless the proposal proves the cycle is collectible by an extension (e.g. linear ownership).

Schema additions

Parameter modes on Type::Fn.

The form-A surface for fn signatures gains mode wrappers:

(fn-type (params (borrow (List Int))) (ret (con Int)))
(fn-type (params (own (List Int))) (ret (own (List Int))))

Internally, this is not a new Type variant. Modes are metadata on Type::Fn — paramModes and retMode fields run parallel to params and ret (see §"Data model" for the JSON schema). The substantive reasons for per-position metadata over a Type::Borrow / Type::Own variant approach:

• Semantic locality. Modes are properties of fn-signature parameter positions, not of types in general. Int does not have a mode; a fn-parameter slot does. Embedding modes in Type would let the schema express forms like (con List (borrow Int)) — syntactically possible, semantically meaningless (you cannot separately own/borrow a list element from the list it lives in). Decision 1 is "schema = data, schema permits exactly what is meaningful"; per-position metadata is the option that holds that line.
• Compositional clarity. A Type value's identity should depend only on the type. Two functions with the same param / ret types but different calling conventions share Type::Fn.params and differ only in param_modes. That is the right factoring: "what data does this carry" is one axis, "how is it transferred" is another. Mixing them under a single hierarchy conflates the two and makes both harder to reason about.
• Future-proof against more position metadata. If later iters add other per-position properties (streaming receiver, captured- by-closure, lifetime witness), they generalise as additional metadata fields on Type::Fn — one consistent hierarchy. The variant approach would force every new dimension into its own Type::* variant (Type::Streamed, Type::Captured, ...) and combinatorics blow up: Type::Borrow(Type::Streamed(T)) versus Type::Streamed(Type::Borrow(T)) raise questions of canonical ordering that don't exist when modes live in a flat metadata vector.

Implicit is the legacy / back-compat state — semantically equivalent to Own but printed bare ((con T), no wrapper). Own and Borrow are explicitly annotated.

JSON canonical hash for every existing fixture stays bit- identical: param_modes is skipped when every entry is Implicit, ret_mode is skipped when Implicit. Existing modules emit the same bytes as before.

Type::Con name scoping (canonical form, since ct.1). Within a .ail.json, a Type::Con.name is interpreted relative to the file's top-level "name" field (the owning module). Bare names (no .) refer to a TypeDef in the owning module's own defs. Cross-module references MUST be qualified <owning_module>.<TypeName> where <owning_module> is a known module in the workspace. Primitives (Int, Bool, Str, Unit, Float) are bare and have no module qualifier. Bare cross-module references are a schema violation (WorkspaceLoadError::BareCrossModuleTypeRef); qualified references whose owner is unknown are also a violation (WorkspaceLoadError::BadCrossModuleTypeRef). The same rule applies to Term::Ctor.type_name.

Class names (ClassDef.name, InstanceDef.class, SuperclassRef.class, Constraint.class) are NOT module-scoped under this rule; they remain workspace-flat with MethodNameCollision enforced at load. Class-name scoping is a future milestone with its own DESIGN amendment.

The legacy (con T) form is treated as (own T) semantically.

(An incidental observation, not a design reason: keeping Type itself unchanged also avoids touching ~250 sites across the typechecker / desugar / codegen that match on Type variants. This is a tiebreaker, not a rationale — the substantive reasons above are what justify the choice.)

New Term variants.

Term::Clone { value: Box<Term> }                     ; `(clone X)` — explicit RC inc
Term::ReuseAs { source: Box<Term>, body: Box<Term> } ; `(reuse-as SRC NEW-CTOR)`

Term::ReuseAs is structured as a wrapper around a body term rather than as a reuse_from: Option<String> modifier on Term::Ctor. Two substantive reasons:

1. Compositional flexibility. Reuse-as is conceptually a wrapper that says "this expression's allocation comes from <source>'s slot". The wrapper form generalises naturally if future iters introduce other allocating constructs (record literals, opaque box wrappers, capability cells) — they all become valid body positions. A modifier on Term::Ctor would have to be replicated on every constructible Term variant the language grows.
2. Source-locality at the head. (reuse-as SRC NEW-CTOR) reads as a single sentence with the source-binder named at the head. The modifier form would scatter the reuse intent across a child position of the constructor's argument syntax, separating the source from the rest of the reuse-as semantics.

The trade-off this accepts: the schema permits Term::ReuseAs { body } where body is not an allocating form (e.g. a literal, a var). Such terms are caught at typecheck via a reuse-as-non-allocating-body diagnostic — structural rejection in the typechecker, not the schema. The principle: prefer composability over schema-level rejection where the typecheck rule is unambiguous.

TypeDef attribute.

TypeDef.drop_iterative: bool                         ; `(drop-iterative)`

All four are skipped during serialisation when absent / false / None so canonical-JSON hashes of every fixture remain stable until the fixture intentionally adopts the feature.

FnDef.suppress. The suppress field on FnDef carries a list of advisory-diagnostic suppress entries; each entry has a code (the diagnostic being suppressed) and a because (a mandatory non-empty reason). See §"Data model" for the canonical schema.

Form-A surface: (suppress (code "...") (because "...")) clause between fn name and (type ...). Multiple clauses allowed; one per entry. Form-B (prose) renders one // @suppress <code>: <because> line per entry above the doc string — lossless, contract metadata.

Skipped from serialisation when empty so existing fixtures keep bit-identical canonical-JSON hashes (regression-pinned by iter19b_empty_suppress_preserves_pre_19b_hashes and iter19b_schema_extension_preserves_pre_19b_hashes). The canonical-form tightening in ct.1 shifted the hashes of two cross-module fixtures (ordering_match.ail.json and test_22b1_dup_a.ail.json); all intra-module fixtures, including the regression-pinned sum.ail.json and list.ail.json, remain bit-identical. The new pins are ct4_migrated_fixtures_have_canonical_form_hashes (locks the post-migration hashes) and ct4_unmigrated_fixtures_remain_bit_identical (re-asserts the existing 13a/19b/22b.1 hashes still hold).

Advisory diagnostics

The advisory-diagnostics arc introduces the language's first advisory typechecker diagnostic and the suppression mechanism that goes with it. Decision 10's mandatory-annotation rule is unchanged: param_modes and ret_mode remain author-required; the typechecker does not infer them. What's new is feedback when an authored annotation is stricter than necessary.

The lint: over-strict-mode. Fires on a fn-param p annotated (own T) when:

1. p's consume_count == 0 (uniqueness pass: the body never consumes p as a whole).
2. For every match arm whose scrutinee is p, no heap-typed pattern-binder has consume_count > 0.

The heap-type filter is load-bearing for soundness: match xs { Cons(h, t) => h } records consume_count(h) == 1, but h: Int is read by-value — no RC traffic, no heap data moved out of xs's allocation. Filtering primitive-typed binders is what lets the lint correctly identify head_or_zero as over-strict (could be borrow) while staying silent on sum_list where t: List is moved out.

Severity: Warning. The first ever Warning-level diagnostic; previously all diagnostics were Error. CLI exit semantics adjusted: ail check, ail build, ail emit-ir exit 1 only on at least one Error. Warnings print but do not abort.

The suppression: mode-strict-because. Authors who want to keep an over-strict annotation deliberately (e.g. RC codegen-test fixtures, fns reserved for planned in-place mutation) attach a Suppress entry naming the diagnostic code and a non-empty reason. The typechecker drops matching diagnostics from the output. Empty because is a hard error (empty-suppress-reason); wrong-code suppresses are silent no-ops (open-set diagnostic registry — a suppress for a code that doesn't fire today may exist defensively for a code that might fire after a future edit).

Why advisory + suppress instead of inference

Three reasons, all anchored in Decision 10's framing:

1. Annotation states intent; inference picks weakest-supporting. These often coincide today but are conceptually different. The annotation captures what the author committed to (e.g. (own T) reserved for a planned mutation that hasn't landed); inference would silently relax it.
2. Annotation is a drift-bremse. Body change that flips the inferred mode produces caller-side breakage at remote sites. Annotation enforces the local-conflict-error pattern instead.
3. Forcing function for LLM authoring. Without mandatory annotation, the LLM never has to commit to ownership intent before writing the body. The advisory lint plus suppress lets us flag accidental over-strictness without weakening the contract.

The suppress mechanism with mandatory-reason mirrors Rust's #[allow(...)]-style escape hatch but sharpens it: the reason is required (not optional), and it becomes part of the contract the next reader sees. CLAUDE.md's "preserve correctness across development cycles" is what this directly serves.

Inference algorithm

Post-typecheck, post-lift_letrecs, pre-codegen pass over the elaborated module. For each Term node that produces or binds a boxed value, the pass computes a uniqueness flag:

• Unique: at this program point, the reference is the only outstanding reference to its referent.
• Shared: there may be multiple outstanding references.

A reference is unique if every path from its allocation to the current program point passes through exactly one binding. The inference is a forward dataflow over the AST. The annotations (borrow / own) provide the inter-fn contract; the inference fills in intra-fn detail.

Codegen contract

Memory layout:

• Every heap allocation has an 8-byte refcount header, followed by the payload. ailang_rc_alloc(size) returns a pointer to the payload; the header is at ptr - 8.
• ailang_rc_inc(ptr): load ptr - 8, +1, store. Non-atomic (single-threaded).
• ailang_rc_dec(ptr): load, -1, store; if zero, recurse-dec child references and free(ptr - 8). For (drop-iterative) types, the recursion is replaced by a worklist loop (via drop-iterative).

Codegen for Term::Ctor / Term::Lam env / closure pair under --alloc=rc calls ailang_rc_alloc(SIZE). The initial RC plumbing stops there — inc/dec instrumentation is added once the inference is wired up. Until then, --alloc=rc deliberately leaks like the pre-Boehm era; this is purely about plumbing.

Mode metadata is load-bearing for codegen

param_modes and ret_mode on Type::Fn are not merely typechecker metadata — codegen consults both to decide where to emit drop calls. They were promoted from "annotation that the typechecker enforces" to "annotation that codegen reads to keep RC correct". Recorded here so the schema metadata's role is explicit:

param_modes — drop-emission gates.

• Iter B: Own-param dec at fn return. When a fn body fall-throughs to a ret (no tail-call), every parameter with param_modes[i] == Own is dec'd before the ret iff its uniqueness consume_count == 0 and the ret value is not the param itself. Borrow and Implicit parameters are skipped: Borrow retains the caller's ownership by contract; Implicit carries no static caller-handed-off-ownership signal (it's the back-compat lane).

• Iter A: arm-close pattern-binder dec. When a match-arm's body terminates without a tail-call, every ptr-typed pattern-bound binder pushed by the arm is dec'd at arm close iff its consume_count == 0 and it is not the arm's tail value, gated on the scrutinee's static ownership. If the scrutinee is a fn-param, only Own-mode scrutinees enable the dec — Borrow and Implicit scrutinees would let the arm dec memory the caller still references.

• Pre-tail-call shallow-dec. When a match-arm's body IS a tail call, both Iter A and Iter B are skipped (the block is terminated). A separate seam in lower_match emits a shallow ailang_rc_dec on the scrutinee outer cell BEFORE the tail call, gated identically on the scrutinee mode plus the requirement that every ptr-typed slot in the active ctor's pattern is in moved_slots[scrutinee].

ret_mode — let-binder trackability.

• Term::App drop at let-scope close. A let-binder whose value is Term::App { callee, .. } is trackable for scope-close drop iff the callee's ret_mode == Own. The signal is the callee's static contract that ownership of the freshly heap-allocated cell flows to the caller. Borrow-returning calls remain non-trackable (the callee retains ownership; the caller holds a view, not an own ref). Implicit-returning calls remain non-trackable (back-compat lane).

The drop fn's symbol resolution for an Own-returning App: synthesise the call's return type, resolve Type::Con { name } to drop_<owner>_<T> (with cross-module qualification through the import map). Falls back to shallow ailang_rc_dec for returns that are not Type::Con (e.g. unresolved type vars on a polymorphic call's pre-monomorphisation site; the monomorphised copies resolve to concrete drop fns).

What this widening does NOT do.

• Does not change the canonical hash. param_modes / ret_mode were already hash-load-bearing when introduced; subsequent work added codegen consumers, not new schema fields.
• Does not introduce a new Type variant. Mode metadata stays flat on Type::Fn (see "Schema additions" above on why).
• Does not cover let-aliases of borrowed values. A let-binder whose value is Term::Var referencing a Borrow-mode param is not yet propagated through; the param-mode gates treat such a binder as "owned" (its current_param_modes lookup misses, default = owned). This is a known carve-out shared by Iter A and the pre-tail-call shallow-dec arm; closing it is a propagation pass through let-bindings that has not shipped yet.

Adjacent extensions for mutability (out of Decision 10's scope)

If future workloads need mutable arrays, hash tables, or other inherently mutable primitives, the answer is not a tracing GC backstop. The answer is a separate ownership/linear extension that gates mutability behind static single-owner discipline. RC

• uniqueness is the universal floor; ownership extends it for specific high-performance primitives without rebreaking the acyclicity invariant.

What this Decision deliberately does not do

• Does not infer everything. An earlier draft of this Decision said "uniqueness is fully inferred" — that was the mainstream RC position. A follow-up discussion replaced it: AILang demands annotations because the LLM author can produce them effortlessly. The compiler does inference plus verification of contracts.
• Does not require existing fixtures to migrate immediately. (con T) is treated as (own T). Existing JSON hashes stay bit-identical until the fixture is intentionally updated.
• Does not commit to atomic refcounts. AILang is currently single-threaded. If/when concurrency arrives, atomic-vs-non- atomic will be a separate decision per allocation kind.
• Does not introduce regions. Regions were considered and rejected; see "Why not other memory models" above.

Decision 11: typeclasses — Haskell-lite, monomorphised, coherent

The design pass for typeclasses. Codified after the Feature-acceptance criterion (this document, above) was committed; the criterion is the explicit basis for the choices below.

AILang ships typeclasses to compress a real LLM-author redundancy: without them, every comparable function must be written per-type (int_eq, string_eq, bool_eq, int_show, string_show, …). With typeclasses behind a monomorphising compiler, the LLM author writes one signature with a class constraint and one method per concrete type, and the compiler emits the same machine code as the per-type version. No runtime cost, no dictionary passing, no vtables.

Choice. A deliberately narrow typeclass design — narrower than Haskell, narrower than Rust traits — calibrated to what an LLM author naturally produces. Five semantic axes are committed:

1. Scope. Multi-method, single-parameter, optional defaults, single-superclass relation. No multi-param classes, no functional dependencies, no associated types.
2. Constraints in signatures. Explicit and mandatory. A function that calls a class method must declare the constraint in its forall block. No constraint inference.
3. Resolution. Orphan-free coherence. An instance C T may be declared only in the module of C or in the module of T. Resolution is global type-directed against a workspace-built registry; coherence makes the lookup unambiguous.
4. Defaults. Opt-in via an explicit default keyword in the class body. Methods without default are abstract-required; methods with default may be overridden or inherited per instance.
5. Class-parameter kind. Concrete types only (kind *). No higher-kinded class params; Functor/Monad-style abstractions over type constructors are not expressible. The LLM-natural pattern is List.map / Tree.map as separate functions per type, which monomorphisation handles directly.

The five axes follow from the Feature-acceptance criterion: each rejected mechanism (multi-param, higher-kinded, FunDeps, assoc types) is one an LLM author does not unprompted produce.

Form-A schema (the JSON authoring surface)

Three additive schema extensions; no existing module becomes invalid.

ClassDef — top-level definition kind, declares a class:

{ "kind": "class",
  "name": "Show",
  "param": "a",
  "superclass": null,
  "methods": [
    { "name": "show",
      "type": { /* full FnSig over `a`, with mode annotations */ },
      "default": null
    }
  ]
}

superclass is either null or { "class": <name>, "type": "a" } where "a" MUST be the same identifier as the class's own param. default is either null (method is abstract-required at instance sites) or an AST body (method is optional with the body as fallback).

InstanceDef — top-level definition kind, declares an instance:

{ "kind": "instance",
  "class": "Show",
  "type": "Int",
  "methods": [
    { "name": "show", "body": <AST> }
  ]
}

type is a concrete type expression, never the class param. methods contains the bodies for every required method plus optional overrides of default-bearing methods.

FnDef.type extension — the existing FnSig schema gains a constraints sibling field next to forall:

"type": {
  "forall": ["a"],
  "constraints": [{ "class": "Show", "type": "a" }],
  "type": ["Fn", [...], "String"]
}

When forall is empty, constraints is also empty (and may be omitted for hash stability of older definitions). Each constraint's type field MUST reference a type variable bound by the surrounding forall.

Method invocation sites do NOT get a new node. A call to a class method is a normal Call node; resolution against a class is the typechecker's job, not the parser's. The LLM author writes show x exactly as for any free function.

Resolution and monomorphisation

Constraint collection (per function body). During typechecking of a body, each method call generates a residual constraint of shape <Class> <Type> where <Type> may still contain type variables. After local typechecking, residual constraints are checked against the function's declared constraints (modulo α-conversion and modulo auto-expansion through superclasses; see below). Any residual not covered by declared constraints fires MissingConstraint.

Instance registry (workspace-global). At workspace load (see crates/ailang-core/src/workspace.rs), all InstanceDef nodes across all reachable modules are collected into a registry keyed by (class-name, canonical-hash-of-instance-type). Registry build performs three checks:

• Coherence. Each instance's module must be either the class's defining module or the instance type's defining module. Otherwise → OrphanInstance.
• Uniqueness. No two entries share a key. Otherwise → DuplicateInstance.
• Method completeness. Each instance specifies every required (non-default) method of its class. Otherwise → MissingMethod.

Registry build is a one-time-per-build pass that fires before any typechecking. Its errors are workspace-load errors, not per-call-site errors.

Resolution at call sites with concrete types. When the typechecker sees a method call where every type variable in the constraint is substituted to a concrete type, it queries the registry. Hit → resolved. Miss → NoInstance.

Resolution at polymorphic call sites. When type variables are still free, the constraint propagates into the surrounding function's constraint context — which the user MUST have declared explicitly (per axis 2). No constraint is implicitly hoisted.

Monomorphisation (post-typecheck, pre-codegen). A pass between typechecking and codegen replaces every call to a Type::Forall-quantified Def::Fn with a call to a synthesised monomorphic FnDef. Two source-body entry points share the same mechanics in one fixpoint:

1. Class-method entry. For each unique (method, concrete-type) pair produced by a class-constraint residual, the pass looks up the resolved instance body via Registry::entries[(class, type-hash)], substitutes the class parameter to the concrete type, and synthesises a top-level FnDef named <method>__<type-surface-name>.
2. Free-fn entry. For each call site to a polymorphic free Def::Fn with a fully-concrete substitution, the pass takes the source body directly from the polymorphic Def::Fn, applies rigid-var substitution on both the type AND the body (the body may contain inner Term::Lams whose param_tys reference the outer Forall vars), and synthesises a top-level FnDef named <name>__<type-surface-name-1>__<type-surface-name-2>__… (concatenated in Type::Forall.vars declaration order; the N-ary case extends the single-type-var class-method shape bit-stably).

Both arms share:

• A fixpoint loop that keeps collecting targets until a round adds nothing new (a synthesised free-fn body may invoke class methods at concrete types, scheduling new class-method targets; a class-method body may invoke polymorphic free fns at concrete types, scheduling new free-fn targets).
• A dedup cache keyed by (kind, base-name, type-hash-or-joined-hashes) where the first component ("class" / "free") guarantees disjoint keying across the two kinds.
• A call-site rewrite walker that rewrites bare polymorphic call sites — class-method-named OR poly-free-fn-named — to their mono symbols before codegen runs. The walker advances a single cursor over interleaved class-method and free-fn slots emitted in synth's traversal order.

After this pass, the IR contains no polymorphism, no class machinery, no polymorphic call sites — only ordinary monomorphic functions and direct calls. Codegen sees no difference between a hand-written show_int and a synthesised show__Int.

Why mono, not virtual dispatch. Monomorphisation makes the call target visible to the optimiser, unlocking inlining and downstream loop transformations that virtual dispatch prevents in principle. On a saturating branch predictor with a monomorphic indirect target, the indirect call itself is comparable in cost to a non-inlined direct call — the win is in what the optimiser can do with the visible target, not in the call instruction. The end-to-end gain shrinks toward zero on larger callee bodies and cold call sites, but the architectural claim — "mono enables optimisations vdisp forbids" — holds across the spectrum (bench/mono_dispatch.py and the corresponding JOURNAL bench-notes entry record the measured ratios).

The separator is __ rather than # or @ because # and @ are invalid in LLVM IR global identifiers (the IR verifier rejects them inside @ail_<module>_<def> mangled names). __ is legal in both LLVM IR and the C ABI used by the runtime glue, and parses unambiguously into <method>__<type-surface-name> because neither component contains __ by project convention.

No runtime dispatch, no dictionary passing. The monomorphisation pass is the ONLY specialiser. Codegen sees only monomorphic Def::Fns and direct calls; the pre-iter-23.4 codegen-time specialiser (lower_polymorphic_call + module_polymorphic_fns + mono_queue) was removed in iter 23.4. A call that cannot be monomorphised — for instance, because a constraint remains unresolved at the entry point — is a static error, not a runtime one. This is the LLVM-friendly form and is consistent with Decision 10's performance commitment.

Defaults and superclasses

Defaults. A ClassDef.methods[i].default is either null (the method is required at instance level) or an AST body. When an InstanceDef does not specify a method that has a default, the typechecker uses the default body for that instance, with the class param substituted to the instance type. Default bodies may call other methods of the same class (the canonical example is default ne x y = not (eq x y)); the called methods resolve at monomorphisation time against the same instance.

Superclasses. A ClassDef.superclass of { "class": "Eq", "type": "a" } declares that any instance C T requires a corresponding instance Eq T to exist. The check fires at workspace-load time: for each InstanceDef whose class declares a superclass, the registry is queried for the matching superclass instance. Missing → MissingSuperclassInstance.

Superclass auto-expansion in constraint contexts. When a function declares Ord a as a constraint, the typechecker treats the constraint context as { Ord a, Eq a } for the purposes of MissingConstraint checking. This is the one deviation from "alles sichtbar": the extension is anchored in the class's own superclass field, so the relation is schema-visible at the class declaration even when the constraint at the function site reads Ord a alone.

Superclass chains are linear (single-superclass relation per class) and auto-expansion closes transitively across the chain.

No deriving. AILang does not auto-derive instances. instance Eq MyType must be written by hand. Auto-derivation may land in a future iteration if the Feature-acceptance criterion holds for it; that discussion is out of scope here.

Diagnostic categories

The typeclass layer introduces three families of diagnostics. Exact wording is fixed; the categories and their triggers are:

Workspace-load (registry-build) diagnostics:

• OrphanInstance — instance is not in class's or type's module.
• DuplicateInstance — two instances match the same key.
• MissingSuperclassInstance — superclass instance absent.
• MissingMethod — instance omits a required method.
• OverridingNonExistentMethod — instance specifies a method not in the class.
• MethodNameCollision — same method name across two in-scope classes, or between a class method and a top-level function.

Class-schema diagnostics (validation of class declarations):

• KindMismatch — class param appears in applied position in any method signature (e.g., f a where f is the class param).
• InvalidSuperclassParam — superclass type differs from the class's own param.
• ConstraintReferencesUnboundTypeVar — a constraint mentions a type variable not bound by the surrounding forall.

Typecheck diagnostics (per function body):

• MissingConstraint — body's residual constraint is not covered by declared (and superclass-expanded) constraints.
• NoInstance — fully concrete constraint has no registry entry.

There is no AmbiguousInstance diagnostic. Coherence (W2) makes every (class, type) key globally unique; resolution is therefore deterministic by construction.

What the typeclass design explicitly does NOT support

Each of the following is rejected by either schema (parser cannot express the construct) or by an enumerated diagnostic. The combination of axis-1 (single-param, no FunDeps, no assoc) and axis-5 (kind * only) covers the space.

• Multi-parameter classes (class Foo a b where ...). Schema rejects: ClassDef.param is a string, not a list.
• Higher-kinded class params (class Functor f). Diagnostic S1 (KindMismatch) at class declaration if any method applies the param.
• Higher-rank polymorphism (forall a. (forall b. b -> b) -> a). Already rejected at parse time per the typeclass-conversation rationale recorded in JOURNAL; constraint-bearing signatures inherit that prohibition.
• Existential / dyn dispatch (exists a. Show a => a, heterogeneous lists like [Show]). Schema does not express existentials. The LLM-natural alternative is sum types.
• Associated types (class Container c where type Element c). Schema does not express type-level methods.
• Functional dependencies (class Convert a b | a -> b). Required only for multi-param classes; entails by axis 1.
• deriving / auto-derivation (data Foo = ... deriving (Eq)). Future iteration, gated separately on Feature-acceptance.
• Numeric literal defaulting. 1 does not auto-resolve to Int under an ambiguous Num a constraint. Bare polymorphic literals with multiple satisfying instances fire NoInstance with a hint to annotate. The LLM-author writes 1 : Int or 1 : Float.
• Orphan-with-warning mode. Coherence is hard. OrphanInstance is always an error, never a warning. The corollary --allow-orphans flag is not provided.

Prelude (built-in) classes

Milestone 22 ships no built-in Prelude classes. An earlier draft committed to a fixed Prelude (Show / Eq / Ord on the primitives), but the implementation work to wire int_to_str as a heap-allocated-string runtime primitive proved substantively separable from the typeclass machinery itself, and the LLM-utility case for primitive Show is weak (LLM-natural form is int_to_str x, not show x through a single-instance class). The user-class end-to-end path is the milestone's typeclass acceptance gate (the user-defined-class fixture: class Foo a + data IntBox + instance Foo IntBox); see docs/specs/2026-05-09-22-typeclasses.md "Amendments" for the substantive rationale.

Milestone 23 amends the above: the prelude now ships the Ordering ADT, the Eq and Ord classes, primitive Eq Int/Bool/Str and Ord Int/Bool/Str instances, and the five polymorphic free-fn helpers ne / lt / le / gt / ge. Show, operator routing through Eq/Ord, and print-rewire remain out of scope per their original substantive reasons (heap-Str ABI, bench rebaseline). Float has neither Eq nor Ord instance per §"Float semantics"; a polymorphic helper invoked at Float fires NoInstance at typecheck with a Float-aware diagnostic cross-referencing this section.

Primitive output goes through io/print_int / io/print_bool / io/print_str directly.

==, <, <=, >, >= REMAIN primitive operators (unchanged from the original draft). Class methods are accessed by name (eq x y, lt x y, …), not via these operators. Routing operators through classes is deliberately deferred — it would require migrating every existing fixture and would re-baseline the bench corpus, which is a new-baseline decision rather than an iter detail.

Num is NOT in milestone 22. Arithmetic operators (+, -, *, /) stay primitive and per-type. Class-based numeric overloading would invoke literal-defaulting which axis-7 already excluded.

What this decision does NOT commit to

• Operator routing. ==, <, etc. stay primitive in milestone 22. Class-routing operators is a future-iteration option, not a Decision-11 commitment.
• Form-B (prose) projection of class/instance. The prose renderer for the new schema nodes is one-way (no parser by design) and was deferred from milestone 22 entirely; the placeholder render in crates/ailang-prose/src/lib.rs is informational only. A future iter ships the full prose projection if/when prose-side authoring needs surface.
• Specific monomorphised-symbol naming format. Milestone 22 uses <method>__<type-surface-name> for primitive type targets (e.g. show__Int) and <method>__<8-hex-prefix> for compound types (where <8-hex-prefix> is the BLAKE3 hash of the canonical type bytes). The __ separator was chosen over # and @ for LLVM IR identifier legality.
• Mode annotations on class methods. Class method signatures ARE full FnSigs and DO carry mode annotations per Decision 10; the convention is borrow for read-only methods. User-defined classes pick modes per method.
• Number of Prelude classes. Milestone 22 ships zero (no Prelude). A future Prelude milestone gates class additions on the Feature-acceptance criterion at the time of proposal.

Roundtrip Invariant

Every well-formed AILang module has both a canonical .ail.json representation and a textual .ailx representation, and the two are exact projections of the same AST. Concretely, both directions of the bijection hold:

1. JSON → text → JSON. For every valid .ail.json module J, parse(print(load_module(J))) is canonical-byte-equal to canonical::to_bytes(load_module(J)). The textual surface is the inverse of the loader composed with the printer; no information is lost across the round-trip.
2. text → JSON → text. For every well-formed .ailx text t, parse(t) is a complete AST, and re-printing print(parse(t)) produces a text that re-parses to the same AST (modulo formatting). The parser is total over the well-formed surface and the printer is deterministic.

Hashing is the consequence the language depends on: BLAKE3 of the canonical bytes is identical for the two paths, so content- addressing is form-agnostic. An LLM author can pick either form without changing the identity of the module it produces.

Float literals are inside the invariant

Float literals carry an IEEE-754 bit pattern, not a decimal approximation. The canonical encoding is {"kind":"float","bits":"<16-lowercase-hex>"} and the surface emits the same bits-hex string. NaN, ±Inf, signed zero, and subnormals all round-trip exactly because the JSON-number path is bypassed (see §"Float semantics", "Form-A serialisation"). Floats are not an exception to the invariant — the bits-hex encoding is the mechanism that keeps them inside it.

Enforcement

The invariant is workspace-CI-enforced by five tests, each operating on the examples/ corpus via dynamic read_dir collection (no hardcoded fixture list, so newly added fixtures inherit the gate automatically):

• crates/ailang-surface/tests/round_trip.rs::print_then_parse_round_trips_every_fixture — for every .ail.json, print then parse produces canonical- byte-equal output. Direction 1 above.
• crates/ailang-surface/tests/round_trip.rs::parse_then_print_then_parse_is_idempotent_on_every_ailx_fixture — for every .ailx text t, parse(t) and parse(print(parse(t))) produce canonical-byte-equal AST. Direction 2 above.
• crates/ailang-surface/tests/round_trip.rs::every_ailx_fixture_matches_its_json_counterpart — for every .ailx with a same-stem .ail.json counterpart, parse of the text yields canonical bytes equal to the JSON counterpart. Pins the hand-authored .ailx corpus against drift between the two forms at the fixture level.
• crates/ailang-core/tests/schema_coverage.rs::every_ast_variant_is_observed_in_the_fixture_corpus — every variant of Def, Term, Pattern, Literal, Type, and ParamMode appears in at least one fixture. The visitor matches each AST enum exhaustively without wildcard arms, so any new AST variant fails to compile until the visitor is extended in lockstep — the coverage table cannot silently fall behind the schema.
• crates/ail/tests/roundtrip_cli.rs::cli_render_then_parse_preserves_canonical_bytes_on_every_fixture — for every .ail.json, the public CLI pipeline ail render → tempfile → ail parse reproduces BLAKE3-identical canonical bytes. Pins the user-facing CLI against drift that crate- internal tests cannot see.

A new fixture or a new AST variant that violates the invariant fails one of these tests; the fix is in render or parse code, never by relaxing the test.

Why this is anchored at top level

The invariant is a property of the language identity, not of any one surface-design Decision. Decision 6 introduces the .ailx surface and lists round-trip-as-property as one of its constraints, but the property is load-bearing for every downstream concern that treats the two forms as exchangeable: content-addressed hashing, the LLM-author's choice of authoring form, the integrity of fixture cross-references, and the prerequisite for empirical cross-model authoring-form studies. Lifting it out of Decision 6 makes the property quotable on its own and reviewable by the architect agent at every milestone close.

Mangling scheme

All AILang functions are mangled to @ail_<module>_<def> — even in the single-module case. Constants likewise (@ail_<module>_<const>). Global string literals carry a short hint for readability: @.str_<module>_<hint>_<idx> (e.g. @.str_sum_fmt_int_0). The entry point is a define i32 @main() trampoline (C / LLVM ABI) that calls @ail_<entry-module>_main(). source_filename exists exactly once per workspace and carries the entry-module name (<entry-module>.ail).

Convention: qualified cross-module references

Cross-module calls use no new AST node. Instead, a Term::Var { name } with exactly one dot in the name is a qualified reference: <prefix>.<def>.

• <prefix> is an import alias (import { module: "X", as: "<prefix>" }) or, when imported without an alias, the module name itself.
• <def> is the name of a top-level definition in the target module.
• Def names MUST NOT contain a dot — the typechecker reports invalid-def-name with ctx: { "reason": "contains-dot" }.
• The workspace loader finds all reachable modules; the typechecker (check_workspace) resolves dotted names through the import map. Diagnostic codes: unknown-module (prefix not imported), unknown-import (module found, def not).

Hash stability: no new AST node, no renamed fields — all previous module hashes stay bit-identical.

Data model

The on-disk JSON-AST is what the toolchain hashes, typechecks, and lowers. This section is the canonical schema. The Rust types in crates/ailang-core/src/ast.rs are the in-memory projection of it; when the two disagree, this section wins, and the drift test crates/ailang-core/tests/design_schema_drift.rs fires. Every additive field is declared with skip_serializing_if so pre-existing fixtures keep bit-identical canonical-JSON hashes — that gating contract is what makes growing the schema cheap.

Module

{
  "schema": "ailang/v0",
  "name": "<id>",
  "imports": [{ "module": "<id>", "as": "<id>" }],
  "defs": [Def...]
}

Def

kind ∈ { "fn", "const", "type", "class", "instance" }. All five are real surface forms.

// fn (the unit that gets a content hash)
{ "kind": "fn",
  "name": "<id>",
  "type": Type,            // typically Type::Fn, optionally wrapped in Forall
  "params": ["<id>"...],   // names bound in body, in type.params order
  "body": Term,
  "doc": "<optional string>",
  "suppress": [Suppress...] // omitted when empty
}

// const (top-level value; codegen emits as a global; body must be pure)
{ "kind": "const",
  "name": "<id>",
  "type": Type,
  "value": Term,
  "doc": "<optional string>"
}

// type (algebraic data type; parameterised)
{ "kind": "type",
  "name": "<id>",
  "vars": ["<id>"...],     // type parameters; omitted when empty (hash-stable when omitted)
  "ctors": [
    { "name": "<id>", "fields": [Type...] }  // nullary ctor: fields = []
    ...
  ],
  "doc": "<optional string>",
  "drop-iterative": true   // opt-in; omitted when false (hash-stable when omitted)
}

// class (typeclass declaration; see Decision 11)
{ "kind": "class",
  "name": "<id>",          // class name (e.g. "Show")
  "param": "<id>",          // single class parameter, kind *
  "superclass": null,       // or { "class": "<id>", "type": "<param>" }
  "methods": [
    { "name": "<id>",
      "type": Type,         // FnSig over the class param
      "default": Term       // optional fallback body; null = abstract-required
    }
    ...
  ],
  "doc": "<optional string>"
}

// instance (typeclass instance; see Decision 11)
{ "kind": "instance",
  "class": "<id>",          // class being instantiated
  "type": Type,             // concrete type expression (never the class param)
  "methods": [
    { "name": "<id>", "body": Term }
    ...
  ],
  "doc": "<optional string>"
}

Suppress (entry in FnDef.suppress):

{ "code": "<diagnostic-code>",   // e.g. "over-strict-mode"
  "because": "<author reason>"   // must be non-empty;
                                 // empty/whitespace fires `empty-suppress-reason` (Error)
}

Term (expression)

{ "t": "lit", "lit": Literal }
{ "t": "var", "name": "<id>" }

// fn application; tail flag triggers musttail under codegen.
// `tail` is omitted when false (hash-stable when omitted).
{ "t": "app",  "fn": Term, "args": [Term...], "tail": false }

{ "t": "let",  "name": "<id>", "value": Term, "body": Term }

// Local recursive let. Always fn-shaped. The desugar pass
// lifts most `letrec` to a synthetic top-level fn; `lift_letrecs`
// finishes the job after typecheck for the residue that captures
// let-bound names. Post-codegen, no `letrec` survives.
{ "t": "letrec",
  "name": "<id>", "type": Type, "params": ["<id>"...],
  "body": Term, "in": Term }

{ "t": "if",   "cond": Term, "then": Term, "else": Term }

// Effect-op invocation. `op` is "<eff>/<op>" (e.g. "io/print_int").
// `tail` triggers musttail (omitted when false).
{ "t": "do",   "op": "<eff>/<op>", "args": [Term...], "tail": false }

// Ctor application; `args` omitted when empty.
{ "t": "ctor", "type": "<id>", "ctor": "<id>", "args": [Term...] }

{ "t": "match", "scrutinee": Term, "arms": [Arm...] }

// Anonymous fn value; free vars captured from enclosing scope.
{ "t": "lam",
  "params": ["<id>"...],
  "paramTypes": [Type...],
  "retType": Type,
  "effects": ["<id>"...],
  "body": Term }

// Sequencing. Semantically `let _ = lhs in rhs`; lhs must be Unit.
{ "t": "seq",   "lhs": Term, "rhs": Term }

// Explicit RC clone. Codegen lowers as
// `call void @ailang_rc_inc(ptr %v)` before returning %v under `--alloc=rc`.
{ "t": "clone", "value": Term }

// Explicit reuse-as hint. `body` must be allocating
// (typically `ctor` or `lam`); `source` must be a bare `var`. Codegen
// lowers as in-place rewrite under `--alloc=rc`.
{ "t": "reuse-as", "source": Term, "body": Term }

In the MVP, do is only a direct call to a built-in effect op (no handler). A lam term constructs an anonymous function value; free variables of its body are captured from the enclosing scope.

Literal:

{ "kind": "int",  "value": <i64> }
{ "kind": "bool", "value": <bool> }
{ "kind": "str",  "value": "<utf-8>" }
{ "kind": "unit" }
{ "kind": "float", "bits": "<16-lowercase-hex>" }

Pattern (the pat field of an Arm; discriminator p):

{ "p": "wild" }                                   // _
{ "p": "var",  "name": "<id>" }                   // x — binds the value
{ "p": "lit",  "lit": Literal }
{ "p": "ctor", "ctor": "<id>", "fields": [Pattern...] } // fields omitted when empty

Patterns are linear: each pattern variable may appear at most once.

Type

// Type-constructor application. `args` omitted when empty
// (hash-stable when omitted, for non-parameterised cases like Int, Bool, ...).
{ "k": "con", "name": "<id>", "args": [Type...] }

// Function type. Decision 10 added paramModes/retMode as
// metadata on Type::Fn — they are NOT separate Type variants, so every
// existing match-arm in the typechecker (unify, occurs, apply) keeps
// working. `paramModes` omitted when every entry is "implicit";
// `retMode` omitted when "implicit" (hash-stable when omitted).
{ "k": "fn",
  "params":     [Type...],
  "paramModes": [ParamMode...],
  "ret":        Type,
  "retMode":    ParamMode,
  "effects":    ["<id>"...] }

{ "k": "var", "name": "<id>" }

// Top-level polymorphism only. `constraints` carries class
// constraints (Decision 11); omitted when empty (hash-stable when omitted).
{ "k": "forall",
  "vars": ["<id>"...],
  "constraints": [{ "class": "<id>", "type": "<id>" }, ...],
  "body": Type }

ParamMode (Decision 10):

"implicit"   — unannotated / back-compat. Treated as `own` by the typechecker.
"own"        — (own T) — caller transfers ownership; callee consumes.
"borrow"     — (borrow T) — caller retains ownership; callee may not consume.

implicit ≡ own semantically; the distinction exists so existing unannotated fixtures continue to serialize without the mode wrapper and keep their canonical-JSON hash.

Pipeline

.ail.json  ─┐
            ├─ load + validate schema
            ├─ resolve names + assign hashes
            ├─ desugar (AST → AST)
            ├─ typecheck (HM, effect rows; mode-strict per Decision 10)
            ├─ lift_letrecs (post-typecheck AST → AST)
            ├─ lower to MIR (SSA-like, named SSA values)
            ├─ emit LLVM IR (.ll)
            └─ clang -O2 *.ll -o binary
                  --alloc=rc → emits inc/dec     (@ailang_rc_inc / _dec; canonical, default)
                  --alloc=gc → links libgc       (@GC_malloc; parity oracle)

Two allocator backends share the same MIR. --alloc=rc is the canonical backend committed to in Decision 10 and the CLI default. The typechecker enforces (own) / (borrow) modes, codegen emits ailang_rc_inc / _dec calls at the points dictated by linearity, and Term::Clone / Term::ReuseAs materialise into actual rc-bumps and in-place rewrites respectively. Boehm-on---alloc=gc is on the path to retirement; see docs/roadmap.md for the active queue.

The desugar pass (ailang-core::desugar::desugar_module) runs before typecheck and codegen in every entry point of ailang-check and ailang-codegen. It is a pure AST → AST rewriter — currently only flattens nested constructor patterns, but is the chosen home for any future surface-smoothing rewrites that should not bloat the core AST or the backends. Critical invariant: CheckedModule.symbols in the check entry point continues to hash from the original on-disk module, not the desugared one, so ail diff and ail manifest report identities that match the canonical JSON the user is editing.

The lift_letrecs pass (ailang-check::lift_letrecs) runs after typecheck and before codegen, but only on the build / run paths — the check subcommand stops at typecheck and never sees a lifted module. It eliminates every Term::LetRec that the desugar pass left in place (the case where at least one capture is Term::Let-bound, so its type is only knowable after inference). The output is a module with synthetic <hint>$lr_N top-level fns appended, ready for codegen. Synthetic FnDefs added by this pass do not appear in CheckedModule.symbols — same invariant as the desugar-pass lifts.

CLI

ail check     <module.ail.json>          — loads, validates, typechecks
ail manifest  <module.ail.json>          — table: name :: type !effects [hash]
ail describe  <module> <name>            — detail of a definition (form-A body)
ail render    <module.ail.json>          — JSON-AST  → form-A text  (exact inverse of `parse`)
ail parse     <module.ailx>              — form-A text → canonical JSON-AST
ail prose     <module.ail.json>          — JSON-AST  → form-B (lossy human prose, no parser)
ail merge-prose <m.ail.json> <m.prose.txt>
                                         — compose the LLM-mediator prompt for the prose round-trip
                                           (see docs/PROSE_ROUNDTRIP.md)
ail deps      <module.ail.json>          — list cross-module references
ail diff      <a.ail.json> <b.ail.json>  — content-addressed def-level diff
ail workspace <entry.ail.json>           — list all modules transitively reachable from entry
                                           (`--json` for machine output;
                                           `manifest --workspace` and `diff --workspace`
                                           extend single-module subcommands to workspaces)
ail builtins                             — list built-in fns and effect ops
ail emit-ir   <module>                   — writes .ll
ail build     <module>                   — full pipeline → binary
ail run       <module>                   — build + execute (tempdir), passthrough exit code

Verification and correctness (across cycles)

1. Snapshot tests for the pretty-printer and IR emit. The diff makes regressions visible immediately.
2. Property tests for the JSON ↔ pretty-print roundtrip.
3. End-to-end tests for examples/ with expected program output.
4. Hash stability: a test ensures the same def always produces the same hash.
5. CI pin of the outputs in tests/expected/.
6. Rustdoc cleanliness: cargo doc --no-deps runs warning-free. Fixing a rustdoc warning is part of the iteration that introduced it, not a follow-up.

Float semantics

Float is IEEE-754 binary64 (LLVM double). One float type ships; no f32 variant. The runtime / codegen contract:

Guaranteed:

• Every individual builtin (+/-/*///neg/</==/...) lowers to a single LLVM IR instruction on the Float arm: fadd/fsub/fmul/fdiv double, fneg double, fcmp olt/ole/ogt/oge double, fcmp oeq double, fcmp une double (for !=). On a fixed (target triple, LLVM version) pair, the bit pattern of the result of any single op is reproducible.
• NaN and ±Inf propagate per IEEE 754 — no silent collapse to zero, no trap. Arithmetic on a NaN operand produces NaN; division by zero produces ±Inf; 0.0 / 0.0 produces NaN.
• -0.0 and +0.0 are distinct bit patterns at the canonical-JSON hash level ({"bits":"0000000000000000",...} vs {"bits":"8000000000000000",...} — distinct def_hashs) but compare equal via == per IEEE (fcmp oeq double returns true for +0 == -0). This asymmetry is the correct IEEE behaviour; it does mean def_hash-equality is finer than ==-equality on Float.
• == returns false whenever either operand is NaN (fcmp oeq is the ordered-equal predicate; ordered = both operands non-NaN).
• != returns true whenever either operand is NaN (fcmp une is the unordered-or-not-equal predicate). This matches Rust f64::ne and IEEE-!= exactly.
• is_nan (fcmp uno double %x, %x) returns true iff x is NaN. Bit-pattern-based NaN detection without dependence on the payload bits.
• int_to_float (sitofp) is exact for |n| < 2^53, round-to-nearest-even otherwise.
• float_to_int_truncate (@llvm.fptosi.sat.i64.f64) is total: NaN → 0, +Inf → i64::MAX, -Inf → i64::MIN, finite-out-of-range saturates, finite-in-range truncates toward zero. Matches Rust as i64 semantics (since 1.45).

Unspecified:

• FMA contraction. LLVM may fold fadd (fmul a b) c into fma a b c. Bit results may differ between an op-emitted-in- isolation pattern and an op-folded-into-FMA pattern.
• Reassociation. The compiler may reorder a chain like (a + b) + c into a + (b + c), producing a bit-different result on numerically sensitive inputs.
• Subnormal flushing modes. If the target enables FTZ (flush-to- zero) or DAZ (denormals-are-zero), subnormal results round to zero; AILang does not enable these flags but does not forbid the target from doing so.
• The exact NaN bit pattern produced by an op. Any quiet NaN bit pattern is conformant; 0.0 / 0.0 may produce 0x7ff8000000000000 on one target and a different qNaN on another.
• The textual rendering of NaN by io/print_float. The libc printf("%g", nan) glue used by the runtime is permitted to emit nan / -nan / NaN etc. depending on libc version and the NaN's sign bit; AILang does not normalise this, since the prose / surface-print paths render NaN as the explicit "NaN" spelling and io/print_float is for human-readable output, not round-trip.

These are the Rust / Swift / standard-LLVM defaults — not research-grade reproducibility guarantees. The stronger guarantee (e.g. Pythonic float.fromhex-level bit reproducibility across ops) would require -ffp-contract=off plus per-op intrinsic selection — out of scope for the milestone; revisit only if a real use case appears.

Form-A serialisation: Float literals carry the IEEE-754 bit pattern as a 16-character lowercase hex string in the canonical JSON: {"kind":"float","bits":"<16-hex>"}. Routing through the JSON string path (not serde_json::Number) preserves bit stability across serde_json versions and lets NaN / ±Inf round-trip through Form-A — JSON numbers cannot represent them.

Pattern matching: Pattern::Lit on Literal::Float is hard- rejected at typecheck (CheckError::FloatPatternNotAllowed). IEEE semantics make Float patterns semantically dubious — NaN never matches via IEEE-==, and bit-exact equality is rarely what an LLM-author wants. Use ordering operators (<, >, ...) and is_nan to discriminate Floats.

float_to_str (Float → Str) is type-installed but codegen- deferred to a follow-up milestone: it requires runtime-allocated Str (the current Str path uses only static @.str_* globals; no malloc-backed dynamic-Str infrastructure). Calling it typechecks but produces a structured CodegenError::Internal.

What is not (yet) supported

Snapshot of the current boundary. Items move out of this list as iterations land; the JOURNAL records when.

• No effect handlers — only the built-in IO and Diverge ops.
• No refinements / SMT escalation.
• No HM inference inside bodies. Top-level def types are explicit; polymorphism is opt-in via Type::Forall { vars, body }. Inside a body, lambdas check monomorphically against their declared type.
• Polymorphic fns must be directly called at the use site. Passing a polymorphic fn as a value (let f = id in f(42)) is not yet supported — it would need one closure-pair global per instantiation, deferred.
• No higher-rank polymorphism. Passing a polymorphic fn to another polymorphic fn (apply(id, 42)) is not supported.
• No recursive let for non-fn values. Plain let x = … in … only sees x inside the body, not inside its own RHS — recursive value bindings would break Decision 10's acyclicity invariant. Recursive fn bindings are supported via Term::LetRec ({ "t": "letrec", ... }); the desugar pass lifts most occurrences to a synthetic top-level fn, with lift_letrecs finishing the residue after typecheck.
• No visibility rules in imports. Every top-level def of an imported module is reachable; there is no pub / priv.

What is supported (and used as the smoke test for the pipeline):

• Int, Bool, Unit, Str, Float as primitive types.
• if, let, function calls, recursion.
• Effects on function signatures, with do op(args) for direct effect ops (io/print_int, io/print_bool, io/print_str).
• Builtins. Arithmetic operators (+, -, *, /) of type forall a. (a, a) -> a (codegen-restricted to {Int, Float}); % of type (Int, Int) -> Int (Int-only — fmod semantics for Float deferred); ordering operators and != (!=, <, <=, >, >=) of type forall a. (a, a) -> Bool (codegen-restricted to {Int, Float}); polymorphic neg : forall a. (a) -> a (codegen-restricted to {Int, Float}; Float arm uses LLVM fneg double for correct -0.0 handling); logical not : (Bool) -> Bool; conversions int_to_float : (Int) -> Float, float_to_int_truncate : (Float) -> Int (saturating, NaN → 0), float_to_str : (Float) -> Str (codegen lowering deferred — symbol installed but errors at codegen pending runtime Str allocation); inspection is_nan : (Float) -> Bool (LLVM fcmp uno); Float bit-pattern constants nan : Float, inf : Float, neg_inf : Float (resolved as bare values, lower to direct hex-float double SSA constants at use site); the IO effect ops (io/print_int|bool|str|float); == : forall a. (a, a) -> Bool; and __unreachable__ : forall a. a.
  • == is polymorphic. The typechecker accepts == at any type whose two sides agree (the rigid a of the Forall is unified by HM at the use site). Codegen monomorphises and dispatches on the resolved AIL arg type: Int → icmp eq i64; Bool → icmp eq i1; Str → call @strcmp(ptr, ptr) then icmp eq i32 0 (@strcmp is declared in the LLVM IR header alongside @printf / @GC_malloc); Unit → constant i1 true (Unit has a single inhabitant; both sides are still evaluated for any side effects); Float → fcmp oeq double. ADT and Fn arg types are rejected at codegen with a CodegenError::Internal mentioning == and the offending type — neither has a canonical structural-equality scheme yet, and the language deliberately does not silently elide the check. != for Float uses fcmp UNE double (NOT one) — one is "ordered and not equal" and would return false for nan != nan, violating IEEE-!=.
  • __unreachable__ is a polymorphic bottom value: a use of __unreachable__ typechecks against any expected type at the use site and codegens to the LLVM unreachable instruction (UB if ever executed). It is the chain machinery's deepest fall-through for matches that the typechecker proved exhaustive, and it is available to user code as an explicit panic primitive ((if cond __unreachable__ ...) for assertions or impossible branches). Reference site is Term::Var { name = "__unreachable__" } / form-A bare __unreachable__.
• ADTs + pattern matching. Sub-patterns of a Ctor pattern may be Var, Wild, another Ctor, or a literal. The desugar pass flattens nested Ctor patterns into a chain of let + match and rewrites every Pattern::Lit (top-level or sub-) to a Term::If on == before typecheck/codegen — see ailang-core::desugar and Pipeline above.
• Literal patterns at top level and inside Ctor sub-patterns (via desugar). (pat-lit 0) and (pat-ctor Cons (pat-lit 0) _) both parse and lower; the rewrite is to Term::If { cond = (== sv lit) }, so any literal kind whose == is supported is authorable. With == polymorphic over Int/Bool/Str/Unit, that covers every lit kind the AST ships — including (pat-lit "hi") over a Str scrutinee, exercised by examples/eq_demo.ail.json.
• Imports + qualified cross-module references via dotted names. Extends to types and constructors: a foreign module's ADT is referenced as (con std_pair.Pair a b), its ctors as (term-ctor std_pair.Pair MkPair x y) and (pat-ctor MkPair x y) inside that scrutinee. Std-library demos (examples/std_*_demo.ail.json) exercise this end-to-end.
• AI-authoring text surface, form (A) (Decision 6). The ailang-surface crate parses .ailx form-A text into a canonical ailang-core::ast::Module and prints any module back as form-A text. ail render and ail describe use it as the sole text projection; ail parse is the inverse direction. Round-trip identity (text → AST → JSON → AST → text) is gated by ailang-surface/tests/round_trip.rs over every shipped fixture.
• Memory management via Boehm conservative GC (Decision 9), with per-fn arena via stack alloca for non-escaping allocations layered on top. Every ADT box, lambda env, and closure pair allocates either via @GC_malloc (escaping; Boehm-managed) or via LLVM alloca (non-escaping; freed at fn return). The decision is made by an escape-analysis pre-pass over the fn body — see Decision 9's "Per-fn arena via stack alloca" subsection. Boehm-only soak tests are unchanged: examples/gc_stress.ail.json and examples/std_list_stress.ail.json still allocate via @GC_malloc because their boxes flow into other fns and escape. The per-fn-arena path is exercised end-to-end by examples/escape_local_demo.ail.json.
• First-class function references. A top-level fn name (or qualified prefix.def) used as a Term::Var is a fn-value.
• Anonymous lambdas with capture. Term::Lam constructs a closure that captures any free variables of its body from the enclosing scope. All fn-values share a single ABI: a ptr to a closure pair { thunk_ptr, env_ptr }. Top-level fns get an auto- generated adapter and a static closure pair (env = null) so they remain passable as values without heap overhead.
• Polymorphism via Type::Forall at top-level def types. Use sites instantiate fresh metavars; unification pins them against the concrete types of the call args. Codegen monomorphises on demand: each unique instantiation emits a specialised LLVM fn mangled @ail_<m>_<def>__<descriptor> (e.g. id__I for id at Int, apply__I_I for apply at (Int, Int)).
• Parameterised ADTs. TypeDef.vars: Vec<String> declares type parameters; Type::Con.args: Vec<Type> carries the type arguments at use sites. Both fields default to empty and are skipped during serialization, so canonical-JSON hashes of every existing definition stay bit-identical (regression test in crates/ailang-core/src/hash.rs). Ctor and match codegen stay inline at every use site — there is no specialised ADT symbol — but LLVM field types are derived per use site by substituting through cdef.ail_fields. The substitution is read off the call's arg types (ctor) or the scrutinee's Type::Con.args (match). An unresolved Type::Var reaching llvm_type is a hard error rather than a silent fallback to ptr.

Pipeline regression smoke tests:

• examples/sum.ail.json → prints 55 (recursion, arithmetic).
• examples/list.ail.json → prints 42 (ADTs + match).
• examples/hof.ail.json → prints 42 (first-class fn-refs, indirect call).
• examples/closure.ail.json → prints 42 (lambda capturing a let-bound var).
• examples/list_map.ail.json → prints 2/4/6 (ADTs + closure + recursive HOF + IO; the dogfood smoke test).
• examples/sort.ail.json → prints sorted [3,1,4,1,5,9,2,6,5,3,5] one-per-line (insertion sort over an 11-element list).
• examples/poly_id.ail.json → prints 42 then "true" (polymorphic identity at Int and Bool; two specialised fns emitted).
• examples/poly_apply.ail.json → prints 42 (polymorphic apply with a fn-typed parameter; apply(succ, 41)).
• examples/box.ail.json → prints 42 (parameterised ADT round- trip: MkBox(42) constructed, then projected by a polymorphic unbox : forall a. (Box<a>) -> a and printed).
• examples/maybe_int.ail.json → prints 7 then 99 (pattern match over Maybe<Int>: or_else(Some(7), 99) then or_else(None, 99)).
• examples/std_list_demo.ail.json → exercises std_list's combinators (length, sum, reverse, take/drop-style uses) end-to-end against std_list's List<a>.
• examples/std_maybe_demo.ail.json → exercises std_maybe combinators over Maybe<Int>, including from_maybe and map.
• examples/std_either_demo.ail.json → first program with three distinct type variables in a single fn (the either eliminator), monomorphised six different ways in the IR.
• examples/std_pair_demo.ail.json → drives every std_pair combinator (fst, snd, swap, map_first, map_second); expected output 7, 9, 9, 7, 8, 18.
• examples/nested_pat.ail.json → first program to use a nested (pat-ctor Cons a (pat-ctor Cons b _)); the desugar pass flattens it into a chain that the existing flat-match codegen consumes. Prints 30 for a 3-element input list.