# cma.1 — Implementation Plan

> **Parent spec:** `docs/specs/0017-cross-model-authoring-form-test.md`
>
> **For agentic workers:** REQUIRED SUB-SKILL: use `skills/implement`
> to run this plan. Steps use `- [ ]` checkboxes for tracking.

**Goal:** Stand up the master mini-spec source and the renderer binary
under a new top-level `experiments/2026-05-12-cross-model-authoring/`
directory, with three green test gates (example roundtrip, token
balance, spec completeness) and the two rendered mini-spec files
checked into the working tree.

**Architecture:** A new top-level dir parallels `crates/`, `bench/`,
`examples/`, `runtime/`, `docs/`. Inside it, `render/` is a standalone
Cargo crate that is **not** a member of the root workspace; it depends
on `ailang-core` and `ailang-surface` by relative path. The renderer
consumes `master/spec.md` (markdown with `{form-only: X}` blocks and
`{example: id}` markers) plus `master/examples/*.ail.json` (curated
AST fixtures) and emits `rendered/json.md` + `rendered/ailx.md`.
Examples are printed in JSON form by emitting the on-disk
canonical-key-order bytes, and in AILX form by `ailang_core::load_module`
→ `ailang_surface::print`.

**Tech Stack:** Rust 2021, `ailang-core::{load_module, canonical,
module_hash}`, `ailang-surface::{print, parse}`, `serde_json`,
`anyhow`. Dev-only: `tiktoken-rs` (used by `token_balance` test).

**Files this plan creates or modifies:**

- Create: `experiments/2026-05-12-cross-model-authoring/README.md` — purpose + how to run the renderer
- Create: `experiments/2026-05-12-cross-model-authoring/render/Cargo.toml` — standalone crate manifest
- Create: `experiments/2026-05-12-cross-model-authoring/render/src/main.rs` — renderer entry point
- Create: `experiments/2026-05-12-cross-model-authoring/render/src/splitter.rs` — directive parser module
- Create: `experiments/2026-05-12-cross-model-authoring/render/src/examples.rs` — example-id → form-projection
- Create: `experiments/2026-05-12-cross-model-authoring/render/tests/splitter_unit.rs` — directive-parser tests
- Create: `experiments/2026-05-12-cross-model-authoring/render/tests/example_roundtrip.rs` — mirrors `round_trip_one` for `master/examples/`
- Create: `experiments/2026-05-12-cross-model-authoring/render/tests/spec_completeness.rs` — borrowed schema-coverage visitor
- Create: `experiments/2026-05-12-cross-model-authoring/render/tests/token_balance.rs` — ±5% form-only block balance
- Create: `experiments/2026-05-12-cross-model-authoring/master/spec.md` — vollumfänglich mini-spec source
- Create: `experiments/2026-05-12-cross-model-authoring/master/examples/*.ail.json` — curated AST fixtures (one per variant cluster; final filename list emerges from Task 4)
- Create: `experiments/2026-05-12-cross-model-authoring/rendered/json.md` — generated; checked in
- Create: `experiments/2026-05-12-cross-model-authoring/rendered/ailx.md` — generated; checked in
- Modify: none. Root `Cargo.toml` is **not** touched (spec §Architecture lines 90–95).

---

## Task 1: Bootstrap experiment dir + Cargo skeleton

**Files:**
- Create: `experiments/2026-05-12-cross-model-authoring/README.md`
- Create: `experiments/2026-05-12-cross-model-authoring/render/Cargo.toml`
- Create: `experiments/2026-05-12-cross-model-authoring/render/src/main.rs`

- [ ] **Step 1.1: Create the directory structure**

Run:
```
mkdir -p experiments/2026-05-12-cross-model-authoring/render/src
mkdir -p experiments/2026-05-12-cross-model-authoring/render/tests
mkdir -p experiments/2026-05-12-cross-model-authoring/master/examples
mkdir -p experiments/2026-05-12-cross-model-authoring/rendered
```
Expected: four empty directories created, no errors.

- [ ] **Step 1.2: Write README.md**

Write `experiments/2026-05-12-cross-model-authoring/README.md`:

```markdown
# Cross-model authoring-form test

Empirical measurement of whether `.ail.json` or `.ailx` is the form a
foreign LLM author reaches for and succeeds with. Single subject for
v1: Qwen3-Coder-Next via IONOS. Two blind cohorts; same four tasks.

Parent spec: `docs/specs/0017-cross-model-authoring-form-test.md`.

## Layout

- `master/spec.md` — canonical mini-spec source (form-agnostic prose +
  `{form-only: X}` blocks + `{example: id}` markers).
- `master/examples/*.ail.json` — AST source-of-truth; each example
  prints to either form via the existing roundtrip machinery.
- `master/tasks/*.task.json` — task definitions consumed by the
  harness. **Authored in cma.2**, not cma.1.
- `render/` — standalone Cargo crate, outside the root workspace,
  builds the renderer binary.
- `rendered/json.md`, `rendered/ailx.md` — projected mini-specs,
  checked into the repo for review.
- `harness/` — **Authored in cma.2**.
- `runs/<date>-<hash>/` — populated by `harness` during a live run.

## Running the renderer

```
cargo run --manifest-path experiments/2026-05-12-cross-model-authoring/render/Cargo.toml -- \
    --master  experiments/2026-05-12-cross-model-authoring/master \
    --rendered experiments/2026-05-12-cross-model-authoring/rendered
```

## Running the tests

```
cargo test --manifest-path experiments/2026-05-12-cross-model-authoring/render/Cargo.toml
```

Three integration tests: `example_roundtrip` (each example loads,
prints to AILX, reparses to the same canonical bytes), `spec_completeness`
(every AST variant in `ailang_core::ast` is exercised by at least
one example), `token_balance` (form-only blocks balanced within ±5%
across the two rendered files).
```

- [ ] **Step 1.3: Write render/Cargo.toml**

Write `experiments/2026-05-12-cross-model-authoring/render/Cargo.toml`:

```toml
[package]
name = "xmodel-render"
version = "0.0.1"
edition = "2021"
publish = false

[[bin]]
name = "xmodel-render"
path = "src/main.rs"

[dependencies]
ailang-core    = { path = "../../../crates/ailang-core" }
ailang-surface = { path = "../../../crates/ailang-surface" }
serde_json = "1"
anyhow     = "1"

[dev-dependencies]
tiktoken-rs = "0.6"
```

- [ ] **Step 1.4: Write skeleton src/main.rs**

Write `experiments/2026-05-12-cross-model-authoring/render/src/main.rs`:

```rust
//! xmodel-render — renders the master mini-spec into two form-specific projections.

use anyhow::Result;

mod splitter;
mod examples;

fn main() -> Result<()> {
    eprintln!("xmodel-render: not yet wired (Task 7)");
    Ok(())
}
```

Create the two empty `.rs` files referenced (will be filled in by
later tasks):

```
touch experiments/2026-05-12-cross-model-authoring/render/src/splitter.rs
touch experiments/2026-05-12-cross-model-authoring/render/src/examples.rs
```

- [ ] **Step 1.5: Verify the skeleton builds**

Run:
```
cargo build --manifest-path experiments/2026-05-12-cross-model-authoring/render/Cargo.toml
```
Expected: succeeds. Two warnings expected (unused `mod splitter;` and
unused `mod examples;`) — these will go once Task 2 and Task 4 wire
the modules in.

---

## Task 2: Directive splitter (TDD)

**Files:**
- Modify: `experiments/2026-05-12-cross-model-authoring/render/src/splitter.rs`
- Create: `experiments/2026-05-12-cross-model-authoring/render/tests/splitter_unit.rs`

The splitter recognises three directives in `master/spec.md`:

- `{form-only: json}` … `{/form-only}` — block visible only in
  JSON-rendered output.
- `{form-only: ailx}` … `{/form-only}` — block visible only in
  AILX-rendered output.
- `{example: <id>}` — single-line marker that the example renderer
  resolves to a fenced code block of the matching form.

Directives appear on their own lines. The parser is line-oriented
(no nested directives, no inline directives).

- [ ] **Step 2.1: Write the splitter unit tests (RED)**

Write `experiments/2026-05-12-cross-model-authoring/render/tests/splitter_unit.rs`:

```rust
use xmodel_render::splitter::{split, Form, Segment};

#[test]
fn plain_text_passes_through_unchanged() {
    let input = "hello\nworld\n";
    let segments = split(input);
    assert_eq!(segments, vec![Segment::Prose("hello\nworld\n".to_string())]);
}

#[test]
fn form_only_block_is_isolated() {
    let input = "before\n{form-only: json}\nschema rule\n{/form-only}\nafter\n";
    let segments = split(input);
    assert_eq!(
        segments,
        vec![
            Segment::Prose("before\n".to_string()),
            Segment::FormOnly { form: Form::Json, body: "schema rule\n".to_string() },
            Segment::Prose("after\n".to_string()),
        ]
    );
}

#[test]
fn ailx_form_only_block_uses_ailx_form() {
    let input = "{form-only: ailx}\ngrammar rule\n{/form-only}\n";
    let segments = split(input);
    assert_eq!(
        segments,
        vec![
            Segment::FormOnly { form: Form::Ailx, body: "grammar rule\n".to_string() },
        ]
    );
}

#[test]
fn example_marker_is_isolated() {
    let input = "before\n{example: empty_module}\nafter\n";
    let segments = split(input);
    assert_eq!(
        segments,
        vec![
            Segment::Prose("before\n".to_string()),
            Segment::Example("empty_module".to_string()),
            Segment::Prose("after\n".to_string()),
        ]
    );
}

#[test]
fn multiple_directives_in_sequence() {
    let input = "p1\n{example: a}\n{form-only: json}\njs\n{/form-only}\n{form-only: ailx}\nax\n{/form-only}\n{example: b}\np2\n";
    let segments = split(input);
    assert_eq!(
        segments,
        vec![
            Segment::Prose("p1\n".to_string()),
            Segment::Example("a".to_string()),
            Segment::FormOnly { form: Form::Json, body: "js\n".to_string() },
            Segment::FormOnly { form: Form::Ailx, body: "ax\n".to_string() },
            Segment::Example("b".to_string()),
            Segment::Prose("p2\n".to_string()),
        ]
    );
}

#[test]
fn unknown_form_name_is_an_error() {
    let input = "{form-only: yaml}\nx\n{/form-only}\n";
    let result = std::panic::catch_unwind(|| split(input));
    assert!(result.is_err(), "split should panic on unknown form name");
}

#[test]
fn unclosed_form_only_is_an_error() {
    let input = "{form-only: json}\nopen\n";
    let result = std::panic::catch_unwind(|| split(input));
    assert!(result.is_err(), "split should panic on unclosed form-only block");
}
```

The test file imports `xmodel_render::splitter::{split, Form, Segment}`,
which forces `src/main.rs` to expose the splitter module via a library
target. Without a library target this fails to compile.

- [ ] **Step 2.2: Add a library target to render/Cargo.toml + lib.rs**

Add to `experiments/2026-05-12-cross-model-authoring/render/Cargo.toml`,
after the existing `[[bin]]` block:

```toml
[lib]
name = "xmodel_render"
path = "src/lib.rs"
```

Create `experiments/2026-05-12-cross-model-authoring/render/src/lib.rs`:

```rust
//! Library surface for xmodel-render's internal modules — exposed so
//! integration tests can reach them.

pub mod splitter;
pub mod examples;
```

Update `experiments/2026-05-12-cross-model-authoring/render/src/main.rs`
to use the library (replace the existing `mod splitter;` / `mod examples;`
lines with):

```rust
//! xmodel-render — renders the master mini-spec into two form-specific projections.

use anyhow::Result;
use xmodel_render::{splitter, examples};

fn main() -> Result<()> {
    let _ = (&splitter::Form::Json, &examples::Form::Json); // silence unused until Task 7
    eprintln!("xmodel-render: not yet wired (Task 7)");
    Ok(())
}
```

(Two `Form` enums avoid silent aliasing later; the bin reconciles them
at wire-up in Task 7. For now, just silence unused-import warnings.)

- [ ] **Step 2.3: Run the tests, verify RED**

Run:
```
cargo test --manifest-path experiments/2026-05-12-cross-model-authoring/render/Cargo.toml --test splitter_unit
```
Expected: **compilation fails** because `splitter::{split, Form, Segment}`
do not exist yet. Output contains "cannot find function `split` in module
`splitter`" or equivalent.

- [ ] **Step 2.4: Write the splitter implementation (GREEN)**

Replace `experiments/2026-05-12-cross-model-authoring/render/src/splitter.rs`
content with:

```rust
//! Directive splitter for master/spec.md.
//!
//! Recognises three directives, one per line, no nesting:
//!   {form-only: json} … {/form-only}
//!   {form-only: ailx} … {/form-only}
//!   {example: <id>}

#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub enum Form {
    Json,
    Ailx,
}

impl Form {
    fn parse(name: &str) -> Option<Form> {
        match name {
            "json" => Some(Form::Json),
            "ailx" => Some(Form::Ailx),
            _ => None,
        }
    }
}

#[derive(Debug, Clone, PartialEq, Eq)]
pub enum Segment {
    Prose(String),
    FormOnly { form: Form, body: String },
    Example(String),
}

/// Split a master/spec.md text into a sequence of segments.
///
/// Panics on malformed input (unknown form name, unclosed form-only block).
/// The renderer treats these as authoring bugs that should fail the test
/// suite rather than silently produce diverging mini-specs.
pub fn split(input: &str) -> Vec<Segment> {
    let mut out: Vec<Segment> = Vec::new();
    let mut current_prose = String::new();
    let mut current_block: Option<(Form, String)> = None;

    for line in input.split_inclusive('\n') {
        let trimmed = line.trim_end_matches('\n');

        // Inside a form-only block: collect until the closing marker.
        if let Some((form, body)) = current_block.as_mut() {
            if trimmed == "{/form-only}" {
                let body_owned = std::mem::take(body);
                out.push(Segment::FormOnly { form: *form, body: body_owned });
                current_block = None;
            } else {
                body.push_str(line);
            }
            continue;
        }

        // {form-only: <name>}
        if let Some(rest) = trimmed.strip_prefix("{form-only: ").and_then(|r| r.strip_suffix("}")) {
            flush_prose(&mut current_prose, &mut out);
            let form = Form::parse(rest)
                .unwrap_or_else(|| panic!("unknown form name in directive: {}", rest));
            current_block = Some((form, String::new()));
            continue;
        }

        // {example: <id>}
        if let Some(id) = trimmed.strip_prefix("{example: ").and_then(|r| r.strip_suffix("}")) {
            flush_prose(&mut current_prose, &mut out);
            out.push(Segment::Example(id.to_string()));
            continue;
        }

        // Stray closer with no opener.
        if trimmed == "{/form-only}" {
            panic!("stray {{/form-only}} without matching opener");
        }

        // Plain prose line.
        current_prose.push_str(line);
    }

    if current_block.is_some() {
        panic!("unclosed form-only block at end of input");
    }
    flush_prose(&mut current_prose, &mut out);
    out
}

fn flush_prose(buf: &mut String, out: &mut Vec<Segment>) {
    if !buf.is_empty() {
        out.push(Segment::Prose(std::mem::take(buf)));
    }
}
```

- [ ] **Step 2.5: Run the tests, verify GREEN**

Run:
```
cargo test --manifest-path experiments/2026-05-12-cross-model-authoring/render/Cargo.toml --test splitter_unit
```
Expected: `test result: ok. 7 passed; 0 failed`.

---

## Task 3: spec_completeness test (RED on empty examples/)

**Files:**
- Create: `experiments/2026-05-12-cross-model-authoring/render/tests/spec_completeness.rs`

This task lifts the AST-variant visitor from
`crates/ailang-core/tests/schema_coverage.rs` verbatim, adjusts the
`examples_dir()` resolver to point at `experiments/.../master/examples/`,
and runs the test to confirm it fails on the empty examples directory.
Task 4 will then add fixtures to drive it green.

- [ ] **Step 3.1: Copy the visitor verbatim from schema_coverage.rs**

Read `crates/ailang-core/tests/schema_coverage.rs` lines 1–399 (the
entire file).

Copy the contents into a new file
`experiments/2026-05-12-cross-model-authoring/render/tests/spec_completeness.rs`
verbatim with one adjustment: replace the body of the `examples_dir()`
function (originally at source lines 324–327) with:

```rust
fn examples_dir() -> std::path::PathBuf {
    // CARGO_MANIFEST_DIR for this test is
    //   experiments/2026-05-12-cross-model-authoring/render/
    // Master examples live one parent up + master/examples/.
    let manifest = std::path::PathBuf::from(env!("CARGO_MANIFEST_DIR"));
    manifest.parent().unwrap().join("master").join("examples")
}
```

Rename the driver test from `every_ast_variant_is_observed_in_the_fixture_corpus`
to `every_ast_variant_is_observed_in_master_examples` so the failure
output names the right scope.

The file ends up structurally identical to schema_coverage.rs except
for `examples_dir()` and the test name. The visitor logic, the
`VariantTag` enum, the `EXPECTED_VARIANTS` constant, and the
`visit_*` functions are unchanged.

- [ ] **Step 3.2: Run the test, verify RED**

Run:
```
cargo test --manifest-path experiments/2026-05-12-cross-model-authoring/render/Cargo.toml --test spec_completeness
```
Expected: the test fails. Failure mode is one of:
- "no fixtures found in `experiments/.../master/examples/`" if the
  visitor pre-checks fixture count (verify behaviour against
  `schema_coverage.rs:329–344` — `list_json_fixtures` returns an empty
  Vec without erroring, but the variant-observation set is empty so
  the assertion at lines 380+ fires with "missing variants: …").
- "missing variants: <full list of 33 variants>" — the body fires
  because `observed` is empty.

Either way: the test is RED. Record the exact failure message in the
implement journal so Task 4 can confirm the change in failure shape
once examples land.

---

## Task 4: master/examples/*.ail.json fixtures

**Files:**
- Create: `experiments/2026-05-12-cross-model-authoring/master/examples/<id>.ail.json` (multiple files; see Step 4.1 for the curated list)

The goal is the smallest set of curated, didactic `.ail.json` files
that together exercise all 33 AST variants the `EXPECTED_VARIANTS`
constant lists (5 Def + 13 Term + 4 Pattern + 5 Literal + 4 Type + 3
ParamMode — note the visitor counts 33 because `Module` itself is
implicit on every file). Each fixture must be a complete, well-formed
AILang module that loads cleanly via `ailang_core::load_module`.

**Common shape constraints** (apply to every fixture):

- Top-level JSON object with `"schema": "ailang/v0"`, `"name":
  "<fixture_id>_module"`, `"imports": []` (unless explicitly noted),
  `"defs": [...]`.
- Canonical key order — fields appear in the order the existing
  workspace fixtures use (the implementer reads any one of
  `examples/*.ail.json` for the precedent).
- Every `Def::Fn` carries mandatory `params`, `param_tys` (with
  ParamMode), `ret_ty`, and `effects` per DESIGN.md §"Data model".
- The fixture must load cleanly via `ailang_core::load_module` (the
  spec_completeness test calls it on every file).
- The fixture must roundtrip via `ailang_surface::print` →
  `ailang_surface::parse` → canonical-byte equality (Task 5 enforces
  this). If a fixture cannot roundtrip, the .ail.json embeds
  something the surface printer cannot faithfully reproduce — fix
  the fixture, not the test.

The implementer reads `crates/ailang-core/src/ast.rs` to confirm the
exact ParamMode variant names (the planner does not pin those
because the recon report did not enumerate them); the structural
template for each fixture's shape can be lifted from a small
existing fixture in `examples/` (e.g. `box.ail.json` for the
Data/TermCtor shape; `list_map_poly.ail.json` for Forall + Match;
the workspace has 136 candidates).

- [ ] **Step 4.1: Author `fn_returns_int.ail.json`**

Single fn `f : () -> Int = 42` (or pick a non-zero integer the
implementer prefers). Exercises: Module, Def::Fn, Term::Lit, Type::Con,
Literal::Int, one ParamMode tag on the (empty) param list — verify
the empty-params shape carries a ParamMode marker; if it doesn't,
fall to `Step 4.13`'s coverage for ParamMode.

- [ ] **Step 4.2: Author `fn_calls_prelude.ail.json`**

A fn that takes two Int parameters and returns their sum by calling
a prelude function (the implementer picks the actual symbol — likely
`int_add` or its current canonical name — by reading
`crates/ailang-check/src/lib.rs` or the prelude module). Exercises:
Term::App, Term::Var, a populated ParamMode set on the params.

- [ ] **Step 4.3: Author `fn_with_lambda.ail.json`**

A fn that returns a lambda value (e.g. `make_adder : (Int) ->
((Int) -> Int) = \x. \y. int_add(x, y)`). Exercises: Term::Lam,
Type::Fn as a return type.

- [ ] **Step 4.4: Author `fn_with_do_seq.ail.json`**

A fn `main : () -> () !IO` whose body is a `do { … }`-style sequence
of two `io_print_int` calls and a unit return. Exercises: Term::Do,
Term::Seq, the IO effect on the fn's effects set, Literal::Unit (if
that variant is the discriminator the visitor recognises; otherwise
it is covered by the unit return value).

- [ ] **Step 4.5: Author `data_simple.ail.json`**

A `data Box a where MkBox a` declaration plus a fn that constructs
a `Box Int` via `MkBox(1)`. Exercises: Def::Data, Term::TermCtor,
Type::Var in the ctor's argument position.

- [ ] **Step 4.6: Author `data_with_match.ail.json`**

A `data List a where Nil | Cons a (List a)` declaration plus a fn
that pattern-matches on a value: `match xs of Nil -> 0 | Cons _ _
-> 1`. Exercises: Term::Match with two arms, Pattern::Ctor (both
nullary `Nil` and binary `Cons`), Pattern::Wildcard (on the two
Cons-field positions).

- [ ] **Step 4.7: Author `match_literal_pattern.ail.json`**

A fn that matches an Int parameter against literal patterns:
`match n of 0 -> "zero" | _ -> "other"`. Exercises: Pattern::Lit and
Pattern::Wildcard. If the visitor treats variable patterns separately
from wildcard, extend one arm to `| n2 -> …` to also cover
Pattern::Var (if not already covered by Step 4.6).

- [ ] **Step 4.8: Author `forall_polymorphic.ail.json`**

A polymorphic identity fn: `fn id : forall a. (a) -> a = \x. x`.
Exercises: Type::Forall, Type::Var in the fn signature.

- [ ] **Step 4.9: Author `class_def.ail.json`**

A `class MyShow a where show : (a) -> Str` declaration. Exercises:
Def::Class, the class method's type signature.

- [ ] **Step 4.10: Author `instance_def.ail.json`**

An instance `instance MyShow Int where show = \x. "an Int"` (or the
real `MyShow` defined in Step 4.9; the implementer ensures the two
files agree on the class name). If AILang's instance form requires
a constraint on a polymorphic instance, add a fn-type with a
Constraint in this fixture; otherwise add a separate trivial fn
`use_eq : forall a. (Eq a) => (a, a) -> Bool = \x y. eq(x, y)` to
exercise the Constraint type-form. Exercises: Def::Instance, the
Constraint variant of Type.

- [ ] **Step 4.11: Author `floats.ail.json`**

A fn returning a specific Float literal: `fn pi : () -> Float =
3.14`. Exercises: Literal::Float, Type::Con (Float). Note: the
authored decimal is what the model writes; the toolchain handles
the bit-hex encoding internally (this is what the spec.md will
teach in Section 9 — see Task 6).

- [ ] **Step 4.12: Author `bool_str.ail.json`**

Two trivial fns: `is_true : () -> Bool = true` and `greeting : ()
-> Str = "hi"`. Exercises: Literal::Bool, Literal::Str.

- [ ] **Step 4.13: Author `param_modes_all.ail.json`**

Three fns, each with one parameter exercising one of the three
ParamMode variants. The implementer reads
`crates/ailang-core/src/ast.rs` to extract the exact mode names
(planner not pinning, recon did not enumerate). Each fn is trivial
(`fn f_<mode> : (Int) -> Int = \x. x`); the load-bearing thing is
the ParamMode tag on each parameter.

- [ ] **Step 4.14: Run spec_completeness, verify GREEN**

Run:
```
cargo test --manifest-path experiments/2026-05-12-cross-model-authoring/render/Cargo.toml --test spec_completeness
```
Expected: `test result: ok. 1 passed; 0 failed`.

If specific variants are still missing, the failure message names
them. Extend the fixture closest in scope (e.g. add a Term::Seq use
to Step 4.4's fixture, add a new pattern arm to Step 4.6's fixture),
do not weaken the test. If a variant is *structurally unreachable*
through any fixture an LLM-author would produce, that is a real
finding worth flagging in the implement journal — but the audit-rt
journal records all 34 variants are observed by the existing
schema_coverage corpus, so the curated subset should reach them
too with one fixture each.

---

## Task 5: example_roundtrip test

**Files:**
- Create: `experiments/2026-05-12-cross-model-authoring/render/tests/example_roundtrip.rs`

This test mirrors `crates/ailang-surface/tests/round_trip.rs::round_trip_one`
for the subset of fixtures under `master/examples/`. It is the local
gate ensuring every example printed by the renderer is faithful.

- [ ] **Step 5.1: Write the roundtrip test**

Write `experiments/2026-05-12-cross-model-authoring/render/tests/example_roundtrip.rs`:

```rust
//! Roundtrip gate for master/examples/.
//!
//! Mirrors crates/ailang-surface/tests/round_trip.rs but scoped to
//! the master examples directory. For each .ail.json fixture: load,
//! print via ailang_surface::print, reparse, canonicalise both sides,
//! assert byte equality.

use std::path::{Path, PathBuf};

fn examples_dir() -> PathBuf {
    let manifest = PathBuf::from(env!("CARGO_MANIFEST_DIR"));
    manifest.parent().unwrap().join("master").join("examples")
}

fn list_examples() -> Vec<PathBuf> {
    let mut paths: Vec<PathBuf> = std::fs::read_dir(examples_dir())
        .expect("master/examples/ should exist")
        .filter_map(|entry| entry.ok())
        .map(|entry| entry.path())
        .filter(|path| path.extension().and_then(|s| s.to_str()) == Some("json"))
        .filter(|path| {
            path.file_name()
                .and_then(|s| s.to_str())
                .map(|name| name.ends_with(".ail.json"))
                .unwrap_or(false)
        })
        .collect();
    paths.sort();
    paths
}

fn round_trip_one(path: &Path) -> Result<(), String> {
    let original = ailang_core::load_module(path)
        .map_err(|e| format!("{}: load_module failed: {e}", path.display()))?;
    let text = ailang_surface::print(&original);
    let parsed = ailang_surface::parse(&text)
        .map_err(|e| format!("{}: surface::parse failed: {e}", path.display()))?;
    let bytes_original = ailang_core::canonical::to_bytes(&original);
    let bytes_parsed = ailang_core::canonical::to_bytes(&parsed);
    if bytes_original != bytes_parsed {
        return Err(format!(
            "{}: roundtrip diverged.\n--- printed AILX ---\n{}\n--- original bytes len {} vs parsed bytes len {} ---",
            path.display(),
            text,
            bytes_original.len(),
            bytes_parsed.len(),
        ));
    }
    Ok(())
}

#[test]
fn every_example_roundtrips_via_ailx() {
    let examples = list_examples();
    assert!(
        !examples.is_empty(),
        "master/examples/ should contain at least one .ail.json fixture"
    );
    let mut failures: Vec<String> = Vec::new();
    for path in &examples {
        if let Err(msg) = round_trip_one(path) {
            failures.push(msg);
        }
    }
    if !failures.is_empty() {
        panic!(
            "{} of {} examples failed roundtrip:\n\n{}",
            failures.len(),
            examples.len(),
            failures.join("\n\n")
        );
    }
}
```

- [ ] **Step 5.2: Run the roundtrip test, verify GREEN**

Run:
```
cargo test --manifest-path experiments/2026-05-12-cross-model-authoring/render/Cargo.toml --test example_roundtrip
```
Expected: `test result: ok. 1 passed; 0 failed`.

If any fixture fails the roundtrip: that's a fixture authoring bug
(the .ail.json embeds something the surface printer can't faithfully
reproduce). Fix the fixture; do not weaken the test.

---

## Task 6: master/spec.md — the mini-spec content

**Files:**
- Create: `experiments/2026-05-12-cross-model-authoring/master/spec.md`

The mini-spec is the load-bearing artifact: this is what the model
sees as its system prompt. It must be vollumfänglich (cover every
authoring construct the model could possibly need) and
form-agnostic in the prose body. Form-specific scaffolding lives in
`{form-only: X}` blocks; code samples are `{example: id}` markers
resolving to fixtures from Task 4.

Token budget guideline: total per-form rendered length ~6k–10k
tokens (well below the 28k DESIGN.md baseline; the rendering will
be measured in Task 8).

**Authoring rules common to every section:**

- Each section is 200–600 words of prose plus example markers as
  specified. The implementer writes the prose; the planner pins the
  *structure* and what each section must cover.
- Any section with a `{form-only: json}` block must also carry a
  `{form-only: ailx}` block of comparable length (the token-balance
  test in Task 8 enforces the aggregate ±5%; per-section symmetry is
  the easiest way to land there).
- Every `{example: <id>}` marker must reference a file authored in
  Task 4. If the marker resolves to a missing fixture the renderer
  fails at runtime in Step 7.3.
- The spec.md is content authoring; quality gates are Task 7's
  end-to-end render, Task 8's token balance, and the orchestrator's
  end-to-end read before any cma.3 live run.

- [ ] **Step 6.1: Write Section 1 — Header**

One-paragraph orientation: "AILang is a small data-as-source
functional language designed for an LLM author. This document is
the complete authoring reference. Every well-formed program is a
module." No form-only blocks. No example markers.

- [ ] **Step 6.2: Write Section 2 — Modules and the on-disk form**

Explain what a module is — a top-level container of `imports` plus
`defs`. Include one `{form-only: json}` block listing schema
requirements: `schema` field equal to `"ailang/v0"`, `name` field, the
`imports` array, the `defs` array, canonical key order. Include one
`{form-only: ailx}` block listing the grammar conventions:
parenthesised tagged-head form, the `(module NAME …)` opener, the
relationship between AILX surface and the on-disk JSON. Close with
`{example: fn_returns_int}` to anchor the simplest concrete module.

- [ ] **Step 6.3: Write Section 3 — Types**

Cover Type::Con (named base types like `Int`, `Bool`, `Str`,
`Float`), Type::Var (polymorphic type variables introduced by
`forall`), Type::Fn (function types with parameter list, return
type, and effect set), Type::Forall (universal quantification at
the fn signature). State that every fn signature is mandatory at
top level — no implicit return-type inference at the binding form.
Form-only blocks: JSON shape of each type variant, AILX surface form
of each. Close with `{example: forall_polymorphic}`.

- [ ] **Step 6.4: Write Section 4 — Mode annotations**

Cover the three ParamMode variants. The implementer reads
`crates/ailang-core/src/ast.rs` to extract the exact mode names
(planner not pinning — recon did not enumerate). State that mode is
mandatory on every fn parameter; the schema rejects unannotated
parameters. Form-only blocks: JSON shape of a fn's `param_tys`
field; AILX surface syntax for mode annotations. Close with
`{example: param_modes_all}`.

- [ ] **Step 6.5: Write Section 5 — Functions**

Cover Def::Fn shape: `name`, type signature, parameter list, body.
Walk through the term forms inside a function body: Term::Var
(referencing a binding), Term::App (function application),
Term::Lam (anonymous lambda), Term::Lit (literal value). Form-only
blocks: JSON shape of each Term variant; AILX surface for each.
Examples: `{example: fn_calls_prelude}` and `{example: fn_with_lambda}`.

- [ ] **Step 6.6: Write Section 6 — Algebraic data types**

Cover Def::Data (declaring a new ADT with its ctors), Term::TermCtor
(constructing a value of that ADT). Note that ctor names are
distinct from fn names. Form-only blocks: JSON shape of `Def::Data`
with `params` (type-var list) and `ctors` (ctor name + arg-type
list); AILX surface for `(data NAME (vars …) (ctor …) …)`. Close
with `{example: data_simple}`.

- [ ] **Step 6.7: Write Section 7 — Pattern matching**

Cover Term::Match and the four Pattern variants: Pattern::Ctor (a
constructor pattern with sub-patterns), Pattern::Var (binding the
matched value to a name), Pattern::Wildcard (`_`), Pattern::Lit
(matching against a literal value). Form-only blocks: JSON shape
of a `match` term with arms; AILX surface for `(match SCRUT (case
PAT BODY) …)`. Examples: `{example: data_with_match}` and
`{example: match_literal_pattern}`.

- [ ] **Step 6.8: Write Section 8 — Effects**

Cover the effect set on a fn type. The two currently wired effects
are `IO` (required to call `io_print_int` and similar) and
`Diverge` (for non-terminating programs). Term::Do introduces an
effectful sequence; Term::Seq is the sequence node. Form-only
blocks: JSON shape of the `effects` field on a fn type; AILX
surface for `!IO` annotations and `(do …)` blocks. Close with
`{example: fn_with_do_seq}`.

- [ ] **Step 6.9: Write Section 9 — Literals**

Cover the five literal variants. Literal::Int (decimal integer
syntax). Literal::Float — note that the model writes the decimal
form (`3.14`); the toolchain encodes the IEEE-754 bit pattern as a
16-lowercase-hex string on parse, per the Roundtrip Invariant. The
model never authors the hex encoding. Literal::Bool (`true` /
`false`). Literal::Str (double-quoted, ASCII-only by Decision 6
Constraint 3). Literal::Unit (`()`). Form-only blocks: JSON shape
of `Lit` variants; AILX surface for each. Examples:
`{example: floats}` and `{example: bool_str}`.

- [ ] **Step 6.10: Write Section 10 — Typeclasses**

Cover Def::Class (declaring a class with method signatures),
Def::Instance (providing an implementation of a class for a
specific type), the Constraint variant of Type (used in
constraint-polymorphic fn signatures like `forall a. (Eq a) =>
…`). Mention the four prelude classes: `Eq`, `Ord`, `Num`,
`Bounded`. State that method invocation is by name; the compiler
monomorphises class methods at the call site. Form-only blocks:
JSON shape of class / instance / constraint; AILX surface for each.
Examples: `{example: class_def}` and `{example: instance_def}`.

- [ ] **Step 6.11: Write Section 11 — The prelude**

Tabulate the names the model can call without importing anything.
The implementer extracts the canonical list at the time of
authoring from the prelude module (likely under
`crates/ailang-check/src/prelude.rs` or `lib.rs`). One row per
function: name, type signature (form-agnostic in the prose; both
forms inside form-only blocks), one-line description. Minimum
content: integer arithmetic (`int_add`, `int_sub`, `int_mul`,
`int_div` or whatever the canonical names are), integer comparison
or its typeclass-routed equivalent (`Eq Int` / `Ord Int`), and
`io_print_int` (plus `io_print_str` only if it currently ships).
State explicitly that `int_to_str` is OUT — it is type-installed
but codegen-deferred pending the heap-Str ABI milestone (roadmap
P1 entry). No example marker (the table itself is the content).

- [ ] **Step 6.12: Write Section 12 — Content addressing**

Short paragraph: AILang has content-addressed identity (Decision
2). Hashes are computed by the toolchain on `ail parse` / `ail
check` from the canonical bytes of each definition. The author
writes no hash literal in either form; the JSON form has no `hash`
field; the AILX surface has no hash production. This is symmetric
across forms and removes hash as a form-distinguishing factor. No
example marker; no form-only blocks (because there is no form
asymmetry to teach).

- [ ] **Step 6.13: Write Section 13 — Out of scope + closing**

Short list of what this mini-spec does not teach: cross-module
imports (MVP tasks live in single modules; the `imports` array is
empty), refinement types (reserved in the schema but pass-through
in MVP), point-free style, operator overloading, syntactic sugar
of any kind. Reason: the four MVP tasks do not need them, and
including them would pad the spec.

Close the file with one paragraph: "When asked to write a module,
return the complete module text and nothing else. No markdown
fences. No prose explanation." This is the directive the harness
relies on to keep response parsing trivial.

---

## Task 7: Wire render/src/main.rs end-to-end

**Files:**
- Modify: `experiments/2026-05-12-cross-model-authoring/render/src/main.rs`
- Modify: `experiments/2026-05-12-cross-model-authoring/render/src/examples.rs`

The renderer assembles the two `rendered/*.md` files from
`master/spec.md` + `master/examples/`.

- [ ] **Step 7.1: Implement the example renderer module**

Replace `experiments/2026-05-12-cross-model-authoring/render/src/examples.rs`
content with:

```rust
//! Per-form rendering of a single master example.
//!
//! - JSON form: read the on-disk canonical bytes verbatim, emit a
//!   ```json``` fenced block.
//! - AILX form: load the example via ailang_core::load_module, print
//!   via ailang_surface::print, emit a ```ailx``` fenced block.

use anyhow::{anyhow, Context, Result};
use std::path::Path;

#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub enum Form {
    Json,
    Ailx,
}

pub fn render_example(examples_dir: &Path, id: &str, form: Form) -> Result<String> {
    let path = examples_dir.join(format!("{id}.ail.json"));
    if !path.exists() {
        return Err(anyhow!(
            "example id `{id}` references missing fixture at {}",
            path.display()
        ));
    }
    match form {
        Form::Json => {
            let bytes = std::fs::read_to_string(&path)
                .with_context(|| format!("reading {}", path.display()))?;
            Ok(format!("```json\n{}\n```\n", bytes.trim_end()))
        }
        Form::Ailx => {
            let module = ailang_core::load_module(&path)
                .with_context(|| format!("load_module on {}", path.display()))?;
            let text = ailang_surface::print(&module);
            Ok(format!("```ailx\n{}\n```\n", text.trim_end()))
        }
    }
}
```

- [ ] **Step 7.2: Implement the renderer in main.rs**

Replace `experiments/2026-05-12-cross-model-authoring/render/src/main.rs`
content with:

```rust
//! xmodel-render — projects master/spec.md into rendered/json.md and rendered/ailx.md.

use anyhow::{anyhow, Context, Result};
use std::path::{Path, PathBuf};

use xmodel_render::splitter::{self, Segment};
use xmodel_render::examples;

fn main() -> Result<()> {
    let (master_dir, rendered_dir) = parse_args()?;
    let spec_path = master_dir.join("spec.md");
    let examples_dir = master_dir.join("examples");
    let spec_text = std::fs::read_to_string(&spec_path)
        .with_context(|| format!("reading {}", spec_path.display()))?;
    let segments = splitter::split(&spec_text);

    let json_out = render_for(&segments, &examples_dir, examples::Form::Json, splitter::Form::Json)?;
    let ailx_out = render_for(&segments, &examples_dir, examples::Form::Ailx, splitter::Form::Ailx)?;

    std::fs::create_dir_all(&rendered_dir)
        .with_context(|| format!("creating {}", rendered_dir.display()))?;
    let json_path = rendered_dir.join("json.md");
    let ailx_path = rendered_dir.join("ailx.md");
    std::fs::write(&json_path, json_out)
        .with_context(|| format!("writing {}", json_path.display()))?;
    std::fs::write(&ailx_path, ailx_out)
        .with_context(|| format!("writing {}", ailx_path.display()))?;
    eprintln!("xmodel-render: wrote {} and {}", json_path.display(), ailx_path.display());
    Ok(())
}

fn render_for(
    segments: &[Segment],
    examples_dir: &Path,
    example_form: examples::Form,
    keep_form: splitter::Form,
) -> Result<String> {
    let mut out = String::new();
    for seg in segments {
        match seg {
            Segment::Prose(text) => out.push_str(text),
            Segment::FormOnly { form, body } if *form == keep_form => out.push_str(body),
            Segment::FormOnly { .. } => {} // drop the other form's block
            Segment::Example(id) => {
                let rendered = examples::render_example(examples_dir, id, example_form)?;
                out.push_str(&rendered);
            }
        }
    }
    Ok(out)
}

fn parse_args() -> Result<(PathBuf, PathBuf)> {
    let mut master: Option<PathBuf> = None;
    let mut rendered: Option<PathBuf> = None;
    let mut args = std::env::args().skip(1);
    while let Some(arg) = args.next() {
        match arg.as_str() {
            "--master" => master = Some(PathBuf::from(args.next().ok_or_else(|| anyhow!("--master needs a value"))?)),
            "--rendered" => rendered = Some(PathBuf::from(args.next().ok_or_else(|| anyhow!("--rendered needs a value"))?)),
            other => return Err(anyhow!("unknown argument: {}", other)),
        }
    }
    let master = master.ok_or_else(|| anyhow!("--master <dir> is required"))?;
    let rendered = rendered.ok_or_else(|| anyhow!("--rendered <dir> is required"))?;
    Ok((master, rendered))
}
```

Note the two `Form` enums: `splitter::Form` is the marker found in
`{form-only: X}`; `examples::Form` is the renderer-side mode for an
example. They are intentionally separate types (different semantics)
and the bin's `render_for` reconciles them. The reconciliation is
trivial because there are only two forms, but keeping the types
separate avoids accidentally feeding one to the other.

- [ ] **Step 7.3: Run the renderer**

Run:
```
cargo run --manifest-path experiments/2026-05-12-cross-model-authoring/render/Cargo.toml -- \
  --master  experiments/2026-05-12-cross-model-authoring/master \
  --rendered experiments/2026-05-12-cross-model-authoring/rendered
```
Expected:
- exit code 0,
- stderr line "xmodel-render: wrote …json.md and …ailx.md",
- `experiments/2026-05-12-cross-model-authoring/rendered/json.md` exists, non-empty,
- `experiments/2026-05-12-cross-model-authoring/rendered/ailx.md` exists, non-empty,
- spot-check: open both files; JSON file's code blocks are
  ```json fences; AILX file's are ```ailx fences; the prose body is
  the same in both.

If the renderer errors on a missing example id: that's a bug in
master/spec.md (a marker references an id with no fixture). Fix the
spec.md or add the fixture, re-run.

---

## Task 8: token_balance test

**Files:**
- Create: `experiments/2026-05-12-cross-model-authoring/render/tests/token_balance.rs`

The test runs the splitter against `master/spec.md`, collects all
`{form-only: json}` block bodies into one buffer and all
`{form-only: ailx}` block bodies into another, tokenises both with
`tiktoken-rs::cl100k_base`, and asserts the totals are within ±5%.

- [ ] **Step 8.1: Write the token-balance test**

Write `experiments/2026-05-12-cross-model-authoring/render/tests/token_balance.rs`:

```rust
//! Token-balance gate for master/spec.md form-only blocks.
//!
//! The mini-spec is the experimental treatment; if the JSON-cohort's
//! form-only scaffolding is meaningfully longer than the AILX-cohort's
//! (or vice versa), the experiment is biased before the model ever
//! sees the prompts. This test catches gross imbalance.

use std::path::PathBuf;
use xmodel_render::splitter::{self, Form, Segment};

fn master_dir() -> PathBuf {
    let manifest = PathBuf::from(env!("CARGO_MANIFEST_DIR"));
    manifest.parent().unwrap().join("master")
}

#[test]
fn form_only_block_token_totals_within_five_percent() {
    let spec_text = std::fs::read_to_string(master_dir().join("spec.md"))
        .expect("master/spec.md must exist");
    let segments = splitter::split(&spec_text);

    let mut json_buf = String::new();
    let mut ailx_buf = String::new();
    for seg in &segments {
        if let Segment::FormOnly { form, body } = seg {
            match form {
                Form::Json => json_buf.push_str(body),
                Form::Ailx => ailx_buf.push_str(body),
            }
        }
    }

    let bpe = tiktoken_rs::cl100k_base().expect("cl100k_base tokenizer should load");
    let json_count = bpe.encode_with_special_tokens(&json_buf).len();
    let ailx_count = bpe.encode_with_special_tokens(&ailx_buf).len();
    assert!(json_count > 0, "JSON-form blocks produced zero tokens");
    assert!(ailx_count > 0, "AILX-form blocks produced zero tokens");

    let max = json_count.max(ailx_count) as f64;
    let diff = (json_count as i64 - ailx_count as i64).unsigned_abs() as f64;
    let ratio = diff / max;
    assert!(
        ratio <= 0.05,
        "form-only block token imbalance: json={json_count}, ailx={ailx_count}, ratio={ratio:.4} > 0.05",
    );
}
```

- [ ] **Step 8.2: Run the test**

Run:
```
cargo test --manifest-path experiments/2026-05-12-cross-model-authoring/render/Cargo.toml --test token_balance
```
Expected: `test result: ok. 1 passed; 0 failed`.

If the test fails: revise master/spec.md to lengthen the shorter
form's blocks (typically by adding examples to the shorter-form's
rules section or factually-equivalent prose). Do not pad with
filler — the rebalancing should be substantive (one form being
shorter usually means it's underexplained). Re-run.

---

## Task 9: Run renderer + commit outputs

**Files:**
- Modify (regenerate): `experiments/2026-05-12-cross-model-authoring/rendered/json.md`
- Modify (regenerate): `experiments/2026-05-12-cross-model-authoring/rendered/ailx.md`

- [ ] **Step 9.1: Regenerate rendered outputs**

Run:
```
cargo run --manifest-path experiments/2026-05-12-cross-model-authoring/render/Cargo.toml -- \
  --master  experiments/2026-05-12-cross-model-authoring/master \
  --rendered experiments/2026-05-12-cross-model-authoring/rendered
```
Expected: exit 0, both rendered files written.

- [ ] **Step 9.2: Full test sweep**

Run:
```
cargo test --manifest-path experiments/2026-05-12-cross-model-authoring/render/Cargo.toml
```
Expected: all four test suites green —
`splitter_unit` (7 passed), `spec_completeness` (1 passed),
`example_roundtrip` (1 passed), `token_balance` (1 passed).
Total: 10 passed; 0 failed.

- [ ] **Step 9.3: Confirm working-tree state**

Run:
```
git status -- experiments/2026-05-12-cross-model-authoring/
```
Expected: the entire `experiments/2026-05-12-cross-model-authoring/`
tree is untracked. The Boss will inspect and decide commit shape at
iter close; the implement step leaves everything unstaged.

---

## Out of scope for cma.1 (mirror of spec)

- `master/tasks/*.task.json` and their reference solutions — cma.2.
- `harness/` Cargo project, the HTTP client, the per-task loop,
  the location-stripping regex, scoring — cma.2.
- Live IONOS run, raw dataset emission, `runs/<date>-<hash>/` — cma.3.
- DESIGN.md §"Decision 6" empirical addendum — cma.3.
- Journal entry under `docs/journals/` summarising the run — cma.3.
- Roadmap edits removing the P2 entry — cma.3.