A programming language you write in plain, canonical English — and that compiles deterministically to Python.

E-- ("English--") is English with the ambiguity removed: a closed grammar and a fixed vocabulary, with exactly one canonical phrasing per construct. It is meant to read and edit like English while still compiling to ordinary, reproducible Python.

LLM-generated code is fuzzy at runtime: non-reproducible, expensive per call, hard to debug. E-- separates the LLM's role from execution — LLM (optionally) writes canonical E-- at authoring time; a deterministic parser compiles the E-- to Python; runtime is pure. Best of both worlds: LLM creativity when you need it, deterministic behavior forever after.

Install from PyPI:

pip install e-minus-minus

Transpile a canonical E-- file to Python:

emm-transpile examples/describe.emm

That's it. No LLM required for canonical E-- with no {{ }} slots. See "Running E--" below for more, and "Resolving {{ }} slots" for the LLM setup when you use free-English input or value slots. The LLM path is optional and lives behind the [llm] extra: pip install "e-minus-minus[llm]".

Developing on E-- itself? Clone the repo and invoke the CLI as a module while your working tree is on PYTHONPATH:

PYTHONPATH=src python -m e_minus_minus.transpiler examples/describe.emm

E-- is licensed under Apache License 2.0 (see LICENSE) — permissive, with an explicit patent grant, so it can be embedded in commercial products freely.

Two clarifications:

• The license covers the E-- tooling. The Python that E-- generates is yours — the output is not encumbered by this project's license.
• The LLM is your own. E--'s normalizer and {{ }} resolution require a language model that you supply; that provider's terms are separate from this project.

Programmatic API:

from e_minus_minus import transpile

python_source = transpile(emm_source)

transpile() is pure: no network, no side effects. Pass a resolve_slot callable to handle {{ ... }} slots (see docs/spec.md and the CLI implementation in src/transpiler.py for the injected-resolver pattern).

E-- is a two-stage pipeline, split so that the unreliable part and the deterministic part never mix:

Free English  --LLM (transpile-time)-->  Canonical E--  --plain parser-->  Python

• Normalizer (LLM, optional). Turns free-form English into canonical E--. This is the only stage that deals with linguistic ambiguity.
• Compiler (deterministic). Turns canonical E-- into Python with an ordinary parser — no LLM, fully reproducible and debuggable.

The LLM runs only at transpile time, never at runtime. Generated Python is always pure and self-contained. The LLM is never allowed to decide program structure; it is used only to fill clearly-delimited value slots written as {{ ... }}, and those resolutions are cached so builds stay reproducible.

Canonical E--:

Set result to [[fibonacci]]( {{the first prime number greater than 5}} ).
Do [[print]](result).

compiles to:

result = fibonacci(7)
print(result)

Markers keep it unambiguous: [[name]] is a function call, a bare word is a variable, "x"/3 are literals, <1, 2, 3> is a list, and {{ ... }} is an English phrase the transpiler resolves once and bakes in.

E-- source files use the .emm extension (English--). The deterministic canonical-to-Python core is implemented; you can transpile and run .emm files from the command line.

Given this canonical E-- source at examples/describe.emm:

Define [[describe]] taking n:
    If n is greater than 10:
        Give back "big".
    Give back "small".

For each n in <3, 42, 7>:
    Do [[print]]([[describe]](n)).

transpile it and print the generated Python to your screen:

python3 src/transpiler.py examples/describe.emm

prints:

def describe(n):
    if n > 10:
        return "big"
    return "small"
for n in [3, 42, 7]:
    print(describe(n))

Write the generated Python to a file instead of the screen:

python3 src/transpiler.py examples/describe.emm -o out.py

Transpile and run it, so you see the program's actual output:

python3 src/transpiler.py examples/describe.emm --run

prints:

See the generated Python and run it in one go with --show (alias -s):

python3 src/transpiler.py examples/describe.emm --run --show

prints the code and its output, separated by comment lines:

# --- generated Python ---
def describe(n):
    if n > 10:
        return "big"
    return "small"
for n in [3, 42, 7]:
    print(describe(n))
# --- output ---
small
big
small

The delimiters are Python comments, so the whole block stays copy-pasteable. --show on its own (without --run) just prints the Python, like the default.

Notes:

• The .emm extension is the convention for E-- source files.
• {{ ... }} LLM value slots are runnable — see "Resolving {{ }} slots" below for the one-time setup. Files with no slots (like examples/describe.emm) need no key and --run works with no model.

A {{ ... }} slot is an English phrase that the transpiler resolves to a Python expression once, at transpile time, using an LLM — then caches the result so later builds are offline and reproducible. Files with no {{ }} slots need no API key and no setup.

To run a slot example end to end:

# 1. create and activate a virtual env
python3 -m venv .venv && source .venv/bin/activate

# 2. install dependencies (the Anthropic SDK)
pip install -r requirements.txt

# 3. set your Anthropic API key
export ANTHROPIC_API_KEY="sk-ant-..."

# 4. transpile and run a slot example
python3 src/transpiler.py examples/primes.emm --show --run

examples/primes.emm is minimal:

For each p in {{the first five prime numbers, as a Python list}}:
    Do [[print]](p).

which transpiles to:

for p in [2, 3, 5, 7, 11]:
    print(p)

and prints 2 3 5 7 11 (one per line). The {{ ... }} slot resolved once via the LLM to the concrete list [2, 3, 5, 7, 11], was cached, and gets used for every subsequent build with no API call.

The first run calls the model (Anthropic Haiku) to resolve each slot, writes the resolved Python expression to .emm_cache.json, and bakes it into the output. Every later run is an offline cache hit — no model call, identical result. The cache file maps the exact slot text to its resolved expression and is meant to be committed, so resolved values stay diffable and reviewable.

Editing a slot's text is a cache miss and re-resolves; deleting the cache forces full re-resolution. Files without {{ }} slots (like examples/describe.emm) never touch the API.

You can also build a slot example inline without any file setup:

printf 'Set year to {{the current year, as an integer literal}}.\nDo [[print]](year).\n' > hello.emm
emm-transpile hello.emm --show --run

The slot at line 1 is at expression position (inside Set year to ...), so the LLM returns a single Python expression — e.g. 2026 — and the compiler splices it in:

Prints 2026. Second run is offline (cache hit).

A {{ ... }} slot can appear at a statement position, not just an expression position — putting it on its own line, at a block's indentation, delegates one or more Python statements to the LLM. Author writes the surrounding structure; the LLM fills the region.

Define [[summarize]] taking numbers:
    {{ compute mean, median and count of numbers into mean_v median_v count_v }}
    Do [[print]](count_v).
    Do [[print]](mean_v).
    Give back mean_v.

At transpile time the statement slot resolves to real Python:

def summarize(numbers):
    from statistics import mean, median
    mean_v = mean(numbers)
    median_v = median(numbers)
    count_v = len(numbers)
    print(count_v)
    print(mean_v)
    return mean_v

Trade-off: [[wikilinks]] or callable references inside a code-slot's resolved Python are opaque to any downstream tool that inspects the E-- source. Author knowingly accepts DAG invisibility inside code-slot regions in exchange for region-level delegation. Use expression slots when you need graph visibility; use code slots when you're delegating structure the LLM knows better than you do.

examples/code_slot_example.emm is a runnable code-slot demo:

Define [[summarize]] taking numbers:
    {{ compute the mean, median and count of numbers into named variables mean_v median_v and count_v }}
    Do [[print]](count_v).
    Do [[print]](mean_v).
    Do [[print]](median_v).
    Give back mean_v.

Set data to <2, 3, 5, 7, 11, 13>.
Do [[summarize]](data).

Transpile and run it:

emm-transpile examples/code_slot_example.emm --show --run

The statement slot on line 2 resolves to real Python that binds mean_v, median_v, count_v — e.g. via from statistics import mean, median plus three assignments. Because it's a code slot, it can add the import on its own line, which a value slot cannot.

Code slots let you delegate imports. A value slot at expression position can only emit a single Python expression, so getting the current year with a value slot produces the awkward __import__('datetime').datetime.now().year. A code slot at statement position sidesteps the constraint:

printf '{{ import datetime and set x to the current year }}\nDo [[print]](x).\n' > hello.emm
emm-transpile hello.emm --show --run

The LLM resolves the statement slot to:

import datetime
x = datetime.datetime.now().year
print(x)

Prints the current year. Same syntax as a value slot; position determines shape.

You don't have to write canonical E-- by hand. The transpiler's first phase normalizes free-English E-- into canonical E-- with an LLM, then compiles the canonical form to Python — one input, two outputs. An English source (examples/describe_en.en) reads like prose:

Define a function called describe that takes a number n. If n is greater than
ten, give back the string "big". Otherwise, give back the string "small".
Then, for each n in the list 3, 42 and 7, print describe of n.

Normalize it to canonical and run the result, saving the canonical form too:

python3 src/transpiler.py examples/describe_en.en --canonical-out out.em --run

out.em holds the canonical E-- (equivalent to examples/describe.emm) and the program prints small / big / small.

Two properties make this safe and cheap:

• The parser is the canonical-detector. Whether a file "is already canonical" is decided by trying to parse it deterministically — no LLM, no heuristic. An already-canonical file needs no API key: normalization short-circuits before any model call. Only genuinely English input hits the model.
• Fixed point + cache. Feeding the canonical output (out.em) back in parses as canonical, so Phase 1 does nothing and reproduces the same outputs. Normalizations are cached in a committed .emm_norm_cache.json (keyed by source text), so re-running English input is an offline cache hit. Setup is the same as for slots: pip install -r requirements.txt and export ANTHROPIC_API_KEY=....

Normalization and {{ }} slot resolution are independent, separately cached LLM touchpoints — a canonical file with all slots cached makes zero live calls.

Early design. The language is specified in docs/spec.md. The deterministic canonical-to-Python core (lexer, parser, emitter) is implemented with a runnable CLI — see "Running E--" above — and {{ }} slot resolution is wired up (Anthropic Haiku + a committed cache; see "Resolving {{ }} slots"). The LLM normalizer (free English → canonical) is wired up at whole-file granularity (see "Writing in free English"); per-region normalization is the next refinement.