ail-interpreter 1.75.0

AIL — Python interpreter for the AI-Intent Language. fn (pure) + intent (LLM) declarations, provenance, purity contracts, confidence-aware match, calibration

AIL — An AI-Authored Programming Language

A programming language where AI is the programmer and humans are the stakeholders.

Install

pip install ail-interpreter
# or: pip install 'ail-interpreter[anthropic]'   # for the Anthropic adapter

The PyPI distribution is ail-interpreter — this wheel is the Python interpreter of AIL, not the language itself. The canonical spec lives in spec/ and a second interpreter lives in go-impl/. (ail and ailang are both unavailable on PyPI for naming-policy reasons.) The Python import name is ail:

from ail import run, ask

and the CLI is ail.

---

The idea

AIL has two kinds of functions:

pure fn word_count(text: Text) -> Number {
    return length(split(text, " "))
}

intent classify(text: Text) -> Text {
    goal: positive_negative_or_neutral
}

entry main(review: Text) {
    words = word_count(review)       // pure fn — no LLM, confidence 1.0
    label = classify(review)         // intent — LLM call, confidence ≤ 1.0
    return join([label, " (", to_text(words), " words)"], "")
}

pure fn is for what the AI can compute — sorting, parsing, arithmetic. Deterministic, no LLM call, no side effects, statically checked.

intent is for what the AI needs to reason about — sentiment, summarization, translation. The runtime dispatches to a language model and returns a (value, confidence) pair.

The language distinguishes the two at declaration time, so you always know which calls are free and deterministic and which need a model.

---

Two ways humans use it

ail ask — the natural interface

export AIL_OLLAMA_MODEL=llama3.1:latest   # or ANTHROPIC_API_KEY=...
ail ask "Count the vowels in 'Hello World'"
# 3

ail ask "factorial of 7" --show-source
# 5040
# (stderr) --- AIL ---
# (stderr) pure fn factorial(n: Number) -> Number {
# (stderr)     if n <= 1 { return 1 }
# (stderr)     return n * factorial(n - 1)
# (stderr) }
# (stderr) entry main(x: Text) { return factorial(7) }

Human types English. An LLM writes AIL. The runtime executes it. The human sees the answer. The AIL is transparent infrastructure — inspectable on demand, invisible by default.

ail run — for programs written explicitly

ail run examples/fizzbuzz.ail --input "20" --mock

When you want to read or write AIL yourself.

---

What's in v1.8

Feature	What it does
pure fn	Statically verified — no intents, no effects, no impurity leaks
intent	LLM-backed, returns (value, confidence)
Provenance	Every value carries its origin tree; queryable via origin_of, lineage_of, has_intent_origin, has_effect_origin
Calibration	Confidence recalibrates from observed outcomes. calibration_of("intent") introspectable from code
attempt	Confidence-priority fallback cascade — cheap pure try first, LLM only as fallback
match	Pattern matching with confidence guards — "positive" with confidence > 0.9 => ...
Parallelism	Independent intent calls run concurrently with no async/await
Effects	perform http.get(url), perform file.read(path), etc.
Evolve	Intents can self-modify (retune, rewrite constraints) with rollback and history
ail ask	Natural-language → AIL authoring loop with parse-error retry

Adapters for Anthropic, Ollama (local), and Mock (tests) ship built-in. A second interpreter in Go (see the project repo) runs the same .ail files with no Python installed at all.

---

Python API

from ail import run, ask, AskResult

# Direct program run:
result, trace = run("path/to/program.ail", input="hello")
result.value        # the entry's return value
result.confidence   # calibrated confidence
result.origin       # full provenance tree

# Natural-language interface:
r = ask("compute the factorial of 7")
r.value             # 5040
r.ail_source        # the AIL the author produced
r.retries           # 0 if first try parsed

See the language reference card for the complete surface.

---

Why this exists

Humans don't write AIL. Humans say what they want in natural language; an LLM writes AIL; the runtime executes it; the result comes back.

The value of the language is in what it guarantees about the code the AI writes:

• Every pure computation is statically separated from every LLM call.
• Every value carries the full chain of operations that produced it.
• Confidence is a first-class runtime property, recalibrated by observation.
• A pure fn that passes parsing is proven to contain no LLM call, no side effect, and no path to one. The model cannot slip an intent past the compiler.
• Independent LLM calls parallelize without the author writing async.

For the full design rationale, see the spec.

---

License

Apache 2.0.