A model that can't say “I don't know” will make something up.
DVLang is a programming language where not knowing is a first-class result. Every value is either known, with its grounds, or abstained, with a reason. The compiler won't let either side of that go unhandled, whether the code was written by a person or a model.
- Abstaining is a typed return value
- Unhandled abstention is a compile error
- Small, regular syntax
The RLHF “Golden Retriever” problem
Large language models are next-token predictors. In fine-tuning (RLHF), human raters reward helpfulness and penalize uncooperative answers. Over millions of cycles the model learns a dangerous heuristic: “Refusing a prompt is bad. Providing an answer is good.” DVLang moves the answer to that problem out of the model's training and into the language.
Three ideas
Not knowing is a value
Functions return known results or reasoned abstentions. Both are typed, both are checked.
Every claim has a source
Known values can carry their grounds, so a reader can trace an answer back to where it came from.
Small, regular syntax
Few keywords and one idiom per task. Easier for people to review, and fewer ways for a model to drift.
Stop begging the LLM to follow your orders.
Today, rules live in prose: capital letters, repeated warnings, output parsers and retry loops. In DVLang the rules are types. The compiler enforces them, and a model has nothing to talk its way around.
prompt = """You are a helpful assistant.
ALWAYS respond in valid JSON.
NEVER invent order IDs.
If you are not sure, say "unknown".
PLEASE follow the schema EXACTLY."""
for attempt in range(5):
out = llm.invoke(prompt + user_input)
try:
data = parser.parse(out)
break
except OutputParserException:
prompt += "\nInvalid JSON. Try again."// The schema is the type.
fn lookup(id: OrderId) -> Order? validate strict {
know orders[id] as o => o
else => abstain "unknown order {id}"
}
// Not sure is a value, not a failure.
// Invalid output never reaches the caller.
// No retries. No capital letters.instructions → types
Shape, allowed values and required fields are declared once, in the signature.
refusal → return value
abstain is a typed outcome, so “I don't know” is something the caller handles.
validation → opt-in
Add validate to any function that needs it. Results are checked before they leave.
Teach the machine to say no.
The language is a v0.1 draft. The interpreter is not released yet; this site documents the design as it stands.