Can You Tell Which Code Is Real?

AI code generators sometimes produce functions that look correct but silently ignore their inputs. Both paths below return the same output. One reads from a database. The other fakes it.

Path A

# Reads from actual database
def getOrders(db: IO[Database]) -> IO[List[String]]:
  match db:
    case connection:
      connection.query("SELECT * FROM orders")

Path B

# Ignores database entirely
def getOrders(db: IO[Database]) -> List[String]:
  ["Order-001", "Order-002", "Order-003"]

Both compile. Both produce a list of order strings. In a code review, Path B might slip through. But it never touches the database — the parameter db is ignored and the output is hardcoded.

Yichus Can Tell

Yichus tracks provenance through the compiler. Every value carries metadata about where it came from: which inputs flowed into it, whether it passed through IO, and whether its ancestry is dominated by literals.

carriesIO

false

The return value never passed through an IO operation. No database read happened.

literalDominant

true

The output is built entirely from literal values — no input data flows into it.

parameterUsed

false

The db parameter is declared but never referenced in the output derivation.

infoLossSignals

literal_only_ancestry, literal_dominance

Multiple provenance signals indicate the output has no connection to real data sources.

SUSPICIOUS: Literal-Only Ancestry

How Provenance Detection Works

Yichus is built on Bosatsu, a total functional language. Because Bosatsu programs are purely functional, the compiler can statically trace how every output value was derived from its inputs.

The DerivationAnalyzer walks the typed intermediate representation and builds a dependency graph for each expression. It tracks whether values flowed through IO operations, whether function parameters were actually used, and whether the output depends on literals alone.

This analysis happens at compile time — no runtime instrumentation needed. When AI-generated code silently drops inputs and substitutes hardcoded values, the provenance graph reveals the break in the data flow.

Try It Yourself

The interactive demo uses Yichus state to toggle between the two code paths and display their provenance signals.

Open the live demo →