karenina.benchmark.trace

trace

Public helpers for formatting stored structured traces.

Classes

Functions

abstention_detection_instruction

abstention_detection_instruction() -> str

Return Karenina's standard instruction for explicit abstention checks.

Source code in src/karenina/benchmark/trace.py

def abstention_detection_instruction() -> str:
    """Return Karenina's standard instruction for explicit abstention checks."""
    return ABSTENTION_DETECTION_SYS

format_trace_messages

format_trace_messages(
    messages: Iterable[Message | Mapping[str, Any]],
) -> str

Format stored message records using Karenina's canonical trace format.

Parameters:

Name	Type	Description	Default
messages	Iterable[Message | Mapping[str, Any]]	Live Message objects or dictionaries from a stored verification result's template.trace_messages field.	required

Returns:

Type	Description
str	The canonical raw trace, including assistant text, tool calls, and tool
str	results.

Source code in src/karenina/benchmark/trace.py

def format_trace_messages(messages: Iterable[Message | Mapping[str, Any]]) -> str:
    """Format stored message records using Karenina's canonical trace format.

    Args:
        messages: Live ``Message`` objects or dictionaries from a stored
            verification result's ``template.trace_messages`` field.

    Returns:
        The canonical raw trace, including assistant text, tool calls, and tool
        results.
    """
    hydrated = [message if isinstance(message, Message) else Message.from_dict(dict(message)) for message in messages]
    return messages_to_raw_trace(hydrated)