grounding

Documentation

Overview

Package grounding is the harness's free, deterministic quality check: did the model invent values not present in the source? No inference, sub-millisecond string ops. It is the keystone label generator for the self-learning loop (conformal calibration, health monitoring, router training all train on it).

Index

• func Check(task core.TaskType, input string, data []byte) (grounded bool, ok bool)
• func CheckFields(task core.TaskType, input string, data []byte) (ungrounded []string, ok bool)

Functions

func Check

func Check(task core.TaskType, input string, data []byte) (grounded bool, ok bool)

Check reports whether structured output is grounded in the source input. ok=false means grounding does not apply (classify/triage are grammar-pinned to a label set, nothing to verify; or the output has no checkable values).

• extract: every leaf string/number value must appear in the source (extraction is verbatim — a value not in the source is a hallucination).
• summarize: every NUMBER in the summary must appear in the source (the clearest invented-fact signal; entity paraphrasing is intentionally not flagged to avoid false positives).

The pipeline ACTS on a false result only for extract (retry/escalate); for summarize it is logged as a quality signal but not actioned.

func CheckFields

func CheckFields(task core.TaskType, input string, data []byte) (ungrounded []string, ok bool)

CheckFields is the per-field variant of Check for extract: it returns the list of TOP-LEVEL field names whose value is not grounded in the source. ok=false means grounding does not apply (non-extract task). Nested values are attributed to their top-level key. Used to name offenders in a targeted corrective re-prompt (lightweight atomic-claim verification).