Groundings and quarantine

Groundings

Every mention, relation and property in a cast has a grounding: where in which text it was read. Groundings are written to groundings.jsonl beside the graph, not into it, and name the record graph's refs; resolution says which entity a ref became.

Field	Holds
source_id	the episode the assertion came from
kind	vertex, edge or property
vertex, ref	for a mention or its property: the type and the mention's ref
edge, source_ref, target_ref	for a relation or its property: the edge and its endpoints' refs
property	for a property: its name
char_start, char_end, quote	the span of the text, and its characters
in_context	set when the span is in a context item: its source_id; offsets then point into that item's text
extractor, model, schema_ref	what produced it: the SchemaCast version, the model, and name@version of the schema
reference_time	the episode's time

The model gives quotes, never offsets; SchemaCast finds each quote in the text, exactly first, then ignoring case, accents, runs of whitespace and typographic quotes and dashes. When a quote occurs more than once, a property takes the occurrence nearest its vertex's or edge's quote. A quote it cannot find is a quote_not_found finding, so every grounding points at text that is there.

Quarantine

quarantine.jsonl holds one record per item left out, per failed episode, and per unmapped entry, in episode order:

record	Means
finding	an item left out; kind is the finding, fragment the item as the model wrote it
episode	the whole episode failed; kind says why
unmapped	something the text states that no type in the slice covers

A finding record carries the model's raw reply and the number of attempts. unmapped entries show where the schema could grow. An entry whose quote is not found in the text is dropped.

In Python, any object with an async write_records(records) method can be the sink; JsonlQuarantineSink appends to a file, MemoryQuarantineSink keeps records in memory.