How a cast works

A cast turns episodes — short texts with their source metadata — into one graph per schema, plus a grounding for every assertion in it. Each episode goes through the same steps, and episodes run concurrently.

Step	What happens
Select	The text is matched against the schema's names, other names and units, and the best matches seed graflo's subschema(), which closes them into a valid slice under a budget. A schema under budget passes whole.
Prompt	The slice is rendered as a type sheet. The text follows inside a delimited block, after a sentence saying it is data, not instructions. The same episode and slice always give the same bytes.
Extract	One model call. The reply is flat: a list of vertices (each with a ref, a type, a quote and name–value properties), a list of edges between refs, and unmapped for what no type covers.
Validate	Every item is checked against the slice and the text. Values are coerced to their field type, an edge given backwards is turned round, and quotes are matched ignoring case, accents and whitespace; each fix is recorded. What is still wrong is a finding.
Repair	If there are findings, the model is shown them, grouped by kind, and asked for its whole reply again; at most twice. The attempt that keeps the most items wins.
Quarantine	What still fails is left out, with the finding and the model's reply, and written to the quarantine sink. The rest of the reply stands.
Resolve	Identities are computed, mentions become edge endpoints, and each assertion gets its grounding. See Identities and mentions.
Assemble	The episodes' graphs are merged into one graflo GraphContainer.
Write	Optionally, the graph is upserted into a database through graflo's writer. See Writing to a database.

Light by contract

One call per episode, at most two repairs, and nothing more: no critic pass, no chunking, no encoder and no vector store. An episode longer than LPGCAST_MAX_EPISODE_TOKENS is refused before any call, not split.

The checks

Finding	The item it removes
unknown_type	a vertex whose type is not in the slice
duplicate_ref	a second vertex with the same ref
identity_insufficient	a vertex whose identity the text does not give (or a mention with no key to look it up by)
undeclared_edge	an edge whose relation the slice does not declare between those types, either way round
dangling_ref	an edge to a ref no vertex has
undeclared_property	a property its type does not declare
duplicate_property	a property given twice
coercion	a value its field type refuses, or a unit other than the declared one
value_not_in_quote	a number written as digits, or a date, that its quote does not state (37.5 on "3.75 kW", 2024-03-01 on "March 2024"); spelled-out and relative values are not checked
quote_not_found	an item whose quote is not in the text
endpoint_invalid	an edge whose endpoint was removed; never shown to the model, it goes when the endpoint is fixed

A repair that returns the same findings is not repeated: the slice does not let the model fix them. A repair request that fails keeps the best earlier reply and records why in the episode's repair_failure.

A failure of the whole episode is recorded with its reason: the selection found nothing (selection_*), the text is over the ceiling (too_long), the provider rejected or dropped the request (provider_error), the reply stopped at the output limit (truncated), the model refused (refused), the reply was not JSON (unparseable) or not the reply shape (wire_shape), or nothing in it survived (nothing_valid). A truncated or refused reply is not re-asked. A model that is misconfigured, or a cache miss while replaying, stops the run instead: every other episode would fail the same way.

Beside OntoCast

OntoCast builds and evolves an ontology from documents and writes RDF; LPGCast casts short, frequent texts into a property-graph schema you already govern.