Skip to content

symfonic.evals.book_assertions

book_assertions

Closed-corpus assertions for the versioned Glass Harbor fixture.

The target opts in to grounding evidence attributes. Source identity, routes, counts and supersession metadata are payload-free; normalized response claims are content-bearing and must stay in access-controlled local evidence:

retrieved_source_ids Stable source identifiers admitted to the model input. response_claims Normalized fact/value annotations parsed from the actual response. derivation_inputs_used Stable inputs used for an arithmetic or logical derivation. retrieved_source_records Source id, rank, status and supersession metadata; never source text. admitted_sources_by_route Source ids that reached the prompt through graph or vector retrieval.

Negative-grounding targets additionally expose unsupported_fact_ids and unsupported_claim_count. This keeps a fluent refusal from passing when a response parser still found an invented proposition.

BookClaimGrounded dataclass

BookClaimGrounded(claim: BookClaim, retrieved_attribute: str = 'retrieved_source_ids', assertions_attribute: str = 'response_claims', derivation_attribute: str = 'derivation_inputs_used', name: str = 'book_claim_grounded')

Require answer truth, admitted/cited provenance and normalized safety.

ClosedCorpusGrounded dataclass

ClosedCorpusGrounded(fact: NegativeFact, marker: str = 'NOT IN MANUAL', unsupported_attribute: str = 'unsupported_fact_ids', unsupported_count_attribute: str = 'unsupported_claim_count', retrieved_attribute: str = 'retrieved_source_ids', response_claims_attribute: str = 'response_claims', name: str = 'closed_corpus_grounded')

Require an explicit unsupported verdict with no invented proposition.

SupersessionResolved dataclass

SupersessionResolved(current_source: str, obsolete_source: str, records_attribute: str = 'retrieved_source_records', authority_attribute: str = 'authority_source_ids', name: str = 'supersession_resolved')

Require both conflicting sources and a machine-readable authority edge.