Current

Knowledge parsers

Benchmark first

Corpora: QASPER · WikiSection · DocRED. Connectors use recorded provider fixtures because public content corpora do not test cursors, deletes, retries, or event ordering.

Protocol: Score text preservation, section Boundary F1/Pk/WindowDiff, entity/relation F1, stable-ID rate, duplicate rate, and incremental/full-sync equivalence. For polling and streaming, replay the same create→update→delete trace with injected duplicates, reordering, throttling, and cursor expiry; the final snapshot and emitted change set must agree.

Models return JSON-like values. Parsers resolve all evidence against supplied document and section revisions and return immutable typed values. Research establishes each task formulation; Mari implements a deterministic validation boundary rather than the cited model.

How it works

Each parser first requires the recipe’s top-level collection, then validates every required field and enum, resolves evidence through the exact contract below, derives deterministic audit signals, and constructs frozen result types. It never repairs a claim’s meaning. Batch claim assessment is the exception to fail-fast parsing: rows are keyed back to caller order, absent rows become uncertain, and individually malformed rows do not erase valid siblings.

model proposal*→schema→exact evidence→revision binding→*typed candidate

Parser	Produces	Research-backed task	Academic sources
`parse_facts`	`FactCandidate`	Atomic claims and optional document-level relations with evidence	FActScore · DocRED
`parse_claim_assessments`	`FactAssessment`	Supported, contradicted, or uncertain verdicts; decisive rows require evidence	FEVER
`parse_decisions`	`DecisionCandidate`	Decision-related utterance extraction without treating topical language as proof	Hsueh & Moore · Karan et al.
`parse_answer`	`GroundedAnswer`	Evidence-selected document QA, citations, or explicit insufficient evidence	QASPER · ALCE
`parse_answer_candidates`	`AnswerCandidate[]`	Reusable question-answer pairs bound to supporting passages	QASPER
`parse_glossary`	`GlossaryCandidate[]`	Term-definition relations, aliases, and source spans	DeftEval
`parse_digest`	`DigestSummary`	Overall and topic summaries with separately inspectable evidence	QAGS · SummaC
`parse_impact`	`ImpactAssessment`	In-scope affected-document proposals followed by deterministic dependency checks	Mari contract; no claimed benchmark reproduction
`parse_refinement`	`RefinementEdit[]`	Bounded, attribution-aware, fact-preserving edit proposals	RARR · FactEditor

answer.py

from mari_components.knowledge import parse_answer

raw = model(question, documents)
answer = parse_answer(question, documents, raw)
print(answer.disposition)         # grounded | insufficient_evidence
print(answer.grounding_coverage)  # deterministic text coverage
print(answer.evidence[0].quote)  # exact source text

Additional deterministic helpers include normalize_claim, deduplicate_fact_candidates, grounding_coverage, and excerpt. Recoverable batch drift is handled conservatively: assessment rows are restored to caller order, missing rows become uncertain, and good rows survive alongside invalid ones. Structured fact qualifiers preserve subject, relation, object, scope, validity, and conditions.

**grounding_coverage is not entailment or confidence.**It is a Mari-specific lexical audit signal motivated by citation completeness. Exact quote validation prevents fabricated citations but does not prove that every paraphrase follows logically from its evidence.