Project: publish an evidence-backed lesson knowledge graph

Create the schema

Define lesson and concept entities with immutable IDs, aliases, teaches and requires predicates, allowed source and target types, and a strict prerequisite acyclicity rule. Add claim IDs, effective and recorded times, evidence revision IDs and review status. The claim contract should make a title change harmless and a false merge reversible.

Ingest an adversarial fixture

Include a renamed lesson, two ambiguous “streams” aliases, duplicate claim delivery, an unreviewed extracted edge, a retracted prerequisite and a three-lesson cycle. Resolve identities with context, quarantine schema violations and preserve all accepted claim history. Produce a validation report with exact counts and reasons rather than dropping difficult rows.

Serve bounded answers

Implement direct and two-hop prerequisite lookup against a versioned snapshot. Return lesson IDs, relation path, claim IDs and current evidence revision. Apply a withdrawal gate to retired pages. Compare graph candidate recall with a text-only lesson search on a small judged query set, including a new isolated lesson that needs a fallback.

Submit a release packet

Provide schema, raw fixture, identity decisions, claim ledger, validation output, snapshot manifest, path traces and an error review. Rebuild the historical graph at a prior time and explain why a retracted edge was visible then but not now. A reviewer should reconstruct one served answer from recorded claims and passage versions.

Implementation

def accepted_claims(claim_rows, relation_types):
    return [claim for claim in claim_rows
            if claim["reviewed"] and relation_types.get(claim["predicate"])
            == (claim["subject_type"], claim["object_type"])
            and not claim["withdrawn"]]

Performance and operating cost

Filtering C claims costs O(C) expected time and O(A) output space for A accepted claims. Indexing, cycle validation and historical snapshots add work, but they make answer paths and corrections reproducible.

Common Mistakes

Do not publish unreviewed extracted claims as facts.
Do not erase retracted claim history.
Do not serve a withdrawn lesson from an old graph snapshot.