Knowledge-graph provenance: claim time, evidence and retraction

Last updated: 5 Oct 20265 min read

tutorial

A graph claim must carry enough history to explain when it was true, when it was known and why it should be trusted.

Keep two times

A lesson can begin requiring a concept on September 1, while an editor records the change on September 4. Effective time and recorded time answer different questions. A recommendation made on September 2 cannot use a claim recorded on September 4, even if the claim is later backdated. Availability boundaries also apply to graph facts.

Attach evidence

Store the exact lesson revision, extraction span or editorial decision that supports each claim. A path to a page is insufficient after that page changes. Evidence may be a source record or review ticket ID inside the content system; public tutorial prose need not expose internal evidence. A confidence score should identify the extractor version and calibration set if it controls automatic publication.

Retract without erasing

When a prerequisite is removed, mark the earlier claim inactive with a retraction event and reason. Do not delete it from historical snapshots used to reconstruct past recommendations. If the original claim was erroneous, record a correction that distinguishes bad historical knowledge from a legitimate curriculum change. Versioned publication prevents half-updated graphs from reaching readers.

Rehearse a correction

Claim C-47 becomes effective at minute 120, is recorded at 150, and is retracted at 240. A decision at minute 130 must not see it; one at 180 may; one at 260 must not. Add an evidence revision at 200 and verify the claim’s support points to the correct page version for each decision.

Implementation

def visible_claims(claim_rows, decision_minute):
    return [claim for claim in claim_rows
            if claim["recorded_minute"] <= decision_minute
            and claim["effective_minute"] <= decision_minute
            and (claim["retracted_minute"] is None
                 or claim["retracted_minute"] > decision_minute)]

Performance and operating cost

A scan of C claims costs O(C) time and O(V) output space for V visible claims. Indexed temporal partitions reduce repeated query work, while immutable claim events increase storage in exchange for reproducibility.

Common Mistakes

Do not backdate a newly recorded claim into an old prediction snapshot.
Do not erase retracted evidence needed for audit.
Do not confuse an erroneous claim with a relationship that later changed.