Where Did This Maintenance Claim Come From?

One document says P7 is overdue for inspection. A later log says the inspection is current. If Botie stores one unqualified status on the asset, whichever import runs last can erase the disagreement.

Keep the claims and their supporting records visible. We can then ask who supplied a claim, when it became usable and whether it still applies. Choosing evidence for review is different from declaring one source correct.

The claims include conflicting, expired, future and unapproved records.

A Claim node can contain its statement, subject and stable claim ID. Separate relationships connect it to the asset and source document. This lets several claims concern the same asset without overwriting each other.

Retain enough source detail to locate the supporting passage or inspection event. A document title alone is often insufficient when revisions change its contents. Prefer a stable document/version identifier plus a specific locator.

Claim identity keeps disagreement inspectable instead of turning import order into an evidence policy.

The requested reader may be allowed to use only particular source documents. That access check must happen before evidence is exposed. A separate quality policy decides which sources are acceptable for the task.

Our workshop receives a permitted-source list and retains only those claim records. Inclusion means the claim passes this supplied policy, not that its statement is verified. Real source permissions belong in the server-side access boundary.

The browser fixture demonstrates selection over invented data and is not an access-control system.

For this exercise, observed_at is the hour the claim became available to this decision process. valid_until is the first hour at which we stop treating it as applicable; null means no stated end. All values use one invented integer-hour clock.

At hour eight, a claim with valid_until eight has expired. A claim first available at hour nine cannot justify an hour-six decision. In real systems, keep event time, ingestion time and validity intervals separate when those meanings differ.

We require availability at or before the requested time and an end strictly after it, when an end exists.

observed_at=2
valid_until=8
for at in [7,8]:
    print(observed_at <= at and at < valid_until)

True
False

Apply the source and time predicates to claim records, then return their identities and source links. If two surviving claims contradict each other, return both for the defined review process. Do not let a dictionary keyed only by asset discard one.

A later stage may resolve conflicts using an approved policy, but that decision should leave an audit trail. A confidence score alone is not proof that an unsupported claim beats a documented one.

The query below uses a property-graph claim model inspired by provenance practice; it is not a PROV-O RDF serialisation.

MATCH (c:Claim)-[:ABOUT]->(a:Asset {id:$asset}),
      (c)-[:WAS_DERIVED_FROM]->(d:Document)
WHERE d.id IN $allowed_sources
  AND c.observed_at <= $at
  AND (c.valid_until IS NULL OR $at < c.valid_until)
RETURN c.id, c.statement, d.id
ORDER BY c.id

An extraction model can misread a sentence or associate it with the wrong asset. Record the source span, extraction version and review status so the graph statement can be checked against the original. A citation pointing to an entire unrelated manual is weak evidence.

When a source is withdrawn, investigate derived claims and downstream summaries that used it. Preserve the minimum audit evidence needed under the retention policy rather than copying sensitive source text into every log.

Provenance gives a route for correction, provided downstream results retain their dependencies.

A graph can contain stale, contradictory or incomplete evidence. The response may need to say that available records disagree, or request inspection rather than assert a status. That behaviour can be tested using deliberately conflicting fixtures.

The coding task stops at eligible claim IDs. It does not infer truth, apply natural-language contradiction detection or call a language model.

A narrow evidence-selection result is useful because its meaning can be inspected and tested.