Skip to content

Provenance-Weighted Event Reconciliation

A process — instantiates Event-Log-Centered Modeling

Resolves conflicting, duplicate, and late event claims by weighting each by the trustworthiness of its source, while keeping the disagreement on the record.

When an event log ingests happenings from many sources, they disagree: two systems report the same event with different details, a late arrival contradicts what was already recorded, one sensor says a value another denies. Provenance-Weighted Event Reconciliation is the process that produces a coherent picture without erasing the disagreement. Its move is to reconcile competing event claims by the provenance of each — how the claim was sourced, by whom, through what path, with what reliability — rather than by recency or arrival order alone. Its defining discipline, and what sets it apart from a merge-and-overwrite, is that reconciliation never deletes the losing claims: the conflict and the residual uncertainty stay on the record, so a downstream reader sees both what was concluded and that it was contested. It answers "which account do we trust, and how sure are we?" — leaving the losers visible for when the trust ranking later proves wrong.

Example

A hospital assembles a patient's medication history from several event streams: the pharmacy dispensing system, a discharge-summary feed from another hospital, and the patient's own app-reported doses. They conflict — the discharge feed says the patient is on 10 mg of a drug; the pharmacy record and a recent app entry say 5 mg. A naïve "latest wins" would take whichever event arrived last, which happens to be the stale discharge summary. Provenance-Weighted Event Reconciliation instead weights the claims by source reliability: a verified pharmacy dispense outranks a free-text discharge summary, which outranks self-report. The reconciled view shows 5 mg — but flags the entry as contested, with the discharge feed's 10 mg preserved and linked, and an explicit uncertainty note.

That preserved conflict is the point. A week later the pharmacy record is found to have been entered against the wrong patient; because the reconciliation kept the losing claim and its provenance rather than overwriting it, the correction is a matter of re-weighting, not archaeology. The clinician always saw a single actionable number and a visible warning that it rested on a resolved conflict — never a false, clean certainty.

How it works

  • Attach provenance to every claim. Each competing event carries where it came from, through what path, and a reliability characterization — the raw material reconciliation weighs.
  • Detect the conflict set. Group claims that purport to describe the same underlying happening, including duplicates and late arrivals that contradict what is already recorded.
  • Weigh, don't overwrite. Resolve to a working answer by provenance weight (and defined tie-breaks), without discarding the claims that lost.
  • Record the residue. Emit the reconciled view plus an explicit record of the conflict and the leftover uncertainty, so the answer travels with its own caveat and can be re-derived if the weights change.

Tuning parameters

  • Provenance weighting scheme — how source reliability is scored and ranked. Steep hierarchies resolve conflicts crisply but entrench trust in a source that may be wrong; flat schemes stay humble but leave more unresolved.
  • Conflict-detection sensitivity — how similar two claims must be to count as describing the same happening. Loose matching catches more real conflicts and risks reconciling genuinely-distinct events; tight matching is safe but lets contradictions slip through as separate facts.
  • Late-arrival policy — how far back a late or out-of-order event may reopen an already-reconciled conclusion. A wide window keeps the record honest as stragglers arrive; a narrow one keeps conclusions stable but can freeze in an early, wrong answer.
  • Uncertainty exposure — how loudly residual conflict is surfaced downstream (silent flag, visible caveat, blocking hold). More exposure protects against false certainty; less keeps consumers simple at the cost of hidden doubt.

When it helps, and when it misleads

It is essential wherever the log has many partly-trusted mouths and the truth is contested — multi-source records, sensor fusion, intelligence and reporting, any integration where "who says so?" matters as much as "what was said." Its signature virtue is refusing false certainty: it delivers a usable answer while keeping the disagreement legible, so a later correction is cheap.

The deep failure mode is that provenance weights can launder bias: rank a systematically-wrong-but-"authoritative" source above a right-but-"informal" one and the process manufactures confident, sourced errors that are hard to challenge precisely because they look adjudicated.[1] The tempting misuse is to collapse the conflict into a single clean value and drop the losing claims — which destroys exactly the audit trail that makes the weighting revisable, and turns an honest "contested 5 mg" into a brittle "5 mg." The disciplines are to keep every losing claim and its provenance on the record, to treat the reliability ranking as a revisable hypothesis rather than ground truth, and to propagate the residual uncertainty downstream instead of hiding it behind a tidy answer.

How it implements the components

Provenance-Weighted Event Reconciliation realizes the archetype's contested-evidence slice — trusting, adjudicating, and keeping disagreement visible — not the storage, derivation, or timing slices:

  • provenance_record — the provenance carried by each claim is both the input it weighs and the evidence it preserves for every claim, winners and losers alike.
  • uncertainty_and_conflict_record — it emits the conflict and the leftover doubt as first-class output, so a reconciled value never poses as an uncontested one.
  • late_and_out_of_order_event_policy — its rule for when a straggler may reopen a settled conclusion is exactly this policy, applied through the provenance lens rather than by arrival order.

It reconciles competing source events against each other; it does not check a projection against the log — that opposite-facing reconciliation is Projection Rebuild and Diff — and it does not itself issue corrections as new events, which is Compensating-Event Correction.

Notes

Reconciliation resolves which recorded claim to trust; it does not create new truth. When the right response is to assert a correcting fact rather than re-weigh existing ones, that belongs to Compensating-Event Correction, which appends a correction event rather than adjudicating between rivals. The two compose: reconciliation surfaces a conflict, a compensating event resolves it in the log's own append-only terms.

References

[1] Provenance (equivalently, data lineage) is the recorded origin and derivation path of a piece of information. It is what lets a reconciliation be by trust rather than by recency — and its abiding risk is that an authoritative-looking origin can carry a systematic error, so provenance weighting must stay a revisable judgement, not a verdict. ↩