Skip to content

Systems and records

Supplier performance data provenance: the events behind every score

By SourceX Editorial · Reviewed by Noah Loul ·

10 sections

Short answer

A supplier performance score should travel with the events it was computed from: purchase order lines with the promised dates in force, goods receipts, inspections and rejections, invoice matches, change notices and approval status, plus the formula version and time window. Supplier performance data provenance means anyone can recompute the score from those linked records and see which event moved it.

Provenance trail

Supplier performance data

This page covers supplier performance data provenance: the events behind every score.Illustrative

Key takeaways

  • A supplier score is a derived record; ship it with its input events, formula version, window and run details.
  • Computed scores such as rolling on-time rates change as windows slide and dates are edited, so snapshots need manifests to be reproducible.
  • PROV-O, OpenLineage, EPCIS and the Data Provenance Standards supply vocabulary for the manifest.
  • Constraints such as approval scope, required sources and change notices explain when the top-scoring supplier was correctly passed over.
  • Automated checks should recompute every score from its events and hold any score that fails.

What underlying events should accompany a supplier performance score?

Every score should ship with the records that produced it and the rule that turned them into a number. For a delivery score that means each purchase order line, the promised date that applied when the supplier confirmed, and each receipt with its date and quantity. For a quality score it means inspections, rejections and complaints. For a price score it means the purchase order price and the invoiced price.

Two further items are easy to forget and decide whether the number can be trusted: the formula version, including weights and exclusions, and the time window the score covers. Without them, the same supplier can show two different scores from the same data and nobody can say which is right.

Treat the score as a derived record, not a fact. It was generated by a calculation, at a time, from specific inputs, and a provenance manifest is simply the list of those inputs and that calculation written down so it can be checked.

Why a score without its events cannot be checked or reused

ERP scores are usually weighted roll-ups that hide their inputs. SAP's MM Vendor Evaluation, for instance, scores vendors from 1 to 100 on main criteria of price, quality, delivery and general service, lets each criterion carry its own weight, and computes the overall score from the weighted criteria. Its automatic subcriteria draw on existing data such as on-time delivery, quantity reliability, adherence to the confirmation date, goods receipt inspections and complaint levels.

Other KPIs compare two documents. SAP's supplier evaluation price KPI is calculated from the difference between the invoiced amount and the purchase order net amount for an item. If the purchase order was amended after the invoice, the KPI depends on which version the calculation read.

Odoo shows the window problem in code. Its On-Time Delivery Rate is a computed field: the quantity received on or before the promised date divided by the quantity ordered, over purchase lines from a trailing period set by a system parameter that defaults to 365 days, with negative one meaning no data. Because it is recomputed rather than stored, the number changes as the window slides and as promised dates are edited, and a snapshot taken last quarter cannot be reproduced from today's database alone.

Provenance models worth borrowing for supplier scores

You do not need a new standard; existing ones already describe derived records, run events and physical supply chain events. Borrow their vocabulary so engineers, auditors and data buyers read the manifest the same way.

W3C PROV-O models provenance with entities, activities and agents: an entity wasGeneratedBy an activity, an activity used entities, and one entity wasDerivedFrom another. A score is an entity generated by a scoring activity that used receipts and order lines. OpenLineage describes jobs, datasets and runs, and requires at least a START event and a COMPLETE, FAIL or ABORT event for each run. GS1's EPCIS organizes physical supply chain events by what, when, where and why. The Data & Trust Alliance's Data Provenance Standards group dataset metadata into Source, Provenance and Use.

Provenance models worth borrowing for supplier scores
ModelCore ideaUse in a score manifest
W3C PROV-OEntities generated by activities, derived from other entities, associated with agentsScore as entity; scoring run as activity; events as used entities; owner as agent
OpenLineageRun events for jobs that read input datasets and write outputsLog each scoring run with start and completion, inputs read and the output table written
GS1 EPCISEvent records with what, when, where and why dimensionsShape receipt and inspection events with item, event time, location and business step
Data Provenance StandardsSource, Provenance and Use metadata for datasetsDescribe the delivered score dataset, its date range and permitted use

What should the score to source event provenance manifest include?

The manifest is one record per published score, with links down to every event that fed it. The fields below are the minimum that lets someone recompute the score and challenge it.

At a glance

What should the score to source event provenance manifest include

Store the manifest next to the score, not in a separate audit system that a data extract will leave behind. When the score moves to a warehouse, a report or a licensed dataset, the manifest goes with it.

Keep the manifest small by storing references, not copies. The input_events field holds record IDs and types; the records themselves stay in the ERP, the quality system and the invoice archive. What the manifest must ensure is that each reference still resolves, so retention rules for purchase orders, receipts and inspections need to be at least as long as the period over which scores are kept.

Version the formula separately from the scores. A change of weights, rounding or the promised-date basis creates a new formula version with an effective date, and scores computed before and after the change are never compared without saying so.

What should the score to source event provenance manifest include?
FieldWhat it holdsFormatValidation rule
score_id and supplier_idThe published score and the supplier it ratesStable IDs from the source systemSupplier ID resolves to one approved-supplier record
score_value and criterionThe number and what it measures, such as on-time deliveryNumber plus criterion nameRecomputation from listed events returns the same value
formula_versionCalculation rule, weights, rounding and exclusionsVersion ID plus effective dateVersion existed on computed_at
window_start and window_endPeriod whose events countISO 8601 datesEvery listed event falls inside the window
computed_at and run_idWhen the score was produced and by which runTimestamp plus job run IDNo listed event is recorded after computed_at
input_eventsIDs of order lines, receipts, inspections, invoices and notices usedList of typed record IDsEach ID resolves; none duplicated
promised_date_basisWhich promised date counts: original, confirmed or latest revisionEnumerated valueSame basis applied to every line
excluded_eventsEvents left out and why, such as customer-caused delaysList of IDs with reason codesEach exclusion has an approver and reason
constraints_in_forceSpec revision, approval status and change notices on each orderLinks with effective datesNo order line falls outside the supplier's approved scope
owner and approverWho owns the formula and who approved exceptionsRole namesPresent for every exclusion and formula change

Which events feed each criterion, and what they change

List the event types per criterion so data engineers know what to extract and reviewers know what to ask for. The table reflects common ERP and quality records; your system's names will differ.

Capture each event as it happened, not as the current row shows it. A purchase order line edited three times holds only its last values; the confirmation and every date change are separate events and need their own timestamps and authors. Where the ERP keeps a change log, extract from it. Where it does not, the supplier's confirmation emails or portal messages are often the only record of the original commitment, and the manifest should say so.

Which events feed each criterion, and what they change
Event typeSource recordKey fields to keepScore it affects
Order line confirmedPurchase order and supplier confirmationLine ID, quantity, need date, confirmed date, spec revisionDelivery baseline
Promised date changedOrder change history or confirmation updateOld and new date, who changed it, reasonDelivery, if the basis is the latest date
Goods receivedGoods receipt or stock moveReceipt date, quantity, order line link, locationDelivery and quantity reliability
Inspection resultQuality inspection lot or receiving inspectionLot, sample size, result, defect codeQuality
Rejection or returnNonconformance report or return to vendorQuantity rejected, cause, dispositionQuality and complaints
Invoice matchedSupplier invoice and match resultInvoiced price and quantity, order version matchedPrice
Change noticeEngineering change or supplier change noticeAffected parts, effective date, approvalContext for exclusions and constraints
Approval status changeApproved supplier list or qualification recordStatus, scope, effective datesEligibility, not the score itself

Constraints that overrule a high supplier score

A high score does not make a supplier the right choice for a given order, and a model trained on scores alone will learn that it does. The manifest should carry the constraints that overrule the score, so the record shows why the top-scoring supplier was passed over when it was.

  • The supplier is not approved for the part, process or specification revision on the order.
  • The customer contract or drawing names a required source, so no other supplier is eligible.
  • An open change notice or quality hold blocks new orders for the affected part.
  • The supplier's quoted lead time cannot meet the need date, however good its history.
  • The best score came from a different part family or plant, so it says little about this order.
  • The score's window holds too few events to support a comparison.

Checks that prove the manifest matches the score

Run these checks automatically each time a score is published, and keep the result with the manifest. A score that fails any check is held, not published with a warning.

  • Recompute the score from input_events using formula_version and confirm the same value within the stated rounding.
  • Resolve every event ID to a record that existed on computed_at, and flag any created or edited afterward.
  • Confirm every event date falls inside the window and every in-window event for the supplier is either used or listed as excluded.
  • Confirm one promised_date_basis was applied to all lines, and that changed dates carry who and why.
  • Confirm each exclusion has a reason code and an approver role.
  • Confirm the run is logged with a start and a completion event, the inputs read and the output written.

Worked example: a passing manifest and a failing one

This example is illustrative; the supplier, counts and dates are invented. A machined-parts supplier gets an on-time delivery score for the four quarters ending in June. The window holds 412 order lines, 398 receipts and 6 excluded lines where the buyer moved the need date. The formula is version 3, using the confirmed promised date.

The passing manifest lists all 412 line IDs, the receipts linked to them, the 6 exclusions each with a reason code and the planner who approved it, and run ID and timestamps from the scoring job. Recomputing from those records returns the published value, and every event predates the run.

The failing manifest for the same supplier a month later shows the same score, but the validator finds 23 promised dates edited after the receipts arrived, all moved later. Under the confirmed-date basis those lines were late; under the latest-date basis they look on time. The score is held, the basis mismatch is recorded, and the edits are kept as their own events, because the pattern of late date changes is itself a useful signal.

The exception in this example is a supplier whose window holds only 9 lines. Its score passes every check but is flagged as low-volume, and procurement compares it only with suppliers that have similar volume.

When a supplier score should stay out of a training or evaluation set

Some scores cannot be backed by their events, and publishing them as data spreads the error. Decide per supplier and per period before anything is extracted.

Bring in the procurement owner of the formula when the basis or weights changed mid-window, a quality engineer when inspection records are incomplete, and a data rights reviewer before supplier names, prices or contract terms leave the company.

  • Drop scores whose input events were purged by a retention rule and cannot be recovered.
  • Split scores whose window spans a formula version change into two periods, each with its own manifest.
  • Hold scores that fail recomputation until the cause is found.
  • Exclude or mask supplier pricing where a supplier agreement or NDA makes it confidential.
  • Mark low-volume scores rather than mixing them with well-supported ones.

Supplier records as licensed AI data

Scores with their events, constraints and exceptions show how procurement decisions are actually made, which is what makes them useful for AI work. They also carry other companies' information. Before external use, separate system access, what supplier agreements and confidentiality terms permit, who owns each record and what a license would grant; removing names does not by itself create a right to license.

Manufacturers and distributors can turn selected supplier records into a licensed AI dataset with SourceX handling rights review, contracting, preparation, delivery and payment. Bulky deliveries stay in the company's own storage or move on encrypted drives, and qualification is settled first by a brief assessment for which no files are sent.

Frequently asked questions

What is supplier performance data provenance?

It is the record of where a supplier score came from: the events it used, the formula and window that turned them into a number, when it was computed and who approved any exceptions. With it, anyone can recompute the score and see which events moved it. Without it, the score is an unverifiable opinion.

Should the manifest use the original or the latest promised date?

Choose one basis, write it into the formula version and apply it to every line. The confirmed date usually reflects what the supplier committed to, while the latest date can hide lateness if dates are moved after the fact. Keep date changes as events either way so the choice can be revisited.

How does this differ from a supplier scorecard?

A scorecard presents scores and reviews them with the supplier. The manifest sits underneath it and proves each score from source records. A scorecard can be built on top of manifests, but a manifest cannot be rebuilt from a scorecard.

Can supplier scores be licensed without supplier consent?

It depends on the supplier agreements, confidentiality terms and the data involved, and counsel should review specific cases. Scores often embed pricing and quality findings that suppliers shared under confidentiality. Masking supplier identities helps but does not settle the contractual question.

What if the ERP recomputes scores and keeps no history?

Start writing a manifest with each published score from now on, and store the input event IDs with it. For past periods, rebuild what you can from purchase order, receipt and inspection history, and mark those scores as reconstructed.

Sources

  • SAP MM Vendor Evaluation scores vendors from 1 to 100 on main criteria of price, quality, delivery and general service, with weights per criterion combining into an overall score. Source
  • SAP's automatic subcriteria include on-time delivery, quantity reliability, compliance with shipping instructions, confirmation date adherence, goods receipt and quality audit and complaints levels. Source
  • SAP's supplier evaluation price KPI is calculated from the difference between the invoiced amount and the purchase order net amount for an item. Source
  • Odoo's On-Time Delivery Rate is computed as products received on time divided by products ordered over a trailing period set by a system parameter defaulting to 365 days, using the promised date, with a negative value for no data. Source
  • The OpenLineage spec defines jobs, datasets and runs and requires at least a START event and a COMPLETE, FAIL or ABORT event per run. Source
  • PROV-O links entities, activities and agents through relations such as used, wasGeneratedBy and wasDerivedFrom. Source
  • The GS1 EPCIS and CBV implementation guideline describes events along what, when, where and why dimensions. Source
  • The Data & Trust Alliance Data Provenance Standards group dataset metadata into Source, Provenance and Use standards. Source

See if your company qualifies

A short company assessment. No data uploads are needed.

See if your company qualifies