AI uses for records
Same split, four reasons: what a split sourcing decision dataset keeps
By SourceX Editorial · Reviewed by Noah Loul ·
9 sections
Short answer
A split sourcing decision dataset stores, for every requirement divided among suppliers, the reason as dated evidence: a capacity confirmation, a concentration or set-aside rule, a dual-sourcing decision, or the quote whose price break made splitting cheaper. Quantities cannot recover it. In an illustrative calculation, a capacity limit, a 60% share cap, a 30% floor and a stock-lot price all produce the same 600/400 split.
Evaluation split
Same split, four reasons
Key takeaways
- Capacity, policy, risk and price can each produce the identical allocation, so a split's reason cannot be inferred from its quantities.
- In the illustrative test, three different constraints yield the same 600/400 split at the same 3.7% premium; only outside evidence separates them.
- SAP quota arrangements, Oracle sourcing rules and the OCDS Lots schema store shares and caps, but none has a field for why.
- Public rules such as FAR 6.202 and 19.502-3 turn split reasons into signed or published documents; private records rarely do.
- A split that costs more with no recorded constraint, made of orders just under an approval limit, is the threshold-avoidance signature.
Why would a buyer split one requirement across suppliers?
Four legitimate reasons explain most splits, and one illegitimate reason hides among them. A buyer splits a requirement because the preferred supplier cannot deliver all of it by the need date, because a written policy caps any one supplier's share or reserves a portion for a class of supplier, because the company keeps a second source qualified on purpose, or because a price break or lot offer makes two orders cheaper than one. The illegitimate reason is breaking a requirement into smaller orders so that each stays under an approval limit.
For an AI team, the split is the easy part to see. Purchase orders show two suppliers, two quantities and two prices. The decision behind them is what an agent has to learn before it can propose an allocation, explain one to an auditor or flag one that slipped past a control, and that decision is rarely stored next to the orders. A split sourcing decision dataset exists to keep it attached.
This page shows why the reason cannot be read off the quantities, what common purchasing records keep and drop, which public procurement rules force the reason into writing, and a decision packet and constraint matrix that make split episodes usable for training and evaluation. It stays on the allocation decision itself; substitutions, disruption tradeoffs and approved-vendor changes are separate questions.
Four reasons, one 600/400 split: the allocation test
The same allocation can come from entirely different reasons, so a model trained on quantities alone learns a guess. We computed the cheapest feasible way to buy one illustrative requirement, 1,000 units needed by day 30, from two qualified suppliers under five different recorded constraints. Every input is invented and labeled as such, and prices are index points rather than currency.
Supplier P quotes all-units price breaks: 100 per unit below 500 units, 95 from 500 and 92 from 800. Supplier Q quotes 98 below 400 units and 96 from 400. With price as the only rule, the answer is a single order of all 1,000 units to P at 92.
Each reason was then added on its own. If P can confirm only 600 units by day 30, the cheapest feasible plan is 600 from P and 400 from Q, 3.7% above the single order. A policy capping any supplier at 60% of a requirement gives the same 600/400 split at the same cost. A risk rule that keeps each qualified supplier at 30% or more also lands on 600/400, because Q's 400-unit break makes buying 400 from Q cheaper than buying 300. And if Q offers a stock lot of exactly 400 units at 85, all or nothing, the 600/400 split becomes 1.1% cheaper than the single order.
Four reasons, one split, and three of them cost the same to the index point, so even the price paid cannot separate them. Only evidence outside the order lines can: P's capacity confirmation, the policy clause, the dual-sourcing decision or Q's lot quote. The table lists every scenario with the evidence that would prove it, and the figure traces the four reasons into the one allocation.
| Scenario | Recorded constraint | Cheapest feasible split, P/Q | Cost vs single order | Evidence that proves the reason |
|---|---|---|---|---|
| S0 Price only | None beyond the two quotes | 1,000/0 | Baseline | Both quotes with their break tables |
| S1 Capacity | P confirms at most 600 units by day 30 | 600/400 | +3.7% | P's dated capacity confirmation against the need date |
| S2 Policy cap | No supplier above 60% of one requirement | 600/400 | +3.7% | The policy clause in force on the order date |
| S3 Risk floor | Each qualified supplier keeps at least 30% | 600/400 | +3.7% | A dual-sourcing decision naming the part, the risk and its owner |
| S4 Price | Q offers a 400-unit stock lot at 85, all or nothing | 600/400 | -1.1% | Q's lot quote and the cost comparison |
| S5 Threshold split | Three orders to P, each under a 35,000-point approval limit | 340, 330 and 330 to P | +8.7% | None legitimate: together the orders exceed the limit |
What purchasing systems and procurement standards keep about a split
Mainstream purchasing records keep the split rule, not the reason for it, and that gap is what the dataset has to close.
SAP's quota arrangements divide a material's requirement among sources by quota. When splitting is active, SAP's help portal gives the order quantity per source as the source's quota multiplied by the requirement quantity, divided by the total of all quotas, and a lot must reach a minimum quantity before it is split. An arrangement can therefore say 60 and 40. It cannot say whether 60 is P's capacity ceiling, a policy cap or a risk decision.
Oracle's supply chain planning documentation treats sourcing rules the same way: each source carries a rank and an allocation percentage, and the percentages of the sources at one rank must total 100. Again the record is a share and a priority, which is enough to repeat an allocation and not enough to explain it.
Public procurement data has the same hole. The Open Contracting Data Standard's Lots extension, at its September 2026 commit, defines 38 fields, including a cap on how many lots one supplier may be awarded (maximumLotsAwardedPerSupplier) and links from awards, items and documents to lots. Searching every field name, title and description for reason, rationale, justif or why returns no match. The schema can show that a tender was divided and capped; it has no place to say why.
| Record | What it stores about the split | What it leaves out | What the dataset must add |
|---|---|---|---|
| SAP quota arrangement | A quota per source, a splitting formula and a minimum lot before splitting | Why each quota was set and when it last changed | A reason code and an evidence link for each quota version |
| Oracle sourcing rule | Rank and allocation percentage per source, totaling 100 within a rank | Whether a percentage reflects capacity, policy or risk | The constraint behind each percentage, with effective dates |
| OCDS Lots extension | Lots, award-to-lot links and caps on lots bid or awarded per supplier | Any field for the reason a tender was divided | The procurement document that states the reason |
| Purchase orders | Supplier, quantity, price and date for each order | That two orders answer one requirement | A requirement identifier shared by every resulting order |
Where the rules turn the reason into a written record
Public procurement forces several split reasons into documents, which shows what a private record ought to hold. The U.S. Federal Acquisition Regulation lets an agency exclude a source in order to establish or maintain an alternative source when doing so would, among other listed grounds, ensure the continuous availability of a reliable source or satisfy projected needs based on a history of high demand. Every such action needs a determination and findings signed by the head of the agency or a designee, and it cannot be made on a class basis (FAR 6.202).
Policy splits are written into the solicitation. When a contracting officer partially sets aside an acquisition for small business, the requirement must be divisible into distinct portions, and the solicitation must identify which portions are set aside and which are not (FAR 19.502-3). In the EU, Article 46 of Directive 2014/24/EU asks a contracting authority that decides not to divide a contract into lots to state its main reasons, and lets it cap the number of lots one tenderer can win if the cap appears in the notice (a paraphrase of the official text).
Private companies carry the same concern without the paperwork. Apple's fiscal 2025 Form 10-K, for example, tells investors that the company obtains certain components from single or limited sources and that this creates significant supply and pricing risks (paraphrased from the filing). The risk is written down in the annual report, yet the purchase order that splits a part between two suppliers seldom points to it.
The FAR text above is the codified version in GSA's repository at commit da52ccb, dated 30 March 2026. FAR subpart 1.4 allows agencies to operate under class deviations, so the text that governs a given purchase can differ; check any deviation that applies before quoting a rule as current.
The split decision packet and constraint matrix
One episode is one requirement, not one order. The packet ties every order that answered the requirement to the constraint that shaped the allocation and to the evidence that existed on the decision date, so a reviewer can tell a recorded reason from one reconstructed afterward.
Twelve elements are enough to grade a model's reasoning. They are listed below, followed by the constraint matrix that tells a reviewer what proves each reason, what disproves it and where the evidence usually sits:
- Requirement: item, total quantity, need date and one requirement identifier shared by every order it produced.
- Candidate suppliers on the decision date, with qualification status and the quote then in force, including its break table.
- Allocation as decided, quantity per supplier, plus any later change with its date and approver.
- Stated reason code: capacity, policy cap, set-aside, risk floor, price or lot, or other with free text; more than one code is allowed.
- Evidence for each reason as a dated document: capacity confirmation, policy clause, dual-sourcing decision or quote.
- Counterfactual: the cheapest allocation without the stated constraint, computed from the quotes, and its cost.
- Premium or saving of the chosen split against that counterfactual, as a percentage.
- Approver of the allocation and the approval limit that applied to each order.
- Threshold check: the requirement's combined value against every limit its individual orders fell under.
- Outcome per supplier: quantity received by the need date, late quantity and rejected quantity.
- Later events that tested the reason, such as a shortfall at the capped supplier or a disruption at the second source.
- Rights and confidentiality status of supplier names, quotes and capacity messages before any outside use.
| Reason | Evidence that proves it | Evidence that disproves it | Where it usually sits |
|---|---|---|---|
| Capacity | Supplier confirmation dated before the order, for less than the requirement | A confirmation dated after the split, or one covering the full quantity | Order acknowledgments, supplier portal messages, email |
| Policy cap or set-aside | A clause in force on the order date that the allocation satisfies | A clause adopted later, or an allocation that breaks it | Procurement policy, solicitation documents |
| Risk floor | A dual-sourcing decision naming the part, the risk and its owner | A second source never qualified, or kept below any useful share | Category strategy, risk register |
| Price or lot | Quotes showing the split costs less than a single order | Quotes showing a cheaper single order and no other constraint | RFQ responses, quote files |
| Threshold avoidance | Same supplier and requirement, orders under a limit that together exceed it | One order, or a documented reason for staging deliveries | Order history, approval logs |
Sourcing decision or a way around an approval limit?
Treat a split that costs more and carries no recorded constraint as a possible control bypass until the record shows otherwise. FAR 13.003(c)(2) states the rule for federal buyers: do not break down requirements that add up to more than the simplified acquisition threshold or the micro-purchase threshold into several purchases below it merely to permit simplified procedures or to avoid a requirement that applies above the micro-purchase threshold. Private approval policies usually carry the same idea under other names.
Scenario S5 shows the signature. Issuing the 1,000 units to P as three orders of 340, 330 and 330 keeps each one under a 35,000-point approval limit, but none of them reaches P's 500-unit break, so the requirement costs 100,000 points, 8.7% more than one order. Same supplier, same requirement, orders just under a limit and a lost price break: that combination is the test a dataset should let a model run.
The opposite error matters as much. The cheapest action, one order to P, would have failed in S1, because P could not deliver 1,000 units by day 30 and the saving would have become a shortage. A dataset that labels every premium as waste teaches the reverse of the threshold mistake. The counterfactual and premium fields exist so a model can say what a split cost and whether a recorded constraint justified the cost.
Grading a model on split episodes
Grade reasoning against the evidence in the packet, not against the allocation the buyer happened to choose. Report results by reason class and by evidence type, because a model can do well on price splits, which leave quotes behind, and badly on risk splits, which often leave only an email.
At a glance
Grading a model on split episodes
Five tasks separate a model that understands splits from one that pattern-matches quantities:
- Explain: given the orders and the evidence, name the reason and cite the document that proves it. Score it correct only when both match; a right label without evidence scores as a guess.
- Discriminate: given four episodes with identical 600/400 splits and different evidence, label each. A model that gives all four the same label has learned the quantities, not the decision.
- Flag: given order histories, find threshold splits. Score precision and recall against the threshold check, and require the model to name the limit and the combined value.
- Allocate: given quotes, constraints and a need date, propose the cheapest feasible allocation and state its premium over the unconstrained answer, scored against the computed counterfactual.
- Abstain: when evidence is missing, say so and name the document that would settle it instead of supplying a plausible reason.
Limits of this design, and requesting split episodes
The allocation test is deliberately small: two suppliers, all-units price breaks, one need date and no freight, lead-time spread or quality difference. Real splits often combine reasons, such as a capacity limit that also happens to satisfy a policy cap, which is why the packet allows several reason codes, each with its own evidence. The FAR and EU rules apply to public buyers; private policies vary and are often unwritten, so every reason in the dataset should be labeled as recorded at the time, linked from another document, or inferred later.
Supplier names, quotes and capacity messages are commercially sensitive and frequently covered by supplier agreements. Masking supplier identities does not settle whether a company may license those records; that turns on its agreements and on counsel's review of the specific facts.
SourceX finds datasets for a request rather than holding them in stock. A team that needs split episodes can describe them at /request-data: the systems involved (ERP purchasing, quota or sourcing-rule history, supplier portals), the years, the reason codes required, the share of episodes that must carry linked evidence and an acceptance test, for example the discrimination task above run on a held-out sample. Whether any company's records can pass that test is decided in review, and nothing here implies that such records are already available.
How the allocation test and schema search were run
The allocation test is a short Python script, split_allocation.py, kept with this article's working files. It tries every allocation of the 1,000 units in 10-unit steps, prices each supplier's quantity at the all-units break it reaches, keeps only the allocations each scenario's constraint allows and reports the cheapest; ties on cost would be printed, and none occurred. Costs are compared as integers, and each premium is rounded half up to one decimal place once. S5 is not an optimization: it divides the requirement into the fewest orders, in 10-unit steps, that each stay under the limit.
The schema search, ocds_lots_fields.py, reads release-schema.json from the OCDS Lots extension at commit 819fa3d, collects every field under its properties and definitions, and matches the substrings reason, rationale, justif and why, ignoring case, in field names, titles and descriptions. It found 38 fields and no match. No records of real companies were used. The SAP, Oracle, EU and Apple statements come from search excerpts of those publishers' own sites and are labeled as paraphrases where the excerpt was not verbatim.
Frequently asked questions
What is split sourcing?
Split sourcing means buying one requirement from two or more suppliers at the same time, usually in planned shares. It differs from switching suppliers, which replaces one source with another, and from a partial delivery, where one supplier ships a single order in pieces.
Can a model infer why an order was split from the quantities?
Not reliably. In this article's illustrative test, a capacity limit, a 60% policy cap, a 30% dual-sourcing floor and a stock-lot price all produced the same 600/400 split, three of them at identical cost. The reason has to come from dated evidence linked to the orders.
How do you spot a split purchase made to avoid an approval limit?
Group orders by supplier and requirement over a short window, add up their values and compare the total with each approval limit the individual orders fell under. Orders that sit just under a limit, together exceed it and lose a price break deserve review. A documented reason, such as staged deliveries, can clear them.
Is dual sourcing always more expensive than a single source?
No. Splitting often gives up a volume break, as in the 3.7% example here, but a supplier's lot offer or break structure can make the split cheaper, as in the 1.1% saving example. Compute the cheapest single-source and split options from the actual quotes before calling a premium waste.
Which records hold the evidence behind a split decision?
Quota or sourcing-rule history holds the shares, and purchase orders hold quantities and prices. The reasons usually sit elsewhere: supplier acknowledgments and portal messages, RFQ files, category strategies, risk registers and approval logs. A usable episode links them through one requirement identifier.
Sources
- FAR 6.202 lets an agency exclude a source to establish or maintain an alternative source on listed grounds, including continuous availability of a reliable source, and requires a signed determination and findings not made on a class basis. Source
- FAR 19.502-3 requires a divisible requirement for a partial small business set-aside and requires the solicitation to identify which portions are set aside and which are not. Source
- FAR 13.003(c)(2) forbids breaking down requirements above the simplified acquisition or micro-purchase threshold into smaller purchases merely to use simplified procedures or avoid a requirement. Source
- FAR 1.404 provides for class deviations from the FAR, so the text governing a purchase can differ from the codified version. Source
- Directive 2014/24/EU Article 46 asks authorities that do not divide a contract into lots to state their main reasons and allows a cap on lots awarded to one tenderer if stated in the notice (paraphrase). Source
- SAP's splitting quota arrangement computes each source's order quantity as its quota times the requirement divided by the total of all quotas, and a lot must reach a minimum quantity before it is split. Source
- Oracle's sourcing rules give each source a rank and an allocation percentage, with percentages for the sources at one rank totaling 100 (paraphrase). Source
- The OCDS Lots extension schema defines lots, links from awards and items to lots, and caps on lots bid or awarded per supplier, with no field for the reason a tender was divided (field audit). Source
- Apple's fiscal 2025 Form 10-K says it obtains certain components from single or limited sources, creating significant supply and pricing risks (paraphrase). Source
- SourceX sources datasets on request rather than holding them in stock, and buyers describe scope, volume, format, timing and licensing needs through its request form. Source
Tell us what data you need
Share scope, volume, format, timing and licensing needs. Datasets are sourced on request.