What the documents look like
Contract and transaction material comparison for marketplace advertising and on-site traffic works across four document families: delivery plans, keyword reports, audience package descriptions, and conversion attribution tables. Delivery pricing adjusts in real time against traffic and conversion, and the underlying materials update daily or hourly, so the same delivery plan can carry different bid, budget, and attribution values on different dates. Comparison therefore has to lock the same entity, the same period, and the same field before any difference between materials can be judged as a contradiction. Workflow routing connects material collection, field checking, human confirmation, and downstream actions; interaction nodes pause the flow at defined points and collect the values the later comparison depends on.
Constraints at the workflow routing step
A form input node enters the interaction state when triggered, records the workflow state, and passes the information the user submits onward after the required fields are completed. A user selection node also enters the interaction state when triggered, and after the user makes a choice the flow executes the corresponding branch. The interaction state is node behaviour; no separate global state-recording switch is documented. Values collected at these nodes must be passed explicitly to later retrieval, judgment, or interface nodes, and each output is compared item by item against its input. Order writes, document backfill, and external status synchronization require a separate business interface or a later process; retrieval and forms do not constitute authorized execution. Libraries built on different vector models cannot be associated together, so identifiers are checked first and non-matching items are excluded rather than running cross-model retrieval to compare missed recall. Optional paths shown in an overall workflow diagram do not mean every request executes all of them.
Configuration and implementation checks
The table below is built around a hypothetical paper exercise sample. Sample content: a hypothetical delivery confirmation sheet states "Entity: Xinglan Commerce; Object: Xinglan Plan – Spring keyword package; Status is pending review by operations lead; Valid range: 1 March 2025 to 31 March 2025." A separate draft sentence awaits judgment: "Xinglan Commerce has obtained delivery execution authorization for the Xinglan Plan – Spring keyword package." The draft sentence is compared against the status field of the source text.
| Configuration item | Hypothetical starting point for this exercise | Reason and comparison method |
|---|---|---|
| Required form input fields | Entity, object, status, valid range | Aligns with the four fields of the sample source text; submitted values are checked one by one |
| User selection branches | Yes / no, two branches | Verifies that the corresponding branch executes after a choice, per the documented behaviour |
| Interaction state recording | Node default behaviour enabled | Checks whether the conversation enters the interaction state after triggering |
| Output receiving node | Comparison judgment node | Passes the collected form values into that node for item-by-item comparison |
| Retrieval scope identifier | Xinglan Plan – Spring keyword package | Checks whether the library identifier and the vector model identifier match |
| External action separation | Separate business interface | Checks whether writes and synchronization are carried by a later process |
Three common errors
- Treating a pending-review status as granted authorization. The sample source text records the status as "pending review by operations lead," while the draft sentence states "has obtained delivery execution authorization." These are mutually exclusive values on the same status field, which makes them contradictory. The error is overwriting the source status with the draft wording.
- Treating a retrieval hit as permission confirmation. A retrieval return shows semantic relevance only; it does not establish entity authorization or entity isolation. The error is substituting a relevance score for an authorization check.
- Treating interaction nodes as automatic execution. Form input and user selection enter the interaction state when triggered and continue after the user completes the interaction; they neither judge nor submit on their own. The error is assuming the node completes approval or writes by itself.
How to verify the setup
- Input the sample source text and the draft sentence, then compare the status field. The source reads "pending review by operations lead" and the draft reads "has obtained delivery execution authorization." The two are mutually exclusive, so the result is a contradiction. If the draft sentence instead reads "pending review by operations lead," the status values agree.
- Input one retrieval request and observe the returned results together with the library identifier. Check whether the vector model identifier of each returned item matches the target library, and exclude items whose identifiers do not match. If the return is empty, record it as an empty result and do not infer either lack of permission or absence of records.
- Input the required form fields and submit, then check the values received by the downstream node. Verify whether entity, object, status, and valid range are passed one by one; if any item is missing, record the gap. The expected outcome is that all four are passed, and no claim is made that this was actually tested.