What the documents look like
Fresh-food concession counters in hypermarkets are often run by on-site supplier staff under a sales-share arrangement. Copy drafting for these counters therefore involves the hypermarket fresh department, the concession supplier and the store manager. The material behind a draft comes from the POS system, supplier delivery notes, the food-safety ledger and store stock counts. Daily documents include the concession contract, the daily sales report, loss sheets and food purchase inspection records. These materials are updated daily or per batch, so product name, specification, origin and sale status in a draft can change with the batch.
The most time-consuming work at the copy-drafting step is reconciling weighed sales, handling loss allocation and checking purchase credentials. The role of forms and interaction at this step is to collect the fields that need checking, then pass them to later nodes or to human judgement. The nodes themselves do not act as a business database.
Constraints at the interactive forms step
When the form input node is triggered, the conversation enters an interactive state, the workflow state is recorded, and execution waits for the user to complete required information and submit. After submission, the collected form information is passed to later nodes. The user selection node also enters an interactive state and executes a branch according to the user's choice. Neither node provides its own business database function. Field validation, business approval, write-back and batch operations need to be implemented in an external system or in later nodes.
Whether a draft sentence is supported by the material depends on whether the material text states the same subject, the same period and the same field. Retrieval similarity is not a category, entity or permission filter, and retrieval itself is not authorization. Libraries built on different vector models cannot be associated together; the model identifier must be checked first and non-matching libraries excluded. Cross-model retrieval is not run to compare missed recalls.
Hypothetical paper exercise sample: the material text reads "The concession contract states that Supplier A has operated chilled pork in the fresh area of Store A since 1 March 2025, under a sales-share arrangement." "The daily sales report of 5 March 2025 states that chilled pork sales at this counter on that day were 8,600 yuan." "The food purchase inspection record of 5 March 2025 states that the quarantine certificate number for the chilled pork arriving that day was A-0305." The draft sentence to be assessed is "This counter's chilled pork is slaughtered and sold fresh daily, and quarantine-passed." The check compares "slaughtered and sold fresh daily" in the draft against the material text. The material states only the operating subject, the share method, the day's sales amount and the quarantine certificate number. It does not state slaughter timing or frequency, so that phrase is unsupported. "Quarantine-passed" corresponds to the certificate number in the inspection record and is a checkable item, but the batch the certificate number points to must still be confirmed as the same as the product the draft refers to.
Configuration and implementation checks
| Configuration item | Setting or check content | Relation to this page's documents |
|---|---|---|
| Form input node | Enters interactive state on trigger, records state and waits for submission | Collects the material fields needed to check draft sentences |
| Required-field list | Lists subject, object, period, field and material source | Without all items, a same-subject same-period comparison cannot be completed |
| User selection node | Executes a branch according to the choice; not a personnel directory | Used to choose between continuing the check and marking for verification |
| Interactive state handling | Execution continues after the interaction completes | No business approval or write-back occurs while waiting |
| Downstream division of work | Field validation, approval and write-back are implemented by an external system | The nodes do not perform database writes |
| Retrieval scope check | Check the vector model identifier first and exclude non-matching libraries | Libraries on different vector models cannot be associated together |
The hypothetical starting value for this exercise is five required fields. That number is chosen because the sample needs one entry each for subject, object, period, field and source to complete one same-field comparison. It is compared item by item against the field list; if any item is missing, the gap is recorded.
Three common errors
- Treating "slaughtered and sold fresh daily" in the draft sentence as already supported by the material. The material text states only the operating subject, the share method, the day's sales amount and the quarantine certificate number. It does not state slaughter timing or frequency, so the phrase is unsupported and the evidence gap must be recorded.
- Using retrieval similarity in place of a permission judgement. Similarity is not a category, entity or permission filter. A retrieval result does not mean the record may be used by the current subject; the authorization scope must be checked separately.
- Treating the user selection node as a business data selector. The node only executes a branch according to the choice. It does not provide a personnel directory or a business record selector. Field validation and approval need to be implemented in later nodes or an external system.
How to verify the setup
- Input the three material sentences and the draft sentence "This counter's chilled pork is slaughtered and sold fresh daily, and quarantine-passed" from the hypothetical sample. Compare each assertion in the draft against what the material states for the same subject, same period and same field. The expected observation is that "slaughtered and sold fresh daily" is marked unsupported, and "quarantine-passed" is marked checkable but requiring confirmation that the batch matches. If every assertion is judged supported, the check was not performed field by field.
- Input a form submission that lacks a material source, and compare it against the required-field list. The expected observation is that the missing item is identified and kept as a gap, and later nodes do not produce a definite conclusion from it. If the missing item is filled in automatically or passed through, the required-field check did not take effect.
- Input a retrieval request that spans different vector models, and compare the library model identifiers against the request scope. The expected observation is that non-matching identifiers are excluded and the actual observation is recorded, with no cross-model association performed. If merged results are returned directly, the model identifier check was not executed.
All of the above is a paper check. No test, deployment or approval is claimed to have taken place.