What the documents look like
Short-video affiliate sales rely on long-term content exposure and product-link conversion rather than live explanation. Copy review therefore centers on four document types: short-video scripts, attached-product lists, commission plans, and violation feedback. Source material comes from the selected-affiliate alliance, short-video playback data, product transaction data, and platform rules. Content operators, creator business development staff, and product distribution specialists handle the same document set on daily and weekly cycles.
Whether a promotional sentence in draft copy holds depends on whether the corresponding document states the subject, object, status, and effective range. A field the document does not state remains unknown. A more complete record in another category does not support treating this category as lacking that field.
Constraints at the workflow routing step
Workflow routing here connects document review with copy drafting into a checkable sequence. The Form Input node is a user-interaction node: when triggered, the conversation enters an interaction state, the workflow state is recorded, required information is collected, and after submission the collected information passes to subsequent nodes. The User Selection node is also a user-interaction node: when triggered, the conversation enters an interaction state, the workflow state is recorded, and after the user makes a choice the node evaluates that choice and executes the "yes" branch. Branch connections must be verified one by one. The recorded interaction state is behavior of the interaction nodes themselves; no separate global state-recording switch is documented.
Collected information must be explicitly passed to later retrieval, judgment, or interface nodes, and outputs must be compared item by item against inputs. Order writing, document backfill, and external status synchronization require a separate business interface or downstream process; the nodes themselves do not perform these business actions. Semantic recall and relevance scores in a retrieval library do not carry permission or entity-isolation duties. The same-vector-model requirement applies only to libraries that can be associated together, and does not extend to external systems.
Hypothetical paper exercise: the source material reads, "A skincare brand ran short-video attached-link promotion for its facial cleansing mousse from March 1 to March 31, 2025, commission rate 15%, and materials must carry the note 'actual results vary by individual.'" The draft sentence to be judged reads, "This cleansing mousse carries 15% commission this month, suits all skin types, and needs no effect disclaimer." Judgment requires the same subject, same period, and same field. The commission rate matches the source at 15%. The effective range covers only March; if "this month" in the draft points to another period, that portion is unsupported. "Suits all skin types" has no corresponding field in the source and is unsupported. "Needs no effect disclaimer" is a mutually exclusive value on the same field as the source's disclaimer requirement and is a contradiction.
Configuration and implementation checks
| Configuration item | Type | Relation to this page's task |
|---|---|---|
| Form Input node | Documented official behavior | Enters interaction state when triggered, collects required information, passes it to later nodes |
| User Selection node | Documented official behavior | Executes the "yes" branch after the user chooses; verify branch connections one by one |
| Interaction-state check | Implementation check | Recorded state is interaction-node behavior; no global state switch is claimed |
| Output handoff check | Implementation check | Compare collected information item by item against inputs of later retrieval, judgment, or interface nodes |
| External action split | Implementation check | Order writing, document backfill, and external synchronization require a separate business interface or downstream process |
| Hypothetical trial starting point: commission-rate field comparison value | Hypothetical starting point | This paper exercise uses the source value of 15% as the comparison value to judge same-field consistency of the draft sentence; not an official default |
Three common errors
- Treating an unstated field as a negative conclusion. The draft sentence "suits all skin types" has no corresponding field in the sample source. The correct conclusion is unsupported, not that the source denies the claim.
- Extending interaction-node behavior into global control. Form Input and User Selection nodes record state and enter interaction state. Inferring a global state switch or submission-count limit from this has no documentary basis.
- Letting retrieval perform permission judgment. Inferring that a record is invisible or a business type is excluded from a relevance score confuses similarity with authorization boundaries. Retrieval does not constitute authorization.
How to verify the setup
- Input the sample source and the draft sentence, then compare the commission-rate field. The source states 15% and the draft states 15%, so the same field is consistent. If the draft states another number, record a contradiction. The observed result is a field-level comparison record; no actual test is claimed.
- Input the sample source and the draft sentence, then compare the effective-range field. The source states March 1 to March 31, 2025. If "this month" in the draft points to that period, the field is consistent; if it points to another period, that portion is unsupported. The source states neither "all skin types" nor "no disclaimer needed," so both are recorded as unsupported, and the disclaimer requirement forms a mutually exclusive value with the source and is recorded as a contradiction.
- Input the paper value [] and check the downstream node design. Confirm the downstream has an explicit branch for empty input; do not claim that a fabricated rule retrieval necessarily returns []. Also check the identifiers of libraries that can be associated together, exclude those with inconsistent vector models, and record actual observations. Do not run cross-model retrieval to compare missed recalls.