What the documents look like
Producer export tax processes combine the offset of domestic sales tax with the calculation of refundable tax, which differs from the handling path for trading-company export refunds. That difference shapes the material available when drafting copy. Export contracts, customs declarations, the exemption-offset-refund declaration schedule, and sales ledger entries together form the basis for judging a draft sentence. Any promotional line that mentions a refund amount, refund progress, or tax preference needs to be checked against those materials for the subject, the object, and the valid scope.
Data updates monthly. When a draft cites a figure or a status, it must state the period it belongs to. Figures and statuses from different periods cannot be mixed.
The four document types carry different roles. The declaration schedule records the filing stage of the exemption-offset-refund calculation. The export contract and customs declaration establish the export event and its declared value. The sales ledger records the internal booking. A draft sentence that treats one stage as another, or that treats a filing status as a completed refund, is not supported by these materials.
Constraints at the context limits step
During drafting, the model must hold the source material, the proposed sentence, and the judging requirement at the same time. The model context is checked against the actual model capability. The reply limit controls only the length of a single answer; the two cannot substitute for each other. The knowledge reference limit is controlled by tokens. When a single material passage is long, it may be truncated, and the question sample should be used to confirm whether the key sentence entered the context in full. The memory turn count affects the input content, so the same question under empty history and under history with prior turns produces different inputs; the check should fix one of the two. When a fixed reply format carries a downstream process, field completeness and parse success need separate verification; text length is not a passing criterion.
Hypothetical paper exercise sample. the source material reads "Party A is a producer enterprise. In June 2024 it exported a batch of self-produced goods. The customs declaration number matches the sales ledger record. The exemption-offset-refund declaration schedule has been prepared by the tax accountant, and its status is pending declaration." Proposed sentence: "This enterprise has fully completed the refund on its exported goods." Judging requirement: the material shows only that the declaration schedule is in pending-declaration status and does not state that the refund was completed. The proposed sentence is not supported and the evidence gap should be recorded.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Reason and comparison object |
|---|---|---|
| AI model | A model that supports a long context and thinking configuration | Compared with a model that supports only a short context, to confirm the source material can enter in full |
| Reply limit | Set long enough to hold the basis and the conclusion | Compared with a truncated reply, to confirm the conclusion sentence is not cut off early |
| Knowledge reference limit | Set by tokens to cover the full sample material | Compared with a smaller value, to confirm the key sentence is not truncated |
| Memory turns | A small number of prior turns | Compared with empty-history input, to confirm the sentence judgment does not depend on extra prior text |
| Reply format | json_schema when fixed fields are required | Compared with plain-text reply, to confirm fields can be parsed downstream |
| Thinking effort | Standard thinking | Compared with no thinking, to confirm the judgment covers the material-status check |
Three common errors
- Treating the reply limit as the full context window. Input is a material passage longer than the reply limit. Compare whether the key sentence of the material enters the request. The observable result is that the key sentence is missing, which shows the two were conflated.
- Using a similarity score in place of a material-status check. Input is the proposed sentence and the source material. Compare the refund status stated in the proposed sentence with the pending-declaration status recorded in the material. The observable result is that the two do not agree, which shows the judgment did not return to the source text.
- Comparing outputs directly between empty history and history with prior turns. Input is the same question under the two history settings. Compare whether the two request inputs are identical. The observable result is that the inputs differ, which shows the comparison premise was not fixed.
How to verify the setup
- Input the source material and the proposed sentence from the hypothetical sample. Compare the proposed sentence's "fully completed the refund" with the material's recorded "pending declaration" status. The condition is met when the judgment output is not supported and the evidence gap is recorded.
- Input the same question as two requests, one with empty history and one with a small number of prior turns. Compare the two request inputs. The condition is met when the inputs agree or the difference is explicitly recorded.
- Input a paper empty result
[]together with the downstream handling design. Compare whether the downstream has a defined handling path for an empty result. The condition is met when the downstream does not treat the empty result as matched content.