Quick answer
Define raw text, interpreted value, status and source location separately. Zero should represent a zero explicitly recorded in the source. Keep missing or unreadable information as an empty value with a status instead of inventing zero. Compare rows before calculating; parseable JSON and fixed fields establish format, not factual accuracy.
Read headings, units and footnotes first
Use an authorized copy and inspect column names, units, merged cells and footnotes. A dash may mean missing or have a business-specific meaning; a blank is not automatically missing either. Define states from this document's explanation. If it has none, mark the interpretation unresolved rather than borrowing conventions from another table.
For image inputs, use the OCR and proofreading guide to inspect digits, labels and row boundaries. An OCR draft, an AI table and the original image are different records. Keep the source and location markers so an early recognition error does not become an accepted fact through repeated formatting.
Separate format constraints from factual review
The official Claude structured-output documentation describes constraining results to a specified structure, including support for null values. It also documents refusal and truncated-output cases. No API request was executed for this article; the source establishes format capabilities.
Our editorial recommendation is to preserve raw, value, status and source. Raw records visible characters; value holds a confirmed number or an empty value; status identifies zero, missing or unresolved information; source locates the page and row. Accept structural compliance and faithful extraction separately. A parseable file should not enter a real ledger solely because software can read it.
Include interpretation rules in the task
- 1Identify the document version and target columns. Extract visible content only: do not estimate, complete values from the web or replace missing data with zero.
- 2Provide row-location markers and request raw characters, units and states. Mark unreadable content unresolved and identify its location.
- 3Use only the source's own footnote to interpret symbols. Do not guess the meaning of an unexplained mark.
- 4Begin with a small sample and request an extraction plus an uncertainty list. Compare it with the original before extending the task.
An example instruction: using only this table, output item, raw, value, status and source for each row. Preserve zero, apply the stated footnote, flag blanks or blurred content for review, and do not total unresolved values. This is an authored instruction, not an observed product result.
Inspect four rows with different states
This fictional fixture uses units of items and defines its footnote as: a dash means not provided; a blank needs review. Row r1, item A, is 12. Row r2, item B, is 0. Row r3, item C, is a dash. Row r4, item D, is blank. These are authored demonstration records, not real orders or cloud-model test data.
- r1: retain raw 12, numeric value 12, provided status and the unit items.
- r2: retain raw 0, numeric value 0 and explicitly provided zero status; do not change it to missing.
- r3: retain the dash, leave the numeric value empty, mark not provided and preserve the footnote location.
- r4: retain the original blank, leave the value empty and mark needs review; do not infer a number from neighboring rows.
Reject the extraction if r3 and r4 both become zero. Even if a spreadsheet calculation runs, the source meaning has changed.
Review extraction before formulas and business use
Compare row count, item labels and source locations first. Check for omitted or duplicated rows and headings mistaken for records. Then inspect digits, symbols, units and states, especially zero versus O, 1, decimal points and minus signs. An uncertainty list should locate the specific cell; a generic reminder to check manually is insufficient.
After accepting the extraction, follow the sample-row formula verification guide to inspect calculations. Rules for missing values in totals come from the business process, not an unannounced model decision. Confirm permission before sending sensitive records to an external AI service. That upload is not local processing, and automatic organization does not constitute a formal audit of financial or business records.
Common questions
What if the source does not explain its dashes?
Keep the original character and mark its interpretation unresolved. Ask the source owner; do not assume zero, missing or not applicable.
Does valid JSON remove the need to inspect cells?
No. Structure does not establish agreement with source numbers, units or states. Check missing rows, duplicates and unreadable characters as well.
Can I replace every blank with zero before calculating?
Only if an explicit business rule allows that transformation and you retain the original values and change record. Otherwise it turns unknown information into a confirmed zero; resolve it first.
