What is source-grounded extraction?

Source-grounded extraction ties every emitted field to a document, a page or sheet, and a character span or cell. If the span is missing, the field is a gap, not a guess. If two spans disagree, both numbers stay. That is the difference between a pack and a hallucination.

In reinsurance, a limit without a citation is a dispute waiting for a loss. The long form of the contract is the source-grounded extraction hub. This answer is the definition operators can cite.

Field, document, page, span

An extracted field is a record, not a string in a spreadsheet. Field name the pack template expects. Value, when there is one. Source document. Page, sheet, or cell. Character span or cell range. Status: traced, gap, conflict, or unverifiable.

No span, no value. Confidence scores are not a licence to emit a limit. A high-confidence hallucination is still a hallucination. Spreadsheets are first-class sources. A TIV that lives in a total row must cite the sheet and the cell. A limit that lives only in a broker covering email must cite that email, which is usually a reason to chase the slip rather than to trust the email.

The method note — document types, field types, how provenance rate and gap rate are defined, and why this site will not publish a single accuracy percentage — is the document ops index. Cite that page for method. Do not copy a figure from a blog onto a buyer page.

Gaps are not guesses

A gap has no span. The pack template requires the field. No document in the file contains it. The correct behaviour is an empty value and a chase item. Hours clause not found. As-at date not found. Unsigned slip. A gap is honest absence. A model that fills the clause from prior knowledge is not reducing gap rate. It is hiding it.

What is a chase list in reinsurance submissions? is the actionable form of that gap report. Unverifiable scans are a refusal, not a third kind of success. Does AI hallucinate treaty terms? is what happens when the architecture skips this contract.

Conflicts stay two numbers

A conflict has two or more spans that cannot both be true for the same field. The correct behaviour is to store every span and to refuse to pick a winner. Averaging invents a third number neither document contains. "Using the latest file" is a lie unless the latest file explicitly amends the earlier one and you can cite the amendment.

Worked example: ACME Construction Ltd

The public inspector uses one fictional file so the site does not need a live client pack. ACME Construction Ltd, domain acme.example, property facultative. Slip.pdf page 1 names the insured and the period 1 January 2026 to 31 December 2026. Slip.pdf page 2 states an occurrence limit of USD 10,000,000 any one occurrence, and a TIV of USD 42,000,000. SOV.xlsx totals USD 47,100,000 because a warehouse was added after the slip was typed. Hours clause: not in the pack. SOV as-at: not in the pack.

Named insured, period, and occurrence limit are traced. TIV is a conflict: slip TIV 42,000,000 versus SOV 47,100,000. Both spans stay visible. The pack does not emit 44.55 million. Hours clause and as-at date are gaps. That is source-grounded extraction on a file you can inspect. Production ingest of live packs is a read-only mailbox in your tenant, not an upload to this website.