DIRECT ANSWER / 01
How does a bounded grant budget cross-check work?
Name the exact budget and narrative versions, import a strict delimited budget table and plain narrative text, extract bounded numeric claims from the prose, confirm each budget-to-claim mapping by hand, recompute the visible arithmetic, record the differences that remain and freeze the reviewed state.
Why this review exists
A total-only check misses the differences that actually get proposals sent back: a role that appears at one rate in the table and another in the prose, a period stated as twelve months in one place and ten in another, a figure that is repeated in three paragraphs and updated in two.
The sequence in practice
Source identity comes first because everything else is only meaningful against it. The bundled sample records the review as Harbor Youth Lab proposal review for North Quay Learning Cooperative, opportunity Harbor Futures Mini-Grant, reference HYF-26-07, with budget version Budget v7 dated 3 September 2026 and narrative version Narrative v11 of the same date.
The budget arrives as a strict delimited table that is previewed before it is committed. You map the columns the review needs, and cells that look like spreadsheet formulas are rejected rather than parsed. The narrative arrives as plain text or Markdown, split into paragraphs, and the extractor proposes bounded numeric candidates from it.
Nothing is matched automatically. Suggestions are offered only where a number is equal within the tolerance you set, and every suggestion still requires a person to confirm the relationship or ignore the claim.
What the sample review contains
Four budget rows: a program coordinator at twelve units of 2000 for 24,000 requested with 6,000 in other funds, workshop materials at four cohorts of 2500 for 10,000, participant transit support at ten units of 750 for 7,500, and an external learning review at 5,000. The totals are 46,500 requested, 6,000 in other funds, no in-kind match, and 52,500 total project cost.
Three narrative paragraphs produce eight extracted claims: twelve months, 24,000, 6,000, ten months, 10,000, 7,500, 5,000 and 46,500. A reviewer confirmed five of them as mappings to budget rows and marked three as contextual or rollup evidence rather than line-item claims.
With those decisions recorded the review reports no open findings and no blocking findings, and every one of the four arithmetic traces agrees at a tolerance of 0.01.
Method and limits
Every check is arithmetic or exact textual comparison on records you supplied, so the same inputs and the same decisions always produce the same result. There is no scoring, no ranking and no model.
A quiet review means the numbers you entered agree with the numbers in the prose you pasted, within the tolerance you chose. It says nothing about eligibility, allowability, reasonableness, the funder's instructions, or whether the proposal should be submitted.
Practical sequence
- Record the exact budget and narrative version labels
- Preview and map the budget table without committing anything
- Paste the narrative and let the extractor propose bounded numeric claims
- Confirm or ignore each claim, then confirm each mapping by hand
- Read the arithmetic traces and the findings, then freeze and export
What remains outside the app
Grantweft performs arithmetic and textual cross-reference on user-supplied material. It does not write proposals or determine eligibility, allowability, accounting treatment, compliance, completeness, accuracy, submission readiness or funding outcome.
The private desk keeps exact source labels, stable budget row identities, paragraph-based claims, confirmed mappings, visible arithmetic, findings, human decisions, review freezes and passive handoff files connected without an account or remote application database.