How it works
From an evidence set to a reviewed finding
The run sheet
- Step 1
Load the evidence set
Drag in contracts, certificates, policies, registers and audit records, one at a time or up to 50 at once. Duplicates are recognised by content, not by file name.
- — PDF, including scans read with OCR
- — Word (DOCX), Excel (XLSX), CSV, HTML, plain text
- — Deduplication by SHA-256 content hash
- Step 2
Processing
A nine-stage pipeline extracts the text, detects sections and headings, stores sentences verbatim, picks out dates, amounts and percentages, and builds searchable passages with vector embeddings.
- — Section and heading detection
- — Sentence-level verbatim storage
- — Spreadsheets kept with their sheet and row positions
- Step 3
Write the requirements
An Evidence Programme is a set of requirements run against a named subject: a supplier, a site, a contract. Start from the Supplier Compliance Review template or write requirements in your own words.
- — Requirements persist between runs
- — Edit wording for future runs; earlier findings keep the wording they were judged against
- — Archive a requirement without losing its history
- Step 4
Search and sort
Keyword and semantic search runs across the whole set, so a certificate is found even when it never uses the words of the requirement. Each finding then separates what DataStruct could support from what it could not.
- — Keyword, semantic and section signals merged
- — Located evidence, unconfirmed citations and related material kept apart
- — Gaps named in plain language
- Step 5
Review
A named reviewer confirms, rejects or marks each finding for follow-up. The decision is attributed and logged. Decisions never carry over to a new run.
- — Confirm, reject or needs follow-up
- — Every decision attributed and audit logged
- — DataStruct records no verdict of its own
- Step 6
Report and rerun
Export a run with its requirements, findings, evidence states, reviewer decisions and sources. Rerun next quarter; every earlier run stays on record with what was found and decided at the time.
- — Programme report: DOCX, XLSX, CSV or JSON
- — Document reports: HTML, DOCX, JSON
- — PDF export is not offered
Anatomy of a finding
Six parts, each one checkable
Requirement
Provide your modern slavery statement and evidence that your board approved it.
Modern slavery statement 2025.pdf · p. 1 · Introduction
This statement is made pursuant to section 54(1) of the Modern Slavery Act 2015 and sets out the steps taken during the financial year ending 31 December 2025.
Gap: board approval of the statement was not located.
The requirement
Written by your team, in your words, and kept between runs.
The locator
File, page and section, or sheet and row range. Where only the sheet is known, only the sheet is named.
The located passage
The exact text DataStruct could support, marked inside its source rather than paraphrased.
The evidence state
In the product’s own words: evidence located, partial, support not confirmed, or no supporting evidence located.
The gap
What was asked for and not found, stated plainly so the reviewer knows what to chase.
The reviewer’s decision
Only a person writes this. It is the one place a finding can turn green.
Where it stops
What DataStruct does not do
- It records no compliance verdict. Findings say what was located; only a reviewer decides whether a requirement is met.
- It does not yet pick the current version of a superseded record. When an old and a new version are both in the set, both can be returned. The reviewer decides which one governs.
- It does not flag contradictions between documents. Detecting disagreement is research on our roadmap, not a feature.
- It can miss evidence, and it can mistake related material for support. That is why every finding goes to a person.
Also in the product
Beyond Evidence Programmes
- Search and ask
- Ask a question across your documents and get an answer whose claims link to the passages they cite. A verification pass flags claims it cannot match to the cited evidence.
- Extraction templates
- Define the fields you need from each document and run them in batches, each value with a confidence score and a link to its source.
- Rule packs
- Deterministic checks you configure: keyword presence and absence, regular expressions, numeric thresholds and missing disclosures, each with a severity.
- Reports and datasets
- Programme reports in DOCX, XLSX, CSV or JSON; document reports in HTML, DOCX or JSON; extraction datasets in CSV, XLSX or JSON.