--- title: "EMBER: Check every claim in a document | Aicadium" description: EMBER checks every claim in a document against credible external evidence, marks how much you can rely on it, and tells you what to fix. --- [![Aicadium](https://landing.aicadium.ai/hubfs/Aicadium_blue_logo.svg)](https://landing.aicadium.ai/?hsLang=en) [Get in touch](https://landing.aicadium.ai/contact?hsLang=en) EMBER: Documents you can stand behind 15 min read What we believe [What we believe 1](https://landing.aicadium.ai/every-claim-checked#believe) [The framework 2](https://landing.aicadium.ai/every-claim-checked#framework) [Field voices 3](https://landing.aicadium.ai/every-claim-checked#voices) [Predictions 4](https://landing.aicadium.ai/every-claim-checked#predict) [Demo 5](https://landing.aicadium.ai/every-claim-checked#demos) [The series 6](https://landing.aicadium.ai/every-claim-checked#series) For every claim that travels from a report into a decision # A claim is yours the moment you use it. A finished document tells you nothing about whether its claims hold. The prose is confident, the citations look tidy, and the reader is left to take the author's word for it. **EMBER** checks each claim against credible external evidence, ranked by how authoritative it is, then marks every claim with how much you can rely on it and tells you what to do about the ones that fall short.  PRESENCE RANKING INSIGHT SOURCING MAPPING YOUR BRAND The numbers behind the problem ## As more content becomes AI-generated, less gets checked, **and** **accountability falls on those who reuse it.** 1 in 4 citations generated by GPT-4 contained substantive errors, and a further 18% did not exist at all A genuine-looking reference is not proof of a true claim. The source can exist and still not say what the document says it says. [Walters and Wilder, Scientific Reports, 2023 (636 references across 42 topics)](https://www.nature.com/articles/s41598-023-41032-5) 66% of AI users rely on its output without evaluating accuracy More than half say AI has already caused them to make mistakes at work. The check is being skipped at the point where it would cost the least.   [University of Melbourne and KPMG, Trust, attitudes and use of AI: a global study 2025 (48,000 people, 47 countries)](https://kpmg.com/xx/en/our-insights/ai-and-technology/trust-attitudes-and-use-of-ai.html) 88% of organisations now use AI regularly in at least one business function, up from 78% a year earlier AI is already in the reports, decks and memos moving through your organisation. The question is no longer whether it is there, but whether anyone has checked what it produced. [McKinsey, The State of AI 2025 (1,993 respondents, 105 countries)](https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai) What we believe ## Three convictions about the documents organisations publish and rely on. CONVICTION I ### A finished document is not a checked document. Fluent prose and a tidy reference list show the writer was confident. They show nothing about whether the claims hold. That is true of the report you are about to send as much as the one you have just been sent. CONVICTION II ### Reading is not verifying. A careful reader can catch a claim that sounds wrong. Nobody can catch the one that sounds right and isn't. Once a document is published or relied on, the organisation carries every claim in it, whether or not anyone saw the evidence. CONVICTION III ### The risk moves with the claim, not with the author. The person who put the statistic in the deck did not write the report it came from. The person presenting it did not put it in. Neither checked it, and both answer for it when someone asks where it came from. A check that either can run, and both can read, is the only kind that closes that gap. The framework ## The check that makes a document **defensible.** **EMBER stands for** **Evidence Mapping and Benchmarking of Existing Research.** A user hands a document to EMBER, and it determines what the document claims, checks each claim against sources outside the document, and returns the original text with a marker on every claim.  EMBER shares its trust infrastructure with [SPARK](https://landing.aicadium.ai/ai-research-you-can-trust?hsLang=en). The difference is the input: SPARK verifies research as it writes it, EMBER verifies a document that already exists. E Evidence Every claim is tested against sources the document did not choose. M Mapping Each verdict sits on the exact sentence it applies to. B Benchmarking Claims are held to a fixed, documented standard, not a reader's instinct. E Existing Check the document you have. It does not write a new one. R Research Any report, memo, guide or deck that makes checkable claims. Five things for documents you can trust ## Don't just read the report. Check the receipts. Most documents expect the reader to take the author's word for it. EMBER holds each claim up against credible external evidence, ranked by how authoritative it is, so the verdict rests on the strength of the sources rather than the confidence of the writer. Five checks run on every document. 01 "What is this document actually asserting?" Claim Extraction EMBER extracts every verifiable factual claim and quotes the exact words it came from. Each claim is checked once, even if the document repeats it. Opinions, framing and statements that cannot be tested are left unmarked. If a passage has no marker, it makes no claim that can be checked. 02 "Does credible evidence back this up?" External Grounding EMBER checks each claim in three ways. It looks at the source the document cites, to see if it says what the document says it does. It then looks for independent evidence, giving more weight to primary documents and official data than to commentary. If nothing outside the document can be found, it informs the reader. 03 "How do I read it at a glance?" Four Trust Tiers Each marker is one of four tiers. A tier describes the claim's relationship to the evidence, not the author. **Verified**: independent sources support the claim as written. **Caution**: support exists but is thin, or the framing needs a caveat. **Contested**: credible sources disagree with the claim or with each other. **Unverified**: no external evidence found, so the claim rests on the document's own word. 04 "Does the cited source say what the document says it does?" Citation Fidelity Where the document cites a source, EMBER checks whether that source supports the sentence it is attached to. A real reference can still be miscited, broken or weakly sourced. These are flagged separately, so a claim can be true and still carry a citation problem. 05 "What do I do about it?" Remediation Brief A prioritised list of what to fix and how. Each item names the claim, what is wrong, what to do about it, and how much effort the fix will take. Factual errors get a correction. Disputed framing gets a clarification. Unverified claims get a request for the source. The sorting is already done, so the list can be worked through as it stands. EMBER in action ## See EMBER run, and what comes back. See It Run Read The Report ### From a document to a verdict on every claim. - 1**Hand over the document to EMBER**Upload the document and ask EMBER to ingest it. It becomes a text file with every word kept verbatim, and every link and footnote recovered. - 2**Type the trigger phrase.**In a fresh chat, "EMBER Verify on \[document\]" is the entire instruction. - 3**EMBER returns two files.**Every run returns the HTML report you read and a fact file that records every claim, verdict and score. - 4**Read the report.**Click any marker to see the evidence EMBER found, how it scored the claim, and what to do if the claim does not hold. ### The report is your document, with the evidence attached. - 1**Executive summary**How many claims were checked and whether the document is safe to rely on. Which parts to quote freely, and which to check first. - 2**Verification summary**How many claims landed in each tier, and how many citations were broken or pointed to the wrong evidence. Click a tier to list its claims. - 3**Remediation brief**The to-do list prioritises the most serious items first. Each entry specifies the claim, the issue, the required action, and the estimated effort. - 4**Quality scorecard**How much of the document was covered, and whether every link and check passed. If coverage is high, the verdicts can be taken at face value. ![Read The Report](https://landing.aicadium.ai/hubfs/ember_report_mock_9x16.jpg) Three Predictions ## What we expect to be obvious by 2027, and uncomfortable to admit today. ### "Has this been checked?" becomes a standard question before a document goes out. Not "who wrote it" or "does it look right", but whether the claims in it have been tested against sources outside the document. The teams that can answer this will find their papers move faster. ### Documents carry a verification status the way emails carry a spam score. A report without a trust summary starts to look like a report with something to hide. The status travels with the document, whether it is leaving the organisation or entering it. ### Verification runs at the point of publication and receipt, not the point of doubt. Today a document gets checked when something feels wrong. Soon it gets checked because it is about to be sent, or because it has just arrived. The check stops being a judgement call and becomes part of how documents move through the organisation. Why this matters now ## Most documents get a skim and a gut check. **EMBER applies the same level of scrutiny to every claim.** Without EMBER A few claims the ones that looked wrong or that you happened to know about - Checking rests on whichever reader has time, and their instinct for what looks off - Citations are trusted because they exist, not because anyone opened them - Claims outside your expertise get taken on the author's word - No record of what was checked, so the next reader starts again - The revised version gets the same skim, or none → With EMBER Every claim tested against the same standard, whoever the reader is - Every checkable claim is extracted and quoted verbatim, so nothing depends on what caught the eye - Each claim tested against sources the author did not choose, ranked by authority - Every citation opened and checked against the sentence it supports - A report that records what was checked, what held, and what to fix - One trigger phrase to run it again on the revised version Voices from the field ## What the first people to run EMBER told us Early feedback from our own team, who put EMBER to work on real documents before anyone outside Aicadium did. “ We ran it on a guide we'd already published, and it caught figures that had quietly gone out of date since we wrote it. The remediation brief told us which ones to correct and which just needed a date on them. The fixes took an afternoon, and the guide went back out with stronger sources than it started with. M Marketing practitioner checking outbound content “ Numbers from reports are sometimes used in briefings and decks, usually without the report attached. EMBER tells me which ones can travel and which ones shouldn't leave the document. For the ones that can, it also gives me the source, so when someone asks where a figure came from, the answer is already there S Senior stakeholder reviewing incoming reports Part of the SPARK series ## SPARK checks the research you generate. EMBER checks the documents you publish and receive. Both sit on the same trust infrastructure: the same scoring, the same tiers, the same insistence that every claim carries its evidence. SPARK came first. People who saw it asked whether the same standard could be applied to documents that already existed, their own and other people's. EMBER is the answer to that question.   | | **SPARK** | **EMBER** | | --- | --- | --- | | You are the | Researcher | Reviewer, of your own document or someone else's | | It works on | Research AI generates for you | Any existing document, yours or someone else's | | Trust tier | Three: Verified, Caution, Contested | Four: Verified, Caution, Contested, Unverified | | Over time | Refreshes and reports what changed | Single use | | Output | Source report and fact file | Annotated original with remediation brief |     The series ## Read more about EMBER ![The document you didn't write is still your problem](https://landing.aicadium.ai/hubfs/placeholder.png) Blog ### The document you didn't write is still your problem Why the recipient carries the liability for every claim they sign off on, and what that means for how documents should arrive. Read → ![When the citation is real but the claim is not](https://landing.aicadium.ai/hs-fs/hubfs/placeholder.png?width=400&height=200&name=placeholder.png) Blog ### When the citation is real but the claim is not The reference exists. The paper is genuine. It just does not say that. Why citation fidelity is the check most readers never run. Read the findings → ![We ran a published industry guide through EMBER. Here's what it caught.](https://landing.aicadium.ai/hs-fs/hubfs/placeholder.png?width=400&height=200&name=placeholder.png) Case Study ### We ran a published industry guide through EMBER. Here's what it caught. What happened when we checked a real guide claim by claim, including the two claims the tool found that both human readers had missed. Read case study → ![Correction or clarification? Two kinds of wrong in every report](https://landing.aicadium.ai/hs-fs/hubfs/placeholder.png?width=400&height=200&name=placeholder.png) Blog ### Correction or clarification? Two kinds of wrong in every report Not every contested claim is a factual error. Some are framing disputes where neither side is wrong. The fix is different, and so is the conversation with the author. Read → ### Ideas like this shape how we work at Aicadium Get updates about our latest AI Transformation projects. Select a form in the module settings to activate the subscribe form. AICADIUM SINGAPORE PTE. LTD. · © 2026 Aicadium [Privacy Policy](https://landing.aicadium.ai/privacy?hsLang=en) | [Terms of Use](https://landing.aicadium.ai/terms?hsLang=en)