The world's most important research — the peer-reviewed studies that contain real solutions to climate breakdown, food insecurity, and inequality — never reaches the people it could help. The gap between academic publication and public understanding is a structural failure. This newsroom exists to close it, using AI as the translation and distribution engine, and experienced journalists as the editorial intelligence that makes it trustworthy.
Zero hallucination by design
An isolated RAG framework means the AI can only use what the researcher provided. No outside facts. No assumed context. Everything in the output traces back to the source paper.
Research matched to global need
Every paper is scanned against the 17 UN Sustainable Development Goals and global coverage gaps. The AI identifies which findings are most undercovered and most urgent for world media.
Journalists decide everything that matters
What gets covered. What angle serves the public. Who to contact. Whether to publish. AI does not touch these decisions. It works for the editor — not the other way around.
The original approach used flat RAG — retrieving isolated sentences from the paper. Version 2.0 uses GraphRAG, which builds a knowledge graph of the paper's entities, claims, and evidence chains. This means the traceability layer doesn't just cite a page number — it maps the full chain of evidence from conclusion back to raw data.
Retrieval-Augmented Generation grounds AI output in a specific, curated knowledge source rather than the model's training data. For journalism, this is the only responsible architecture — because it means the AI cannot invent, assume, or embellish. It can only work with what you give it.
Standard LLMs hallucinate — they produce fluent, confident text that is factually wrong. A 2025 BBC study found over half of AI answers to news queries had significant issues including factual errors and fabricated quotations. RAG with a strictly isolated knowledge source removes this failure mode entirely, because the model has nowhere else to look.
Academic papers are not flat documents. They contain linked claims: a conclusion in section 5 depends on a method in section 2, which cites data in an appendix. Flat RAG retrieves the conclusion without the chain. GraphRAG retrieves the whole chain — so the traceability view shows the editor not just where the claim appears but what it depends on. This is the difference between citation and verification.
Three-tier claim classification
Every claim in the output is tagged as drawn from: Core findings (highest confidence), Methodology (conditional), or Limitations (requires explicit qualification in the news piece). The editor sees these tags in the side-by-side view before approving.
Contested findings flag
When the paper's own data presents tension — a finding that contradicts prior literature it cites, or a limitation that qualifies a headline result — the RAG system flags this to the editor. It does not resolve the dispute. It routes it to human judgment.
The UN SDG Media Compact now has almost 400 member organisations across 160 countries reaching a combined audience of two billion people — all actively seeking content that connects news to the Goals. This newsroom doesn't just reference the SDGs. It uses them as a real-time signal for where verified, factual journalism is most needed and least available.
Rather than scanning academic papers for SDG connections after the fact, the SDG intelligence layer runs continuously. It tracks which of the 17 Goals has the least verified news coverage in the past 30–90 days, broken down by region. This becomes the primary signal the editor uses when commissioning academic sources — turning research selection from reactive to strategic.
Which Goals get underreported
SDGs 2 (Zero Hunger), 6 (Clean Water), 14 (Life Below Water) and 16 (Peace & Justice) consistently receive a fraction of the media coverage that SDG 13 (Climate) attracts. The gap signal prioritises these.
Coverage by geography
A story on SDG 3 (Good Health) is well-covered in Western Europe but severely underreported across Sub-Saharan Africa. The SDG layer identifies the regional gap, not just the topical one.
Drives which academics to approach
The editor doesn't wait for papers to arrive. The SDG gap map shows which research topics are most needed right now — then the editor finds the right academic to match.
Each agent has a strictly bounded job. None of them make editorial decisions. All of them report to the editor. The isolation is intentional — it prevents scope creep where an agent starts making judgments it has no authority to make.
Research Agent — SDG gap scanner + signal detector
Scans the academic paper against the live SDG coverage gap map and historical global news patterns. Identifies the 3–5 most impactful narrative angles. Flags which regions and media markets need this research most. Outputs a ranked editorial brief — not a story, a brief.
GraphRAG · SDG gap index · NewsWhip signal dataLanguage Condensation Agent — jargon removal + style translation
Strips academic density and rewrites for five audience registers: Reuters wire style, Economist analysis, BBC World Service radio, plain-language community brief, and social-native Reel script. The RAG constraint means every claim in every version traces to the source paper. No outside colour added.
Isolated RAG · Five style guides · Confidence layer tags visible Limitations-section claims flagged in all outputsTraceability Agent — sentence-level citation mapping
Maps every sentence in every output format back to the exact page, paragraph, and claim tier in the original paper. Produces the side-by-side view the academic sees before final approval. Also generates a NewsGuard compliance report showing the evidence chain for every published claim.
GraphRAG citation graph · Three-tier confidence display · NewsGuard exportDistribution Agent — relevance-scored pitching
Monitors breaking global news. When a relevant event occurs, it matches pre-translated packages to partner news desks by SDG beat, geographic coverage area, and language preference — then pitches only to editors for whom this is directly relevant. Not a broadcast. A targeted, contextual pitch.
Partner beat database · Breaking news feed · Multi-language packages No pitch sent without relevance score above thresholdNone of the agents decide what gets covered. None contact sources directly. None approve final copy. None bypass the academic's review of the traceability map. None publish without the editor's sign-off. Every agent output is a draft routed to human judgment.
The original approach produced three formats. Version 2.0 produces five — adding a community plain-language brief and a BBC-style radio script, which are the formats that reach the communities most directly affected by SDG issues. All five derive from the same RAG-grounded, traceability-mapped source material.
| Format | Audience | Style register | Length | Distribution channel |
|---|---|---|---|---|
| Digital wire copyFor editors at global outlets | Global news desks | Reuters style | 400–600 words | Active pitch via distribution agent |
| Analysis featureFor quality press and magazines | Informed general readers | Economist style | 900–1400 words | Long-form partners, newsletters |
| Radio scriptNew — regional broadcasters | Radio audiences, Global South | BBC World Service | 3–4 min read | Regional broadcast partners |
| Community briefNew — NGOs, schools, local orgs | Affected communities directly | Plain English | Sub-800 words | NGO partner network, SDG Compact |
| Social packageReel script + caption + thread | Social media audiences | Platform-native | 60–90s script + captions | Own channels + partner reposts |
All five formats are generated from a single RAG-grounded condensation of the source paper. The language condensation agent applies the five style guides in parallel — it does not rewrite the story five times independently. This means the factual core is identical across all formats, and the traceability map covers all five simultaneously.
Reaching the communities SDGs are about
SDG stories about hunger, water access, and poverty need to reach the communities experiencing those issues. Wire copy reaches editors in capital cities. Radio reaches rural communities without reliable internet. Community briefs reach local NGO workers and teachers. These are the audiences with the highest stake in the research.
One review, five sign-offs
The academic sees the traceability map for all five formats simultaneously. They approve the factual core once — the style differences between Reuters copy and a community brief are not their concern. Their sign-off is that every claim in every format is accurate to their research.
The original active pitching concept is strong but needs precision. A news desk that receives an unsolicited research package when a climate crisis hits will ignore it unless it's immediately, specifically relevant to their beat, geography, and audience. The distribution agent's job is to know the difference.
Breaking news signal monitored continuously
The distribution agent tracks global news feeds for events that match the SDG tags and regional markers attached to every pre-translated package in the library.
NewsWhip · AP wire monitor · UN news feedPartner beat database — who covers what, where
Every media partner in the network has a profile: SDG beats they cover, geographic regions, language, audience type (general, specialist, community), and their preferred format (wire, radio script, briefing). The pitching agent matches against all four dimensions before sending anything.
Partner CRM · SDG beat tags · Format preference matrixRelevance scoring — threshold before any pitch is sent
A package is only pitched to a partner if its relevance score — calculated from SDG match, geographic overlap, recency, and breaking news context — exceeds a set threshold. Below threshold: the match is flagged for the editor to review manually before deciding whether to pitch.
Relevance scoring model · Editor review queue for borderline matches No automated pitch below relevance thresholdEditor approves all pitches above threshold
Even high-scoring matches are reviewed by the editor before the pitch goes out. The agent prepares the pitch; the editor sends it. This keeps the relationship between this newsroom and its partners in human hands.
Wire copy goes to news desk editors. Radio scripts go to broadcast producers. Community briefs go to NGO programme officers and school coordinators. Social packages go to digital editors and social media managers. The same package, delivered in the right format, to the right person, at the right moment — is the difference between pickup and deletion.
The editor is not a gatekeeper at the end of an AI pipeline. They are the intelligence that shapes the pipeline from the start. The SDG gap map informs their commissioning. The research brief informs their story selection. But the judgment — what matters, why it matters now, who needs to hear it — is entirely theirs.
SDG gap map drives what gets covered
The editor uses the live SDG coverage gap analysis to identify which Goals, in which regions, are most underserved by factual journalism right now. This shapes which academic researchers they approach — making editorial selection strategic rather than reactive.
Finding and briefing the researcher
The editor identifies researchers who have already published peer-reviewed work on the selected SDG gap. They interview the academic not just to understand the paper, but to extract the human story — the real-world application the data points to — before anything goes to the AI.
Translating research intent before AI touches it
The editor's conversation with the academic determines the core message. What did this research actually find? What should a policymaker, a teacher, a farmer do differently because of this? This framing is set by humans before any agent is given the paper.
Every publish decision, every pitch, every correction
The editor approves all five format outputs. They approve every distribution pitch. They issue all corrections — publicly, on every platform, at the same prominence as the original. The accountability chain ends with them, always.
The RAG system can only be as good as the understanding of the research that was fed into it. If the editor's briefing conversation with the academic is shallow, the research agent's brief will be shallow, and the language condensation agent will faithfully produce shallow wire copy. The depth of the journalism flows from this conversation — not from the AI. This is the role the technology cannot touch.
The RAG isolation and GraphRAG traceability are architectural truth protections. These rules are human truth protections. They apply regardless of how good the technology is, because technology does not bear reputational accountability. You do.
The Eight Rules
This timeline shows a single story moving through the complete system — from the editor's commissioning decision through to live distribution across five formats. It assumes the academic paper has already been delivered. The whole cycle, with an experienced editor and a functional agent stack, runs in roughly one working day.
SDG gap review — editorial commissioning Editor
Editor reviews overnight SDG coverage gap analysis. Identifies which Goals and regions are underserved. Selects the academic paper for today's cycle based on this signal.
Academic briefing call Editor + Researcher
Editor interviews the researcher to extract the core message — the human implication of the data. This conversation shapes what the research agent prioritises. Cannot be skipped or delegated.
Research agent runs — SDG angle brief Research Agent
GraphRAG builds the paper's knowledge graph. Research agent scans against SDG gap map and news history. Outputs ranked editorial brief with 3–5 narrative angles and target markets.
Editor selects angle and target markets Editor
Reviews the research brief. Selects the primary angle. Identifies which of the five formats to prioritise. Approves the brief before condensation begins.
Language condensation — all five formats in parallel RAG
Isolated RAG condenses the paper into five simultaneous drafts. Confidence layer tags every claim. Limitations-section findings flagged across all formats automatically.
Traceability map generated Traceability Agent
Every sentence in every format mapped to exact page, paragraph and claim tier in the source paper. Side-by-side view prepared for academic review. NewsGuard compliance report generated.
Academic reviews traceability map Researcher
Researcher reviews side-by-side view. Confirms every claim accurately represents their research. Raises any concerns about qualification of preliminary findings. Approves the factual core.
Editor final review and editorial additions Editor
Editor reviews all five formats. Adds context the RAG cannot provide — current events relevance, additional sourcing, editorial voice. Makes the public interest case. Final copy is edited, not just approved.
Publication — wire + social package live Publish
Wire copy and social package published first. Reel filmed and posted. Distribution agent begins matching against partner beat database.
Editor approves distribution pitches Editor
Distribution agent presents scored pitch list. Editor reviews, approves or adjusts each pitch. Pitches go to relevant news desks, radio producers, NGO partners.
Long-form analysis + community brief published Publish
Economist-style analysis and community plain-language brief go live. Radio script delivered to broadcast partners. All five formats now in distribution.
Monitoring, corrections, next commissioning signal Editor
Monitor partner pickup and audience response. Address any corrections immediately across all formats. Review SDG gap signal for tomorrow's commissioning decision.
A traditional team of three journalists cannot produce five publication-ready formats from a complex academic paper in one working day while maintaining full source traceability and academic sign-off. The AI agents compress the production work. The editor focuses entirely on judgment, editorial depth, and relationships. That combination — machine speed, human standards — is what makes this newsroom different.