Illustrative Scenario: A Fact-Checking Network’s First 90 Days on Retelnist

Context and Challenge

A mid-sized fact-checking network (roughly a few dozen contributors spanning editors, researchers, and regional stringers) set out to modernize how it managed claims, evidence, and publication workflows. The work was high-stakes and time-sensitive: election cycles, public health updates, and breaking news regularly spiked demand. Yet the operational backbone had grown organically—good enough for steady periods, brittle under pressure.

Several challenges converged:

• Fragmented intake and triage: Tips arrived through email, messaging apps, web forms, and social media. Each channel had its own “unwritten rules,” leading to inconsistent prioritization.
• Evidence sprawl: Screenshots, archived pages, datasets, and interview notes lived across drives and chat threads, making it hard to verify what was already collected.
• Inconsistent documentation: Different teams used different templates (or none), complicating peer review and editorial oversight.
• Limited auditability: When a reader asked why a claim was rated a certain way, reconstructing the reasoning took too long and depended on institutional memory.
• Burnout risk: Peak periods meant rushed decisions and overtime. Editors spent too much time coordinating instead of reviewing.

The network’s goal for the quarter was clear: reduce time-to-decision without compromising rigor, and create a repeatable process resilient to surges.

Approach and Solution (Days 1–90)

1) Onboarding and Workflow Mapping (Days 1–14)

The first two weeks focused on translating an existing editorial philosophy into a structured operational system. Rather than starting with tools and features, the network documented how work should flow end-to-end:

• Claim intake → triage → assignment
• Evidence collection → source evaluation → analyst conclusion
• Peer review → editorial approval → publication
• Post-publication monitoring → corrections/updates → archive

A key early decision was to define a single “unit of work”: the claim. Every artifact—screenshots, transcripts, datasets, messages—would attach back to that claim record. This created a shared mental model across contributors.

To avoid disrupting ongoing coverage, onboarding was staged:

• A pilot desk (a small subset of editors and researchers) adopted the full process first.
• The remainder of the network continued with existing methods while receiving short training sessions and a simplified intake view.

2) Standardized Intake and Triage (Days 15–30)

The next step addressed the front door: claims arriving from too many places. The network introduced a unified intake layer with consistent required fields, including:

• Claim text (verbatim where possible)
• Origin and context (platform, date observed, relevant audience)
• Potential harm level (public safety, civic integrity, financial harm, reputational damage)
• Urgency (time-bound events vs evergreen misinformation)
• Preliminary evidence pointers (known sources, related debunks)

To keep triage reliable, the network agreed on a lightweight scoring rubric. The rubric wasn’t meant to replace judgment; it provided shared language for prioritization. Editors could override the score, but overrides required a short note—improving transparency without slowing work.

A practical improvement emerged quickly: the network separated “fast-check” items (clear, low-ambiguity claims with readily available evidence) from “deep-dive” items (complex chains of causality, technical topics, or conflicting sources). This prevented simple claims from waiting behind investigations that naturally took longer.

3) Evidence Management and Review Readiness (Days 31–60)

With intake stabilized, attention shifted to the core of fact-checking: evidence.

The network implemented a consistent evidence approach built around three principles:

1. Traceability: Every supporting or refuting element had to be linked to a specific artifact and description, not just a vague reference.
2. Source evaluation: Evidence was tagged by type and assessed for credibility, relevance, and recency.
3. Reusability: If an evidence artifact applied to multiple claims, it could be referenced without duplicating files or losing context.

To reduce back-and-forth during review, contributors adopted a “review-ready” checklist:

• Claim is precisely stated and scoped
• Key terms defined (to prevent semantic disputes)
• Evidence summary includes both supporting and contradicting points
• Reasoning steps are explicit
• Verdict aligns with documented reasoning
• Sensitivity notes flagged (legal, privacy, safety concerns)

Peer review was formalized as a distinct stage rather than an informal request in chat. Reviewers could leave structured comments tied to specific reasoning steps, which improved clarity and reduced repetitive clarification questions.

4) Publication, Corrections, and Quarterly Review Prep (Days 61–90)

In the final month, the focus shifted to making outcomes measurable and post-publication work less reactive.

The network defined a small set of operational signals to monitor weekly:

• Volume of new claims received
• Share of claims triaged within a target window (tracked as an approximate internal goal)
• Time from assignment to review-ready status
• Number and type of review cycles (first-pass acceptance vs revisions)
• Corrections and updates, categorized by cause (new evidence, wording clarity, rating adjustment, data error)

Rather than using these signals to judge individuals, the network used them to identify friction points. For example, repeated revisions often traced back to ambiguous claim framing, not poor research. That insight fed back into training and templates.

A final deliverable for the first quarterly review was a process playbook: a concise guide to how claims move, what “done” means at each step, and how exceptions are handled (e.g., time-critical claims, sensitive investigations, or multi-language coordination).

Results After the First 90 Days

Because the network avoided publishing hard numbers without a longer baseline, outcomes were evaluated as qualitative improvements with approximate internal indicators. The shift was still meaningful across daily operations:

• Faster, calmer triage: Editors reported fewer “where should this go?” discussions. The rubric and unified intake reduced ambiguity and helped prioritize high-impact claims quickly.
• Improved review quality: Reviewers spent less time asking for missing context because claim framing and evidence summaries became more consistent.
• Better continuity across shifts: When a claim moved between contributors, the record contained enough reasoning and artifacts to maintain momentum without restarting.
• Stronger accountability and auditability: The network could reconstruct how a verdict was reached, including what evidence was considered and what uncertainties remained.
• Reduced rework: Standard templates and the review-ready checklist cut down on avoidable revisions, especially on recurring topics.
• More predictable quarterly reporting: The operational signals enabled a clearer first quarterly review discussion focused on process improvements rather than anecdotes.

One subtle but important result was cultural: contributors began to treat documentation not as bureaucracy, but as part of editorial integrity—making it easier to defend decisions and correct them when necessary.

Key Takeaways

• Define the unit of work early. Anchoring everything to a single claim record prevents evidence and discussions from scattering across channels.
• Standardize intake without over-engineering it. A few required fields and a simple scoring rubric can dramatically improve consistency while preserving editorial judgment.
• Separate “fast-check” from “deep-dive.” Different claim types need different service levels; mixing them in one queue slows both.
• Make review readiness explicit. A checklist and structured peer review reduce clarification cycles and improve reasoning transparency.
• Track signals that reveal friction, not performance theater. Operational metrics work best when used to improve workflows rather than rank people.
• Treat corrections as feedback loops. Categorizing why updates happen turns inevitable change into a tool for strengthening future work.

Over the first 90 days, the network moved from a collection of well-intentioned individual practices to a cohesive, repeatable system—one designed to scale under pressure while preserving the rigor that fact-checking demands.