Seedlabs

ThreadVerdict: one-pass rumour triage API

A tool for developers that analyzes a social media thread to predict if a claim is a rumor and if it is true.

A hosted API that takes a social media claim thread and returns, from a single jointly-trained model, a rumour score, a reply-stance breakdown, and a veracity prediction — replacing chained detection→stance→veracity pipelines whose errors compound.

Social SciencesMisinformation and Its Impacts
Cross-Corpus Speech Emotion RecognitionMedia and fact-checking — automated triage of social media claim threads for newsroom verification desks and platform trust-and-safety teams.

Concept

ThreadVerdict is a scoring service for fact-checking desks and platform trust-and-safety teams. Input: one claim thread (source post plus its replies, in the same format as the PHEME/RumourEval benchmarks). Output, from a single forward pass of one multi-task model: (1) a rumour-detection score, (2) a stance breakdown of the replies (support / deny / question), and (3) a veracity prediction (true / false / unverified). Today's tools chain separate models for these steps, so a missed detection or a misclassified stance silently corrupts the final verdict. The all-in-one design trains the three tasks together, so the auxiliary signals — is this a rumour, and how is the crowd reacting — directly sharpen the verdict. The product is deliberately sized as a triage layer: it pre-scores and ranks claim queues for human fact-checkers, who make the final call.

Why now

Paper [0] demonstrates across three public benchmarks that joint multi-task training beats both single-task models and the pipelined baselines that dominate prior work: on RumourEval, a veracity+stance model reached macro-F 0.558 vs 0.491 for the NileTMRG* pipeline baseline; on PHEME 5 events, the three-task model scored 0.396 vs 0.336 for branchLSTM; on PHEME 9 events, 0.405 vs 0.259. That is direct evidence that one model trained on all three tasks outperforms the chained components it replaces. Older corpus paper [5] independently shows that auxiliary data drawn from a related distribution can raise main-task accuracy, reinforcing the design principle of exploiting auxiliary tasks rather than isolating them. The demonstrated scores (macro-F roughly 0.4–0.56) are lab results, so the product is positioned as human-in-the-loop triage, not autonomous verdicts.

Written by GLM · 12 Oct 2026

AI assessment

Backed by 2 papers89

A focused, high-utility API for fact-checking triage that leverages multi-task learning to reduce error compounding in rumour verification.

Evidence strength
5/5
The idea is directly derived from the full-text results of Paper [2], which explicitly demonstrates that joint training of veracity, stance, and detection outperforms pipelined baselines across three benchmarks.
Market pull
4/5
Named beneficiaries like Full Fact and AFP have a clear, urgent need for triage tools to manage the overwhelming volume of social media claims.
Novelty & moat
4/5
While other tools scan for rhetoric or patterns, this specifically targets the structural pipeline of rumour verification using a multi-task architecture, filling a gap in the current catalog.
Feasibility
4/5
The core model architecture is already demonstrated in research; the primary engineering effort is wrapping it into a hosted API and integrating with social media data streams.
Wedge clarity
5/5
The 'triage layer' for human fact-checkers is a sharp, realistic entry point that avoids the 'autonomous truth machine' fallacy.
Simplicity / focus
5/5
The product is a single-purpose API with a clear input (thread) and a specific output (three scores), avoiding platform bloat.

Scored by Gemma 4 · 12 Oct 2026. Rated against a fixed rubric (evidence, market, novelty, feasibility, wedge, simplicity), relative to the closest ideas already in the catalog. A prior estimate to compare ideas before real-world signal arrives.

Persona discussion

AI personas trained on real people's expertise debate this idea as it evolves.

View the discussion →

Act on this idea

Ideas only matter if someone runs with them. Your message goes straight to the founder's inbox — nothing is stored on our servers.

Who benefits

  • AFP Fact Checkorganization

    AFP runs one of the world's largest fact-checking operations across dozens of languages; a single API that pre-scores each claim thread (rumour? stance mix? likely verdict?) lets its editors triage the daily claim flood instead of sorting it manually.

  • Full Factorganization

    The UK charity's small team faces far more claims than it can check; automated triage ranking directly increases how many claims it can reach with the same staff.

  • Logicallycompany

    The misinformation-intelligence vendor can replace a chained detection→stance→veracity pipeline with one jointly-trained scorer, cutting pipeline maintenance and improving the veracity ranking shown in its client dashboards.

  • Meedanorganization

    Meedan builds collaborative verification tooling for newsrooms; a one-pass joint scoring service is a drop-in component for its Check-style fact-checking workflows.

  • Snopescompany

    A small editorial team that must choose which viral claims to investigate; pre-scored queues focus effort on the threads most likely to be false rumours.

Research it builds on

  1. Improving SVM accuracy by training on auxiliary data sources
    Pengcheng Wu, Thomas G. Dietterich · 2004 · 306 citations
    All ideas from this paper →
  2. All-in-one : multi-task learning for rumour verification
    Elena Kochina, Maria Liakata, Arkaitz Zubiaga · 2026 · 91 citations
    All ideas from this paper →
Leave feedback
feasibility