Source Quorum: provenance and corroboration checks for news-reading trading agents

No company I found offers a per-headline check of a story's origin and independent corroboration, with a hold-or-allow verdict and a suggested delay, built to run right before an LLM agent calls its order tool. I also found no commercial service that replays fabricated headlines through a client's agent and reports which fakes caused trades and whether several agents moved together. The research (Fake News Real Crashes, TradeTrap, AutoRedTrader, Adversarial News) makes the risk concrete and gives the drill a credible basis. Today's options are institutional news data from RavenPack/Bigdata.com and Dataminr, which is not delivered as a pre-trade verdict; general guardrails from Lakera, NeMo and Future AGI, which catch prompt injection but not true-sounding fake news; and hobby-scale crypto confirmation feeds. A newcomer could win with crypto bot platforms and retail AI trading-assistant vendors that cannot afford RavenPack, by offering an MCP or tool-call integration that deduplicates syndicated copies so twenty reprints of one wire story don't count as confirmation, and a timestamped audit log, which a vendor page says helps with SEC 15c3-5 and FINRA 3110 duties. The main risks are that RavenPack or Bigdata.com could add a verdict endpoint fairly easily, that timing when a story first appeared on each independent wire requires paid real-time feed access, and that many vendors may skip a paid check and simply require human approval or two-source confirmation, since a delay costs them speed in the seconds after news breaks.

• Fake News, Real Crashes (arXiv 2610.26438)

  Research, not a product. Shows that fabricated headlines push LLM trading agents into the same trade at once, and that cues like a named source or urgency tag decide whether a fake works. It shows the problem is real and is likely the basis for this idea.

• TradeTrap

  Academic setup that injects coordinated fake news into agents' news and social feeds while leaving everything else untouched. One agent's Sharpe ratio fell from 4.34 to 3.20. It overlaps conceptually with the Injection drill but is not sold as a service.

• AutoRedTrader (arXiv 2605.09185)

  Automated red-teaming tool that generates finance-specific fake news and refines attacks using the agent's responses. It reports up to 26.67% attack success and is the closest prior work to the Injection drill. It is research; I could not confirm whether its code or dataset is public.

• Adversarial News and Lost Profits (arXiv 2601.13082)

  Headline attacks using look-alike characters and hidden text. A one-day attack cut annual returns by up to 17.7 percentage points in a backtest. Research only.

• Crypto Intel API (DEV Community project)

  Free developer API that reports crypto hacks, depegs, delistings and lawsuits only after two or more independent sources confirm them, with a confidence score, source list and source trust weights. Closest small product to the Quorum API, but it covers whole events, not individual headlines, has no hold-or-allow verdict, and is a solo hobby project with no independent coverage.

• RavenPack / Bigdata.com

  Describes source credibility scores based on publisher reliability and topic, plus tracking how information spreads between sources. Its news analytics include novelty and relevance scores used by quant clients. It is the incumbent best placed to add a pre-trade verdict endpoint, but today it sells institutional data rather than per-headline checks for agents.

• Dataminr

  Breaking-news alerts from social media, historically pitched as filtering credible market-moving posts. A 2026 review says verification still falls to the user and that it publishes no credibility score.

• ChainGPT AI Crypto News API

  AI-written crypto news feed that says it cross-checks facts internally. It sells news content, not a provenance score for third-party headlines.

• Qveris financial news API guide

  Recommends checking source type, freshness, corroboration and policy approval before news can trigger an action, and says one headline should never authorize an irreversible trade. Best practice that matches the idea, not a product that does it.

• APITube

  News data API whose blog covers cross-referencing sources and notes that copies of one wire story are not corroboration. I did not confirm whether it offers a verdict or verification endpoint.

• CLEANROOM (lablab.ai hackathon)

  Hackathon design that keeps an untrusted LLM news-reading step, limited to validated structured output, separate from the trading logic. A defensive architecture, not a scoring service.

• Future AGI and other general LLM guardrail vendors

  Lists guardrail products for fintech AI, including Lakera, NeMo Guardrails, Protect AI and AWS Bedrock Guardrails. They check inputs and tool calls for prompt injection and policy violations but do not check whether a news item is true or corroborated.

• Aleta Index

  Old sponsor listing on a now-defunct festival site describing a source-credibility index and an API that flags misinformation risk in trades. I could not confirm it still exists.

Checked 2026-10-20