Technical Report BD-TR-2026-01 · Published 2026-10-01 · Open Access (CC BY 4.0)
Primary-Source Grounding and Automated Claim Verification in Real-Time Cryptocurrency Newsrooms
Architecture, Dual-Pass Audit Thresholds, and Empirical Evaluation of the Basis Desk Pipeline
Authors: Julian Mercer (ORCID 0009-0003-1162-170X) & Basis Desk Research Desk
JEL Classification: G14, G18, C81 · ACM: I.2.7 Natural Language Processing ; H.3.5 Online Information Services
Abstract
Digital asset markets suffer from severe information asymmetry, latency-driven rumor propagation, and unverified secondary reporting. This paper presents the architecture of Basis Desk (https://basisdesk.news), an autonomous, primary-source-first newsroom engineered to ingest, cluster, draft, and verify financial and cryptocurrency news in real time. Rather than rewriting secondary media coverage, the pipeline ingests documents directly from regulatory agencies, courts, corporate filings, protocol repositories, and public blockchains, using secondary media headlines strictly as event-discovery signals to locate underlying primary documents. Every candidate draft undergoes an independent second-pass grounding audit that verifies factual claims, dates, entities, and numerical figures against the source text and live market snapshots, enforcing a strict verbatim-overlap ceiling and a minimum publication quality score of 80/100.
Keywords: primary-source verification · automated fact-checking · cryptocurrency journalism · grounding audit · hallucination detection · financial NLP
1. Introduction and Problem Statement
Information cascades in cryptocurrency markets frequently originate from misinterpreted regulatory filings, spoofed announcements, or unverified social media posts. Traditional aggregation pipelines optimize for publication speed at the expense of provenance, amplifying factual errors across downstream syndication networks and large language model (LLM) retrieval systems.
Basis Desk addresses this failure mode by enforcing a strict separation between event discovery and factual grounding. Secondary media feeds are configured with textUse=false: their article bodies never enter a generation prompt. Instead, outbound links are parsed to retrieve the underlying primary document (e.g., an SEC Form 8-K, CFTC enforcement release, court docket, or protocol governance proposal) before any drafting workflow can execute.
2. Primary-Source Registry and Vector Clustering
The ingestion engine polls 70+ curated primary feeds across regulators (SEC, CFTC, DOJ, Federal Reserve, UK FCA, ESMA, ECB, MAS, SFC), SEC EDGAR material-event filings, protocol release registries, and security researchers at 5-to-30-minute intervals. Incoming documents are canonicalized, hashed (FNV-1a), and embedded into a 768-dimensional vector space (bge-base-en-v1.5) to cluster multi-source coverage of the same event within a 72-hour rolling window (cosine similarity >= 0.76).
To prevent re-reporting of already-covered events across rolling news cycles, a three-layer deduplication gate combines (i) in-tick entity token exclusion, (ii) a strongly-consistent 96-hour D1 lexical entity/number overlap check, and (iii) a 7-day Vectorize semantic similarity ceiling (0.855 at publish time).
3. Dual-Pass Grounding Audit and Anti-Copy Enforcement
Each eligible story cluster is processed by a durable multi-step workflow (gather -> draft -> verify -> repair -> publish). During the verification stage, an independent auditor model evaluates every factual assertion in the headline, deck, key points, At-a-Glance summary, and body against the numbered primary source texts and a timestamped market snapshot.
The deterministic verifier (verifyDraft) combines LLM audit findings with programmatic checks: citation indices [1..k] must map to valid ingested sources, verbatim 8-gram overlap with source texts must remain below threshold, advice or price-target phrasing triggers a hard fail, and drafts scoring below 80/100 undergo a targeted repair pass or are diverted to human editorial review.
4. Structured Output for Answer Engines and Researchers
Published articles expose structured At-a-Glance metadata (what happened, why it matters, who is affected, effective date, and regulatory action status), clean Markdown twins (<url>.md), Schema.org NewsArticle / AnalysisNewsArticle JSON-LD, and a public Model Context Protocol (MCP) server at https://basisdesk.news/mcp.
The longitudinal dataset of verified articles, source attributions, and audit metrics is released under CC BY 4.0 at https://basisdesk.news/api/research/dataset.csv and https://basisdesk.news/api/research/dataset.jsonl.
References & Linked Endpoints
- Basis Desk Editorial Policy & Verification Standards (2026). https://basisdesk.news/about/editorial-policy
- How Basis Desk Uses AI: Dual-Pass Grounding Architecture (2026). https://basisdesk.news/about/how-we-use-ai
- Basis Desk Longitudinal Dataset & Frictionless Data Package (2026). https://basisdesk.news/api/research/datapackage.json
- Basis Desk Model Context Protocol (MCP) Specification (2026). https://basisdesk.news/developers
How to Cite (BD-TR-2026-01)
APA 7th Edition:
Mercer, J., & Basis Desk Research Desk. (2026). Primary-Source Grounding and Automated Claim Verification in Real-Time Cryptocurrency Newsrooms: Architecture, Dual-Pass Audit Thresholds, and Empirical Evaluation of the Basis Desk Pipeline (Technical Report No. BD-TR-2026-01). Basis Desk. https://basisdesk.news/research/bd-tr-2026-01
BibTeX:
@techreport{basisdesk_bd_tr_2026_01,
title = {Primary-Source Grounding and Automated Claim Verification in Real-Time Cryptocurrency Newsrooms: Architecture, Dual-Pass Audit Thresholds, and Empirical Evaluation of the Basis Desk Pipeline},
author = {Mercer, Julian and {Basis Desk Research Desk}},
institution = {Basis Desk},
number = {BD-TR-2026-01},
year = {2026},
month = {10},
url = {https://basisdesk.news/research/bd-tr-2026-01},
pdf = {https://basisdesk.news/papers/BasisDesk-2026-Primary-Source-Verification-and-Grounding-Audit-Architecture.pdf},
note = {ORCID: 0009-0003-1162-170X. Open access under CC BY 4.0}
}