Research · KVAi Lab

The papers behind Newjee.

Newjee is built on published methodology, not opinions. The papers below are produced by the KVAi Lab — the AI laboratory of Kakashi Ventures Accelerator — and define the formal apparatus the platform computes on. Listed from the most recent preprint backwards.

  • Foundational · 2024

    From Fact-Checking to Framing Measurement: A KPI-Driven Framework for Media Analysis and Information Disorder

    Federico Bottino

    Foundational white paper · KVAi Lab

    Foundational white paper for Newjee. Argues that meaningful media analysis requires moving beyond binary fact-checking toward measurable language-framing analysis. Building on the information-disorder typology of Wardle (First Draft Coalition; Wardle and Derakhshan, 2017), the propaganda-technique inventory of Da San Martino et al. (2020), and the framing taxonomy of Semetko and Valkenburg (2000), the paper introduces a measurement framework that treats news content the way product analytics treats user behaviour: a stream of structured signals that admit reproducible KPIs — frame distribution, manipulation indicators, source trust, propagation velocity, contradiction load — rather than discrete true/false verdicts. Sets the conceptual foundation that every subsequent KVAi Lab paper operationalises.

    • Foundational
    • Information disorder
    • Framing measurement
    • KPIs
    • Newjee thesis
  • Preprint · 2026

    Bloc-Conditional Event States: Spectral Decomposition, Permutation Inference, and Cross-Coverage Divergence for Threat-Intelligence Analysis

    Maryam Fooladi, Federico Bottino, Manuel Peruzzo, Nicholas Dosio

    Preprint · KVAi Lab · Draft, 30 April 2026

    Content-level measurement of cross-bloc framing divergence in news coverage of contested events. Each editorial bloc is represented as a density matrix ρ on a 15-dim framing-feature space (5 Semetko–Valkenburg frames · 4 dimension features · 6 Da San Martino manipulation indicators); trace distance D(ρ_state, ρ_mainstream) measures cross-bloc divergence. Three inferential layers — bloc-permutation null, within-bloc bootstrap, leave-one-outlet-out — test partition-construction circularity, sample stability, and outlet-level robustness against an enumerated within-mainstream baseline. The leading rank-2 subspace decomposes into a frame-axis and a manipulation-axis pair on both case studies (Hormuz blockade 2026, n = 16; Navalny death 2024, n = 14). The work positions the bloc-difference operator as a content-level analogue to the source-level signals on which threat-intelligence workflows currently rely.

    • Density matrices
    • Spectral decomposition
    • Trace distance
    • Permutation inference
    • Bootstrap
    • Threat intelligence
    • Hormuz
    • Navalny
  • Accepted · 2026

    A Multi-Layer AI Framework for Information Disorder

    Federico Bottino, Maryam Fooladi

    LREC 2026 — Workshop on Information Disorder

    Multi-layer extraction architecture that turns raw news signals into the 15-dim epistemic vector at the heart of Newjee. Establishes the Frame · Dimension · Manipulation block decomposition and the per-feature evidence-span requirement that grounds every score in verbatim text. The architecture separates strategic narrative interpretation, operational extraction, regulatory mapping, and human-AI collaboration into distinct layers with structured interfaces — replacing the monolithic-LLM approach that fails on information-landscape analysis.

    • Multi-layer architecture
    • Information disorder
    • Epistemic vectors
    • Framing analysis
    • Pipeline
  • Accepted · 2026

    Beyond Sentiment: Comparing Traditional NLP and LLM-Based Multi-Dimensional Analysis for Political News Evaluation

    Maryam Fooladi, Federico Bottino · KVA Accelerator

    LREC 2026 — Political NLP workshop

    Comparative study of RoBERTa-based sentiment classifiers vs. LLM-based multi-dimensional analysis on political news. Identifies the "neutral collapse" — RoBERTa labels ~70% of political articles as neutral, with 23% of those carrying negative-probability scores above 0.30 — and demonstrates that LLM-based frameworks expose framing direction and intensity, sensationalism, emotional appeal, and ideological lean that single-axis sentiment hides. Applied to a corpus of 50 political articles from 17 international outlets.

    • Sentiment analysis
    • LLMs
    • Political news
    • Multi-dimensional framing
    • NLP evaluation

Want to publish with us?

KVAi Lab collaborates with researchers and labs working on information disorder, computational journalism, narrative dynamics, and epistemic measurement. Active directions: a balanced-corpus replication of the Bloc-Conditional Event States rank-2 finding (n ≥ 10 events), the cross-domain bridge to adoption-signal density on stance-typed RKHS, and a Scientific Perimeter Validation Layer (claim ↔ recognized scientific evidence) integrated with a Wikipedia/Wikidata consensus baseline.

Contact KVAi Lab — hello@kakashi.ventures
Ask me to verify a claim ✓