The Instrument — GFIS: A Verifiable Agentic Research Pipeline. Methods Note (System State as of 31 July 2026)

Abstract

Methods note (instrument paper), bilingual: full Hungarian and English texts included. System state as of 31 July 2026. This note documents GFIS (Gestalt Field Intelligence System), the coverage-driven, iterative, multi-source research pipeline behind the evidence metrics cited in the author's 2026 research essays ("triangulated claims", "coverage", "adversarial analyses"). It provides the exact, code-level operationalization of these metrics (triangulation via within-dimension similarity matching with domain-based source independence; dimension-ratio coverage with composite dimension scores; two-stage confidence calibration with abstention; rival-model counter-hypothesis generation), the pipeline stages and quality layers (CRAG/RAGAs gates, prompt-injection screening, entailment verification, loss manifest with algorithm fingerprinting), a 2026 version timeline mapping earlier publications to the system versions that produced their metrics, and a dedicated limitations section (process metrics vs. phenomenon validation, LLM-judge noise, correlated blind spots, single-developer validation). Provenance and cutoff: the note documents the system state as of 31 July 2026 (repository commit e865151); every run, code change, configuration and metric referenced predates that cutoff and was produced during the author's period as an independent researcher, on private infrastructure. The note is a private work describing pre-existing private intellectual property and represents no institutional position. This note supplies the methodological background for the author's earlier deposits: The Silent Agent (10.5281/zenodo.21738075), Attention Exchange (10.5281/zenodo.21738332), The Real Impact of AI Slop (10.5281/zenodo.21738336).