rnel: a Python library for evidential reasoning that keeps contradiction, ignorance and retracted evidence apart

Resumo

A single probability of truth cannot tell two confident sources that contradict each other from a fair coin, and a pipeline that removes discredited evidence can silently remove it twice. This article presents rnel, an open-source Python library for evidential reasoning that keeps these cases apart. It implements Subjective Logic, the Refined Neutrosophic Evidential Logic (RNEL) tuple, in which contradiction between sources, undetermination, an ill-posed claim and lack of evidence are separate components, the priority-product operators of the logics that RNEL extends with their base-rate-calibrated versions, ranking and decision layers, and an experimental ledger of signed evidence in which reports can be retracted per source and over-retraction is refused or flagged. The core depends only on the Python standard library; optional modules add neural evidential models and a zero-shot reader that builds a tuple from a claim and its evidence texts. Theoretical results are encoded as automated tests (450 tests pass), and a verification script checks the signed-evidence theorems numerically. Verification also found a non-monotone decision rule in version 0.1.0, in which a report supporting a claim could make the answer refuted; the article documents the fix. Four separation cases, the design, the interface and the limitations are described. The library is released under the MIT licence.