🧪 This platform is in early beta. Features may change and you might encounter bugs. We appreciate your patience!
How we built a machine-readable taxonomy of human reasoning patterns.
Traditional fallacy taxonomies are flat lists organized by a single criterion (e.g., formal vs. informal). This limits their analytical power because real-world flawed reasoning rarely fits a single category. A conspiratorial argument might simultaneously employ a logical fallacy (false cause), exploit a cognitive bias (confirmation bias), use a propaganda technique (scapegoating), and commit a statistical error (cherry picking).
TellDear addresses this by organizing reasoning aspects along six orthogonal dimensions, each capturing a distinct facet of how reasoning can go wrong — or go right:
Structural and content errors in deductive and inductive reasoning. Includes formal fallacies (invalid logical forms like affirming the consequent) and informal fallacies (relevance, ambiguity, and presumption errors like ad hominem or straw man). Based on Aristotle's Sophistical Refutations, Hamblin's taxonomy, and modern critical thinking literature.
Rhetorical and psychological techniques designed to persuade through deception rather than rational argument. Covers classic propaganda (IPA's seven techniques), digital-age disinformation tactics (firehose of falsehood, astroturfing), and interpersonal manipulation strategies (DARVO, gaslighting). Sources include RAND Corporation, EU DisinfoLab, and the Institute for Propaganda Analysis.
Systematic psychological deviations from rational judgment, organized by domain: decision-making biases (loss aversion, sunk cost), social biases (fundamental attribution error, in-group bias), memory biases (peak-end rule, illusory truth), attention biases (frequency illusion, salience bias), and probability biases (gambler's fallacy, base rate neglect). Based on Kahneman & Tversky's heuristics-and-biases program.
Errors in data analysis, interpretation, and presentation. Covers research methodology pitfalls (p-hacking, publication bias, underpowered studies), data visualization deceptions (truncated axes, misleading aggregation), and probability reasoning errors (Simpson's paradox, Berkson's paradox). Grounded in Ioannidis (2005), Gigerenzer's risk literacy research, and ASA guidelines on p-values.
Normative templates for rational argumentation from Walton, Reed & Macagno's compendium of 60+ schemes. Includes argument from expert opinion, argument from analogy, argument from consequences, practical reasoning, and others. Each scheme has critical questions that must be answered for the argument to succeed — failure to answer constitutes a fallacious use.
Meta-level patterns in how discussions and debates are conducted. Covers bad-faith tactics (goalpost moving, tone policing, sealioning), digital discourse phenomena (concern trolling, just asking questions), and structural argument patterns (motte and bailey, Kafka trap). These are not flaws in individual arguments but in the process of argumentation itself.
The TellDear knowledge graph is a directed labeled graph G = (V, E, L) where vertices represent reasoning concepts and edges represent typed relations between them. The graph serves as the terminological box (TBox) of a lightweight ontology.
| Class | Count | Description |
|---|---|---|
dimension | 6 | Top-level organizational axes |
category | varies | Sub-groupings within dimensions |
formal_fallacy | 27 | Structurally invalid argument forms |
informal_fallacy | 66 | Content-based reasoning errors |
propaganda | 50 | Persuasion through deception |
bias | 109 | Systematic cognitive deviations |
stat_error | 79 | Data analysis errors |
scheme | 26 | Argumentation templates |
discourse | 94 | Discussion-level patterns |
Edges in the graph are typed with one of six relation labels:
is_a — Taxonomic hierarchy (e.g., ad_hominem is_a informal_fallacy)sub_type_of — Specialization within a categorytriggers — Causal enablement (e.g., confirmation_bias triggers cherry_picking)correlates_with — Cross-dimensional associationpipeline_next — Sequential processing dependencytool_supports — Tooling for detection or analysisThe Atomic Instruction Dataset is the operational core of TellDear. For each reasoning aspect, AID provides a sequence of 2–4 binary verification steps — yes/no questions that can be answered independently to determine whether the aspect is present in a given text.
This approach is inspired by the Atomic Skills methodology in NLP evaluation, where complex capabilities are decomposed into minimal, independently testable units. The binary constraint ensures that each step has a clear, unambiguous answer, reducing inter-annotator disagreement and enabling reliable automated assessment.
The TellDear analyzer implements a systematic verification protocol for each aspect:
The protocol supports two execution modes: heuristic (client-side keyword matching for instant results) and deep analysis (LLM-powered evaluation for nuanced semantic understanding).
A key innovation of TellDear is the systematic mapping of cross-dimensional relationships. These lateral connections reveal how reasoning flaws interact and reinforce each other:
Cognitive biases create psychological conditions that make certain fallacies more likely. For example, confirmation bias triggers cherry picking — a person who seeks confirming evidence will naturally select only supporting data points.
Many propaganda techniques are formalized versions of informal fallacies deployed at scale. Appeal to fear is both a logical fallacy and a propaganda technique; the distinction lies in whether it is an individual reasoning error or a deliberate persuasion strategy.
Statistical errors can both result from and reinforce cognitive biases. Survivorship bias leads to sampling bias in data collection, which produces misleading results that further entrench the original bias.
These relationships are formally represented as triggers and correlates_with edges in the knowledge graph, enabling automated chain-of-reasoning analysis that goes beyond detecting individual flaws.