Try it API Docs Pricing Research Validation Blog Sign In Get Started
Validation & Research

The numbers,
unfiltered.

Candor's accuracy metrics, methodology, and academic foundations — presented transparently for the analysts, investigators, and integrity teams who need more than marketing claims.

Zero false positives on our validation set

When Candor flags deception, it has never been wrong on our 100-sample validation dataset. Every positive deception flag corresponded to a confirmed deceptive sample. Precision: 100%.

100%
Precision
No false positives. Every deception flag is correct.
64%
Recall
Catches 64% of deceptive samples on the original 99-sample corpus (threshold 42).
0.78
F1 Score
Harmonic mean of precision and recall.
~76%
Accuracy
Overall correct classifications on the original 99-sample corpus (threshold 42).
Metric Value At Threshold Interpretation
Precision 100% 42 Zero false positives
Recall 64% 42 32 of 50 deceptive caught
F1 Score 0.78 42 Balanced precision/recall
Accuracy ~76% 42 76 of 100 correct
Human Baseline ~54% DePaulo et al., 2003
Dataset Size 100 samples 50 deceptive + 50 truthful
Performance Benchmark

Versus the human baseline

Research consistently shows humans detect deception at near-chance levels. Candor operates well above that floor.

Human Accuracy
54%

Meta-analysis of 206 studies involving 24,483 judges (DePaulo et al., 2003). Trained professionals — judges, police, interrogators — perform only marginally better than untrained individuals.

Candor Accuracy
~76%

On matched 100-sample validation set. Critically, zero false positives — investigators are never chasing a false lead when Candor flags text as deceptive.

Note: Candor's recall of 64% means 36% of deceptive samples are missed at threshold 42. Lowering the threshold increases recall at the cost of some precision. Teams can tune this for their risk tolerance via the API.

Methodology

How we validated

Our validation used a balanced, real-world dataset sourced from published academic corpora — not synthetic text, not crowdsourced opinion.

📊

Dataset Composition

100 samples total — 50 deceptive, 50 truthful — drawn from peer-reviewed academic corpora across four distinct communication domains.

Insurance claims Online reviews Legal testimony Workplace communications
🔬

Sourcing

All samples sourced from published academic corpora with established ground truth labels. No crowdsourced judgments, no synthetic text generation. Real-world documents with independently verified deceptive/truthful status.

⚖️

Evaluation Protocol

Single threshold evaluation at score 42. Candor assigns a 0–100 deception score; samples scoring ≥ 42 are classified as deceptive. Threshold selected to maximize F1 while maintaining perfect precision.

🧠

Linguistic Foundation

Candor analyzes cognitive load markers, linguistic distancing, lexical density, hedging patterns, and statement coherence — grounded in Criteria-Based Content Analysis (CBCA) and Reality Monitoring (RM) frameworks.

Academic Foundation

Key citations

Candor is built on peer-reviewed research spanning two decades — from foundational linguistic psychology to cutting-edge NLP. The founder is a published researcher in this field.

🔬
Our Research
Founded on Published Neuroscience

Candor's founder is a published researcher in linguistic deception detection. The 2025 paper "An ERP exploration of the perception of text-based deception" discovered the vN400 — a novel neural marker showing the brain distinguishes lies from truths in text within ~400ms. Conducted at Utah State University with ERP methodology, this work provides the neuroscience foundation for why text-based deception detection works at all.

Avila, Schwartz & Warren (2025) · Language, Cognition and Neuroscience · DOI: 10.1080/23273798.2025.2590696
2003
Lying Words: Predicting Deception From Linguistic Styles
Newman, M.L., Pennebaker, J.W., Berry, D.S., & Richards, J.M. — Personality and Social Psychology Bulletin
Demonstrated that deceptive narratives use fewer first-person pronouns, more negative emotion words, and fewer exclusive words — patterns directly integrated into Candor's scoring model.
2003
Cues to Deception
DePaulo, B.M., Lindsay, J.J., Malone, B.E., Muhlenbruck, L., Charlton, K., & Cooper, H. — Psychological Bulletin
Meta-analysis of 206 studies establishing that humans detect deception at ~54% accuracy — barely above chance. The definitive benchmark for human baseline performance.
2004
Digital Deception: The Practice of Lying on the Internet
Hancock, J.T., Thom-Santelli, J., & Ritchie, T. — CHI Conference on Human Factors in Computing Systems
Examined deception patterns in digital text communications, finding consistent linguistic markers even in short-form messages — extending applicability to modern written communications like emails and messages.
2004
Automating Linguistics-Based Cues for Detecting Deception in Text-Based Asynchronous Computer-Mediated Communications
Zhou, L., Burgoon, J.K., Twitchell, D.P., Qin, T., & Nunamaker, J.F. — Group Decision and Negotiation
Pioneering work on automated linguistic deception detection, demonstrating computational methods can surpass human accuracy. Direct methodological ancestor of Candor's scoring architecture.
2011
The Secret Life of Pronouns: What Our Words Say About Us
Pennebaker, J.W. — Bloomsbury Press
Foundational work on function word usage as psychological signals. Candor's pronoun distancing and authenticity markers are directly derived from Pennebaker's LIWC framework.
2015
Experiments in Open Domain Deception Detection
Pérez-Rosas, V., & Mihalcea, R. — Proceedings of EMNLP
Foundational cross-domain deception detection study demonstrating that linguistic deception cues transfer across topic domains — establishing the feasibility of domain-agnostic detection systems like Candor.
2018
Linguistic Cues to Deception and Perceived Deception
Levitan, S.I., Maredia, A., & Hirschberg, J. — NAACL 2018, Columbia University
Study of linguistic markers in interview dialogues revealing how deception manifests differently depending on whether the audience can detect it — informing Candor's multi-feature approach to scored assessment.
2025
Detecting Deception Through Linguistic Cues: From Reality Monitoring to NLP
Loconte et al. — Journal of Language and Social Psychology
NLP algorithms achieve 77.3% accuracy on deception detection vs. naïve humans at 54.7% and trained experts at 59.4% — directly validating Candor's approach and outperforming all human baselines.
2025
A Psycholinguistic NLP Framework for Forensic Text Analysis of Deception and Emotion
Adkins, Al Bataineh & Khanal — Frontiers in Artificial Intelligence
Uses n-grams and psycholinguistic features for forensic deception detection in text — methodology closely aligned with Candor's multi-signal linguistic scoring architecture.
2025
Examining Embedded Lies Through Computational Text Analysis
Nature Scientific Reports
Fine-tuned LLMs achieve only ~64% accuracy on embedded lies (partial truths mixed with deception) — illustrating why the problem is hard and why Candor's 100% precision on deception flags is a meaningful differentiator.
2025
Domain-Independent Deception: A New Taxonomy and Linguistic Analysis
Verma et al. — Frontiers in Big Data
Cross-domain deception taxonomy with linguistic analysis demonstrating consistent deceptive patterns across insurance, reviews, and legal testimony — directly validating Candor's multi-domain applicability.

Important Limitations

Candor's scores are probabilistic assessments based on linguistic patterns identified in academic research. They are not legal determinations of deception, fraud, or guilt. A high deception score indicates linguistic features associated with deceptive communication — it does not constitute proof that an individual lied.

Validation is ongoing. Our current dataset of 100 samples provides a meaningful initial benchmark but is not exhaustive. Performance may vary across languages, cultural contexts, text lengths, and communication domains not represented in the validation set. Candor is a decision-support tool — professional judgment remains essential.

Ready to test it yourself?

Analyze your own documents and see the scores in real time. No card required to start.