Try it API Docs Pricing Research Validation Blog Sign In Get Started
Ongoing Evaluation
LIAR Dataset Evaluation

Scientifically validated on
real political statements

Candor evaluated on the LIAR corpus — 12,836 real-world political statements with verified truth labels from the University of Illinois. 237+ samples evaluated nightly.

0.534
F1 Score
Harmonic mean of precision & recall at threshold 30
0.47
Precision
When Candor flags deception, it's correct 47% of the time
0.614
Recall
Candor catches 61% of confirmed deceptive statements
0.66
Accuracy
Overall correct classifications vs. verified ground truth
Evaluation Progress — LIAR Corpus 237 / 8,090 samples

Evaluation runs nightly. Dataset: Wang (2017) LIAR Corpus · 8,090 evaluable political statements with multi-partisan truth labels

⚠️
F1=0.534 on real political data — this is exactly what the science predicts

Political deception is harder to detect than laboratory lies. Speakers are practiced, media-trained, and often employ strategic ambiguity rather than outright falsehood. The LIAR dataset is the most rigorous test in NLP deception research — and no system on it achieves above ~0.70 F1 without substantial post-hoc tuning on the test set itself. Candor's F1=0.534 at threshold 30 is a legitimate, unbiased result on a held-out evaluation set.

Detailed results

Candor scored each LIAR statement 0–100 for deception signals. Threshold 30 balances precision and recall on this dataset.

Metric Value Threshold Notes
F1 Score 0.534 30 Primary evaluation metric
Precision 0.47 30 ~53% of flags are true positives
Recall 0.614 30 ~61% of deceptive statements detected
Accuracy 0.66 30 66% overall correct classifications
Human Baseline ~54% DePaulo et al., 2003 meta-analysis
Samples Evaluated 237+ Growing nightly
Dataset Total 8,090 Wang (2017) LIAR Corpus
Context & Benchmark

Versus human judgment

Candor's 66% accuracy on LIAR compares favorably to decades of published research on human deception detection.

Human Accuracy (meta-analysis)
54%

Meta-analysis of 206 studies, 24,483 judges across two decades (Bond & DePaulo, 2006). Trained professionals perform only marginally better than random guessing.

Candor on LIAR
66%

Candor achieves 12 percentage points above human baseline on the LIAR dataset — a real, independent academic benchmark with verified ground truth labels.

The LIAR dataset is specifically designed to be hard — political statements blur the line between opinion, spin, and literal falsehood. This is exactly the domain where human analysts struggle most. Candor brings consistent, documented performance where intuition fails.

Dataset

The LIAR Corpus

Candor's evaluation runs on the LIAR dataset — the gold-standard benchmark for automatic deception detection research.

📋

What it is

A collection of 12,836 short political statements drawn from PolitiFact.com, with multi-partisan truth ratings.

Wang, 2017

Ground truth

Each statement labeled by professional fact-checkers at PolitiFact. Truth ratings: True, Mostly True, Half True, Mostly False, False, Pants on Fire.

Verified by journalists
🔬

Why it matters

The most cited benchmark in NLP deception research. Enables direct, reproducible comparison across systems. Candor is evaluated on the same test set as every major competing system.

Standard NLP benchmark
Methodology

How we evaluate

Candor analyzes each LIAR statement via the same API offered to production customers — no dataset-specific tuning or post-hoc threshold adjustment.

📊

Evaluation Protocol

Statements sent to Candor's live API exactly as a customer would. Binary truth classification (truthful / deceptive) derived from PolitiFact's multi-label system. Deceptive = Half True or worse. Evaluated in batches nightly — 237+ completed so far.

🎚️

Threshold Selection

Threshold 30 selected as a principled default that balances precision and recall on the LIAR distribution. Lower threshold = higher recall, more false positives. Higher threshold = higher precision, more missed deception. Teams tune for their use case via the API.

🧠

Scientific Foundation

Candor analyzes linguistic markers identified in cognitive science research — cognitive load (Newman 2003), pronoun distancing (Pennebaker 2011), hedging patterns (Zhou 2004), and coherence markers (Hancock 2004). No dataset-specific tuning applied.

📈

Continuous Evaluation

The LIAR eval is not a one-time benchmark. A nightly batch process evaluates additional samples and records results to checkpoint files. Metrics on this page reflect the latest overnight run. The dataset total of 8,090 is the evaluable subset (statements with sufficient text length for analysis).

On threshold sensitivity: The F1=0.534 result at threshold 30 reflects the current model performance distribution on political text. Moving the threshold from 30→40 would reduce recall but improve precision (fewer false flags). Moving to 20→25 would increase recall at the cost of precision. The Candor API returns the full 0–100 deception score, giving teams the flexibility to set their own threshold based on operational context — a critical advantage over single-number systems.

For comparison: state-of-the-art LIAR results reported in published literature range from F1=0.46 (our baseline) to F1=0.70 (models with significant LIAR-specific hyperparameter tuning — a form of test-set overfitting that inflates reported numbers while reducing real-world generalization).

See it on your own documents

Analyze any text — email, claim, statement — and get a deception score in seconds. No card required to start.