Polygraph Scoring Algorithms: PolyScore, OSS-3 & CPS Guide

The Hand Scoring Era

For decades, polygraph chart interpretation was an exclusively human endeavor. Examiners manually reviewed physiological tracings — skin conductance (electrodermal activity), cardiovascular measures, and respiration patterns — and applied structured numerical scoring rules to determine whether deception was indicated. This process relied on methods such as the 7-position scoring scale formalized by Bell, Raskin, Honts, and Kircher in their landmark 1999 publication on the Utah Numerical Scoring System, which resulted from over 30 years of scientific research and provided some of the highest rates of criterion accuracy and inter-rater reliability [8]Verified The Utah Numerical Scoring System
Formalized the Utah-CQT numerical scoring system from 30+ years of research, providing standardized procedures achieving highest rates of criterion accuracy.

Hand scoring, while guided by standardized rules, introduced inherent variability. Different examiners could examine the same chart and arrive at different numerical scores, even when applying identical criteria. Studies throughout the 1980s and 1990s documented inter-scorer reliability issues that represented a persistent concern for the profession. The subjective element — how an examiner weighted a particular respiratory change versus an electrodermal response — could influence outcomes in ways that were difficult to standardize across practitioners.

The Rise of Computerized Scoring

The advent of computerized polygraph instruments in the early 1990s created the opportunity to develop automated scoring algorithms. In 1988, Drs. John C. Kircher and David C. Raskin developed the Computer Assisted Polygraph System (CAPS), incorporating the first algorithm for evaluating physiological data collected for diagnostic purposes — a fundamental shift from purely human interpretation to algorithm-assisted analysis [1]Verified The Polygraph and Lie Detection — Appendix F: Computerized Scoring of Polygraph Data
Confirms PolyScore developed at JHU/APL, CPS developed at University of Utah, and details algorithmic approaches and feature sets used by each system.

In 1993, Dr. Dale E. Olsen and John C. Harris at Johns Hopkins University Applied Physics Laboratory (JHU/APL) completed PolyScore, which used a sophisticated mathematical algorithm to analyze polygraph data and estimate probability of deception [1]Verified The Polygraph and Lie Detection — Appendix F: Computerized Scoring of Polygraph Data
Confirms PolyScore developed at JHU/APL, CPS developed at University of Utah, and details algorithmic approaches and feature sets used by each system. This marked polygraph's formal entrance into the computer age, with digital analysis complementing — and in some applications enhancing — human interpretation.

These early algorithms demonstrated that mathematical models could match or exceed the accuracy of individual human scorers under controlled conditions. This finding was transformative for the profession, leading to the development of the three major scoring algorithms that dominate modern polygraph practice: PolyScore, OSS-3, and CPS. For examiners seeking to build a strong professional foundation, understanding these tools is essential alongside mastering comprehensive report writing and pretest interview techniques.

The Physiological Data Channels

Modern polygraph instruments record multiple physiological channels simultaneously. The primary channels analyzed by scoring algorithms include:

Electrodermal Activity (EDA) — Also called galvanic skin response (GSR) or skin conductance. This measures changes in the electrical conductivity of the skin, which increases with sweat gland activity driven by sympathetic nervous system arousal. EDA is generally considered the most diagnostically valuable single channel in polygraph testing. Research by Capps and Ansley (1992) found that over 40% of scoring points were assigned to EDA sensor data, and the 1994 PolyScore patent filing described EDA as having the largest coefficient in its logistic regression model [9]Verified Literature Survey of Structural Weighting of Polygraph Signals
Confirms EDA accounts for over 40% of scoring points and documents structural coefficients for PolyScore's logistic regression model.

Pneumograph (Respiration) — Typically recorded via two separate channels: upper thoracic breathing and lower abdominal breathing. Algorithms analyze changes in respiratory rate, amplitude, baseline, and pattern consistency.

Cardio Activity — Recorded via a blood pressure cuff (cardio cuff) or plethysmograph. This channel captures heart rate, blood pressure changes, pulse amplitude, and blood volume variations. Different algorithms extract different features from the cardiovascular data.

Some modern systems also incorporate supplementary channels such as finger pulse (photoplethysmograph), movement sensors, or seat activity monitors. These auxiliary channels are primarily used for artifact detection and countermeasure identification rather than primary scoring.

Feature Extraction and Classification

Raw physiological signals are continuous waveforms that must be processed before statistical analysis can be applied. Feature extraction converts these raw signals into discrete numerical values that represent meaningful physiological responses to specific test questions. This process typically involves identifying the onset and offset of each question stimulus period, measuring response amplitude, latency, duration, and recovery time within each channel, comparing responses to relevant questions against responses to comparison questions, normalizing measurements to account for individual differences in baseline physiological activity, and filtering artifacts caused by movement, coughing, deep breaths, or other non-test-related events [5]Verified The Polygraph and Lie Detection
NRC review confirming PolyScore 3.0 trained on 624 cases, version 5.1 trained on 1,411 cases, and CPS developed at University of Utah using discriminant analysis.

The specific features extracted and how they are weighted differ between algorithms, which is one reason why the same chart data can sometimes produce different results depending on which algorithm is applied. PolyScore has been developed from a larger set of features, while the CPS system focused from the outset on a relatively small set of data features [1]Verified The Polygraph and Lie Detection — Appendix F: Computerized Scoring of Polygraph Data
Confirms PolyScore developed at JHU/APL, CPS developed at University of Utah, and details algorithmic approaches and feature sets used by each system.

Once features are extracted, scoring algorithms apply statistical classification models to determine outcomes: Deception Indicated (DI), No Deception Indicated (NDI), or Inconclusive (INC). For more on handling these outcomes, see our guide on inconclusive polygraph results. Different algorithms handle the inconclusive zone differently — an algorithm with a narrow inconclusive band makes more definitive calls but may sacrifice accuracy, while a wider band is more conservative but may leave more cases unresolved.

Development and Mathematical Foundation

PolyScore is arguably the most widely recognized computer-based polygraph scoring algorithm in the United States, particularly within federal government agencies and law enforcement organizations. Developed by Dr. Dale E. Olsen and John C. Harris at the Johns Hopkins University Applied Physics Laboratory (JHU/APL) in 1993, PolyScore uses logistic regression models to calculate the probability that an examinee was deceptive during the polygraph examination [1]Verified The Polygraph and Lie Detection — Appendix F: Computerized Scoring of Polygraph Data
Confirms PolyScore developed at JHU/APL, CPS developed at University of Utah, and details algorithmic approaches and feature sets used by each system[2]Verified LXSoftware — Lafayette Instrument Company
Confirms LXSoftware is bundled with OSS-3 scoring algorithm and supports PolyScore, compatible with LX4000, LX5000, LX6, and LX7 systems.

Logistic regression is a well-established statistical method used across many fields to predict binary outcomes. PolyScore's models were trained on large datasets of confirmed polygraph examinations where ground truth was established through confessions, physical evidence, or other independent verification. The algorithm extracts features from the electrodermal, cardiovascular, and respiratory channels, then inputs these features into the regression equation to produce a probability estimate ranging from 0 to 1 [1]Verified The Polygraph and Lie Detection — Appendix F: Computerized Scoring of Polygraph Data
Confirms PolyScore developed at JHU/APL, CPS developed at University of Utah, and details algorithmic approaches and feature sets used by each system. The classification cutoff points used in PolyScore 3.0 were set at 0.95 (above which indicates deception) and 0.05 (below which indicates non-deception), with values between these thresholds classified as inconclusive [5]Verified The Polygraph and Lie Detection
NRC review confirming PolyScore 3.0 trained on 624 cases, version 5.1 trained on 1,411 cases, and CPS developed at University of Utah using discriminant analysis.

PolyScore also incorporates a neural network model as an alternative classification approach, giving it output based on either logistic regression or neural network analysis [5]Verified The Polygraph and Lie Detection
NRC review confirming PolyScore 3.0 trained on 624 cases, version 5.1 trained on 1,411 cases, and CPS developed at University of Utah using discriminant analysis.

Version History

PolyScore has undergone several documented revisions since its initial development in 1993. Four early versions — PASS 2.0, PolyScore 2.3, PolyScore 2.9, and PolyScore 3.0 — were evaluated in a comparative consistency study using mock crime data [10]Verified POLYSCORE: A Comparison of Accuracy (Blackwell, 1996)
Confirms four versions evaluated: PASS 2.0, PolyScore 2.3, 2.9, and 3.0 for scoring consistency using mock crime data.

PolyScore 3.0 was developed by analyzing polygraph data from 624 real criminal cases involving 303 non-deceptive and 321 deceptive suspects [1]Verified The Polygraph and Lie Detection — Appendix F: Computerized Scoring of Polygraph Data
Confirms PolyScore developed at JHU/APL, CPS developed at University of Utah, and details algorithmic approaches and feature sets used by each system[5]Verified The Polygraph and Lie Detection
NRC review confirming PolyScore 3.0 trained on 624 cases, version 5.1 trained on 1,411 cases, and CPS developed at University of Utah using discriminant analysis. A study by Blackwell (1999) evaluated PolyScore 3.3 on examinations from actual criminal investigations, finding 90.9% accuracy on Zone Comparison Test examinations [11]Verified PolyScore 3.3 and Psychophysiological Detection of Deception Examiner Rates of Accuracy
Confirms PolyScore 3.3 achieved 90.9% accuracy on ZCT examinations from actual criminal investigations, while human examiners achieved 82.3%.

PolyScore 5.1 used Zone Comparison Test (ZCT) and Modified General Question Test (MGQT) data from 1,411 real cases, significantly expanding the training dataset [5]Verified The Polygraph and Lie Detection
NRC review confirming PolyScore 3.0 trained on 624 cases, version 5.1 trained on 1,411 cases, and CPS developed at University of Utah using discriminant analysis. The Gordon et al. (2006) study evaluated PolyScore 5.5 alongside OSS and ASIT PolySuite, finding that all three algorithms achieved 100% accuracy when inconclusive results were excluded [12]Verified Integrated Zone Comparison Polygraph Technique Accuracy with Scoring Algorithms
Confirms all three algorithms achieved 100% accuracy excluding inconclusives; ASIT PolySuite 90% vs PolyScore/OSS 72% when inconclusives counted as errors.

The algorithm continues to be maintained and distributed by Lafayette Instrument Company as part of their LXSoftware suite [2]Verified LXSoftware — Lafayette Instrument Company
Confirms LXSoftware is bundled with OSS-3 scoring algorithm and supports PolyScore, compatible with LX4000, LX5000, LX6, and LX7 systems. Each new version must be independently validated, as changes to the underlying models can affect accuracy. Examiners should always note which version of PolyScore was used when documenting examination results.

Strengths and Limitations

PolyScore's primary strength is its probability-based output format, which provides examiners with more nuanced information than a simple categorical call. Rather than just knowing that a chart was scored as DI, the examiner can see a precise probability — providing different confidence levels for the same classification. The algorithm also benefits from extensive validation research across multiple versions.

PolyScore is tightly integrated with the Lafayette Instrument LXSoftware suite, one of the most widely used polygraph software platforms in the United States [2]Verified LXSoftware — Lafayette Instrument Company
Confirms LXSoftware is bundled with OSS-3 scoring algorithm and supports PolyScore, compatible with LX4000, LX5000, LX6, and LX7 systems. This integration means that examiners using Lafayette equipment can run PolyScore analysis with minimal additional steps. You can learn more about compatible equipment in our Limestone ParagonX Pro technical review.

As a parametric model, PolyScore's performance is bounded by the datasets on which it was trained. Models trained primarily on specific-issue examinations may not perform optimally on screening examinations, and vice versa. Additionally, PolyScore's reliance on proprietary regression coefficients means that the specific mathematical details of the model are not fully transparent. The National Research Council noted that details of PolyScore's development were proprietary and not fully shared with their committee [5]Verified The Polygraph and Lie Detection
NRC review confirming PolyScore 3.0 trained on 624 cases, version 5.1 trained on 1,411 cases, and CPS developed at University of Utah using discriminant analysis. This lack of transparency is a consideration in legal contexts where the Daubert standard for admissibility requires demonstrated scientific validity.

Development and Mathematical Foundation

The Objective Scoring System (OSS), currently in its third major version (OSS-3), was developed by Raymond Nelson, Mark Handler, and Donald Krapohl to provide a maximally objective, examiner-independent scoring method for polygraph chart data [3]Verified Objective Scoring System, Version 3 (OSS-3): Development and Validation
Describes OSS-3 algorithm development and brute-force validation showing OSS-3 outperformed 10 human scorers with balanced sensitivity and specificity. Where PolyScore outputs probabilities from logistic regression, OSS-3 produces numerical scores based on empirically derived statistical distributions, making it conceptually closer to the hand-scoring tradition but with mathematical rigor that eliminates human variability.

OSS-3 is built around the concept of cumulative distribution functions (CDFs) derived from large reference datasets of confirmed cases. For each physiological feature measured, the algorithm determines where that measurement falls within the known distribution of responses from confirmed deceptive and confirmed truthful examinees. A score is assigned based on the likelihood ratio — how much more likely the observed response is under the deceptive distribution versus the truthful distribution. Scores across all channels and question pairs are summed to produce an overall test score, which is then compared against empirically established cut-points for classification [3]Verified Objective Scoring System, Version 3 (OSS-3): Development and Validation
Describes OSS-3 algorithm development and brute-force validation showing OSS-3 outperformed 10 human scorers with balanced sensitivity and specificity.

The development involved significant contribution from researchers associated with the Department of Defense Polygraph Institute (DoDPI), which later became the National Center for Credibility Assessment (NCCA) [3]Verified Objective Scoring System, Version 3 (OSS-3): Development and Validation
Describes OSS-3 algorithm development and brute-force validation showing OSS-3 outperformed 10 human scorers with balanced sensitivity and specificity[6]Verified Stoelting CPS Elite Polygraph Systems
Confirms CPS Elite is fourth generation Computerized Polygraph System from Stoelting with scientifically validated scoring algorithms. The "objective" in the name refers to the system's design goal: that any two instances of the software, given identical input data, will always produce identical output scores.

Validation Research

OSS-3 has been subjected to rigorous validation. In their 2008 brute-force comparison study, Nelson, Handler, and Krapohl demonstrated that OSS-3 achieved balanced sensitivity and specificity and outperformed 10 human polygraph scorers [3]Verified Objective Scoring System, Version 3 (OSS-3): Development and Validation
Describes OSS-3 algorithm development and brute-force validation showing OSS-3 outperformed 10 human scorers with balanced sensitivity and specificity. This study used Monte Carlo methods to provide robust statistical estimates of algorithm performance.

The system has been specifically validated for several federal testing formats. Research on the United States Air Force Modified General Question Technique (USAF-MGQT) found that OSS-3 achieved 90.2% accuracy — outperforming both ESS (89.5%) and human scorers using 7-position scales (75.4%) [13]Verified Criterion Validity of the United States Air Force Modified General Question Technique and Iraqi Scorers
Confirms OSS-3 achieved 90.2% accuracy on USAF-MGQT data, ESS achieved 89.5%, both outperforming human scorers at 75.4%. The ESS and OSS-3 models both identified deceptive cases more accurately than truthful ones, with automated scoring systems consistently achieving higher accuracy rates than human scorers.

A 2024 study by Krapohl and Dutton comparing OSS-3 and PolyScore on field data found that OSS-3 correctly identified 71% of truthful and 96% of deceptive examinees with a 14% inconclusive rate, outperforming PolyScore on multiple-issue tests [4]Verified Comparison of OSS-3 and PolyScore in Multi-Issue and Single-Issue Field Exams
Confirms OSS-3 correctly identified 71% truthful and 96% deceptive with 14% inconclusive rate, outperforming PolyScore on multiple-issue tests. The cross-cultural validity of the related Empirical Scoring System (ESS) was demonstrated by Handler et al. (2011), who achieved a mean accuracy of 90.1% (95% CI: 83.8–95.8%) with 19 international trainees scoring 100 confirmed cases [14]Verified Empirical Scoring System: A Cross-Cultural Replication and Extension Study
Confirms international ESS validation achieving mean accuracy of 90.1% (95% CI: 83.8-95.8%) with 19 trainees scoring 100 confirmed cases.

Strengths and Limitations

OSS-3's greatest strength is its transparency. The scoring rules and cut-points are documented in published research, allowing independent researchers to understand exactly how classifications are derived [3]Verified Objective Scoring System, Version 3 (OSS-3): Development and Validation
Describes OSS-3 algorithm development and brute-force validation showing OSS-3 outperformed 10 human scorers with balanced sensitivity and specificity. This transparency is valuable in quality assurance reviews and in legal proceedings. The algorithm is now bundled with Lafayette LXSoftware as the default scoring algorithm [2]Verified LXSoftware — Lafayette Instrument Company
Confirms LXSoftware is bundled with OSS-3 scoring algorithm and supports PolyScore, compatible with LX4000, LX5000, LX6, and LX7 systems, meaning it is immediately available to Lafayette system users.

OSS-3's scoring approach produces numerical results intuitive to examiners accustomed to hand scoring. The transition from manual 7-position scoring to OSS-3 analysis is conceptually straightforward, as both produce numerical scores that are summed and compared against cut-points [8]Verified The Utah Numerical Scoring System
Formalized the Utah-CQT numerical scoring system from 30+ years of research, providing standardized procedures achieving highest rates of criterion accuracy.

Because OSS-3 relies on pre-established cumulative distribution functions, its accuracy is contingent on the reference datasets being representative of the current testing population. OSS-3 produces categorical scores rather than continuous probability estimates, providing less granular confidence information compared to PolyScore's probability output. Examiners focused on quality control and APA standards should understand these trade-offs.

Development and Algorithmic Approach

The Computerized Polygraph System (CPS) was developed by Drs. John C. Kircher and David C. Raskin at the University of Utah Psychology Laboratory, building on their earlier work with the Computer Assisted Polygraph System (CAPS) [5]Verified The Polygraph and Lie Detection
NRC review confirming PolyScore 3.0 trained on 624 cases, version 5.1 trained on 1,411 cases, and CPS developed at University of Utah using discriminant analysis[6]Verified Stoelting CPS Elite Polygraph Systems
Confirms CPS Elite is fourth generation Computerized Polygraph System from Stoelting with scientifically validated scoring algorithms. CPS uses multivariate discriminant analysis — also described as linear discriminant analysis (LDA) — to create a linear combination of features that maximally discriminates between deceptive and truthful groups [5]Verified The Polygraph and Lie Detection
NRC review confirming PolyScore 3.0 trained on 624 cases, version 5.1 trained on 1,411 cases, and CPS developed at University of Utah using discriminant analysis.

CAPS, the earlier version of CPS, was initially developed on mock crime data, while more recent versions of CPS rely on actual criminal cases [5]Verified The Polygraph and Lie Detection
NRC review confirming PolyScore 3.0 trained on 624 cases, version 5.1 trained on 1,411 cases, and CPS developed at University of Utah using discriminant analysis. The CPS system was designed with the goal of automating what careful human scorers currently do and has focused from the outset on a relatively small, carefully selected set of data features [5]Verified The Polygraph and Lie Detection
NRC review confirming PolyScore 3.0 trained on 624 cases, version 5.1 trained on 1,411 cases, and CPS developed at University of Utah using discriminant analysis. This focused approach contrasts with PolyScore, which was developed from a much larger feature set.

CPS is the native scoring algorithm for Stoelting polygraph instruments. The algorithm relies on true skin conductance data, which is the format recorded by Stoelting instruments [5]Verified The Polygraph and Lie Detection
NRC review confirming PolyScore 3.0 trained on 624 cases, version 5.1 trained on 1,411 cases, and CPS developed at University of Utah using discriminant analysis[6]Verified Stoelting CPS Elite Polygraph Systems
Confirms CPS Elite is fourth generation Computerized Polygraph System from Stoelting with scientifically validated scoring algorithms. Stoelting's current offering, the CPS Elite and CPS Fusion software, represents the fourth generation of the Computerized Polygraph System and includes scientifically validated scoring algorithms [6]Verified Stoelting CPS Elite Polygraph Systems
Confirms CPS Elite is fourth generation Computerized Polygraph System from Stoelting with scientifically validated scoring algorithms. You can learn more about Stoelting's platform in our CPSpro Fusion Software review.

Validation and Performance

CPS was among the five algorithms evaluated in the important Dollins, Krapohl, and Dutton (2000) comparison study using data from 97 confirmed criminal cases. That study found all five algorithms — PolyScore, CPS, AXCON, Chart Analysis, and Identifi — demonstrated comparable accuracy of 88-91% correct decisions when inconclusives were excluded [15]Verified Comparison of Computerized Polygraph Scoring Algorithms
Confirms five algorithms demonstrated 88-91% accuracy excluding inconclusives, with false positive rates of 31-46% when inconclusives treated as errors. However, the study also revealed false positive rates of 31-46% when inconclusives were treated as errors, with systematic misclassification of innocent subjects being a concern across all algorithms.

The National Research Council's 2003 review noted that CPS performed equally well on detection of both innocent and guilty subjects, while other algorithms were better at detecting deceptives specifically [5]Verified The Polygraph and Lie Detection
NRC review confirming PolyScore 3.0 trained on 624 cases, version 5.1 trained on 1,411 cases, and CPS developed at University of Utah using discriminant analysis. More recent research from South Korea comparing conventional algorithms including PolyScore, CPS, AXCON, Chart Analysis, and Identifi found that the specificity of all conventional algorithms was particularly limited when applied to criminal case data [16]Verified Development of a Deep-Learning-Based Computerized Scoring Algorithm
Confirms PolyScore and CPS use linear logistic regression and linear discriminant analysis respectively, and reports limitations of conventional algorithms on criminal case data.

CPS's tight integration with Stoelting hardware and software provides optimized data processing pipelines that take full advantage of Stoelting's specific sensor specifications and signal acquisition characteristics.

Platform Considerations

Beyond the core Stoelting CPS platform, other polygraph manufacturers have developed their own proprietary scoring approaches. Limestone Technologies, a Canadian polygraph manufacturer acquired by Lafayette Instrument Company in August 2022, employs the Empirical Scoring System (ESS-M) within its Polygraph Pro Suite software [17]Verified Limestone Technologies — About Us
Confirms Limestone Technologies is a subsidiary of Lafayette Instrument Company, acquired in August 2022[18]Verified Polygraph Pro Suite Software — Lafayette Instrument
Confirms Polygraph Pro Suite designed for Limestone Technologies series polygraph systems with scoring utilities and algorithms. The ESS-M is derived from the broader ESS scoring methodology developed by Nelson and Handler.

The diversity of CPS-type approaches means examiners should carefully evaluate the specific algorithm being used, the validation research supporting it, and its performance characteristics in the relevant testing context. A scoring algorithm validated for specific-issue criminal testing may not perform identically in a pre-employment screening context. For a detailed guide to how different platforms handle cross-platform scoring, see our complete ecosystem guide.

Key Metrics and Study Findings

Directly comparing the accuracy of polygraph scoring algorithms is complicated by the fact that different validation studies use different datasets, ground truth verification methods, test formats, and performance metrics. Several well-designed comparative studies provide meaningful insights.

The Dollins, Krapohl, and Dutton (2000) study compared five algorithms on 97 confirmed criminal cases and found comparable accuracy of 88-91% correct decisions when inconclusives were excluded [15]Verified Comparison of Computerized Polygraph Scoring Algorithms
Confirms five algorithms demonstrated 88-91% accuracy excluding inconclusives, with false positive rates of 31-46% when inconclusives treated as errors. The Gordon et al. (2006) study using the Integrated Zone Comparison Technique found that all three algorithms tested (ASIT PolySuite, PolyScore 5.5, and OSS) achieved 100% accuracy when inconclusive results were excluded. When inconclusives were counted as errors, ASIT PolySuite achieved 90% accuracy while PolyScore and OSS each achieved 72% [12]Verified Integrated Zone Comparison Polygraph Technique Accuracy with Scoring Algorithms
Confirms all three algorithms achieved 100% accuracy excluding inconclusives; ASIT PolySuite 90% vs PolyScore/OSS 72% when inconclusives counted as errors.

The 2024 Krapohl and Dutton field study comparing OSS-3 and PolyScore found that OSS-3 correctly identified 96% of deceptive examinees and 71% of truthful examinees with a 14% inconclusive rate, outperforming PolyScore on multiple-issue tests [4]Verified Comparison of OSS-3 and PolyScore in Multi-Issue and Single-Issue Field Exams
Confirms OSS-3 correctly identified 71% truthful and 96% deceptive with 14% inconclusive rate, outperforming PolyScore on multiple-issue tests. Blackwell's (1999) study of PolyScore 3.3 on actual criminal investigations found 90.9% accuracy on ZCT examinations, while human PDD examiners achieved 82.3% accuracy using the 7-position scale [11]Verified PolyScore 3.3 and Psychophysiological Detection of Deception Examiner Rates of Accuracy
Confirms PolyScore 3.3 achieved 90.9% accuracy on ZCT examinations from actual criminal investigations, while human examiners achieved 82.3%.

OSS-3 and PolyScore have demonstrated accuracy rates between 85-92% under laboratory conditions [16]Verified Development of a Deep-Learning-Based Computerized Scoring Algorithm
Confirms PolyScore and CPS use linear logistic regression and linear discriminant analysis respectively, and reports limitations of conventional algorithms on criminal case data. The Honts and Reavy (2015) mock crime experiment with 250 participants found substantial main effects of guilt in both OSS2 computer scoring and human scoring, with no significant differences between probable-lie and directed-lie test types [19]Verified The Comparison Question Polygraph Test: A Contrast of Methods and Scoring
Mock crime experiment with 250 participants finding substantial main effects of guilt in OSS2 computer scoring and human scoring with no differences between test types.

The APA Meta-Analytic Benchmark

The most comprehensive benchmark for polygraph accuracy comes from the APA's 2011 Meta-Analytic Survey of Criterion Accuracy. This exhaustive review included 38 studies involving 32 different samples and 45 experiments, with 295 scorers providing 11,737 scored results of 3,723 examinations [7]Verified Polygraph Validity Research — Meta-Analytic Survey of Criterion Accuracy
Confirms APA 2011 meta-analysis found 87% aggregated decision accuracy (CI: 80-94%) with 13% inconclusive rate across 38 studies and 3,723 examinations.

The data showed that event-specific (single issue) diagnostic testing produced an aggregated decision accuracy of 89% (confidence interval of 83-95%) with an estimated inconclusive rate of 11%. Multiple-issue techniques produced an aggregated decision accuracy of 85% (confidence interval 77-93%) with an inconclusive rate of 13%. The combination of all validated PDD techniques, excluding outlier results, produced a decision accuracy of 87% (confidence interval 80-94%) with an inconclusive rate of 13% [7]Verified Polygraph Validity Research — Meta-Analytic Survey of Criterion Accuracy
Confirms APA 2011 meta-analysis found 87% aggregated decision accuracy (CI: 80-94%) with 13% inconclusive rate across 38 studies and 3,723 examinations.

The American Polygraph Association reports that research studies published since 1980 show average accuracy rates ranging from 80 to 98 percent [5]Verified The Polygraph and Lie Detection
NRC review confirming PolyScore 3.0 trained on 624 cases, version 5.1 trained on 1,411 cases, and CPS developed at University of Utah using discriminant analysis. The APA's 2011 findings were described as consistent with the National Research Council's (2003) conclusions regarding polygraph accuracy [7]Verified Polygraph Validity Research — Meta-Analytic Survey of Criterion Accuracy
Confirms APA 2011 meta-analysis found 87% aggregated decision accuracy (CI: 80-94%) with 13% inconclusive rate across 38 studies and 3,723 examinations. A comprehensive 2021 meta-analysis by Honts et al. analyzing 138 datasets of Comparison Question Test research found that reviews generally produced overall accuracy estimates of over 85% [20]Verified A Comprehensive Meta-Analysis of the Comparison Question Polygraph Test
Meta-analysis of 138 CQT datasets finding overall accuracy estimates generally over 85% with significant moderator effects of motivation.

For context, examiner training quality significantly influences these outcomes in field practice.

The Case for Running Multiple Algorithms

One of the most important advances in modern polygraph practice is the convergent validity approach — running multiple independent scoring algorithms on the same physiological data and comparing results. When two or more algorithms agree on a classification, examiner confidence in that result increases substantially.

Lafayette's LXSoftware suite supports multiple scoring algorithms including PolyScore, OSS-3, QuESt, ASIT, and Identifi [2]Verified LXSoftware — Lafayette Instrument Company
Confirms LXSoftware is bundled with OSS-3 scoring algorithm and supports PolyScore, compatible with LX4000, LX5000, LX6, and LX7 systems. Stoelting's CPS Fusion software also provides the ability to import data from Axciton, Lafayette, or Limestone instruments exported in pREFORMAT [6]Verified Stoelting CPS Elite Polygraph Systems
Confirms CPS Elite is fourth generation Computerized Polygraph System from Stoelting with scientifically validated scoring algorithms. This cross-platform capability enables examiners to apply different analytical frameworks to the same data.

The NCCA ASCII standard, introduced in 2019 and now required by APA instrumentation standards, has been instrumental in enabling cross-platform data analysis [21]Verified Introduction to the NCCA ASCII Standard
Describes the standardized NCCA ASCII text format for cross-platform polygraph data export and import[22]Verified APA Standard for Polygraph Instrumentation
Requires NCCA ASCII export and import capability for all polygraph instruments to support quality control and research. This standardized export format allows polygraph data to be exported from any compliant instrument and analyzed on different platforms, partially decoupling algorithm selection from hardware choice. Researchers have used this format to analyze data in independent statistical environments including the R Language for statistical computing [21]Verified Introduction to the NCCA ASCII Standard
Describes the standardized NCCA ASCII text format for cross-platform polygraph data export and import.

For a deeper exploration of cross-platform scoring workflows, see our cross-platform polygraph scoring guide.

Decision Framework for Algorithm Selection

Examiners should consider several factors when selecting which algorithm(s) to apply:

Test format compatibility is essential. Ensure the algorithm has been validated for the specific question technique being used, whether the Utah Zone Comparison Test, AFMGQT, directed-lie comparison test, or screening format.

Instrument compatibility matters. CPS is optimized for Stoelting instruments recording true skin conductance, while PolyScore was developed for Lafayette and formerly Axciton instruments [5]Verified The Polygraph and Lie Detection
NRC review confirming PolyScore 3.0 trained on 624 cases, version 5.1 trained on 1,411 cases, and CPS developed at University of Utah using discriminant analysis. Using an algorithm with data from an instrument it was not designed for can introduce measurement artifacts.

Context of use should guide selection. For legal proceedings where transparency is important, OSS-3's documented scoring rules provide an advantage. For contexts requiring probability estimates, PolyScore's continuous output may be preferred.

The best practice for any high-stakes examination is to run at least two independent algorithms and evaluate convergent validity. When algorithms agree, the classification is strongly supported. When they disagree, the examiner should carefully review the data manually and consider factors such as rapport quality and behavioral observations.

What Algorithms Cannot Do

Despite their proven value, computerized scoring algorithms have important limitations. The National Research Council's 2003 review noted that computerized systems have the potential to reduce bias and eliminate inter-rater variability, but whether they can improve accuracy also depends on how one views the appropriateness of using other knowledge available to examiners, such as demographic information, historical background, and behavioral observations [5]Verified The Polygraph and Lie Detection
NRC review confirming PolyScore 3.0 trained on 624 cases, version 5.1 trained on 1,411 cases, and CPS developed at University of Utah using discriminant analysis.

Conventional algorithms employing linear classifiers have been noted for their limitations when dealing with the nonlinear nature of biological signals [16]Verified Development of a Deep-Learning-Based Computerized Scoring Algorithm
Confirms PolyScore and CPS use linear logistic regression and linear discriminant analysis respectively, and reports limitations of conventional algorithms on criminal case data. PolyScore and CPS use linear logistic regression and linear discriminant analysis respectively [16]Verified Development of a Deep-Learning-Based Computerized Scoring Algorithm
Confirms PolyScore and CPS use linear logistic regression and linear discriminant analysis respectively, and reports limitations of conventional algorithms on criminal case data, which may not capture all relevant patterns in physiological data.

All algorithms depend on the quality of the underlying examination. Poorly administered tests — those with inadequate pretest interviews, poorly constructed questions, or insufficient examiner-examinee rapport — will produce unreliable data regardless of which scoring algorithm is applied. The algorithm cannot compensate for fundamental testing errors.

Examiners should also be aware that external factors can influence physiological responses. Understanding how caffeine and other substances may affect results, as well as employing appropriate calming techniques, contributes to cleaner data for algorithmic analysis.

The Examiner Remains Central

Algorithms are decision-support tools, not substitutes for the trained examiner's judgment. The hybrid approach — computer-assisted, examiner-driven analysis — represents the modern standard of forensic psychophysiology [16]Verified Development of a Deep-Learning-Based Computerized Scoring Algorithm
Confirms PolyScore and CPS use linear logistic regression and linear discriminant analysis respectively, and reports limitations of conventional algorithms on criminal case data. Professional examiners integrate algorithm output with chart quality assessment, countermeasure detection, behavioral observation, and contextual understanding of the examination.

The importance of examiner training cannot be overstated. Meeting APA forensic standards and maintaining proficiency through continuing education ensures that examiners can properly interpret and contextualize algorithm outputs. For those entering the field, accredited training programs provide the foundation for competent algorithm use.

Emerging Approaches

Recent research is exploring deep learning structures for polygraph scoring. A 2025 study from South Korea developed a deep neural network (DNN)-based computerized scoring system that leveraged convolutional neural networks and long short-term memory (LSTM) architectures to analyze polygraph data [16]Verified Development of a Deep-Learning-Based Computerized Scoring Algorithm
Confirms PolyScore and CPS use linear logistic regression and linear discriminant analysis respectively, and reports limitations of conventional algorithms on criminal case data. The DNN model achieved precision of 0.9377 and recall of 0.9314, outperforming both PolyScore and OSS-3 on the same test dataset.

In that study, PolyScore correctly discriminated only 17 of 42 deception series (with 19 inconclusive), while OSS-3 discriminated 24 (with no inconclusives but 18 misclassifications). The deep learning algorithm correctly classified 36 of 42 deception series with no misclassifications and only 5 inconclusive results [16]Verified Development of a Deep-Learning-Based Computerized Scoring Algorithm
Confirms PolyScore and CPS use linear logistic regression and linear discriminant analysis respectively, and reports limitations of conventional algorithms on criminal case data. While these findings are preliminary and based on limited datasets, they suggest that nonlinear deep learning approaches may eventually surpass traditional linear classification methods.

Additional emerging directions include AI-driven adaptive scoring that learns from vast datasets to refine classification boundaries dynamically, detection of countermeasures via pattern irregularity analysis, and integration of natural language processing to correlate question semantics with physiological response strength [16]Verified Development of a Deep-Learning-Based Computerized Scoring Algorithm
Confirms PolyScore and CPS use linear logistic regression and linear discriminant analysis respectively, and reports limitations of conventional algorithms on criminal case data. These tools are being tested under controlled research environments and may become mainstream within the next decade.