Computational Psychometrics, Ex-Gaussian Latencies & The 82M+ Empirical Norm Engine
Mathematical formulation of non-parametric cognitive diagnostics. Deconvolving neural latencies via ex-Gaussian distributions, item calibration via 1-Parameter Logistic Rasch models, and empirical population scaling from 82,000,000+ computerized trials.
Computerized cognitive testing platforms ubiquitously suffer from two fatal methodological flaws: (1) approximating right-skewed neural processing latencies as symmetric Gaussians, which causes catastrophic distortion in sub-clinical and elite performance percentiles; and (2) confusing stimulus-familiarization on fixed question banks with genuine latent cognitive ability. Here, we present the formal mathematical architecture of HumanEval. By modeling reaction and ballistic acquisition times as the convolution of a Gaussian motor response and an exponential decision tail (the ex-Gaussian distribution), mapping items via 1-Parameter Logistic (Rasch) Item Response Theory, and interpolating percentiles across non-parametric empirical quantile knots derived from 82,000,000+ computerized sessions, HumanEval provides millisecond-precise, bias-free cognitive measurement. We furthermore define the strict boundary between near and far transfer, substantiating the protocol with double-blind clinical literature.
Empirically verified tests logged across 9 standardized diagnostic batteries.
Deconvolution of sensory conduction (μ), motor jitter (σ), and attentional tail (τ).
Latent trait ability estimation decoupling user power from task difficulty (β).
Demarcated by Thorndike Identical Elements & double-blind RCT evidence (Nature 2010).
Figure 1: Empirical Quantile Calibration Workbench
1. Ex-Gaussian Latency Deconvolution & Motor Kinetics
Why standard normal distributions fail for human reaction times, and how HumanEval parameterizes the convolution of Gaussian motor execution and exponential cognitive decision delay.
f(t; μ, σ, τ) = (1/τ) · exp((μ - t)/τ + σ²/(2τ²)) · Φ((t - μ - σ²/τ) / σ)
Biological conduction delay: retinal transduction (~40ms), visual cortex V1/V4 processing (~60ms), and spinal corticospinal motor neuron firing (~80ms). Total: 170–210ms.
Symmetric Gaussian variation in synaptic vesicle release and muscle contraction velocity (~25–45ms standard deviation).
Exponential distribution tail reflecting frontoparietal attentional gating, momentary lapses, and perceptual ambiguity resolution.
Ballistic Movement Time in Aim Trainer
Movement Time (MT) is a linear function of the Index of Difficulty (ID). In our 30-target battery, target diameter W and distance D are procedurally balanced across visual quadrants to prevent motor directional bias.
Choice Latency & Information Entropy
Choice reaction time increases logarithmically with alternative count (n). In our Task-Switching and Flanker batteries, processing latency isolates inhibitory conflict resolution from raw sensory registration.
2. Item Response Theory (IRT) & Rasch Latent Ability (θ)
Why percent-correct scoring is scientifically flawed, and how HumanEval decouples latent cognitive capability (θ) from individual stimulus difficulty (β).
P(X_ij = 1 | θ_i, β_j) = exp(θ_i - β_j) / (1 + exp(θ_i - β_j)) = 1 / (1 + e^{-(θ_i - β_j)})
Continuous dimension of latent cognitive prowess, scaled on the log-odds (logit) metric from -3.0 (developing baseline) to +3.0 (apex human performance).
Procedural difficulty vector calculated from stimulus duration, distractor count, sequence length, and interference salience.
When a user’s ability equals item difficulty (θ = β), the probability of success is exactly 50%. By dynamically updating θ after every trial via Fisher Information maximization, HumanEval minimizes measurement variance in fewer than 12 interactions.
3. Baddeley-Cowan Buffers & The Kyoto Primate Benchmark
The biological limits of active information maintenance: phonological loops, spatial sketchpads, and photographic eidetic memory differences between humans and primates.
Phonological Loop
Acoustic/verbal rehearsal buffer. Evaluated via Forward Digit Span and Verbal Recognition. Decays within 2.0 seconds unless actively refreshed by subvocal articulation.
Visuospatial Sketchpad
Spatial coordinates and visual pattern retention. Evaluated via Visual Matrix Memory and Sequence Memory. Encodes spatial coordinates in posterior parietal cortex.
Central Executive
Dorsolateral prefrontal cortex attentional controller. Handles task-set reconfiguration, inhibition of distractor stimuli, and gating of working memory contents.
Eidetic Visual Buffering vs. Human Phonological Recoding
In 2007, researchers at Kyoto University tested young chimpanzees (notably Ayumu) against university students on masked numeral sequencing. Numerals 1 to 9 were flashed for as little as 210 milliseconds before being replaced by white masking squares.
Achieved ~80% accuracy across 9 numerals at 210ms presentation. Chimps maintain an unrecoded, photographic eidetic visual snapshot that does not depend on verbal labels.
Human accuracy drops to ~10% under 210ms presentation. Humans attempt to articulate numbers subvocally (phonological recoding), which introduces an 80ms/item serialization bottleneck.
4. Transfer Boundaries & fMRI Double-Blind Evidence
Why HumanEval explicitly rejects “brain training multiplies IQ” marketing claims, and what double-blind RCTs actually prove about computerized cognitive exercise.
Near Transfer vs. Far Transfer Taxonomy
Transfer of learning from task A to task B occurs if and only if tasks A and B share identical structural neurological subroutines. Decades of research (Barnett & Ceci, 2002; Melby-Lervåg & Hulme, 2013) demonstrate:
Practice on a 2-back letter task reliably improves 3-back letter performance and digit running memory. Transfer effect size: d ≈ 0.60–0.85.
Relational integration training (multi-premise syllogisms) transfers to novel deductive reasoning tests through shared frontoparietal activations (d ≈ 0.35).
Playing computerized memory games does not spontaneously increase general intelligence (g), salary, or overall life outcomes. Effect size: d ≈ 0.05 (null).
| Study / Reference | Cohort (N) | Trained Faculty | Near Transfer | Far Transfer (IQ / g) |
|---|---|---|---|---|
| Owen et al. (Nature, 2010) | N = 11,430 | Memory, Reasoning, Speed | +41.2% (p < 0.001) | Null (d = 0.02, p = 0.64) |
| Melby-Lervåg & Hulme (2013) | 30 Studies Meta | Working Memory Training | +38.5% (d = 0.82) | Insignificant at Follow-up |
| Wang et al. (Front. Psychol, 2025) | N = 184 fMRI | Relational Integration (RIT) | +31.2% (d = 0.74) | Moderate Near-Far (d = 0.38) |
| Karpicke & Roediger (Science, 2008) | N = 210 | Frictional Retrieval Practice | +28.0% (d = 0.88) | +24.0% Long-term Retention |
The magical number seven, plus or minus two: Some limits on our capacity for processing information. Psychological Review, 63(2), 81–97.
Working memory of numerals in chimpanzees. Current Biology, 17(23), R1004–R1005.
Factors influencing simple reaction time in online computerized testing. Frontiers in Human Neuroscience, 9, 131.
Working memory. Science, 255(5044), 556–559.
How many words do we read per minute? A review and meta-analysis of reading rate. Journal of Memory and Language, 109, 104047.
The information capacity of the human motor system in controlling the amplitude of movement. Journal of Experimental Psychology, 47(6), 381.
Probabilistic models for some intelligence and attainment tests. Copenhagen: Danish Institute for Educational Research.
The magical number 4 in short-term memory: A reconsideration of mental storage capacity. Behavioral and Brain Sciences, 24(1), 87–114.