N‑Back
Each stimulus is compared with the one shown N steps earlier. Position, audio, or both modalities may be active. Repeated practice reliably includes task learning: becoming better at this task and its strategies.
Evidence, without the hype.
This page separates what DualNBackApp measures inside the task from hypotheses about transfer beyond it. It documents every current scoring and workload rule.
Important limitation. DualNBackApp is not a medical device. The Performance Check is not a diagnosis. Better N‑Back performance does not by itself demonstrate higher general intelligence or better everyday functioning. Adaptive Load is a product difficulty system, not a medical algorithm.
What the task measures
Accuracy
hits / (hits + misses + false alarms). It summarizes how cleanly one round was completed. Correct rejections do not inflate it.
d′ discrimination
d′ combines the hit rate and false-alarm rate. It helps distinguish responding to real matches from pressing too freely. It is not IQ, “brain level,” or a clinical norm.
Reaction time
Reported only for valid hits. It is descriptive and is not used to advance load when accuracy or d′ is worsening.
d′ formula: dprime_v1
Position and audio are calculated separately when active. Dual uses pooled opportunity counts—not the mean of two d′ values. A single-modality combined value equals that modality’s value. At least four targets and four non-targets are required; otherwise the UI says “Not enough data.”
target_count = hits + misses
non_target_count = false_alarms + correct_rejections
corrected_hit_rate = (hits + 0.5) / (target_count + 1)
corrected_false_alarm_rate = (false_alarms + 0.5) / (non_target_count + 1)
d′ = Z(corrected_hit_rate) − Z(corrected_false_alarm_rate)
non_target_count = false_alarms + correct_rejections
corrected_hit_rate = (hits + 0.5) / (target_count + 1)
corrected_false_alarm_rate = (false_alarms + 0.5) / (non_target_count + 1)
d′ = Z(corrected_hit_rate) − Z(corrected_false_alarm_rate)
The log-linear correction avoids infinite Z-scores when a rate is exactly 0 or 1. Historical sessions are recalculated only when all four opportunity counts exist; their original Accuracy is never overwritten.
Adaptive Load: adaptive-load-v1
The visible level always remains an integer from 1N to 6N. Adaptive Load changes one workload dimension between rounds. The progression is a product hypothesis, not a scientific fact.
| Profile | Main change | Parameters |
|---|---|---|
| Standard | Baseline | 2500 ms stimulus; 2300 ms response; no controlled lures |
| Faster | Tempo | 2250 ms stimulus; 2100 ms response |
| High Interference | Controlled non-target lures | 20% planned lure rate; tempo retained |
| Endurance | Round length | 8 additional stimuli; other parameters retained |
Two consecutive valid rounds advance a profile when combined d′ ≥ 1.5 and false-alarm rate ≤ 0.20. Two consecutive rounds roll back one profile when d′ < 0.8 or false-alarm rate > 0.35. Tutorial, Academy, refresher, benchmark, incomplete, corrupted, and insufficient-opportunity rounds are excluded. After Endurance, the existing integer-N confirmation is used; accepting N+1 resets the profile to Standard and adds a stabilization round. Classic mode never changes these parameters automatically.
Controlled lures
A lure is a planned non-target resembling a match (currently an N+1 repeat). It remains a non-target; responding counts as a false alarm. The generator is deterministic with a seeded fixture, preserves planned real targets, and prevents an inserted lure from becoming an N-match.
Performance Check protocols
| Task | Score | Interpretation and limitation |
|---|---|---|
| Digit Span Backward | Longest backward span | Task-specific span; affected by strategy and repeat practice. |
| Stroop | Accuracy, median correct RT, interference | Interference is shown only with sufficient accurate trials; device/input affects timing. |
| Corsi | Longest spatial span | Task-specific visuospatial span; not a general memory score. |
| Go/No-Go | Hits, omissions, commissions, median valid RT | Measures performance in this brief protocol; not an attention diagnosis. |
Current versions: performance-check-v1 and performance-check-scoring-v1. Tests remain separate: there is no Brain Score, cognitive age, or percentile without a normative sample. Compare mainly with yourself and avoid frequent retesting merely to raise a number. Different protocol versions are labeled and must not be drawn as one continuous trend.
Learning and transfer
Task learning
Improvement on the practiced N‑Back setup is the most direct conclusion supported by in-product data.
Near transfer
Some research reports transfer to untrained N‑Back or closely related working-memory tasks. Effects vary with protocol and study design.
Far transfer
Evidence for broad effects on intelligence or daily performance remains mixed. A training curve or rising d′ cannot establish such transfer.
Experiment Journal
Sleep, energy, caffeine, tags, and notes stay local. Associations are observational, not causal. Small samples should not be interpreted as patterns.
Versions and changelog
- App: 1.3.0; data schema 3.
- Round score: performance-accuracy-v1.
- d′: dprime_v1.
- Integer N: integer-n-v1.
- Adaptive Load: adaptive-load-v1, initial conservative hypothesis thresholds.
- 2026-08-21: introduced d′, pooled Dual scoring, load profiles, controlled lures, Classic control, optional local journal fields, and this methodology page.
Bibliography
- Stanislaw & Todorov (1999), Calculation of signal detection theory measures.
- Hautus (1995), Corrections for extreme proportions.
- Jaeggi et al. (2008), initial report of transfer from working-memory training.
- Soveri et al. (2017), multilevel meta-analysis of N‑Back training.
- Pappa et al. (2020), systematic review and meta-analysis of updating training.
- Redick & Lindsey (2018), N‑Back versus complex-span training.