Reflexes, Judgment, and Composure Don't Correlate: A Dissociation Analysis From 243 NeuroRank Combines

Abstract

Cognitive-skill instruments for competitive gaming implicitly assume some players are simply "more cognitively gifted" across the board, a single latent factor pushing every dimension up or down together. We tested that assumption against 243 completed NeuroRank combines (general, FPS, and MOBA genres) by computing pairwise Pearson correlations between three of our eight scored dimensions: raw reaction speed, decision quality (Flanker interference-based executive control), and consistency (tilt-recovery composure under failure). All three pairwise correlations were statistically indistinguishable from zero (raw speed vs. decision quality, r = -0.005; raw speed vs. consistency, r = 0.051; decision quality vs. consistency, r = 0.025; all n = 243, all |t| < 0.8). Yet each dimension independently and significantly predicted overall combine score (r = 0.53, 0.38, and 0.48 respectively, all p < .001, n = 220). This pattern, strong individual predictive value with near-zero mutual correlation, is inconsistent with a single unitary "gaming aptitude" factor and consistent with a multidimensional model in which reflex speed, executive control, and pressure regulation are separable cognitive systems that each contribute independent variance to overall performance. We discuss the finding against Diamond's (2013) review of dissociable executive functions, Baumeister's (1984) work on choking under pressure, and Ericsson, Krampe, and Tesch-Römer's (1993) account of expertise as an accumulation of independently trainable components. We also disclose two measurement caveats: 23 of 243 profiles lack a computed overall score, and our N (243 profiles) slightly exceeds the count of sessions the platform's own session-state machine records as completed (218), a discrepancy we could not fully reconcile from the aggregate endpoints available to us and are reporting rather than resolving speculatively.

Introduction and Hypothesis

NeuroRank scores every combine across six dimensions, drawn from an eight-dimension cross-genre catalog: raw speed, consistency, aim precision, working memory, decision quality, tracking accuracy, flick speed (FPS only), and prioritization (MOBA only). These feed a genre-weighted composite, the overall score, that anchors the archetype classification and the scouting report a player receives at the end of their combine.

A natural question, and one every player who has ever compared their own dimension breakdown to a friend's has probably asked informally, is whether these dimensions actually measure independent things, or whether they are mostly redundant expressions of one underlying trait. If a single "cognitive horsepower" factor drove all of them, we would expect players who score high on one dimension to reliably score high on the others: fast reflexes would predict good judgment, which would predict staying composed under pressure. If, instead, the test battery is measuring genuinely separable cognitive systems, a player could plausibly have lightning reflexes and poor judgment, or excellent judgment and a tendency to fall apart after a bad round, in any combination.

This question matters for reasons beyond intellectual curiosity. Scouting reports, coaching recommendations, and even a player's own read of their "development ceiling" are all built on the premise that dimension scores carry independent diagnostic information. If the dimensions were highly intercorrelated, the eight-dimension battery would be a more elaborate way of measuring one thing, and a shorter test could do the same job. If they are independent, a full battery is doing real work, and a player's profile shape, not just their overall number, is the more informative artifact.

We focus this analysis on three dimensions that are computed independently of each other by construction (see Methodology) and that are present for every combine regardless of genre: raw speed (from the Reaction module's simple-RT trials), decision quality (from the Composure module's Flanker interference cost, or from Reaction go/no-go accuracy when Composure data is unavailable), and consistency (from the Tilt module's post-failure recovery accuracy and rational betting rate, or from Reaction RT coefficient of variation as a fallback). Our hypothesis, informed by the broader literature on dissociable executive functions and by the platform's own design intent (three distinct psychophysical paradigms, three distinct scoring pipelines), was that these three dimensions would show low mutual correlation while each retaining predictive value for the overall composite score they help compute.

Data and Methodology

Sample

We queried NeuroRank's production Supabase profiles table directly (243 rows returned, one row per completed combine profile) and cross-checked platform-level counters against the /api/admin/stats endpoint, which reported 477 total sessions, 218 completed sessions (46% completion rate), and 243 total profiles as of query time. All profiles carry a createdAt timestamp; the sample spans 2026-02-02 to 2026-06-14, roughly four and a half months of live platform activity.

Genre breakdown: general, n = 121; FPS, n = 89; MOBA, n = 33 (121 + 89 + 33 = 243). Per the platform's genre-to-module mapping, general combines run reaction, aim, tracking, memory, composure, and tilt; FPS combines swap memory for flicker; MOBA combines swap aim for sequencing.

Scoring pipeline (as implemented, not reinterpreted)

Every dimension is computed server-side in scoringService.js and converted to a percentile against a fixed norm table in normData.js before being stored. Concretely, for the three dimensions this paper focuses on:

• Raw speed is the percentile of a player's mean simple-RT trial time within the simpleRT norm table.
• Decision quality is the percentile of the Composure module's composure-delta metric (accuracy and RT retention under Flanker-style distractor interference) within the composure norm table. When Composure data is unavailable for a session, the platform falls back to Reaction module go/no-go blended accuracy against the goNoGoAccuracy table. This fallback path means "decision quality" is not a single uniform construct across every row in our sample, a caveat we return to below.
• Consistency is the percentile of the Tilt module's post-failure recovery accuracy and rational high/low bet rate within the tiltRecovery norm table, with a fallback to Reaction RT coefficient of variation within the consistency table when Tilt data is unavailable.

These three dimensions are computed by three structurally separate code paths reading three structurally separate trial types (simple-RT trials, Flanker trials, and Tilt bet-under-pressure trials), which is what makes a low correlation between them a meaningful empirical question rather than a foregone conclusion of shared inputs.

Overall score is a genre-weighted linear combination of all dimensions present for a given genre (for example, FPS weights raw speed at 1.3x and general weights it at 1.0x, reflecting differing per-genre emphasis). Because raw speed, decision quality, and consistency are each summands inside the overall-score formula, a correlation between any one of them and overall score is expected by construction to some degree; we report those correlations for completeness but do not treat them as evidence of an external validity claim. The correlations between the three dimensions themselves, by contrast, are not mechanically guaranteed by the scoring formula and are the core empirical result of this paper.

Statistical approach

We computed Pearson product-moment correlations for all pairwise combinations of the three dimensions (using every profile with numeric values, n = 243) and for each dimension against overall score (n = 220, reflecting the 23 profiles in our sample that lack a computed overall score). Significance was assessed via the standard t-transformation of r (t = r × sqrt(df / (1 - r²))), reported with degrees of freedom.

Known data-quality caveats (disclosed, not resolved)

Two issues surfaced during this analysis that we are reporting rather than papering over:

1. Profile count exceeds completed-session count. /api/admin/stats reports 218 completed sessions but 243 total profiles, a gap of 25. The platform's session-state machine (active → completed) should produce exactly one profile per completed session. We could not fully reconcile this discrepancy from the aggregate endpoints available to us. Two known mechanisms could contribute: profiles migrated into the current Supabase schema from a legacy SQLite store predating the sessions table, and an admin-only "autocomplete" tool that force-completes an existing session by filling any modules the player did not finish with synthetic placeholder trials for QA purposes. We cannot identify which, if any, rows in our 243-profile sample originated from the latter path, so we cannot exclude them. This means a small, unknown number of rows in our sample may contain partially synthetic trial data rather than fully organic play.
2. Missing values are not missing at random across genre. Working-memory scores are present for only 10 of 89 FPS profiles and aim-precision scores are present for only 5 of 33 MOBA profiles, despite neither module being part of the standard FPS or MOBA module set (FPS substitutes flicker for memory; MOBA substitutes sequencing for aim). These small pockets likely reflect players who switched genre mid-account history or legacy-schema rows, not a representative FPS or MOBA sub-sample. We report the values that exist for transparency but do not treat the FPS working-memory or MOBA aim-precision means as genre-representative, and we exclude them from the core dissociation analysis (which uses only raw speed, decision quality, and consistency, present across all genres).

Results

Dimension means by genre

Genre	n	Raw speed	Consistency	Decision quality	Overall score
General	121	45.9 (SD 19.8)	58.9 (SD 14.6)	69.5 (SD 24.1)	65.6 (SD 8.0, n=113)
FPS	89	48.8 (SD 22.7)	62.5 (SD 15.8)	64.9 (SD 23.6)	60.0 (SD 11.2, n=79)
MOBA	33	47.1 (SD 16.4)	66.5 (SD 11.9)	71.4 (SD 21.4)	71.8 (SD 6.0, n=28)

(Source: aggregate query against the profiles.data.scores JSONB field, grouped by profiles.data.genre. Platform-wide equivalents for these fields are also visible in /api/admin/stats's avgDimensions object: rawSpeed 47, consistency 61, decisionQuality 68, drawn from all genres pooled.)

A first descriptive observation: MOBA profiles show the highest mean consistency (66.5) and the highest mean overall score (71.8) of the three genres, while FPS profiles show the highest mean raw speed (48.8) but the lowest mean overall score (60.0). This is consistent with, though not proof of, the genre-weighting scheme, FPS weights raw speed more heavily but still requires decision quality and consistency to reach a high composite, and with self-selection, MOBA's longer, planning-heavy matches may draw a different play style than FPS's short high-arousal rounds.

Pairwise correlations among the three dimensions (n = 243)

Pair	r	t	df	Interpretation
Raw speed × decision quality	-0.005	-0.08	241	No relationship
Raw speed × consistency	0.051	0.79	241	No relationship
Decision quality × consistency	0.025	0.39	241	No relationship

None of the three pairwise correlations differ significantly from zero. The largest, raw speed × consistency at r = 0.051, would need to be roughly six times larger before reaching conventional significance at this sample size. In practical terms: knowing a player's raw-speed percentile gives you no useful information about where their decision-quality or consistency percentile is likely to fall, and vice versa in all three directions.

Each dimension independently predicts overall score (n = 220)

Dimension	r vs. overall score	t	df	p
Raw speed	0.532	9.28	218	< .001
Consistency	0.478	8.04	218	< .001
Decision quality	0.376	5.99	218	< .001

All three dimensions are strong, statistically significant predictors of the composite score they feed into. As noted in Methodology, this predictive relationship is partly guaranteed by construction (each dimension is a summand in the weighted composite) and should not be read as evidence of external predictive validity on its own. What it does establish is that none of the three dimensions is "dead weight" in the composite: each carries enough independent signal that players spread widely on overall score even when we already know their raw-speed percentile alone (or their consistency percentile alone, or their decision-quality percentile alone).

Quartile view: overall score climbs steadily with either raw speed or consistency, independently

To make the independent-contribution point concrete without relying only on a correlation coefficient, we split the sample into quartiles by raw speed and, separately, by consistency, and computed mean overall score per quartile.

Raw speed quartile	Range	n	Mean overall score
Q1 (lowest)	1-30	55	56.4
Q2	30-45	55	64.6
Q3	45-60	55	67.4
Q4 (highest)	60-85	55	69.2

Consistency quartile	Range	n	Mean overall score
Q1 (lowest)	5-60	55	59.7
Q2	60-65	55	64.0
Q3	65-75	55	66.6
Q4 (highest)	75-99	55	67.3

Both quartile ladders climb monotonically, and both span a roughly 10-13 point range in overall score from bottom to top quartile. Because raw speed and consistency do not correlate with each other (r = 0.051, above), a player can occupy the top quartile of one ladder and any quartile of the other. The two ladders are, in effect, two separate routes to a high overall score, not the same route measured twice.

A secondary, purely descriptive note on archetype concentration by genre

While outside this paper's core hypothesis, one pattern in the same dataset is worth flagging for future analysis: archetype assignments are far more concentrated in FPS and MOBA than in general-genre combines. In our sample, 68.5% of FPS profiles (61 of 89) were classified as "The Fragmaster," and 54.5% of MOBA profiles (18 of 33) were classified as "The Flex," versus a more even spread in general-genre combines (the largest single archetype there, "The Balanced Operator," accounts for 41.3% of 121 profiles, with five other archetypes making up the remainder). We flag this as a candidate topic for a future data paper (added to our topic queue) rather than analyzing it further here, since it speaks to the discriminative power of the archetype classifier per genre rather than to dimension independence, the question this paper set out to answer.

Discussion and Limitations

The central result, three dimensions with near-zero mutual correlation that each independently predict a shared composite, argues against treating NeuroRank's overall score as a proxy for a single "cognitive gifted-ness" trait, and in favor of treating the dimension breakdown itself as the more informative artifact a player receives. This is consistent with Diamond's (2013) review of executive functions, which describes inhibitory control, working memory, and cognitive flexibility as separable systems with distinguishable developmental trajectories and neural correlates rather than expressions of one general executive capacity. It is also consistent with Baumeister's (1984) account of choking under pressure, which frames performance degradation after failure as governed by self-focused attention processes distinct from baseline skill or raw processing speed, exactly the kind of dissociation our consistency dimension (tilt recovery) is designed to isolate from raw reaction speed. And it fits the broader picture from Ericsson, Krampe, and Tesch-Römer (1993) of expertise as an accumulation of separately trainable component skills rather than a single trait that rises or falls in lockstep; if that is right, we should expect exactly the pattern we found, individually meaningful, mutually uncorrelated components.

The finding also stands in useful contrast to Spearman's (1904) general-intelligence ("g") tradition, which posits a single dominant factor underlying performance across cognitively demanding tasks. Whatever the merits of that model for the domains it was built on, our data do not support an analogous unitary factor for the specific reflex-judgment-composure triad measured here, at least not one strong enough to produce measurable shared variance at n = 243. We are not claiming to have tested or refuted a general-intelligence model in any strong sense; we are reporting that in this specific instrument, on this specific sample, these three constructs behave as though they are independent.

Methodologically, the Reaction module's own internal logic (Donders, 1868/1869) is worth noting as the reason raw speed is measured the way it is: Donders' subtractive method, isolating simple detection-and-response time from more complex choice and go/no-go processes, is the basis for treating simple RT as a relatively pure index of processing and motor speed, uncontaminated by the decision-making load that the Composure (Eriksen & Eriksen, 1974, Flanker paradigm) and Tilt modules are specifically designed to probe. That the three modules draw on distinct paradigms with distinct theoretical lineages is, in a sense, the a priori reason to expect the dissociation we found; the contribution of this paper is confirming that the dissociation actually shows up in a live combine population rather than assuming it from paradigm design alone.

Several limitations bound how far these results should be read. First, and most importantly, both data-quality caveats disclosed in Methodology apply: an unknown, likely small number of profiles may include admin-autocompleted synthetic trials for modules a player did not finish, and we cannot identify or exclude those rows from this analysis. Second, our sample of 243 is adequate for detecting the moderate-to-large effects reported here but is not large enough to rule out small true correlations (say, r = 0.10-0.15) between the three dimensions; our null results should be read as "no detectable relationship at this sample size," not as proof of exact independence. Third, the fallback scoring paths (Composure-unavailable sessions falling back to Reaction accuracy for decision quality; Tilt-unavailable sessions falling back to Reaction RT variability for consistency) mean "decision quality" and "consistency" are not perfectly uniform constructs across every row; we did not have a field in the aggregate data available to us that flags which fallback path, if any, a given profile used, so we could not stratify by measurement source. Fourth, this is an observational, cross-sectional snapshot; it says nothing about whether these dimensions remain independent as an individual player improves over repeated sessions, a question our archetype-stability item in the topic queue is aimed at for a future data paper once we have a larger retest cohort. Finally, the genre imbalance (121 general, 89 FPS, 33 MOBA) means the MOBA-specific descriptive statistics in particular rest on a comparatively small base and should be treated as preliminary.

References

• Donders, F. C. (1868/1869). On the speed of mental processes (Translated by W. G. Koster, 1969, in Acta Psychologica, 30, 412-431).
• Eriksen, B. A., & Eriksen, C. W. (1974). Effects of noise letters upon the identification of a target letter in a nonsearch task. Perception & Psychophysics, 16(1), 143-149.
• Diamond, A. (2013). Executive functions. Annual Review of Psychology, 64, 135-168.
• Baumeister, R. F. (1984). Choking under pressure: Self-consciousness and paradoxical effects of incentives on skillful performance. Journal of Personality and Social Psychology, 46(3), 610-620.
• Ericsson, K. A., Krampe, R. T., & Tesch-Römer, C. (1993). The role of deliberate practice in the acquisition of expert performance. Psychological Review, 100(3), 363-406.
• Spearman, C. (1904). "General Intelligence," objectively determined and measured. American Journal of Psychology, 15(2), 201-292.