Same Test, Narrower Outcomes: Archetype Concentration by Genre Across 223 NeuroRank Combines

Abstract

NeuroRank classifies every completed combine into one of six genre-specific archetypes using a fixed-priority rule set applied to the same six-dimension percentile battery, regardless of whether a player took the general, FPS, or MOBA combine. We asked whether this classification concentrates differently across genres: do FPS and MOBA cohorts cluster into fewer distinct archetypes than the general cohort, given the same six-slot archetype space in every genre? Using 223 completed combines with genre-consistent scoring schemas (general n=116, FPS n=79, MOBA n=28), we found a clear concentration gradient. The general cohort used all six available archetypes, with the modal archetype (The Balanced Operator) capturing 61.2% of profiles (Herfindahl-Hirschman Index, HHI=4429, normalized Shannon entropy=0.612). The FPS cohort used only four of its six archetypes, with the modal archetype (The Fragmaster) capturing 78.5% (HHI=6334, normalized entropy=0.414). The MOBA cohort used only three of its six archetypes, with the modal archetype (The Flex) capturing 78.6% (HHI=6429, normalized entropy=0.366). Bootstrap 95% confidence intervals on the modal share barely overlap between general and either specialist genre. While assembling this sample we discovered and disclose a scoring-pipeline anomaly: 15 FPS/MOBA profiles (10 FPS, 5 MOBA), all created between 2026-02-06 and 2026-03-03, carry the general-genre scoring schema (a workingMemory dimension and general-pool archetype names) despite being tagged with a specialist genre. We exclude these from the core analysis and report them as a disclosed, unresolved data-quality finding rather than folding them into either genre's totals.

Introduction and Hypothesis

NeuroRank assigns every completed combine to one of eighteen total archetypes, six per genre (general, FPS, MOBA), through a fixed-priority classifier: a profile's percentile scores are tested against each archetype's activation rule in a fixed order, and the first rule that matches wins, with an unconditional catch-all archetype (The Balanced Operator, The Fragmaster, The Flex) as the final fallback in each genre's rule list. Critically, every genre draws from the same size archetype space, six slots, so a naive expectation is that, absent any real difference in how players distribute across cognitive profiles, the three genres should look similarly concentrated: a similar modal-archetype share, a similar number of the six slots actually used.

This month's data-paper queue posed the question directly: does the FPS cohort cluster into fewer distinct profiles than the general cohort? There is a substantive reason to expect yes. The general combine samples the broadest possible cross-section of the platform's traffic, casual visitors, curious lurkers, and dedicated competitive players alike, arriving with no particular self-selected skill package. The FPS and MOBA combines, by contrast, are opt-in choices made by players who already identify with a specific competitive genre and, plausibly, have already trained a specific narrow skill package for it (aim and flick speed for FPS, prioritization and decision timing for MOBA) rather than a broad, balanced cognitive profile. If genre self-selection concentrates the underlying skill distributions being fed into the classifier, the classifier's output distribution should concentrate too, funneling more players into the same one or two archetype buckets and leaving other buckets in that genre's rule list empty.

We test this with three complementary measures of concentration applied to each genre's observed archetype-count distribution: the modal (top-1) archetype's share of the cohort, the Herfindahl-Hirschman Index (HHI), a sum-of-squared-shares statistic imported from industrial-concentration economics that ranges from a theoretical minimum near 1667 for six perfectly equal categories to 10000 for total concentration in one category, and Shannon entropy normalized against the six-category maximum, which ranges from 0 (total concentration) to 1 (perfectly even use of all six archetype slots). We also report, more simply, how many of each genre's six defined archetypes appear at all in the observed sample.

Data and Methodology

Sample construction and the schema-mismatch caveat

We queried NeuroRank's production Supabase profiles table directly: 248 total rows, of which 10 carry the flagged marker (per the anti-cheat/bot exclusion introduced for issue #231) and are dropped from all quality aggregates here, exactly as the platform's own /api/admin/stats endpoint does. That leaves 238 clean profiles.

Building the genre-level cohorts for this analysis, we found that not every clean profile's stored score object matches its declared genre. NeuroRank's genre-to-module mapping means an FPS combine should never produce a workingMemory score (FPS substitutes the flicker module for memory) and a MOBA combine should never produce an aimPrecision score in the same slot pattern general combines use (MOBA substitutes sequencing for aim). Cross-checking every profile's score-object keys and archetype name against its declared genre surfaced 15 profiles that fail this check: 10 tagged genre: "fps" carrying a workingMemory score and a general-pool archetype name (for example "The Anchor," "The Clutch Reactor," "The Raw Talent," names that do not exist in the FPS archetype pool at all), and 5 tagged genre: "moba" in the same pattern (general-pool names like "The Strategist" and "The Balanced Operator"). Every one of these 15 mismatched profiles has a createdAt timestamp between 2026-02-06 and 2026-03-03. By contrast, every FPS profile with a correctly genre-matched schema starts at 2026-03-13, and every correctly matched MOBA profile starts at 2026-03-19, both comfortably after the last mismatched row. General-genre profiles show no equivalent mismatch anywhere in the sample.

This is a clean date boundary, not scattered noise, and it lines up with a second, independent signal: the completed-sessions table's genre breakdown for the same window (223 completed sessions total: general 116, FPS 79, MOBA 28) matches the count of correctly-schema-matched profiles in each genre exactly, while the 15 mismatched profiles have no corresponding entry in that count. We read this as evidence that these 15 rows originated outside the current session-to-profile pipeline entirely, consistent with the mechanism this platform's July data paper already disclosed as a live possibility (legacy-schema rows predating the current sessions table, or admin-tool-generated rows), rather than as evidence that the live scoring pipeline is currently misassigning genres. We did not have access to a field distinguishing which of those two mechanisms, if either, produced these 15 rows, so we report the finding rather than resolving it, and we exclude all 15 from the concentration analysis below because comparing an FPS cohort against a rule set built for a different genre's dimension set would not be a meaningful test of this paper's hypothesis.

The resulting analysis sample is 223 profiles: general n=116, FPS n=79, MOBA n=28, matching the completed-sessions genre breakdown for the same period exactly.

Archetype rule sets

Each genre draws from a disjoint six-archetype pool defined in the platform's classifier (nextjs/src/lib/archetype-classify.ts, mirrored from the legacy server/utils/archetypes.js):

• General: The Anchor, The Clutch Reactor, The Fragger, The Strategist, The Raw Talent, The Balanced Operator (fallback)
• FPS: The AWPer, The Entry Fragger, The Rifler, The Lurker, The FPS Raw Talent, The Fragmaster (fallback)
• MOBA: The Carry/ADC, The Support, The Jungler, The Mid/Assassin, The MOBA Raw Talent, The Flex (fallback)

Every genre's fallback archetype ("Balanced Operator," "Fragmaster," "Flex") is reached only when none of the five more specific rules above it in priority order match, so a high fallback share is itself informative: it means most profiles in that genre are not distinctive enough on any single dimension combination to trip a specialist rule.

One further methodological note: a drift-normalized classification engine (classifyArchetypeDynamic) replaced the legacy static-cutoff engine on 2026-07-13 for live combine completions. Because our FPS and MOBA valid-schema samples run only from 2026-03-13 to 2026-05-17 and 2026-03-19 to 2026-04-19 respectively, every profile in those two cohorts was classified under the legacy engine. The general cohort spans 2026-02-02 to 2026-07-28 and so includes 3 profiles (of 116) classified after the 2026-07-13 engine change; those 3 profiles split 2 Raw Talent and 1 Balanced Operator, consistent with the rest of the general distribution, so we do not think this small, late slice materially affects the general-cohort concentration figures reported below, but we flag it for completeness.

Concentration metrics

For each genre's observed archetype-count distribution we report: the modal (top-1) archetype's share of the cohort with a 95% confidence interval from 5,000 percentile bootstrap resamples (resampling profiles within genre, with replacement); the Herfindahl-Hirschman Index, HHI = 10000 × Σ(share_i²), a standard concentration measure imported from antitrust economics (Rhoades, 1993) and mathematically identical to Simpson's diversity index (Simpson, 1949) up to the 10000 scaling constant; and Shannon entropy in bits, normalized by log2(6) so that 1.0 represents perfectly even use of all six archetype slots and 0.0 represents total concentration in one slot.

Results

Archetype distribution by genre

Genre	n	Archetype	Share
General	116	The Balanced Operator	61.2%
		The Raw Talent	25.0%
		The Strategist	6.0%
		The Anchor	3.4%
		The Clutch Reactor	2.6%
		The Fragger	1.7%
FPS	79	The Fragmaster	78.5%
		The Lurker	10.1%
		The FPS Raw Talent	7.6%
		The Rifler	3.8%
		The AWPer	0.0%
		The Entry Fragger	0.0%
MOBA	28	The Flex	78.6%
		The MOBA Raw Talent	14.3%
		The Jungler	7.1%
		The Carry/ADC	0.0%
		The Support	0.0%
		The Mid/Assassin	0.0%

General is the only genre in which all six defined archetypes appear at least once. FPS uses 4 of its 6 slots (2 empty: The AWPer, The Entry Fragger). MOBA uses 3 of its 6 slots (3 empty: The Carry/ADC, The Support, The Mid/Assassin).

Concentration metrics

Genre	n	Distinct archetypes used (of 6)	Modal share	95% CI (bootstrap)	HHI	Normalized entropy
General	116	6	61.2%	51.7%-69.8%	4429	0.612
FPS	79	4	78.5%	68.4%-87.3%	6334	0.414
MOBA	28	3	78.6%	64.3%-92.9%	6429	0.366

All three concentration measures move in the same direction and agree with each other: general is the least concentrated cohort by every measure, FPS is more concentrated than general, and MOBA is the most concentrated of the three, though its 95% CI (64.3%-92.9%) is wide due to its comparatively small sample and overlaps meaningfully with FPS's CI. The general cohort's modal-share CI (51.7%-69.8%) and the FPS cohort's modal-share CI (68.4%-87.3%) overlap only in a narrow 68.4%-69.8% band, so while we would not claim a large-margin statistical separation at this sample size, the two distributions are distinguishable in the direction the hypothesis predicted. The HHI figures make the same point on a scale where general's 4429 sits closer to the four-equal-categories reference point (2500) than to the near-single-category ceiling (10000), while FPS's 6334 and MOBA's 6429 sit past the two-equal-categories reference point (5000), indicating that a large majority of both specialist cohorts' variance is absorbed by essentially one dominant label plus a distant runner-up.

Dimension means by genre

To look for a candidate explanation of the concentration gradient beyond the classifier's rule structure itself, we computed mean and population standard deviation for every scored dimension in each of the three valid-schema cohorts.

Genre	n	Raw speed	Decision quality	Consistency	Aim precision	Tracking accuracy	Working memory	Flick speed	Prioritization	Overall score
General	116	49.3 (SD 27.2)	52.3 (SD 28.8)	51.9 (SD 29.4)	56.7 (SD 15.4)	50.2 (SD 27.2)	51.4 (SD 29.2)	-	-	51.0 (SD 13.2, n=108)
FPS	79	49.6 (SD 30.7)	49.8 (SD 29.9)	64.7 (SD 27.0)	46.8 (SD 16.2)	51.4 (SD 26.9)	-	51.0 (SD 20.8)	-	51.9 (SD 14.3)
MOBA	28	50.4 (SD 26.7)	54.1 (SD 28.6)	72.0 (SD 20.2)	-	51.2 (SD 28.6)	55.2 (SD 21.8)	-	69.0 (SD 19.4)	59.1 (SD 10.1)

The dimension with the largest genre-to-genre gap is consistency: the FPS cohort's mean consistency (64.7) sits nearly 13 points above the general cohort's (51.9), and the MOBA cohort's mean consistency (72.0) sits over 20 points above general's. Raw speed, decision quality, and tracking accuracy are all within a few points of each other across genres and show no comparable gradient. We flag this descriptively rather than causally: several of the specialist archetype rules in both the FPS and MOBA pools gate on a consistency threshold either directly (LURKER requires consistency >= 75; AWPER requires consistency >= 68) or as an exclusion condition on the Raw Talent rule (both require consistency < 55 to qualify as a development-stage profile). A cohort whose consistency scores cluster in a broad middle band, high enough to rule out both Raw Talent's low-consistency requirement and any consistency-gated exclusion, but not reliably above the 68-75 range the top specialist rules require, would systematically fall through to the fallback archetype more often. We did not test this mechanism directly (doing so would require simulating classifier output against permuted score vectors, outside the scope of this descriptive data paper) and we note it only as a plausible, testable, but unconfirmed contributor to the concentration pattern reported above, not as an established cause.

Fallback-archetype share as a specific diagnostic

Because every genre's most-common archetype in this sample happens to be that genre's fallback rule (the catch-all that fires only when no specialist rule above it matches), the modal share doubles as a direct measure of how often the specialist rules fire at all. In the general genre, five specialist rules combined account for 38.8% of profiles (100% minus the 61.2% Balanced Operator fallback share). In FPS, five specialist rules combined account for just 21.5% of profiles. In MOBA, five specialist rules combined account for just 21.4% of profiles. Put differently, a general-genre player is roughly 1.8 times more likely to trip one of the five specialist archetype rules than an FPS or MOBA player is to trip one of theirs, despite all three rule sets using the same rule-count structure (five specialist rules plus one fallback) and the same underlying six-dimension percentile scale.

Discussion and Limitations

The concentration gradient we found, general least concentrated, FPS more, MOBA most, is consistent with the genre self-selection hypothesis stated in the introduction: players who choose the FPS or MOBA combine are choosing a test that measures a narrower, more specialized skill package (aim precision and flick speed for FPS; prioritization and decision quality for MOBA) than the broad six-dimension general combine measures, and a narrower underlying skill package naturally produces fewer distinguishable cognitive shapes for the same fixed classifier to sort into. This finding sits alongside a broader tradition in games research of building typologies of players from behavioral or motivational data, most influentially Bartle's (1996) taxonomy of MUD player types, though that work classified players by self-reported motivation rather than measured cognitive performance, and did not examine how strongly a given typology concentrates within a self-selected sub-population the way this analysis does.

This platform's own July 2026 data paper is relevant context here: that analysis found raw speed, decision quality, and consistency to be essentially uncorrelated with each other across the pooled cohort (all pairwise |r| < 0.06, n=243). If the six scored dimensions are genuinely close to statistically independent within a genre, as that finding suggests, then a specialist archetype rule that requires three dimensions to simultaneously clear separate thresholds is mechanically harder to trip than a single elevated dimension would suggest in isolation, since the probability of three independent events co-occurring is smaller than any one of them occurring alone. Every genre's specialist rules share this three-condition structure (with the exception of each genre's Raw Talent rule, which uses two), so this alone does not explain why FPS and MOBA concentrate more than general specifically. What the consistency-mean shift documented above adds to this picture is a candidate genre-specific mechanism layered on top of that general conjunction-rarity effect: if FPS and MOBA cohorts' consistency scores cluster in a band that satisfies neither the high-consistency specialist gates nor the low-consistency Raw Talent exclusion, the conjunction becomes harder to trip in those two genres specifically, which is directionally consistent with, though not proof of, the concentration gradient we observed.

There is a more mundane alternative explanation worth taking seriously: it is possible the FPS and MOBA rule sets are simply miscalibrated (too strict, or requiring dimension combinations that are individually rare in the live cohort) rather than the underlying FPS and MOBA player populations being genuinely more homogeneous than the general population. We cannot fully distinguish these two explanations from the aggregate data available to us. The Discussion in this platform's own codebase notes that the drift-normalized classification engine deployed 2026-07-13 was built specifically because population replay under the legacy static-cutoff engine (the engine underlying every profile in our FPS and MOBA samples) showed cutoff-based rules over- or under-firing relative to the live cohort's actual score distribution. That the dynamic engine's stated purpose is re-anchoring rule thresholds to a trailing cohort median is itself indirect evidence that at least some of what we are measuring here as "concentration" may partly reflect legacy rule-threshold miscalibration rather than a pure population-homogeneity effect. A natural follow-up, once enough FPS and MOBA combines have accumulated under the dynamic engine specifically, would be to rerun this same concentration analysis on a dynamic-engine-only sample and see whether the gradient persists, shrinks, or reverses; we have added that comparison to this platform's topic queue for a future data paper.

Several further limitations bound these results. First, the MOBA sample (n=28) is small, and its bootstrap confidence interval reflects that: a modal share anywhere from roughly two-thirds to over 90% is statistically consistent with our data, so the MOBA point estimate (78.6%) should be treated as considerably less precise than the FPS or general estimates. Second, this is an observational snapshot of self-selected genre choice; we cannot separate genre-driven skill specialization from genre-driven differences in who chooses to play NeuroRank's FPS or MOBA combine at all (for instance, if MOBA combine-takers skew toward a narrower, more dedicated competitive sub-population than general combine-takers, some of what we attribute to "narrower skill package" could instead be "narrower player population," a distinction our data cannot resolve). Third, and most importantly, the schema-mismatch finding disclosed in Methodology remains unresolved: we do not know with certainty why 15 profiles carry a genre tag inconsistent with their scoring schema, only that the pattern is temporally clean and does not appear in the completed-sessions table, and we would rather report that honestly than speculate past what the aggregate data can support. Fourth, our concentration metrics (HHI, normalized entropy, modal share) are all descriptive statistics of the same underlying categorical distribution and are highly correlated with each other by construction; we report all three primarily to show they agree, not because they constitute three independent lines of evidence.

References

• Bartle, R. (1996). Hearts, clubs, diamonds, spades: Players who suit MUDs. Journal of MUD Research, 1(1).
• Simpson, E. H. (1949). Measurement of diversity. Nature, 163, 688.
• Rhoades, S. A. (1993). The Herfindahl-Hirschman Index. Federal Reserve Bulletin, 79, 188-189.
• 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.