Altamens Research · Independent public-data reanalysis · 6 August 2026
Difficulty-balanced forms were more consistent than random forms in simulations using 59 participants and an independent open arithmetic dataset. The improvement was meaningful but modest: balancing reduced score variance, not all score noise.
The source paper reported a robust problem-size effect: arithmetic sentences with larger operands tend to take longer. The reanalysis reproduced the paper's participant-level Kendall correlations exactly to rounding.
2 + 9 = 11.Five repeated five-fold participant cross-validations estimated item difficulty only on training participants. The difficulty measure combined participant-centred log response time and smoothed error rate. Held-out participants were simulated on 1,000 forms per condition. Random forms used shuffled item order; quota-balanced forms controlled operation and truth status; difficulty-balanced forms additionally controlled cross-validated difficulty quartiles. Participant bootstrap intervals used 10,000 resamples.
Question composition changed short correct-count scores in this dataset. The tested balancing method reduced form variance. Meaningful variation remained.
No clinical validation, population norms, intelligence claim, training effect, or direct reliability estimate for Altamens.
These are simulations from a U.S. undergraduate laboratory true/false task, not direct observations from the Altamens four-choice mobile test. Device latency, day-to-day state, fresh response noise, repeated testing and the live item bank were not measured. The analysis was protocol-guided but not formally preregistered.
The accompanying package contains reproducible scripts, simulation rows, aggregate tables, item calibrations, figures and a full Markdown report.