Eidetic Recall Training Duration and Cognitive Load Limits in Mentat Development: An Archival Cohort of 156 Candidates, 10110–10240 AG

1. Introduction

When the Butlerian Jihad ended in 108 BG and the Orange Catholic Bible forbade machines in the likeness of a human mind, the work those machines had done passed largely to the Mentats, human computers trained to hold vast bodies of data and to render projections from them. Comparative work has shown where that substitution succeeds and has observed that Mentat curricula devote most of their hours to bounded-data synthesis.

Mentat schools have long assumed that more training yields a better Mentat. Curricula of fifteen years or longer appear in the records of several Great Houses, and household registers treat an extended programme as a mark of quality, yet the assumption has never been tested against examination data. A ceiling matters practically, since candidates are costly to train and one trained past his optimum may serve worse than one released earlier.

Any such test must contend with sapho juice, the lip-staining drug that amplifies Mentat computation and whose sustained use is standard practice. A companion analysis found that most of the drug's benefit to synthesis is realised within about seven years of use, whereas dependency builds over roughly the first decade toward a ceiling. Because use usually begins during training, long-trained candidates carry the heaviest cumulative exposure and most established dependency, so a late decline attributed to training may belong partly to the drug.

We write at approximately 10250 AG, some decades after the Jihad of Muad'Dib. We test three propositions: that recall and synthesis plateau with training duration; that some candidates decline beyond their plateau; and that such decline carries a signature of interference between recall pathways, distinguishable from general fatigue.

2. Methods

The records came from 23 training programmes active between 10110 and 10240 AG: 17 Mentat schools and 6 House-sponsored trainers, the latter drawn from the Atreides Household Archive on Caladan, the Harkonnen Administrative Records on Giedi Prime and House service contracts lodged with the Landsraad Public Archives. The Harkonnen ledgers hold both Tleilaxu-supplied Mentats and household-trained candidates; only the latter were retained, and 13 entered the sample. Three programmes recorded supplementary prana-bindu instruction by Bene Gesserit teachers, which we treat as a programme characteristic. Of 160 candidate files, 4 were Harkonnen-ledger files of Tleilaxu-trained, or "twisted", Mentats; these were excluded before analysis because their conditioning follows a different and largely undocumented regimen. The analytic sample is therefore 156 candidates.

The files record examinations conducted under the Collegium of Mentat Examiners' protocols as set out in its 10214 AG monograph, and we adopt the archive's own usage for them. Recall was measured by digit span, the longest random digit sequence reproduced without error, which the monograph lists as one of several recall examinations; we chose it as the most consistently recorded. Synthesis was measured as time to answer the query the registers call a complexity-8 query, a structured-dataset problem of fixed difficulty under the same protocols. Projection accuracy was the proportion of 40 closed-outcome scenarios projected correctly. Some programmes also re-examined candidates after eight hours of continuous work (the fatigue condition).

Each candidate was assigned to one stage by total training duration at the close of the record: Stage 1 (under 3 years, n=24), Stage 2 (3 to under 6 years, n=31), Stage 3 (6 to under 10 years, n=48), Stage 4 (10 to under 15 years, n=39) and Stage 5 (15 years or more, n=14), so that a candidate at a boundary falls in the higher stage. Stage comparisons use each candidate's final examination. Recall was analysed by one-way ANOVA and synthesis time by Welch's ANOVA, because its variance fell sharply with training; adjacent stages were compared by Welch contrasts with Holm correction. Projection accuracy, a count of correct outcomes out of 40 scenarios, was modelled by mixed-effects binomial regression with a random intercept for each candidate; the stage effect was tested by likelihood ratio. A second-order polynomial regression of final digit span on years of training, with standard errors clustered by programme, estimated the shape of the curve. Weekly training hours (median 38, range 22–61) were tested as a programme-level covariate.

Within-person analyses used the 97 candidates with at least three annual examinations: 44 in Stage 3, all 39 in Stage 4 and all 14 in Stage 5. For the 61 whose series reached a visible plateau (16 of 44 in Stage 3, 35 of 39 in Stage 4 and 10 of 14 in Stage 5), we fitted saturating curves to recall and synthesis separately and computed the time to 95% of plateau. A candidate was coded as showing negative returns if, after more than a year of further training, the final examination fell below his own best earlier result on any of three criteria. The first was a recall loss of at least 3 digits, about twice the standard error of measurement estimated from same-week re-examinations. The second was a synthesis time at least 15% slower. The third was a fall of at least 5 points in fatigue-condition accuracy while unfatigued accuracy held steady, a pattern that isolates interference from general decline. Negative-return status was modelled by logistic regression on training years, then with years of sapho use added (recorded for 89 of the 97).

Each recall error in the final examination was classified from the error logs as random or systematic, the latter being substitution of a related item from another learned dataset. The proportion of systematic errors was compared by binomial regression adjusted for stage. Interview notes and instructors' logs, available for 11 of the 18 decliners, were read for descriptions of recall difficulty. Career outcomes came from 47 service records with a standardised performance rating.

3. Results

Recall and synthesis improved steeply for ten years, then levelled (Table 1). Digit span differed across stages (F(4,151)=69.5, p<.001, η²=.65). It rose by 17.5 digits (51.2%) from Stage 1 to Stage 2, by 16.7 digits (32.3%) to Stage 3 and by 5.7 digits (8.3%) to Stage 4; all three contrasts survived correction (Stage 3 vs 4, t(75.6)=2.38, p=.020). Stage 5 scored 1.8 digits (2.4%) below Stage 4, a difference well within chance (t(19.9)=0.43, p=.67).

Synthesis time followed the same course (Welch F(4,53.0)=52.2, p<.001). It fell by 89 s (47.6%) between Stages 1 and 2, by 47 s (48.0%) between Stages 2 and 3, and by 17 s (33.3%) between Stages 3 and 4 (t(85.0)=3.96, p<.001). Stage 5 candidates were 7 s slower than Stage 4 (+20.6%), but the difference was not significant (t(17.7)=0.93, p=.36). Projection accuracy differed across stages (likelihood-ratio χ²(4)=118.6, p<.001), rising from 38% in Stage 1 to 74% in Stage 4 (odds ratio per stage over Stages 1–4, 1.63, 95% CI 1.47–1.81). It stood at 73% in Stage 5, indistinguishable from Stage 4 (odds ratio 0.95, 95% CI 0.65–1.38, p=.79).

In the polynomial model, final digit span = 24.2 + 7.61 × years − 0.282 × years² (R²=.60). The quadratic term was clearly required (F(1,153)=51.7, p<.001), and the fitted curve peaks at about 13.5 years and 75.5 digits. Weekly training hours left both coefficients unchanged. Among the 61 candidates whose series levelled, individual curves reached 95% of plateau at 10.2±2.1 years for synthesis and 11.7±2.8 years for recall, a combined mean of 10.9±1.8 years, with individual values spread across roughly 7 to 15 years. The 16 Stage 3 candidates who levelled did so earliest, at 8.6±0.7 years, against 11.5±1.3 years in Stage 4 and 12.4±1.5 years in Stage 5.

Negative returns were recorded in 18 of the 97 serially examined candidates (18.6%): 3 of 44 in Stage 3, 7 of 39 in Stage 4 and 8 of 14 in Stage 5 (57.1%, 95% CI 28.9–82.3%). Among the eight Stage 5 decliners, five lost recall (mean 4.6 digits), six slowed in synthesis (mean +24%) and three met the fatigue-selective criterion (mean loss 6.1 points). Four other Stage 5 candidates were still improving at their final examination. The odds of negative returns rose with training duration (odds ratio 1.37 per year, 95% CI 1.17–1.61, p<.001). Once years of sapho use were added, the training coefficient shrank to 1.21 (95% CI 0.98–1.49, p=.08), while sapho years retained an independent association (1.14, 95% CI 1.01–1.29, p=.04); the two exposures were strongly correlated (r=.71).

Decliners made far more systematic errors. Of their recall errors, 18.3±7.2% were substitutions from a neighbouring dataset, against 4.1±2.3% among the 79 non-decliners (odds ratio 5.1, 95% CI 3.6–7.2, p<.001). Instructors' logs and interview notes describe candidates struggling to isolate one body of data in the computation state, a related but wrong dataset surfacing in place of the one sought; one instructor recorded a Stage 5 candidate's intrusions as "false memories".

Career records did not settle the question. Among the 17 of the 47 service records with ten or more years of training, the six Mentats trained fifteen years or more were rated 0.52 standard deviations below the eleven trained 10 to under 15 years (95% CI −1.62 to 0.58, t-based with small-sample standard error, p=.33). The 30 records from shorter programmes were too heterogeneous in role to compare.

4. Discussion

The firmest finding is a plateau: recall and synthesis gain little after about eleven years, and the fitted curve tops out between twelve and fifteen. The late decline is a weaker claim. Stage 5 means fall slightly, but no contrast approaches significance, and fourteen candidates allow only large effects to be detected. We therefore present decline as a hypothesis the data make plausible and do not treat it as established.

Two observations keep that hypothesis alive. Decline is concentrated in particular candidates while some Stage 5 candidates keep improving, which points to individual optima hidden by group averages. Decliners also substitute related material far more often than peers, a signature that fits interference between densely overlapping recall pathways better than it fits general exhaustion, and the fatigue-selective cases point the same way. The association is cross-sectional, and interference remains a proposed mechanism.

Sapho is the most serious rival explanation. Long training and long sapho use travel together, and adding sapho years weakened the training effect below significance. A saturating drug benefit cannot account for the pattern, because the companion study places most of sapho's gain within about seven years of use, well before the plateau near eleven years of training. The plausible confound is cumulative exposure and the dependency that builds through the first decade of use. Whether established dependency itself degrades recall is untested, and our data cannot say which exposure matters more. Separating them will need the rare programmes that delayed or withheld sapho.

The practical implication holds even if decline proves illusory. Because individual plateaus spread across roughly seven to fifteen years, a fixed long curriculum will over-train some candidates, and the serial examinations most schools already keep could identify each plateau as it arrives.

5. Limitations

The archive is selective: files survive where Houses and schools chose to keep them, and failed or withdrawn candidates are probably under-represented. With fourteen Stage 5 candidates, the decline rate is imprecise and the stage contrasts under-powered.

Stage comparisons are between groups, whereas negative returns are coded within persons for the serial subset only. Curve-fitting excluded 36 candidates whose series had not visibly plateaued, 28 of them in Stage 3, which probably biases times to plateau downward. Sapho use was recorded inconsistently and was missing for 8 serial candidates, and melange co-exposure was not recorded at all. Digit span captures only one facet of Mentat recall, and examination standards varied across programmes and decades in ways that clustering absorbs only in part.