Mentat Cognition as Post-Jihad Substitution for Prohibited Computation: Service Records of 47 House-Retained Mentats, 10150–10230 AG, Against 23 Pre-Jihad Technical Fragments

1. Introduction

The Butlerian Jihad (201–108 BG) ended with the destruction of thinking machines and left the Imperium one governing rule, the Orange Catholic Bible's prohibition on building machines that imitate human thought. The Mentat is the best-known answer to the loss of those machines. The Imperial glossary describes Mentats as people trained for the highest accomplishments of logic, human computers; reading them as a deliberate substitute for the vanished machines is the framing of this paper.

Mentats appear in the record as retained servants of the Great Houses. They act as household and strategic advisers, sometimes also answer for a House's security, and in Houses with large CHOAM holdings work as commercial analysts. Thufir Hawat served House Atreides as adviser and Master of Assassins for decades; Piter de Vries, the "twisted" Mentat in Harkonnen service, was the Baron's Mentat adviser. No Mentat in the primary record processes data in bulk for a whole planet.

Earlier comparisons asked whether Mentats are, overall, as fast as the machines were (Reyes-Okafor, 10229 AG). We propose a narrower hypothesis. Bounded, sequential, decision-support cognition can be taken over by a trained person, whereas computation that depends on massive parallel handling of data cannot. If so, Mentats should approach the machine benchmark on advisory tasks, fall behind as scale grows, and be absent from planetary-scale work.

We write around 10240 AG, some decades after Muad'Dib's Jihad. The machine-side material is late and contested; we treat it as a benchmark of what was claimed for the machines.

2. Methods

Mentat performance data come from registers of retained Mentats kept at Kaitain for 10150–10230 AG (Imperial Archives, Kaitain, 10150–10230 AG); the series shows the registry still receiving House returns after the imperial seat moved to Arrakis under Muad'Dib. Atreides household service logs for 10150–10191 AG supplement them. The registers note each commission's subject, the scale of material supplied, and when analysis began and the finding was delivered. The final set comprises 1,200 task records from 47 Mentats serving 19 Houses. Because most Mentats worked in more than one category, the per-category counts in Table 1 sum to 88.

Before any timing was extracted, twelve task categories were defined from the registers' vocabulary and assigned to classes by two fixed criteria: whether one person could carry out the task as a single chain of inference, and the scale band of the material. Six categories were classed as bounded, three as intermediate and three as parallel. The registers contain no Mentat commission in any parallel category. Houses gave that work to allocation committees whose clerks used calculating aids of fixed programme, Ixian and Richese devices tolerated because they cannot revise their own goals. For those categories we drew 76 committee episodes from CHOAM allocation minutes (CHOAM Directorate, 10150–10230 AG), analysed separately and never counted as Mentat data.

The machine benchmark rests on 23 fragments of pre-Jihad technical writing held in the Ixian Consortium Archive (accession IXA-R-17). All survive only as later copies; internal evidence places their composition around 350–201 BG, with an uncertainty of at least a century. From them we extracted 31 capability claims and assigned each to a category, following the reconstruction protocol of Marn (10236 AG). Unlike the 247 incidental references recovered from three prohibition-era archives by an earlier study, these are capability statements written on the machine side, which a faction advertising its assets might have exaggerated.

Twenty-two claims gave a numeric time and were used as stated. Five qualitative claims described only the manner of computation and served for classification alone. The remaining four were speed terms such as "rapid" or "near-instantaneous", which a rule written before extraction coded as 60 seconds; a sensitivity analysis recoded them as 10 seconds. Each category's reference time was the geometric mean of its timed claims.

The outcome for each Mentat task was the log of Mentat time over reference time (positive values mean the Mentat was slower), analysed within each category by a linear mixed model with a random intercept for Mentat. Because the reference is itself an estimate from few claims, we added to each category's variance the sampling variance of its reference mean, using the log-scale spread among claims within a category pooled across the nine Mentat categories (SD 0.30 on 12 degrees of freedom). We report the exponentiated mean as a time ratio with this widened 95% confidence interval and tested equivalence by two one-sided tests against a margin, set in advance, of a factor of two (ratios 0.5–2); any equivalence shown is equivalence to the stated claims only. Parallel references rest on coded vague terms, one upper bound and one pair of figures, from which no claim-level variance can be estimated, so committee comparisons are given as order-of-magnitude ranges. Finally, for threat assessments and negotiation analyses whose outcome was later recorded, two readers independently classified each wrong judgment as arising from a false or withheld premise or from faulty inference, and disagreements were resolved by discussion.

3. Results

Table 1 gives the nine categories with Mentat records. Bounded ratios (850 tasks) lay between 0.9 and 1.4, and no bounded interval excludes 1. Equivalence within a factor of two was shown in five of the six categories (two one-sided tests, p ≤ .025 in each). Household revenue audit, with the highest point estimate, has an upper limit above 2, so its result is indeterminate. The 350 intermediate tasks ran three to five times slower; every interval lies above 1 and extends well past 2, and equivalence could not be shown for any of them.

For the three parallel categories, Table 2 sets committee times against the fragment claims. Allocating a planet's water, power and labour for one planning cycle took committees one to three days, against claimed machine times of two and ten minutes, a ratio of roughly 10² to 2 × 10³. Scheduling production across several interlocking systems took two to three weeks, about 2–3 × 10⁴ times the coded fragment time. Reconciling a fief-wide census with its tithe rolls occupied ten to fourteen weeks; since the fragment claims only that it took under an hour, the ratio is a lower bound of at least 1.7 × 10³. The minutes record these durations without comment or complaint.

Recoding the four vague speed terms as 10 seconds altered three categories. Commercial contract appraisal, which rests partly on one such term, moved to a ratio of 2.4 (95% CI 1.55–3.71) and no longer meets the margin. Multi-front coordination rose to about 9, and production scheduling to roughly 1–2 × 10⁵. Equivalence then holds in four of the six bounded categories, and the bounded–parallel gap widens. With a pooled claim spread half again as large (SD 0.45), equivalence likewise held in four.

Of the 340 threat assessments and negotiation analyses, 188 have recorded outcomes, and the retaining House judged 41 of these (22%) wrong. The two readers agreed on 37 of the 41 classifications (Cohen's κ = 0.77). In 29 of the 41 (71%; exact 95% CI 54–84%) the error traced to a false or withheld premise, and in the remaining 12 to faulty inference from sound material.

Curricula survive from five Mentat schools (Dorvane, 10226 AG). They allot between 67% and 75% of instructional time to bounded synthesis, pattern recognition and sequential logic, and between 8% and 12% to the handling of large parallel bodies of data.

4. Discussion

The results support the split hypothesis, with the qualification that the benchmark is what was claimed for the machines. On the advisory work Houses commission most often, a trained Mentat matched those claims within a factor of two. Beyond a single chain of reasoning the margin opened, and at planetary scale no House attempted a Mentat solution. The prohibition may have endured because the substitute covers the functions whose loss those with power would have felt at once.

A mechanism for the ceiling is available in the training literature. An archival cohort of Mentat candidates found that recall and synthesis gains level off after roughly a decade of training, with signs of interference between learned datasets thereafter. A mind working sequentially from memory should scale well with the depth of an inference and poorly with the breadth of its data, the pattern in Table 1. The school curricula point the same way: instruction concentrates on bounded synthesis and gives little time to parallel work. Time allocation alone cannot show whether the schools judged such work beyond training.

Most wrong Mentat judgments in the threat and negotiation categories came from bad inputs, and far fewer from bad reasoning. The canonical case is Hawat's belief, built on planted Harkonnen evidence, that Lady Jessica was the traitor within House Atreides. A Mentat is most vulnerable where an adversary controls what it is told, a weakness a House can at least probe by questioning its adviser.

Committees in the parallel categories show how the Imperium absorbed what it could not replace. They worked hundreds to tens of thousands of times more slowly, and the minutes pass over this in silence; that Houses treated it as an ordinary cost of government is a hypothesis for ledger study. The aids belong to a tolerated class whose boundary, an earlier analysis of Ixian inspection practice argues, is set by what inspectors can verify about a device, so the limits of inspection and of Mentat competence together fix how large-scale work is done. Lastly, the architectural evidence suggests pre-Jihad installations were separately held federations. If the fragment figures describe their pooled capacity, the civilisation-level comparison made here is the right one; set against a single installation the gap would narrow, though not by an order of magnitude.

5. Limitations

The machine benchmark is the weakest element: its claims cannot be checked against any surviving device, and their authors had reasons to overstate. Our intervals carry the spread among claims, but that spread is estimated from 21 timed claims, and a claim spread twice the pooled value would leave no bounded category formally equivalent. The bounded parity is therefore a statement about the claims made for machines, which may differ from what the machines could do.

Service registers record what a clerk chose to note. Start times may lag the real beginning of analysis, and failed or abandoned commissions are probably under-recorded, which would flatter Mentat performance. Surviving House archives are not a random sample, and the Atreides logs supply a disproportionate share of detailed entries. Outcomes were recorded for only 188 of 340 eligible commissions, and Houses may have noted outcomes more readily after a conspicuous failure, so the error rate and the premise share may not describe unrecorded cases.

Sapho juice, the lip-staining stimulant many Mentats take, is noted for 19 of the 47 but irregularly and without dose, so we could not model it; nor can the registers reliably identify twisted Mentats. The committee comparisons measure an institution against a machine, not what a single Mentat could achieve at planetary scale.