Evermind or Federation? Thinking-Machine Architecture of the Butlerian Jihad (201–108 BG) Reconstructed from 247 Technical Fragments in Three Jihad- and Prohibition-Era Archives

1. Introduction

An Imperial historian of the thinking machines works under a peculiar handicap. The Butlerian Jihad (201–108 BG) ended more than ten thousand years ago, and its victors set out to erase the technical knowledge that had made the machines possible. The Orange Catholic Bible, compiled afterwards by the Commission of Ecumenical Translators (C.E.T.), made the ban on machines made to resemble human minds its central commandment (Ossery, 10212 AG), and the statutes that followed ordered the destruction of plans, manuals and designs. What survives is the paperwork of prohibition itself: laws that had to specify what they forbade, campaign records describing enemy machines, and Ixian commentaries marking where the permitted machine arts stop.

These records mention redundancy, damage, delay and command only because enforcement or combat required it. This paper asks whether their incidental detail better fits a centralized or a distributed design. Our centralized model draws on the late-tradition compilations about the machine side of the war, which describe a single sovereign intelligence, the evermind (named Omnius in those texts), whose copies on many worlds were periodically brought back into agreement with one authoritative state. Such material is contested and here frames a hypothesis only; the primary record speaks of the war chiefly as men with machines subjugating other men. Our distributed model is the alternative that primary picture implies, a federation of installations held by different operators, each with local authority and coordinating with its peers.

One term needs definition. Late-tradition sources use cymek for a human brain housed in a machine body. Cymeks are therefore mind-housings, not computational nodes, and we treat the few fragments that mention them separately from the architectural evidence.

2. Sources and Method

The first archive series is the Salusa Secundus Military Archives series JH-47 (files 1–47), a group of engagement records dated to 201–108 BG that survive only as later recopies. The same 47 files were examined in a companion study, sharing this paper's first author, that coded who fought whom; here we coded only passages in which commanders describe what machines did. Neither architectural coder worked on that study, and the first author, who co-coded its faction variables, took no part in the architectural coding.

A second series, held by the Landsraad Public Archives as series JIHAD-TECH, gathers 156 prohibition statutes from 1–480 AG that contain technical specification clauses. The third series comprises 44 boundary commentaries in the Ixian Consortium Archive (series IC-ARCH, 9870–10180 AG). They analyse what the proscriptions forbid, to keep Ixian work on the permitted side of the line; none describes building a prohibited machine. The series is restricted.

Three searchers independently read every document in the three series for technical content. Rhetorical or purely regulatory passages were set aside. Of the 247 technical references finally retained, 193 were found by all three searchers, 38 by exactly two and 16 by one only. An incidence-based Chao-type capture–recapture estimator applied to these frequencies (Dace, 10205 AG) puts the number of retrievable references in the surviving holdings at about 249, so the search probably missed very few. The legal series supplied 132 references, the engagement series 61 and the Ixian series 54.

Each reference was first assigned to one of four criteria. Two coders, one trained in Mentat-school data synthesis and one in Imperial military history, agreed on 216 of the 247 assignments (Cohen's κ = 0.81), and the remaining 31 were settled by consensus. The final counts were redundancy 73, failure tolerance 68, latency 54 and control 52. The two models' predictions for each criterion were written down before scoring and are summarized in Table 1. Following the fragment-reconstruction protocol of Marn (10224 AG), each reference was then scored for fit to each model on an ordinal scale: 0 contradicts the prediction, 1 is neutral or uninformative, 2 is moderate confirmation and 3 strong confirmation. Agreement on these scores was substantial (quadratic-weighted κ = 0.78); disagreements on 39 references were resolved by consensus and are revisited in the Limitations of the Record. The rubric was frozen after a pilot of 49 references; the share supporting the federation model was 59% in the pilot and 60% in the remaining 198.

Since the scores are ordinal and every reference is scored under both models, we treat a score of 2 or more as support and compare the models with McNemar's test on the paired outcomes. The 247 references came from 176 of the 247 documents read; the median contributing document supplied one reference and none more than four. Because references from one document are not independent, we repeated the main comparison with each document casting a single vote, the direction held by most of its discordant references, ties discarded.

3. Redundancy and Failure Tolerance in the Statutes and Engagement Records

Most of the redundancy evidence comes from the statutes (compare Reyes-Okafor, 10221 AG). A lawgiver who forbids a machine must say what counts as having destroyed it, and the JIHAD-TECH clauses repeatedly require each reckoning centre to be seized and broken separately, on the stated ground that taking one gave no command over the others and that none answered to another. A smaller group of clauses orders every centre broken lest a survivor restore the rest. That wording fits interchangeable replicas as well as independent peers, so these clauses were scored 1 under both models unless their context specified how the centres related. Of the 73 redundancy references, 55 (75%) supported the federation model and 10 (14%) the evermind model. Discordant references split 49 to 4 in the federation's favour.

The engagement records supply most of the failure-tolerance evidence. One recopied file records that an installation went on functioning after its principal control nexus had been wrecked, and its author concluded that authority must have been lodged elsewhere as well. Of 68 failure-tolerance references, 45 (66%) supported the federation model and 9 (13%) the evermind model, and the discordant split was 41 to 5. No file describes a whole enemy system collapsing after one decisive strike.

4. Latency and Control

Delay is mentioned less directly. Some engagement records note that machine formations held back until separate elements had reached agreement, a pattern closer to negotiation among peers than to replicas waiting on a master. Of 54 latency references, 28 (52%) supported the federation model and 10 (19%) the evermind model; the discordant split was 25 to 7, a weaker margin than for redundancy.

Among the control references the two models are no longer separable. Of 52 control references, 20 (38%) supported the federation model and 17 (33%) the evermind model, with a discordant split of 16 to 13 (exact binomial p = .71). The data are compatible with local hierarchy inside a federated whole, but an evermind delegating command to subordinate installations would be expected to leave a similar mixture, so the command structure remains undetermined.

5. The Ixian Boundary Commentaries and the Cymek References

Ixian commentators approach the old machines from the opposite direction: their task is to show that a permitted device lacks the autonomous-cognition capacity the proscriptions target, and to do so they characterise the forbidden machines in functional terms. One commentary distinguishes the permitted class by its inability to coordinate with peers without a human operator, implying that peer coordination was a defining property of what had been banned. Of the 54 Ixian references, 31 (57%) supported the federation model. The corresponding figures for the statutes (80 of 132, 61%) and the engagement records (37 of 61, 61%) are almost identical, and support did not differ by archive (χ²(2) = 0.18, p = .91).

Cymeks appear in nine passages. Two are marginalia added to JH-47 files at recopying, in a hand later than the operative text; the other seven occur in Ixian commentaries that cite late-tradition compilations as their authority. None belongs to the primary record, none describes how cymeks connected to the installations, and all nine were excluded from the 247 scored references.

6. Discussion

Taken together, 148 of 247 references (60%; 95% CI 54–66%) supported the federation model and 46 (19%; 95% CI 14–24%) the evermind model. Of the 160 discordant references, 131 favoured the federation (82%; exact 95% CI 75–88%; McNemar χ²(1) = 65.0, p < .001). The advantage is concentrated in redundancy and failure tolerance, is smaller for latency and is absent for control. With each document casting one vote, the federation led in 97 of 121 documents with a discordant majority (80%; exact binomial p < .001). The fragments support an architecture federated for redundancy and failure tolerance; its command structure is undetermined. The late-tradition picture of a single authoritative state fits the fragments poorly.

The most important argument concerns the nature of the sample. A purge shapes what survives, so the pattern might merely reflect what the censors left. The three series were written by soldiers, legislators and Ixian engineers across more than ten millennia, yet they are only partly independent. The JIHAD-TECH statutes are the prohibition authorities' own texts, the Ixian commentaries analyse what those statutes forbid, and the engagement files reached us through the post-prohibition recopying tradition. A shared censoring hand cannot be excluded. What weighs against a single planted signature is that the engagement records and the Ixian commentaries differ in purpose and date: campaign reports explaining a battle's course and engineering opinions written ten millennia later would each have needed separate and matching alteration. Cross-archive agreement is therefore supporting evidence, weaker than three independent witnesses would give.

This result bears on the companion finding that most determinable Jihad-era engagement records show human commands with machine assets on both sides. A federation of installations held by different operators is the architecture such a war would require, since rival factions cannot share a single sovereign intelligence. Whether the prohibition then functioned as a settlement among exhausted factions is a further interpretation that the architecture results do not test; we note only that it is compatible with them.

Two later strands of evidence fit this picture without adding weight to it. Reconstructions from the same class of fragments find Mentats furthest short of the machines on large-scale parallel work, which cooperating installations would supply, and the Ixian boundary is still policed by inspection for autonomous decision-making capacity.

7. Limitations of the Record

Purge bias has no known direction. Censors who judged networked designs most dangerous would have destroyed their documentation first, biasing the sample toward the evermind model; censors who found a single intelligence simpler to erase would have biased it the other way. Cross-archive agreement cannot exclude a bias shared by every institution that inherited the prohibition.

The two models are stylized. Hybrids, such as federated installations with regional masters, would score moderately under both, and the indeterminate control results allow for such mixtures without showing them.

Of the 39 disputed references, 22 were federation-favouring discordant references on which one coder's reading would have moved them to the evermind side. In the remaining 17, both coders' readings lay on the same side of the support threshold under both models, so no resolution of the dispute could change their classification. If every one of the 22 is therefore resolved against the federation model, the discordant total stays at 160 and the split narrows from 131:29 to 109:51. The federation still holds 68% of discordant references (McNemar χ²(1) = 21.0, p < .001), but the margin is materially smaller.

The engagement series is recopied, the Ixian series is self-selected by a Consortium with an interest in where the boundary lies, and the legal series postdates the war by one to six centuries. The reconstruction describes how the machines appeared to those who fought, outlawed and policed them.