Two Machine Cultures under One Rule: Control Depth, Operator Mediation and Volume in 251 Richese and Ixian Product Lines, 10150–10230 AG

1. Introduction

Two entries in a CHOAM trade-classification register for 10208 AG describe instruments sold for one task, measuring distance on rough ground. The Richese hand ranger holds its lens train in focus with two nested feedback loops in a control assembly of about 55 millilitres, and asks its holder to touch the focus ring every thirty-five minutes or so. The Ixian survey head nests six loops in a cabinet of about 27 litres and runs some forty hours on a stored programme before an operator sets it again. Both were classified under the same rule (CHOAM Directorate, 10150–10230 AG), and they are not the same kind of machine.

We write c. 10232 AG, some decades after Muad'Dib's Jihad. Ix and Richese both work at the edge of the proscription on machines in the likeness of a human mind, and most of what is documented about them comes through inspection. The Ixian exception is recorded as an inspection standard (see the study of Ixian exception-making), and the wider enforcement record holds no Richese counterpart to the Ixian disclosure of clearance applications (see the study of enforcement mechanisms). Product filings are the only systematic view of what Richese builds.

The standard defines a permitted class with three conditions. A device must execute a fixed programme or direct operator commands; any sensory feedback must serve only to hold values the operator has set; and the device must be unable to revise its goals or change its behaviour through accumulated experience. The first condition offers two routes, a stored programme or an operator issuing commands during the run. We argue that the houses took different routes and that most of what separates their lines follows from the choice. Richese keeps the operator in the loop and spends its skill on volume; Ix stores the programme and spends its skill on depth, at a price that falls on the third condition.

2. System Description

Powered product lines, as filed for trade, are the system studied. Our sources are the CHOAM trade-classification filings of 10150–10230 AG, each carrying a specification sheet and a summary of the control schematic (CHOAM Directorate, 10150–10230 AG; Tervane, 10226 AG). We scored 251 lines: 164 Richese lines in 41 product families and 87 Ixian lines in 29. A further 71 Richese lines are passive optics with no feedback loop and were not scored. Lines in a family share a design, so every interval below comes from 5,000 resamplings of whole families within each house.

Five quantities were scored. Control depth, d, is the number of nested closed feedback loops holding a value derived from an operator-set value, counted from the schematic summary by the Ixian coding protocol (Delvane, 10219 AG); a second reader's recount of 40 lines agreed exactly on 32 and within one loop on all. The control branch, P, is 1 where the sheet's principal operating sequence is a stored programme and 0 where the operator issues it during the run. The state indicator, S, is 1 where any loop contains a part whose setting is changed by the unit's own operating history, such as a wear compensator. Volume, V, is the smallest enclosure in litres holding sensing, programme store and drivers; the operator-input interval, τ, is the rated operating time in hours between inputs the operator must make.

Field data come from two ledgers: CHOAM warranty-return ledgers of unscheduled returns (CHOAM Directorate, 10171–10230 AG) and Ixian field-service logs of operator-input registers and service events (Ixian Consortium, 10171–10230 AG). They cover 91 scored lines first listed by 10215 AG, 52 Richese lines in 13 families and 39 Ixian lines in 11, and give the interval at which operators actually made inputs, exposure in unit-years and the cause of each return as assigned by the depot. Table 1 summarises the lines.

3. Analysis and Model

The two lines barely overlap. A random Ixian line has more loops than a random Richese line with probability 0.94 (95% interval 0.89–0.98), a larger control assembly with probability 0.997 (0.989–1.00) and a longer rated interval with probability 0.95 (0.92–0.98). Only 26 of 164 Richese lines (15.9%) reach the Ixian lower quartile of four loops, and 11 of 87 Ixian lines (12.6%) fall below it.

We fitted two linear models by least squares. The first treats the logarithm of the rated interval as a sum of structural terms: log10 τ = α + βd + γP + δS + ε. The second relates log control volume to loop count and make: log10 V = α′ + bd + cI + ε′, with I equal to 1 for Ixian lines.

Each added loop lengthens the rated interval by a factor of 2.0, a stored programme by 8.7 and a history-dependent element by 1.6 (Table 2). The three terms explain 92.5% of the variance of the log interval against 59.4% for house alone. Given them, Ixian make adds nothing we can detect: its multiplier is 0.91 (0.65–1.32, p ≈ .58), which excludes a house effect larger than about a third in either direction.

The structural terms therefore account for the whole of the gap. The geometric-mean interval of Ixian lines is about 46 times that of Richese lines, or 1.66 log10 units; depth supplies 0.98 of them (59%), the stored-programme branch 0.63 (38%), history-dependent elements 0.07 (4%) and the house −0.01 (−1%). The houses do not follow two engineering laws; they sit at two points on one.

Volume behaves differently. Each added loop multiplies control volume by 2.1 (1.9–2.4), and Ixian make multiplies it by a further 33 (17–64). At equal depth a Richese control assembly thus occupies about 3% of the space an Ixian one needs. This premium belongs to the house, and it matches Richese's reputation for miniaturisation. Earlier work by one of us attributes much of it to optical position sensing in small lens-train housings (Ferrand-Aix, 10224 AG), an account we do not test.

Richese can build to a given depth in 3% of the space, yet 138 of its 164 lines have fewer than four loops. State offers a reason: across all lines the odds of a history-dependent element rise by a factor of 1.7 (1.3–2.5) with each added loop, and Ixian make adds nothing distinguishable once depth is known (odds ratio 2.3, 0.7–10.9). Among Richese lines of four or more loops, 5 of 26 (19.2%) carry such an element and 10 (38.5%) store their programme, against 3 of 138 (2.2%) and 21 (15.2%) among shallower lines. Depth cannot be bought without moving toward the third condition, though we cannot show what the makers intended.

4. Validation Against Field Data

We validated the model on lines it had not seen. The interval model was refitted to the 160 lines without field logs (112 Richese, 48 Ixian) and used to predict the logged interval of the 91 lines with logs. The refit's depth, branch and state terms (0.28, 0.94 and 0.14 log10 units) are close to the full fit.

Logged intervals fell within a factor of two of prediction for 61 of 91 lines (67%, 95% interval 59–76%; Table 3 splits them by house and state) and within a factor of three for 76 (84%); the typical error was a factor of 1.8 (1.6–2.0). Predictions ran about 15% long on the median (logged over predicted 0.85, 0.69–1.02).

Refitting the interval model to the logged intervals themselves reproduces depth and branch (0.32, 0.27–0.37, log10 units per loop in the field against 0.30 on the sheets of the same lines; 0.87, 0.74–1.00, for the stored-programme branch against 0.95) but not state (0.02, −0.21 to 0.26, against 0.32).

The shortfall also shows line by line. Logged intervals were 0.89 (0.86–1.02) of the rated interval for the 78 lines without a history-dependent element and 0.51 (0.39–0.64) for the 13 with one, drawn from 9 families; the ratio of the two is 0.54 (0.42–0.66). The sheets credit such an element with a longer interval, and service does not honour the credit.

Validation here is narrow: it tests the interval model only, on lines first listed by 10215 AG, in two ledgers that record differently, and only 2 of the 13 lines with a history-dependent element are Richese.

5. Failure Modes

The two routes fail in different places. Table 4 sorts the 327 logged returns by the line's control branch. Direct-command lines (48 lines, about 13,400 unit-years) returned at 1.07 per 100 unit-years, and 118 of their 143 returns (82.5%) were operator misadjustment or the entry of grit and moisture. Stored-programme lines (43 lines, about 8,700 unit-years) returned at 2.11 per 100, 1.97 times as often (1.41–2.69), and 95 of 184 returns (51.6%) were trim drift or a fault in the programme store.

A sheet that asks for an input every half hour admits every operator error: misadjustment made up 50.3% of direct-command returns. Fine tolerances in sub-litre lens-train housings leave little margin against grit and moisture (Marbruck, 10221 AG), and ingress made up 32.2%.

The Ixian route fails by drift. Among the 13 lines with a history-dependent element, trim drift made up 47.0% of returns (39 of 83), and these lines returned at 3.05 per 100 unit-years against 1.26 for the other 78 (ratio 2.42, 1.77–3.71); bench work reports such long-service drift (Orlenko, 10228 AG). Drift also costs more under the third condition. A trim whose setting moves with wear is the element an inspector cannot at once tell from a device that has learned, and the study of Ixian exception-making found most referred cases to involve feedback mechanisms whose settings shift with wear.

Branch and house nearly coincide among logged lines, with only 6 Ixian lines on the direct-command branch and 10 Richese lines on the stored-programme one, so the return rates cannot separate them. The ledgers also record differently, warranty claims within a term against all service. The overall rates, 2.78 returns per 100 unit-years for Ix (158 in about 5,700) and 1.03 for Richese (169 in about 16,400), give a ratio of 2.7 (2.0–3.7) that probably overstates the difference.

A third risk is convergence, but the gap has not narrowed over eight decades of listings. The ratio of mean loop counts was 2.2 (1.7–3.0) for lines first listed in 10150–10179 AG and 3.3 (2.3–4.6) for 10210–10230 AG, and neither house's loop count trends detectably (−0.5% per decade, interval −8.0 to 6.9, for Richese; +2.6%, −3.1 to 7.0, for Ix). Were Richese to deepen, its frontier lines show the cost: state arrives with depth.

Sheets and filings are the makers' statements, and loop counts come from schematic summaries. Families, not lines, are the independent units, so the field comparison rests on 24 families. The reasons offered for the volume premium and the choice of depth are inferences from the pattern, not observations.

6. Conclusion

Under one permitted class, the Richese and Ixian lines are different kinds of machine, and the difference is structural. Richese works the direct-command branch of the first condition (81% of its lines): the operator is the programme, loops are few, state is rare, and the house's skill goes into volume, an assembly of given depth taking about a thirty-third of the space an Ixian one needs. Ix works the stored-programme branch (86%): loops are deep, state is common, and the skill goes into unattended running. These choices account for the whole of the 46-fold gap in rated operator-input interval, with no house effect we can detect.

In service the Ixian route is paid for in drift: lines with history-dependent elements delivered about half their rated interval and returned 2.4 times as often as the rest. The Richese route is paid for at the operator and the seal. A buyer should discount the rated interval of a state-bearing line by about half, and a house that deepens its loops should expect the state, and the drift, that go with depth.