Substrate Density, Stability and Portability in a Disconnected Matrix Approximation: A Scaling Model Calibrated on the Aleph

1. Introduction

An environment built in the matrix is rarely self-sufficient. Constructs, architectures and the data they reference are refreshed continuously from the wider network, so the local hardware stores only a working portion of the whole. Our model treats this as the central economy of networked design: when external state updates are available, the local substrate need carry only a fraction of the environment's state. The assumption follows earlier Hosaka comparisons of networked and self-contained architectures (Kessack, 2047), but in this paper it is a modelling premise, and we test its consequences directly.

Disconnection removes that economy. A disconnected environment must hold every piece of its own state, and must keep that state coherent under load with no outside reference. The question we address is how much substrate density such an environment needs to stay stable, and how that requirement changes as the environment grows richer.

One device makes the question concrete. The Aleph, known from accounts of the later Sprawl period, is a biochip unit, built on biosoft (biological computing substrate of the kind developed by Maas Biolabs), that holds an approximation of the entire matrix and can operate without a network connection. We write from the years after the Wintermute–Neuromancer merger, in the 2050s, when the Aleph is documented and the consolidated matrix it approximates has been partly reconstructed. That reconstruction sets the scale of the complexity the device must represent.

2. System Description

Our description of the Aleph draws on two sources. The provenance and transfer ledger held in the Tessier-Ashpool corporate archive (Tessier-Ashpool S.A., 2049–2053) gives the unit's mass, footprint and support requirements in broad terms. The Maas biochip specifications summarised by Nakada-Ross (2049) give the storage and switching density of the biosoft generation to which the unit appears to belong. Neither source is complete, and the figures below are ranges.

Physically, the Aleph is transportable but hardly portable. It is moved as freight with a dedicated power supply and associated support equipment, and users jack in to it through dermatrodes. There is no indication that it can run for long on its own supply, and the ledger records its installation at fixed sites. Any analysis of portability therefore concerns the size of the substrate and power plant, and does not concern pocket-scale hardware.

Functionally, the unit hosts a navigable approximation of the matrix and supports resident personality constructs, recorded minds that can be entered and addressed inside its environment. Those constructs, like the ROM constructs of the earlier period, do not receive outside updates while the unit is disconnected. The environment is, in engineering terms, a closed system: all state change must be generated and reconciled internally.

3. Model

We built a state-maintenance simulation in which an environment of complexity C is held on biosoft of density D. Complexity was indexed in reference units on nine log-spaced levels. The reference unit is deliberately large: level one is a single corporate core architecture of the size catalogued by Oyelaran-Hesse (2051), many times richer than a construct-scale environment, and level nine is our estimate of the Aleph's complexity, three orders of magnitude higher, so that adjacent levels differ by a factor of about 2.4. The level-nine estimate rests on the same post-merger topology survey, which found most of the consolidated matrix's addressable state concentrated in roughly a thousand architectures of that class. We did not derive it by summing the two precursor intelligences: Tanaka-Reyes and Achterberg's reconstruction of the Berne and Rio records before the merger found that the consolidated entity's reported conduct did not follow from either half, so an estimate built up from the precursors would have little warrant. The level-nine figure is therefore the most uncertain input to the model.

Across the nine complexity levels we ran twelve densities on a log-spaced grid set from pilot runs, adjacent grid points differing by a factor of 1.8, with five replicates per cell, for 540 runs in total. Each run applied sustained internal load for an 18-month simulated window with no external update. A run was scored as degraded at the first point where internal state inconsistency exceeded the tolerance used in Maas acceptance testing (Nakada-Ross, 2049). Runs that had not degraded by the end of the window were right-censored.

Time to degradation was fitted with a Weibull accelerated-failure-time model, with log density and log complexity as covariates. Each doubling of density lengthened expected time to degradation by a factor of 2.9 (95% CI 2.5–3.4, p < .001). A tenfold increase in complexity at fixed density multiplied expected time to degradation by 0.008 (95% CI 0.005–0.012, p < .001), roughly a 125-fold reduction. The Weibull shape parameter was 1.6 (95% CI 1.4–1.8). The stability threshold at each level was defined as the density at which predicted degradation probability within 18 months fell to 10%.

Because the thresholds are predictions from a single model with log density and log complexity as its only covariates, log threshold density is exactly linear in log complexity, and the scaling exponent is the ratio of the two log-scale coefficients. We therefore report that ratio directly, with a delta-method interval, rather than regressing the predicted thresholds on complexity, which would only restate it. The exponent was 1.37 (95% CI 1.28–1.46), so required density grows faster than complexity. Running the same simulation with a modelled update feed, calibrated on the networked baselines in Kessack (2047), and deriving the exponent in the same way gave 1.02 (95% CI 0.96–1.08), indistinguishable from linear. The difference of 0.35 (95% CI 0.24–0.46) excludes zero. At level one, disconnected operation required 3.4 times the networked density (95% CI 3.1–3.8).

The gap widens with scale. At level nine the ratio of disconnected to networked substrate is 3.4 × 1000^0.35, or roughly 38-fold. Combining the interval endpoints for the level-one ratio and the exponent difference gives a conservative plausible range of 16–91. Under the Maas assumption that power draw scales with active biosoft volume at fixed technology, the same ratio applies to the power plant. This is the portability cost: a disconnected environment at the Aleph's complexity needs tens of times the substrate and supply of a networked node of equal richness.

As an internal check, 300 runs fell below their level's threshold and 240 at or above it. Of the former, 245 degraded within the window (81.7%), with a median time to degradation of 1.5 months (interquartile range 0.4–5.0) among those that did. Of the latter, 3 degraded (1.3%) and 237 were censored at 18 months. These proportions are in line with the fitted model given the grid spacing: cells one grid step below threshold degrade in well under half of runs, while those several steps below typically fail within weeks.

4. Validation Against Field Data

The run-level agreement reported above is internal to the simulation. Our first external test used fourteen update-feed outages in Hosaka networked environments recorded between 2047 and 2053 (Hosaka Cognitive Systems Division, 2054). In each, the feed was lost long enough for degradation to be logged. Outages that ended without logged degradation were not recorded in the summaries, so the test concerns timing among failures only and says nothing about how often the model would wrongly predict failure. Applying the disconnected model to each environment's recorded density and complexity, predicted and observed log times to degradation correlated at r = 0.71 (95% CI 0.29–0.90, p = .004). The interval is wide, as fourteen episodes allow, but the association is clearly positive.

A second test concerns the Aleph itself. The Tessier-Ashpool ledger implies a period of disconnected operation that we estimate at between 20 and 30 months, with no degradation event recorded. From the archival figures we place the unit's density at 1.3 to 1.9 times the model threshold for its complexity. For a unit in that range the fitted parameters give a probability of degradation within 30 months of about 7% at mid-range density (1.6 times threshold), and 5–12% across the full range; the shorter 20-month estimate would lower these figures. The 30-month figure extrapolates the fitted hazard beyond the 18-month simulated window and should be read with corresponding caution. The absence of a recorded failure is compatible with this prediction, though the ledger was not kept for engineering purposes and could have omitted a minor event.

5. Failure Modes

Under-provisioning is the most direct failure. Below threshold, degradation arrives within months, and the Weibull shape parameter above one indicates a hazard that rises with time. The simulation scores only the first degradation event, so whether an under-built unit can recover was not modelled.

Divergence from the live matrix is a second, slower failure that the stability model does not capture. A disconnected approximation is frozen at its last synchronisation, while the post-merger matrix keeps changing, not least through the autonomous fragments that some communities call loa, whose behaviour, as Okonkwo and Kessack showed, varies along recurring lines that the vernacular names partly track. A unit that is internally stable can still become an increasingly poor map of the network it copies. We have no field measure of this drift and leave it unquantified.

Resident constructs raise a related problem. Kessack and Nakada-Ross's repeated-activation testing of a read-only construct found highly repeatable positions within its recorded domain and far weaker agreement across activations at prompts outside it. Inside a disconnected unit, every construct operates without outside update, so stability of the environment does not guarantee coherent behaviour from its inhabitants when they meet novel conditions.

Security is where disconnection looks most attractive. An environment with no network path cannot be reached by an icebreaker, so containment no longer depends on ICE. The Turing Registry has discussed disconnection as a containment measure for autonomous systems (Lindqvist-Amadi, 2054). Our results qualify that appeal. The attack surface moves from the network to physical custody of the unit and its power supply, and the superlinear cost means rich contained environments are expensive to build. Offline construct hosting on deck-side storage (Ferrand, 2046) sits at the opposite, minimal end of the same curve.

A final open question is whether the exponent reflects a fundamental property of disconnected state maintenance or a feature of current biosoft. Our model fixes the Maas substrate generation; a different substrate could shift the exponent. We therefore claim only that, under stated substrate assumptions, the cost of richness is superlinear.

6. Conclusion

Disconnected environments pay for their independence in substrate. Under our model, required density grows with complexity at an exponent of 1.37, against a near-linear 1.02 for networked environments, and at the Aleph's scale the resulting substrate and power burden is nearly forty times that of a networked equivalent. Validation against Hosaka outage records and the Aleph's own archival history supports the model within the limits of small and incomplete field data.

These results are consistent with the Aleph's standing as an outlier. It is the richest disconnected environment on record, and the ledger places it, at least for part of its history, among the holdings of Tessier-Ashpool, an owner with resources few other actors command and so able to meet a cost that rises faster than richness itself. The simulation cannot test that attribution. The Aleph is thus evidence of what extreme resources can buy, and the superlinear cost predicts that devices without such backing will stay minimal. We propose, as a hypothesis for any successors, that designers without comparable means would settle on narrow environments, particularly for containment. Testing that hypothesis requires a second documented device, which the record does not yet provide.