Methodology

The scenario was written from year N to year N+1 without a predetermined ending. Each period was drafted using only what was plausible given the previous one, then revised for internal consistency rather than for narrative satisfaction.

This guards against hindsight bias. Scenarios written backwards from a conclusion tend to assemble a tidy chain of causes toward an outcome the author already preferred. Writing forwards produces messier, less quotable, more useful results. It also leaves room for developments that do not matter in the end.

Vague claims cannot be tested and are useless for planning. Wherever possible the scenario names dates, orders of magnitude for datacentre load, likely employment effects, and the way a policy takes effect.

Granularity is not confidence. A specific claim is easier to argue with, and being argued with is the point. Where a number is a rough order of magnitude rather than a projection, the text says so.

Four constraints are tracked separately throughout, because they move at very different speeds and conflating them produces most bad AI-policy analysis.

Automation of execution and automation of direction are tracked as separate variables. Execution covers drafting, coding, triage, translation, and routine analysis. Direction covers deciding what is worth doing, what risk is acceptable, and whose complaint counts.

Most employment analysis collapses these. Doing so overstates near-term displacement in judgement-heavy roles and understates the erosion of the entry-level ladder, where routine execution work was the training mechanism for future direction-setters.

Seven parameters drive the scenario's branching, and are exposed directly in the sandbox. They were chosen because they are consequential, partly independent of each other, and at least partially subject to regional decisions.

Automation intensity and reskilling throughput matter most together. Small changes in either produce the largest differences in social outcomes. Capability pace matters less on its own. A fast frontier with strong transition capacity produces better outcomes than a slow frontier without it.

Throughout, the project asks a narrow question: who can actually stop a system, and who can compel disclosure when it fails? Ownership of buildings, hardware, and even models is treated as secondary to that.

This framing is why sovereign compute programmes are assessed here by their audit, portability, and continuity provisions rather than by their announced capacity.

This is a scenario, not a forecast, and it carries no probability. It should not be cited as an estimate of what will happen.

It is one path among many. It assumes no great-power conflict, no pandemic-scale disruption, and continuity of leading-edge fabrication in East Asia. Any of those assumptions failing would dominate everything modelled here.

Employment figures referenced are exposure estimates, not displacement projections. Exposure means a task profile that overlaps with current capability; it does not mean a job will be eliminated.

Personas are composites written to make system dynamics legible. They are not accounts of real individuals, and no organisation described is a real one.

The sandbox produces written qualitative states from parameter bands. It contains no simulation, no fitted model, and no quantitative output, deliberately.