Plain-language reference for the methodology. Read this first if the grid is confusing.
What you're actually looking at
A labour-market projection 2025–2030 tested under 5 different framings of the same question. Each framing was run 1,000 times. The default Aggregate view shows only outcomes that survive across all 5 framings — the project's most stress-tested output. This is a thinking instrument, not a forecast.
How to read the grid
Each row is an outcome. Each column is a six-month period — H1 2025, H2 2025, H1 2026 (the orange "NOW" line), H2 2026, and so on out to H2 2030. The 5-year-resolution long-horizon Civilisation grids use 2026–2060 instead. Filled cells mean the outcome shows up by that period in the runs where it fires at all. Empty means it hasn't arrived yet.
Row opacity does the heavy lifting: a row that fires in 950/1000 runs looks fully solid; 500/1000 looks distinctly translucent; 250/1000 looks ghostly. The fill colour just marks presence; the opacity tells you how often the outcome fired at all. The earlier Fast/Baseline/Slow track shading was dropped — it carried little signal once we moved from one-axis pace projections to nine-axis verticals where most outcomes are driven by domain-specific axes, not by AI pace alone.
What "470/1000" means
For each projection labelled "1000 runs", the number on the right tells you how many of the 1,000 simulated worlds produced that outcome. 470/1000 means 470 of 1,000 ran into it. The smaller range below (e.g. [44–50%]) is the 95% confidence interval — given a sample of 1,000, the underlying probability is most likely somewhere in that band. Outcomes in 850+ of 1,000 runs are treated as near-certain; 55–85% as high-probability; 20–55% as context-dependent.
Where the 1,000 runs come from
Each run picks a combination of nine variables — AI capability pace, robotics pace, macro environment, regulatory tone, sector concentration, energy supply, capital markets, sectoral structure, and decomposed China-parity status — plus a possible AI-incident profile. There are roughly 20,000 unique combinations; the model samples 1,000 strategically: 700 random (uniform across each variable), 200 corner-stress (extreme combinations like recession + tight capital + spiked energy), and 100 plausibility-weighted (April 2026 priors). For each run, a calibrated scoring function computes which outcomes follow.
Why it's called "cascade"
Each year's projection becomes the fixed ground for the next year's projection. You're not predicting 2028 directly — you're asking: given that 2026 played out this way, and 2027 followed from it, what does 2028 look like? This builds discipline. You can't retrofit later years to make a story work, because earlier years have already been locked. In v4 the cascade is also confidence-weighted — uncertain year-1 outcomes don't pretend to be certain when they roll into year 2.
What the model can't see
Each run produces two honesty checks. The named uncertainty is the single thing most likely to break that run's projection — collected across 1,000 runs, it becomes a register of recurring fragility points. The missed loop is a category of consequence the model's variables architecturally cannot surface — corporate tax base erosion, the informal care economy, the lived texture of middle-income displacement, things like that. Pooled, missed loops become the project's blind-spot register. The model is a frame; the missed-loop register is the model admitting which other frames it isn't.
Axes shape what's findable
Every projection in this site uses a chosen set of axes (typically 8–10) to vary across simulated worlds. That choice determines what the model can surface — and what it can't. Outcomes driven by dynamics outside the axis design simply don't appear, no matter how many runs we do. The missed-loop register partially addresses this by forcing each run to name one such blind spot, but the choice of axes is itself a judgment call that shapes every output. Treat the convergence map as "what's structurally robust under this framing" — not as a complete picture of what's coming.
What the convergence numbers mean (and don't)
"470/1000" is honest about its sample size, but it is not the same as "47% probability in the world." It means: under the chosen distribution of scenarios, with the chosen scoring function, 470 of 1,000 simulated worlds produced this outcome. The numbers are internally consistent, not independently validated. There is no fitted statistical model behind them — the scoring function is calibrated on judgment plus documented April–June 2026 anchors. Treat high-convergence outcomes (say 800+ of 1,000) as robust under the framing; treat low-convergence outcomes as conditional on specific combinations. The independence check (avg Jaccard ≈ 0.6) suggests runs vary meaningfully under identical parameters — so convergence reflects something real about the model, not just generator coherence. Whether it reflects something real about the world is a different and harder question.
Why these axes (and not others)
The axes determine what the model can find. With nine axes per vertical (plus decomposed sub-fields for China parity and AI incident), the parameter space is wide enough to surface interesting combinations — but every axis is itself a judgment about what dimensions matter. Different axes would produce different convergence maps over the same underlying reality. The missed-loop register is the partial corrective: each run explicitly names a category of consequence the model architecturally cannot surface, and those get pooled into a blind-spot register. Read the missed loops as the model admitting which other framings it isn't. Read the convergence map as what this framing produces. The framing is the unit of analysis, not the truth.