# Part IV - What Actually Happens ← [Index](../README.md) · [Part III](../03-domains/) --- Three periods, each with a different rate-limiting factor. Each page synthesizes the substrate, games, domains, macro channel, indicators, and uncertainties that bear on that window - this part is a *schedule*, not a separate argument. | Period | Rate-limited by | The question it answers | |---|---|---| | [2026–2028](2026-2028.md) | Capital and reliability | Does the revenue justify the capex? | | [2028–2032](2028-2032.md) | Institutions and organizations | How fast can the world absorb it? | | [2032–2040](2032-2040.md) | Physics and atoms | Does the growth rate actually change? | | [Scenarios](scenarios.md) | Joint resolution of the chain | Best, base, worst, and the named branches between | ## The shape of it The story moves outward through the stack. First the technology has to work reliably enough to sell (2026–28). Then organizations have to restructure around it, which is slower than the technology and always has been (2028–32). Then it either crosses into the physical world or it doesn't, and that determines whether this decade is remembered as an important one or a decisive one (2032–40). Note that the *capability* frontier is not the rate-limiting factor in any of the three periods. That is the central claim of the timeline, and the main thing that would falsify it is [recursive research acceleration](../06-uncertainties/recursive-self-improvement.md) - subject to the three governors (verification, physical supply chain, financial) named there. ## Why the rate-limiters arrive in this order The sequence is not arbitrary; each constraint can only bind after the previous one has released. Capital and reliability come first because nothing else is even testable until the technology works well enough to sell and someone has paid for the compute to run it - a diffusion problem cannot exist before there is something to diffuse at a price someone will bear. Institutions come second because organizational absorption only becomes the bottleneck once the commercial product exists; before that, slow institutions are indistinguishable from absent demand. Atoms come last because the physical economy is downstream of both: robots need the cheap cognition the first period buys and the deployment infrastructure - insurance lines, liability norms, trained integrators - the second period builds. Reverse any two stages and the story breaks: institutions cannot restructure around an unreliable product, and physical capital cannot be justified before the institutional demand signal exists. This ordering is also why the three pages weight evidence differently. The first period is scored mostly on *financial* observables (revenue vs. capex, financing mix), the second mostly on *statistical and legal* ones (productivity prints, insurance lines, posting ratios), the third mostly on *physical* ones (delivered $/hour, autonomy ratios, interconnection). A reader who finds a 2027 claim being defended with a productivity statistic, or a 2035 claim defended with a demo video, has found a category error worth flagging. **Failure mode of the periodization itself.** The clean three-stage structure is a modelling convenience, and its most likely failure is *overlap*, not inversion: institutional adaptation is already visibly underway during the capex test, and structured-environment robotics crosses commercial thresholds inside the middle period. If by 2029 the three constraints are all binding simultaneously in different sectors - capital in frontier training, institutions in professional services, atoms in logistics - the periods have not failed as analysis, but the *dates* attached to each page carry less information than the constraint logic does. Treat the constraint identification as the durable claim and the calendar boundaries as the soft one. ## What each period now carries Rounds 1–7 deposited material that these pages have to hold, not just link to: | Thread | Lands hardest in | |---|---| | Capex test, financing mix, correction form | [2026–2028](2026-2028.md) | | Apprenticeship gap as measurable fact; junior-hiring test | [2026–2028](2026-2028.md) → [2028–2032](2028-2032.md) | | J-curve, incident → regulation, insurance as deployment frontier | [2028–2032](2028-2032.md) | | Two-economy price split as political fact | [2028–2032](2028-2032.md) | | Regional binding constraints (five regions) | All three; decisive splits by 2032–40 | | Demography revising aggregate labor down, composition standing | [2028–2032](2028-2032.md) and [2032–2040](2032-2040.md) | | Robotics fork; science automation; macro feedback loops | [2032–2040](2032-2040.md) | | Taiwan as the underlying modelling assumption | All three; named explicitly | Cross-check claims against [Part VII](../07-indicators/) (observables), [Part V](../05-probabilities/) (stated probabilities), and [Part VI](../06-uncertainties/) (where the schedule breaks). For best / worst / other joint worlds in one place, see [scenarios](scenarios.md) - stress tests of the chain, not a replacement for it. ## Why the schedule stops at 2040 The horizon is a claim, not a convenience. Past the robotics fork, the branches diverge too far for a single narrative to mean anything - a 2045 in the high-growth branch and a 2045 in the disappointment branch share almost no facts that matter, so any page covering both would be written at a level of abstraction that cannot be wrong, and a page that cannot be wrong is excluded by the [file conventions](../RESEARCH.md). The honest form for post-2040 content is conditional (already present inside [2032–2040](2032-2040.md) as branch descriptions), not calendrical. If a future round finds the 2032–40 fork resolving early and cleanly, *that* is the trigger for drafting a fourth period - the schedule extends when a branch is chosen, not when time passes. The same logic runs backward as a reading instruction: the further a page's dates sit from the present, the more its value lives in the *conditionals* (what determines which branch) and the less in the *dates*. The 2026–28 page should be read as a forecast; the 2032–40 page as a decision tree with probabilities on the branches. Scoring should respect this - marking the third period "wrong" because an event slipped two years would miss what the page actually claims, and [scoring](../08-method/scoring.md) rule 1's mechanism-over-outcome principle exists for exactly this case. ### Evidence weight by period | Period | Primary evidence family | |---|---| | 2026–28 | Financial: revenue vs capex, financing mix, junior hiring | | 2028–32 | Statistical + legal: TFP, insurance lines, price split, first incident | | 2032–40 | Physical: delivered $/hour, autonomy ratios, wafer share outside TW | A claim in the wrong evidence family is a category error - e.g. defending a 2027 forecast with a productivity print, or a 2035 growth claim with a demo video. --- **Next:** [2026–2028](2026-2028.md) → … → [2032–2040](2032-2040.md) → [Scenarios](scenarios.md) → [Probabilities](../05-probabilities/)