Base Rates - the reference classes underneath the estimates#
Every probability in Part V starts from a reference class and then gets adjusted. These pages state the classes and the adjustments, so both can be attacked separately.
The discipline that matters: name the class before looking at the case. A reference class picked after forming a view is a rationalization with a citation attached. The companion rule, enforced in Scoring: the class is fixed at authoring - swapping to a friendlier class after the fact is a logged revision, never an interpretation.
Six classes carry the corpus's priors. They are numbered globally because other pages cite them by number:
| # | Class | Page | What it anchors |
|---|---|---|---|
| 1 | General-purpose technology diffusion | Cycles | The 2029–33 TFP window; Part V row 2 |
| 2 | Infrastructure capex booms | Cycles | The correction-not-death shape; Capital |
| 3 | Regulatory response cycles | Social response | Game 2 event-driven coordination |
| 4 | AI winters - failure archaeology | Winters | Why this cycle rhymes with capex booms, not 1974 |
| 5 | Labor displacement episodes | Social response | Concentration-not-aggregate; Game 4 |
| 6 | Demographics | Social response | The counterweight; displacement estimates as upper bounds |
The classes are not equally strong, and the corpus leans on them unequally. Classes 2 and 3 have the most instances and the most consistent structure; class 4 has only two full instances and is the most abused analogy in the field, which is why it gets the longest treatment; class 1's headline trend (each cycle faster) rests on five points and two offsetting adjustments that roughly cancel - a weak reason to believe the base case, stated as such where it is used.
The failure mode of the whole apparatus: reference classes discipline priors, they do not generate truth. A genuinely unprecedented mechanism - recursive capability feedback is the candidate (Uncertainty 1) - belongs to no class, and forcing it into one manufactures false confidence in both directions. Where the corpus argues from no-precedent, it says so instead of borrowing a class that does not fit.
How classes compose in practice#
Most Part V rows are conjunctions of classes, not pure draws from one. Row 6 (correction) is class 2's boom pattern × class 4's "not a full winter" filter. Row 4 (agreement) is class 3's event-driven window × the privacy-class exception if harms stay diffuse. Row 2 (TFP) is class 1's diffusion lag × class 6's demographic counterweight on labor politics. When a re-score moves a row, name which class's adjustment changed - otherwise the ledger cannot be audited.
Score classes before cases. If the GPT-diffusion analogy (class 1) is wrong, half the timeline moves; if only software's specifics are wrong, one domain page moves. That is why dependencies routes misses to parent claims, and why this hub exists separately from domain detail.
Related: Part V - Probabilities · Steelman · Scoring