Musk's Economic Model: Reconstruction, Critique, and a Scenario Framework

On this document. This is the full analysis, in the form it was assembled — 23,000 words, twelve sections, four appendices. The short version, and the one most people want, is Seven Futures for the AI Economy: a distillation of what follows that stands on its own.

On the verification gate. The document was drafted with its own source-checking protocol (Appendix C) unexecuted, and gated its circulation on executing it. That gate has now been run: all eleven post-cutoff claims were put to five independent research passes against primary sources. The results are recorded in Appendix B, and the corrections are applied inline throughout — marked [corrected] where a figure was wrong as stated and [updated] where it was merely stale. Six claims failed and four moved. No result changed sign, and no mechanism was overturned; two findings strengthened the argument. §0.1’s evidence tags remain load-bearing: [O] observed, [E] estimate, [I] inference, [A] assumption, [F] forecast. Read the probabilities as stated priors rather than measurements, and §12 for what the framework still says it gets wrong.

v3 delta (2026-08-13). Three changes beyond v2’s inline corrections, each entering through the framework’s own machinery rather than around it. (1) E7 propagated. v2 corrected the METR horizon evidence in prose but left §6’s probabilities unconditioned on it — “corrections applied inline” was true of the text and false of the numbers. v3 re-anchors Layer 1 (0.20/0.55/0.25 → 0.18/0.47/0.35; LR log at §6.2) and recomputes every terminal mechanically. (2) The tree gains a terminal. Cross-bloc gate divergence — carried since D.6.4 as “substantially the residual’s content” while §5.3 called it “the likely path” — is promoted to S6, split world (≈ 0.08), via divergence conditionals on the two leaves where gates move under live capability. (S5 stays retired; the v1 label is not resurrected.) The essay mapping shifts accordingly: Split World now names S6, not the residual. (3) The I-gate becomes directional. External feedback on the essay, adjudicated this revision, supplied the framework’s missing mechanism — institutions as accelerated consumers of AI, not only frictions on its deployment. v3 splits the gate into I-out (permission to automate services — the existing gate, unchanged semantics) and I-in (permission to run institutional inference on persons), adds the sovereign carve-out to §3.2’s min-function, amends R6/R12/R15 and M14 accordingly, installs overlay O3 — computational governance with a promotion test (§6.4), adds dashboard row ι-4 and falsifiers M17–M18, and re-derives A.3’s wedge prize (now ~5–12% of GDP, lower half better supported — closing §12.3’s owed item). Netting note, logged per §10.2: E7, R15’s ratchet audit, and the feedback’s control asymmetry all push probability toward S4a by three different arguments that may be one intuition wearing three justifications; only the τ-1 update is applied to the numbers — the feedback’s push enters structurally (O3, ι-4, M17–M18) and its probability expression waits at O3’s promotion test.

Contents

§0 — Front Matter and Conventions

Source interview: “The full-length interview with Elon Musk” — The Economist, Zanny Minton Beddoes, Tesla Gigafactory Texas (2026).

Document status: v3, stand-alone analysis artefact. Supersedes the v1 and v2 analyses of the same interview; the v3 delta is summarized in the front-matter note above and logged transaction-by-transaction in §0.1. Reconstructs the economic model Musk articulates, tests it against data and theory (§§1–2), builds the framework the divergences demand (§§3–6), derives implications (§§7–8), synthesizes (§9), and installs its own error-correction and retirement machinery (§§10–12). Appendices A–D are attached: A (classification memo) is delivered and its share vector adopted in §3.2; B (evidence ledger) is running; D (jurisdictional decompositions) is delivered at sketch-grade. Appendix C has been executed (2026-08-13); the gate is open, six items were corrected and four updated, and the sign-off is at C.3.

0.1 Evidence conventions

Every load-bearing claim carries a tag:

Tag Meaning Discipline
[O] observed — primary-source-verifiable measurement Appendix B ledger
[E] estimate — measured with material uncertainty or contested identification caveats stated inline
[I] inference — derived within the framework from tagged inputs derivation cited
[A] assumption — chosen, not derived; stated to be attacked §10 re-conditions
[F] forecast — falsifiable, dated where possible §11 registers the kill condition

Complexity budget: one parameter in, one out. Transaction ledger: §3 retired the scalar σ (now emergent; R1’s worked propagation keeps an illustrative aggregate σ̄ ≈ 0.5 for arithmetic only — a display convenience, not a reinstated parameter); §5 retired scalars μ, λ, δ (replaced by the layer vector, Λ, and the incidence identity); §6 retired the v1 scenario grid and S5-as-scenario (replaced by the tree and the overlay promotion rule); §10 retired fixed-percentage-point update rules (replaced by LR bands); §7 logged the one net addition (row G-2). v3 transactions: terminal S6 added (a re-partition — the divergence content D.6.4 already carried, promoted out of the residual; no new causal axis); overlay O3 added (overlay count 2→3 — v3’s one structural addition, priced against the I-gate split doing double duty as its mechanism); dashboard row ι-4 added bundled with ι-2/Λ-B (correlated politics — no new effective evidence unit); falsifiers M17–M18 (falsifiers are free by house rule: every mechanism ships one). The §0 registries below are navigation, not parameters.

0.2 Standing corrections (carried from review; applied throughout)

  1. Canaries: the youth-displacement estimate was cited as ~13–16% [E], never as settled. [updated] The current figure is ~19% (revision of 12 August 2026, ADP data through June 2026); 13% and 16% are successive data vintages of one specification, not competing specifications — 13% from the August 2025 version (data through July 2025), 16% from the 13 November 2025 revision (through September 2025). Two further corrections matter more than the number. The authors now explicitly describe the result as “early, descriptive indicators… rather than causal estimates,” and with firm-time controls the divergence becomes significant only from 2024. And the §174 amortization objection is addressed inside the paper (footnote plus robustness checks excluding technology firms and computer occupations) rather than being an established external rebuttal — no peer-reviewed methodological critique exists. Single-study dependence (§12.2) is now partly resolved in both directions: LinkedIn-based work replicates the pattern in the US and UK, while Johnston & Makridis (QCEW) find employment increases in more-exposed cells and Humlum & Vestergaard find near-zero displacement in Danish registers.
  2. “AI investment ~5% of GDP” is treated as an upper bound [E]: hyperscaler capex arithmetic supports ~2–2.5%; the broader complex plausibly 3–4%.
  3. All post-cutoff sources (2026-dated Fed/BIS/headcount data) have now been verified against primary sources — Appendix C executed, results in B.2. The interview itself remains the one unverified input: no participant in the check was able to close it against a recording or transcript, so §1’s reconstruction still rests on the quotes as given. This matters less than it appears — §§3–12 analyze the position, not the person, and the position is held far beyond one speaker.
  4. All dated windows are [F], illustrative unless a derivation is attached.

0.3 The named objects (reader’s map)

Classes D / P / I / X (constraints, not sectors — §3.1) · Gates: P-gate (embodiment, opened by κ), I-gate (permission, opened by insurance economics + law — directional since v3: I-out = permission to automate services, the original gate; I-in = permission to run institutional inference on persons; sovereign carve-out where permitter = operator, §3.2); X never opens, D needs none · Clocks: capability (months), buildout (years), institutional (decades), none — I-out moves by statute and insurance (decades, ι-1 the fast hand); I-in moves by legitimacy, scandal, and administrative law (faster, lumpier — ι-4) · Flows F1–F4: licensure flight, liability unbundling, actuation, X-manufacture · Wedge prize: ~5–12% of GDP priced by permission on D-feasible tasks, lower half better supported (§3.4, A.3) · Funnel (§1): exposure → feasibility → viability → adoption → autonomy · Waves: A (power-bill politics, 2026–28), B (cohort politics, 2028–32; ADM-scandal sub-current, §5.3) · Overlays: O1 macro-cyclical, O2 geopolitical, O3 computational governance (v3, §6.4) · Policy vector Λ = {ρ, χ, g, φ, ν} · Incidence channels: π, b, Ω, χ · Dashboard: 21 rows, ~12 evidence units (§10) · Falsifiers: K1–K3 framework kill conditions, M1–M18 module falsifiers (§11).

Terminals (masses as re-conditioned in v3, 2026-08-13 — LR log at §6.2; §10 re-conditions quarterly): S1a plateau 0.09 · S1b valuation bust 0.09 · S2 Baumol-institutional 0.28 (modal) · S3 compressed industrial revolution 0.13 · S4a recursive-but-bottlenecked 0.19 · S4b unbound takeoff 0.09 · S6 split world 0.08 · residual 0.07.

Results registry:

R One line
R1 Deceleration paradox: S2 growth is front-loaded; a mid-2030s slowdown confirms the capability story under complementarity
R2 Terminal scarcity theorem: every trajectory terminates in X + residual-I; money collapses onto claims over position and permission
R3 Rent sink: fixed-supply X absorbs a share of all productivity gains; land share of income rises in every growth scenario
R4 Baumol policy trap: one monetary instrument vs class-divergent inflation whose inflating classes are policy-inert in trend (X-asset price levels are rate-sensitive — the level/trend distinction is stated in §4.1)
R5 Capability straddle: the credit system loses in both capability tails; safe only in the orderly middle
R6 Fiscal scissors: the labor-tax base erodes exactly when transition-funding need peaks; one counter-current logged (AI tax enforcement partially closes Blade 1 — §4.2)
R7 Stack recapitulation: scarcity migration operates inside the AI stack; rents migrate from models to compute, indemnification, distribution
R8 Clock-speed distribution: incidence defaults to ownership because Ω is the fastest channel and χ the slowest
R9 Default posture: fiscal constraint + security competition jointly manufacture accelerate-capability / restrict-deployment — the S4a machine
R10 Reversal symmetry: each class’s dominant asset risk is the reversal of its own scarcity mechanism; the safest asset is the terminal tax base
R11 Corridor paradox: the cross-scenario-robust asset self-erodes via capitalization, regulation, and supply response; its residue is X
R12 Absorption machine: the modal labor outcome is absorption into protected-sector inefficiency; cost disease is the de facto welfare state — predicted, not yet observed (L-2), and contingent on the state not self-automating its administrative layer (§8.1)
R13 Unfunded tuition: junior work bundled spot output with option manufacture; AI unbundles it and nothing re-funds the option
R14 Mistimed dividend: the demographic cushion and the automation shock are geographically anti-correlated
R15 Deployment ratchet: unmodeled feedbacks found during construction suppressed deployment, never capability — the residual bias has a sign, but it holds for power-dispersing deployment only; two logged counter-classes: actuarial import (exogenous, D.6.1) and institutional self-strengthening deployment (endogenous, §9.3)

Reading paths. Full argument: in order. Analysts: §3 → §6 → §10. Capital: §4 → §7. Policy: §5 → §8. Skeptics: §11 first, then §3.

§1 — Musk’s Model Reconstructed

Status: the object under analysis. Consumes the interview; supplies the P-chain that §2 tests, §§3–8 dismantle and rebuild, and §9.4 scores. Introduces the document’s first named instrument — the funnel. Results numbering begins at §3; this section asserts nothing of its own beyond the reconstruction.

1.1 The chain

Stripped of rhetoric, the interview asserts eight propositions in sequence:

# Proposition His timeline
P1 AI capability exceeds the sum of human intelligence ~2031
P2 Digital task automation reaches “Stockfish level” in all cognitive work — “no way to compete”; already “beats 90% of professional software engineers” “very soon”
P3 Humanoid robots extend automation to physical tasks (“end effectors”; “lots of bots”) before 2036
P4 Economy = digital intelligence × actuators → output “quasi-infinite” ~2036
P5 Binding constraints are electricity and chips — not labor, capital, or institutions now
P6 Wages disappear as an income channel → “universal high income”; “Treasury should simply issue people checks” transition
P7 Output growth > money growth → “deflation will be the issue, not inflation” transition
P8 “Money won’t matter in 2036”; taxation irrelevant 2036

Load-bearing structure: P4 requires P2 and P3 and the negation of every constraint other than P5; P6–P8 require P4. Supporting claims: China already out-generates US+EU+India combined, heading to ~4× US production; outside China the constraint is power (hence orbital data centers), inside China it is chips, though lithography is “closer than most people realize”; “work is going to be optional” (the gardening analogy); p(doom) 10–20%, “even if there was a stop button, we probably shouldn’t press it,” “enjoy the ride”; “I’m very good at predicting the future… maybe not exactly in the timing that I thought.”

1.2 The funnel

The instrument that organizes everything that follows. A task passes through five stages between “AI could touch this” and “no human is involved”:

  1. Exposure — AI could in principle affect the task.
  2. Technical feasibility — AI performs it under benchmark conditions.
  3. Economic viability — cheaper after integration, oversight, error, and liability costs.
  4. Adoption — organizations actually run it in production.
  5. Autonomy — no material human oversight or sign-off remains.

Musk’s evidence lives at stages 1–2. His conclusions live at stage 5. The economy — output, prices, wages, employment — lives at stages 3–4.

Claim Evidence stage Claim stage The gap
P1 1–2 (capability curves) 2 none — a legitimate capability claim; the tree prices it (§6), the dashboard monitors it (τ-rows)
P2 2 (benchmarks) 5 (“no way to compete”) the entire funnel
P3 2 (demos) 4 (deployment at scale) plus a production-throughput problem that isn’t funnel-shaped at all — it’s κ
P4 — 5, across all tasks the funnel, economy-wide
P5 3–4 (a genuine viability/adoption observation) 3–4 none — the one funnel-aware claim in the chain; credited in §2.1 and absorbed as κ in §3
P6–P8 — 5, macro stage-5 endpoints asserted with no stages in between — and the funnel is also a time structure, so skipping it deletes the transition

The Stockfish tell. Chess is the one domain where the funnel has zero length by construction: no integration cost, no liability, instant verification, no adoption friction, total autonomy the moment feasibility arrives. “Stockfish level” is exactly true — in domains shaped like chess. The analogy doesn’t support the conclusion; it selects the one world in which the conclusion is trivially true and projects it onto a world of sign-offs, permits, and insurance. The rebuttal at stage 3 is empirical and cited in §2.3.

1.3 Internal tensions — all three are funnel-collapses

  1. Fatalist on doom, planner on upside. He holds p(doom) at 10–20% with resignation (“no stop button”) while holding the economics with near-certainty. Both follow from treating capability as the only variable: if stages 3–4 don’t exist, there are no independent actors between the model and the outcome — nobody to press a button, and nobody to slow the abundance. The framework’s core claim is the opposite: stages 3–4 are crowded with actors — insurers, guilds, rate commissions, central banks, legislatures — and §§5–8 are about nothing else.
  2. “The constraint is electricity” vs. “money won’t matter.” P5 concedes funnel friction for the physical layer he can see from a factory floor; P8 requires funnel friction to vanish for the institutional layers he can’t. And the direct contradiction stands: a binding physical constraint is precisely the condition under which prices do their work.
  3. Interest payments exceed defense spending; deflation is coming — stated separately, never connected. Fisher debt-deflation is a transition phenomenon; a stage-5 endpoint claim with no intervening stages has no transition in which anything could connect. §4 restores the stages and finds the corridor (S3) where the connection detonates.

Steelman, preserved for scoring: “money won’t matter for subsistence” — the defensible narrow reading, which the framework partially endorses (§3.2/R2: subsistence = D∪P; everything above it lives in X and I, where money intensifies). Scored with the rest of the chain at §9.4.

§2 — Where the Data and the Theory Diverge

Status: empirical motivation and routing table. In v1 this section was the analysis; in v2 each divergence names the module built to handle it. Carried from v1 with the §0.2 standing corrections applied and one formal retirement (the v1 rates test). Consumes §1; feeds everything.

2.1 What checks out — credit before critique

  • China electricity [O, verified]: ~10,583 TWh (2025) vs US + EU + India ≈ 9,398 combined (US 4,520 · EU-27 2,797 · India 2,082, Ember/OWID 2026 vintage); the 4× projection is roughly the population ratio, directionally plausible. One distinction to preserve: 10,583 TWh is generation in the Ember dataset; Ember’s report text gives 10,573 TWh of demand. The margin is +1,180 TWh either way.
  • Energy as the binding constraint outside China [O]: interconnection queues, 3–5-year transformer leads, turbine backlogs — not GPU supply.
  • Early labor damage [E, §0.2 caveats]: [updated] ~19% relative employment shortfall (was ~13–16% at drafting), ages 22–25, most-exposed occupations, concentrated where AI automates rather than augments; adjustment through hiring, not wages; older workers stable. The hiring channel and the older-worker stability both verify; the magnitude has grown with each vintage. Mechanically consistent with what §8 will derive (R13), and still the framework’s single most load-bearing contested datapoint — now explicitly descriptive rather than causal, per its own authors.
  • Investment macro-significance [E]: [corrected] the St. Louis Fed contributions are separate quarterly figures, not an H1 aggregate: 1.30pp in Q1 2025, 1.16pp in Q2, 0.48pp in Q3, and 0.97pp across the first three quarters combined (0.90pp excluding data centres). The drafting text read them as two estimates of one half-year, which they are not. Contributions track investment growth, not levels; the “~5% of GDP” figure is an upper bound per §0.2. The four categories are information-processing equipment, software, R&D, and data-centre construction — proxies for AI investment, not a measured AI-only series.
  • P5 itself: the chain’s one stage-3/4 observation. Musk is a reliable observer of physical constraints — the verdict v1 reached and §9.4 retains. §3 absorbs the observation as κ.

2.2 Divergence 1 — the demand side is booming; the supply side hasn’t shown up

AI appears in GDP as capex (demand), not TFP (supply): three years post-ChatGPT, no aggregate TFP inflection [O]. Three readings — productivity J-curve (payoff 2028–35), task-model arithmetic (~0.5–0.7% TFP per decade), genuine discontinuity — were rivals in v1. In v2 they are parameter settings, not worldviews: the J-curve is ι_int (§3.3’s intangibles shadow, §7.3), the task arithmetic is a row of the §3.2 worked table (the Acemoglu corner, g_D ≈ 2%), and the discontinuity is Layer 1 of the tree (recursive = 0.35 [A], as re-anchored in v3).

Formal retirement. v1 adjudicated with “real long rates are the cleanest test — the bond market voting against P4.” That argument is retired here. The contamination lemma (§4.3) shows r is non-monotonic in AI outcomes: concentration deepens the X-savings-sink, so the rentier terminal (S4a) plausibly prints low rates. Low rates cannot vote against takeoff; they cannot distinguish “no takeoff” from “takeoff whose gains pool.” Replacement instruments: the confirmation basket and the rates×land cell (F-3). Routing: →§3.2, §4.3, §6.*

2.3 Divergence 2 — tasks ≠ jobs ≠ economy (the funnel, applied)

The “beats 90% of software engineers” claim is stage-2 evidence for a stage-5 conclusion. The stage-3 observation exists [O]: the METR RCT found experienced developers 19% slower with AI tools on their own repositories — while believing they were faster. The effect size verifies exactly (16 developers, 246 randomised issues, 95% CI +2% to +39%; forecast 24% faster beforehand, still believed 20% faster afterwards). [updated] Its scope does not extend to current tools: this measures early-2025 assistants, and METR’s February 2026 follow-up — which hints at a speedup — is one METR itself calls an unreliable signal, since developers increasingly refused to work without AI and the pay rate fell from $150 to $50 an hour. Cite it as a stage-3 existence proof for its vintage, not as a current productivity estimate. The funnel’s middle stages are not friction on the truth; they are where the truth lives.

The deeper divergence is Baumol: when tasks complement each other, growth is set by the bottleneck, not the frontier. v1 hung this entire question on one scalar (σ ≷ 1) and a historical prior (every automation wave migrated scarcity rather than eliminating it — agriculture and manufacturing had their “Musk moments,” and the residual became the economy). v2 replaces the scalar with the class engine and sharpens the question: “does scarcity migrate again?” becomes “does the I-gate hold?” — a political question, not a technical one, and an observable one (ι-1, ι-2). The historical prior stops being destiny and becomes calibration. Routing: →§3, and K1/K2 (§11) are its kill conditions.

2.4 Divergence 3 — the robot leg is far behind the timeline

P3 carries half the chain on the weakest data — and the verification pass found the data weaker still. [corrected] Global humanoid production in 2025 was not “tens of thousands.” No primary global production count exists at all: the only independent figure is IDC’s estimate of ~18,000 units shipped, and the only audited producer is AGIBOT at 5,000 cumulative units; Unitree self-reports >6,500 produced against >5,500 delivered. Tesla discloses no 2025 Optimus production count whatsoever — its target fell from 10,000 to ~5,000 to ~2,000 across the year, actual output ran in the hundreds, and Gen-3 mass production slipped to 2026. Against a ~3.5B workforce, the corrected figure is single-digit thousands built, not tens of thousands. Replacing even the labor-substituting slice of a ~$68T capital stock (BEA current-cost net stock of private fixed assets, end-2024) at feasible production growth pushes “physical abundance” well past 2036; dexterous manipulation lacks an internet-scale training corpus — the data must be generated, slowly and expensively. This correction moves in the framework’s favour: the P-gate is more firmly shut than the drafting estimate assumed, and the shipped-versus-produced-versus-ordered-versus-capacity confusion that defeated five independent research passes is itself the κ-2 measurement problem the dashboard was built to police. Framework status: this is the P-gate, S3-conditional (~0.13 plus recursive branches, v3 masses), and the framework’s least-instrumented parameter (§12.1). The indicator is production numbers — Musk’s own factories are the best data source, and his shipped units, not his demos, are the row (κ-2). Routing: →§3.4 (F3), §6.

2.5 Divergence 4 — the monetary claim is one-fifth right

v1’s three objections (relative vs. aggregate prices; Fisher; “Treasury checks” is a distribution rule, not a macro policy) are upgraded to modules:

  • P7 claims A-type deflation; the taxonomy (§4.1) shows R-type is the certainty, C-type the near-term financial danger, and A-type a policy choice under inflation targeting — the consistent version of P7 requires an aggressively activist central bank, which refutes P8 on its own.
  • Fisher risk is non-monotonic and lives specifically in the S3 corridor, amplified by the policy trap (R4).
  • “Treasury issues checks” collides with the fiscal scissors (R6): the labor-tax base erodes exactly when the need peaks, and the check-size fight is the politics (Wave B, §5.3) — v1’s “he treats the transition as an implementation detail; it is the entire problem,” now with the mechanism attached. Routing: →§4, §5.3.

2.6 Divergence 5 — “money won’t matter” fails on first principles, now by derivation

v1 asserted that money coordinates scarcity and something always stays scarce. v2 derives it: the terminal scarcity theorem (R2) — X never opens by construction, I opens only politically, so every trajectory including full takeoff terminates in money as claims over position and permission. The steelman (“won’t matter for subsistence”) survives exactly as far as §3.2 says: D∪P covers subsistence in the open-gate terminals; above subsistence, the price system concentrates. Routing: →§3.2 (R2), §9.4 (scored).

2.7 Divergence 6 — timeline base rates, now structural

The record: directionally strong, chronically time-compressed (FSD “next year” annually from 2015; robotaxis; Mars). v1’s prior — update on his direction, apply a 2–4× multiplier to his dates — is retained, with a v2 upgrade: the multiplier is not a personality quirk; it is the funnel’s length. A stage-1/2 forecaster experiences stages 3–4 as inexplicable delay; the multiplier is short where the world is chess-shaped and long where gates exist — which is why his hardware timelines outperform his institutional ones. Applied to P1, the multiplier lands capability-parity mid-2030s+, consistent with the tree’s recursive prior [A]. The multiplier alone dissolves most of P6–P8 as near-term policy: a 20-year transition is governable; a 5-year one is not. Routing: →§6 (Layer 1), §9.4.

§3 — The Four-Class Engine

Status: core analytical module. Replaces v1 §3.2’s scalar σ and the v2.0 four-class sketch. Evidence tags per §0 conventions: [O] observed · [E] estimate · [I] inference · [A] assumption · [F] forecast. Everything downstream — the scenario tree (§6), the capital and labor implications (§§7–8), the dashboard (§10), and the falsifiers (§11) — references this section.

The single-parameter σ of v1 did three jobs at once: cross-class substitution (can digital cognition replace physical execution?), within-class demand response (does cheaper code mean more code-consumption or less code-expenditure?), and institutional mediation (does the legal system permit substitution at the liability-bearing layer?). These are governed by different mechanisms, move at different speeds, and answer to different actors. This section separates them.

3.1 Classification: classes are constraints, not sectors

The classification rule

Every economic task is scored on four dimensions:

Dimension Low end High end Relieved by
Digitizability requires situated action pure information processing capability (τ) — months
Embodiment intensity no physical capital needed robots, infrastructure, dexterity buildout (κ) — years
Verification/liability intensity errors cheap, output verifiable errors costly, legally attributed institutions (ι, λ) — decades
Positionality value from function value from scarcity itself never

A task’s class is determined by its last unrelieved constraint — the one that still binds after all faster-moving constraints are relieved:

  • D-class: capability is the last constraint. Once τ is adequate, nothing else blocks substitution. Routine code, content, translation, analysis, support.
  • P-class: embodiment is the last constraint. Capability may already suffice in simulation; actuators, energy, and capital vintages bind. Construction, logistics, physical care, manufacturing residual.
  • I-class: institutional permission is the last constraint. The task may be fully digitizable — the signature is what cannot be automated without legal change. Clinical sign-off, fiduciary duty, adjudication, permitting, audit attestation.
  • X-class: scarcity is the value. No constraint-relief is defined. Located land, status, attention, originals, political access.

The ordering matters because it fixes each class’s clock: D moves at capability speed (months), P at buildout speed (years), I at institutional speed (decades), X not at all. This is the formal basis for the four-clocks synthesis (§9).

Worked example of the rule: medical diagnosis is fully digitizable (D-feasible on the first dimension), requires no embodiment, but carries high liability intensity. Its last unrelieved constraint is institutional → I-class. This is why diagnostic capability benchmarks [O] tell you almost nothing about diagnostic employment: the binding constraint was never the analysis; it is the signature.

The I-rent wedge

A refinement the decomposition forces: the I-constraint does not always appear as its own expenditure line. Often it appears as a rent wedge — a gap between a task’s technical cost and its permitted price.

A radiology read is technically D-class work. It is priced at I-class rates because a licensed read is legally required. The difference between the AI-read cost (approaching zero) and the billed price is I-rent: value-added attributable not to the task but to the permission structure around it. [I]

Two consequences:

  1. Decomposition must be done by constraint, not by expenditure line. A sector’s “I-share” is the sum of rent wedges on its D- and P-feasible tasks, plus any residually human institutional work.
  2. The I-class can erode without any sector shrinking. Flow F2 (§3.4) collapses rent wedges in place — prices fall, the sector’s measured size falls, and no separate “institutional sector” ever appears in the statistics. This is what to look for; a naive sectoral analysis would miss it entirely.

Where permission holds, note the complementarity: AI raises I-incumbent productivity (the doctor sees more patients with diagnostic support) without reducing I-employment, because the constraint is the signature, not the throughput. Capability makes the incumbent more valuable, not less — one reason I-protection is economically comfortable for its holders and therefore politically durable. [I]

The I-gate is directional

v2 treated the I-gate as a scalar: permission either flows or it doesn’t. That missed a structural fact the framework’s own appendices carried unassembled (D.4’s fiat-F2; A.6’s government-share caveat): the gate has two directions, and they move under different politics on different clocks.

  • I-out — permission to automate services: replacing licensed humans, collapsing rent wedges, F2 as previously specified. The existing gate, semantics unchanged. Clock: statute and insurance — decades, with ι-1 as the one fast hand.
  • I-in — permission to run institutional inference on persons: eligibility recomputation, anomaly flagging, audit selection, risk scoring — the state and other institutions as consumers of AI, deployed on the governed rather than for them. Clock: legitimacy, scandal, and administrative law — faster and lumpier than statute (the same jurisdiction can expand ADM for years, then lose an entire program to one court ruling — E15–E17).

The asymmetry that makes the split load-bearing: control-enhancing applications face stronger adoption incentives than power-dispersing ones. An AI that cheapens benefit administration has an obvious institutional buyer; an AI that dissolves a licensing monopoly threatens its own approver. A polity can therefore hold I-out shut while I-in opens wide — forbidding the automation of doctors while its health authority scores claims, flags anomalies, and allocates investigations. The EU AI Act is the statutory trace: it prohibits certain citizen-facing scoring and manipulation while merely regulating most high-risk institutional use [O]. Classification note: proposals to treat “opacity” (knowing how one is scored) as a fifth terminal scarcity fail the classification rule — opacity is relieved by institutions (audit rights, explanation mandates), so inferential permission is I-class by last unrelieved constraint; the split above, not a new class, is the correct integration. Consequences propagate to §3.2 (the sovereign carve-out), §5.3 (the ADM-scandal sub-current), §6.4 (overlay O3), §8.1 (R12’s contingency), §9.3 (R15’s endogenous counter-column), and §11 (M14’s attribution requirement, M17–M18). Portability per D.7: the I-in mechanism binds hardest where administrative states are strong and courts weak — the inverse of M2’s scope limit.

Two worked decompositions

Housing — the framework’s standing test asset:

house price = structure (P) + land (X) + entitlement (I)

In supply-constrained metros, structure is roughly 30–40% of price; land plus capitalized entitlement is 60–70%. In unconstrained markets the proportions invert. [E] The entitlement component is a pure I-rent wedge capitalized into the land price: the value of permission to build, separable from the land itself wherever upzoning events let us observe the jump. [I]

Healthcare (~17–18% of US GDP [O]):

Component Class Price dynamic under the model
Diagnosis, documentation, coding, administrative cognition D (much of it wedge-priced at I rates) deflates if and only if F2 collapses the wedge; otherwise held at I prices
Procedures, nursing, physical care P tracks wages; robot-gated far out
Sign-off, prescribing authority, liability bearing I (rent wedge + residual human work) inflates under protection (cost disease)
Concierge access to scarce specialists X inflates with top incomes

The payoff of decomposition: “healthcare is exposed to AI” is an ill-posed claim. The D-component could deflate 90% while the sector’s total cost rises — because the I-wedge and P-wages dominate expenditure, and X-access inflates. Sector-level analysis averages over opposing price dynamics and mispredicts systematically. Task-level constraint scoring is the correct resolution. (Full NAICS→class mapping and decomposition rules: Appendix A, with the honest caveat that all share estimates there are US-shaped and rough. [E])

3.2 Growth accounting

The two equations

With expenditure shares s_i and class productivity growth g_i, i ∈ {D, P, I, X}:

g_Y ≈ s_D·g_D + s_P·g_P + s_I·g_I + s_X·g_X              (aggregation)

d ln s_i ≈ (1 − σ)·(π_i − π̄)                              (share drift)

Class growth rates are gated by different mechanisms:

  • g_D = f(τ, β) — high and rising; the capability frontier.
  • g_P = f(κ) — low near-term; rises only after the robot crossing.
  • g_I = min(g_technical, g_permitted) — a min function. Capability sets one ceiling, law the other; the lower binds. Under current institutions, g_permitted ≈ historical services productivity, ~0.5%/yr. [E] Sovereign carve-out (v3): the min binds only where permitter ≠ operator. Where the state operates on itself — its own administrative layer, the D-feasible slice A.6(iv) flags — g_permitted collapses toward g_technical, because the approver and the deployer are the same actor (D.4’s fiat-F2 generalized from China to every sovereign’s internal deployments). The carve-out is exactly the I-in direction of §3.1’s split, and it is why the informational state can run ahead of the economy it governs.
  • g_X ≡ 0 — by construction. No productivity growth is defined over zero-sum goods.

Under σ < 1 (complementarity), the share-drift equation says classes with rising relative prices — the slow-growth classes — gain expenditure share. The fast class deflates itself out of the weighted average.

One structural observation before the arithmetic: AI’s distinguishing macro feature is not the level of g_D but the dispersion of the g_i — 15%+ against ~0.5% against zero. Under complementarity, the aggregate discount from growth-rate dispersion scales with its variance. No prior automation wave produced anything close to this spread, which is why the complementarity penalty is larger here than in any historical episode used to calibrate intuitions. [I]

The worked table

Initial shares (US-shaped; derived in Appendix A.2): s_D = 0.28, s_P = 0.37, s_I = 0.20, s_X = 0.15. [E] Non-D baseline: g_P = 1.5%, g_I = 0.5%, g_X = 0 → non-D contribution ≈ 0.66pp. First-order growth at t = 0, as a function of g_D:

g_D g_Y (first-order) Reading
2% ~1.2% The Acemoglu corner — his ~0.5–0.7%-per-decade TFP arithmetic is this row, expressed as a parameter value
10% ~3.5% Weak S2 (≈ v1’s point estimate)
15% ~4.9% Strong S2
25% ~7.7% S3 corridor

The nesting is the point: the framework does not compete with the task-model literature or with takeoff models — it contains both as parameter settings. Disagreement about the future becomes disagreement about a row of this table plus a value of σ, which is arguable and monitorable, rather than a clash of worldviews.

Result R1 — the deceleration paradox

Under σ < 1, the share-drift equation bites: with an illustrative aggregate σ̄ ≈ 0.5 (a worked-example scalar standing in for the class-pair elasticity structure — the retired v1 σ is not reinstated; §0.1 ledger) and g_D − ḡ ≈ 10pp, s_D falls roughly 1.4pp/yr. Propagating: growth in the strong-S2 row decays from ~4.9% toward ~3.7% within about seven years with τ held constant. [I]

S2 growth is front-loaded, not flat. A mid-2030s GDP deceleration is a prediction of the capability story under complementarity — not evidence against it.

The practical rule this generates: never adjudicate τ from the GDP print. When 2033 growth comes in below 2030 growth, the capability question is answered by the horizon curve (dashboard τ-rows), not the national accounts. The predictable misreading — “the AI story is dead” — is pre-registered here as a model confirmation under S2.

Result R2 — the Musk condition and the terminal scarcity theorem

“Quasi-infinite output” (P4) requires the slow-class shares to vanish, since g_Y converges to the growth rate of whatever still absorbs expenditure. Decompose the requirement by class:

  • P → conquerable by actuation. An engineering-and-capital problem. Slow, but no principled barrier.
  • I → conquerable only by political-legal dissolution of the permission structure. Orthogonal to capability; no value of τ opens this gate.
  • X → unconquerable by construction. Positional goods are zero-sum; productivity growth over them is undefined.

Therefore:

The terminal scarcity structure of any AI trajectory — including full takeoff — is X plus whatever survives of I. The terminal role of money is claims on positional and institutional scarcity.

This derives, rather than asserts, the v1 verdict on P8. Money does not stop mattering; it collapses onto the two classes that cannot deflate. Musk’s defensible steelman — “money won’t matter for subsistence” — is precisely the claim that D∪P covers subsistence, which the model endorses in S3/S4. Everything above subsistence lives in I and X, where the price system does not relax; it concentrates.

Result R3 — the rent sink

X has fixed supply and income-elastic demand at the top of the distribution (ε ≈ 1.2–1.5 [A]). Any growth in aggregate real income therefore pushes X-demand against fixed X-supply, and a fraction of all productivity gains — whatever their class of origin — is absorbed as X-price inflation and capitalized into land and positional assets.

Observable implication [F]: the land share of national income rises through the decade in every scenario with positive growth. Note that this extends a pre-AI trend [O: 2010s land-value and superstar-city dynamics] — so S2 predicts an acceleration of an existing pattern, not a break, which makes the test cleaner.

Employment corollary

Class employment share ≈ expenditure share ÷ relative productivity. Since I-class productivity is near-static under protection, I-employment rises at least as fast as I-expenditure; since D-productivity explodes, D-employment falls faster than D-expenditure. The hollowing middle is thereby derived: employment migrates from the fast class to the protected class, with the unprotected-but-unautomated middle (§8’s squeeze zone) compressed between them.

3.3 Demand and pass-through

The growth accounting is silent on employment within a class until demand response and pricing behavior are specified. Sectoral approximation, with z productivity, π the pass-through of cost savings into prices, ε the absolute demand elasticity, and R reinstatement from new tasks:

d ln Q ≈ ε·π·d ln z                      (output response)
d ln L ≈ (ε·π − 1)·d ln z + R            (employment response)

Three canonical patterns:

Pattern Condition Historical instance Outcome
Agriculture ε·π < 1 food, 1900–2000: productivity up 20×+, employment share ~40% → ~1.5% [O] prices collapse, demand response insufficient, employment exits
Jevons/textile ε·π > 1 weaving 1790–1840; computing 1970–2010 [O] cheapness unlocks so much demand that output — sometimes employment — expands
Rent capture π low concentrated platform markets [I] gains retained as margin; prices, demand, and employment barely move; profits rise

The third pattern is the under-discussed one, and it connects this module to market structure (§5.1): whether AI produces consumer abundance or producer rents is a π question, and π is set by competition at the access and distribution layers — not by capability. A world of miraculous models behind concentrated distribution is a rent-capture world: modest deflation, modest displacement, spectacular margins.

The η_D decomposition

Aggregate D-class demand elasticity is a weighted average of two components with opposite signs relative to the η = 1 threshold:

  • Routine substitution (η < 1): existing D-consumption — reports, code for existing systems, support tickets — bounded by organizational absorptive capacity and human attention. Current evidence is consistent with η < 1 here [I]: displacement without offsetting volume-driven rehiring; negative net headcount in IT services [E, pending verification].
  • New-category creation (η effectively unbounded): consumption that did not exist at prior prices — disposable personal software, always-on expert advice, generated media. No prior market existed, so any positive price creates demand from zero.

Aggregate η_D therefore drifts upward over time as the new-category weight grows. The framework’s D-class labor pessimism is front-loaded and could partially reverse in the 2030s [F, explicitly flagged as the module’s most uncertain forecast].

Measurement: Jevons Meter B

The naive estimator — volume growth ÷ price decline — is biased upward twice: the demand curve shifts outward as capability improves (quality confound), and a material fraction of current volume is subsidized experimentation rather than production demand. [I]

Specification: constant-capability price series (the price of a fixed model-quality tier over time) × paid production volume (enterprise revenue at that tier), estimated within task categories. Add a saturation check: per-capita D-consumption plateauing — the TV-hours pattern — is the non-homothetic signature of η falling with quantity, and would cap the new-category offset. Dashboard row η-1 (§10).

3.4 Boundary flows

Class membership is not fixed. The political economy of the transition lives almost entirely in four flows across class boundaries — and each flow has a leading indicator that runs ahead of the statistics.

F1: D→I — licensure flight, with the guild refuge loop

Displaced or threatened cognitive occupations seek statutory protection: scope-of-practice rules, mandatory human sign-off, liability requirements. The flow is self-reinforcing:

D-displacement → credential flight + lobby formation → protection tightens (σ_I ↓)
     → I expenditure share grows (Baumol drift) → returns to protection rise
     → lobby strengthens → further tightening

The loop’s political meaning: λ has an organized beneficiary class, not merely a diffuse public. Throttling is not a taste that fades; it is a rent that compounds. This makes restrictive-λ paths stickier than opinion-poll models suggest. [I]

Predicted flight order [F]: occupations adjacent to active displacement with existing professional bodies move first — accounting, financial advice, therapy, engineering certification. Software engineering, which lacks a guild structure, absorbs Wave 1 undefended; its undefendedness is not an anomaly but the model’s control case.

Leading indicator: professional-school and licensing-exam applications, running ~2–3 years ahead of the legislation the applicants will eventually demand. Dashboard row ι-2.

F2: I→D — liability unbundling, with insurance as the hinge

The I-class erodes when liability can attach to systems rather than licensed humans — i.e., when the rent wedge (§3.1) collapses. The hinge actor is the insurance industry, not the legislature: the moment an insurer prices AI professional-liability coverage below the human-equivalent premium, the economics of the human signature collapse even where statute still requires it, and statute follows insurance economics with a lag. [I]

This is also the channel through which capability attacks the I-class indirectly, resolving the v2.0 overstatement that I is “capability-immune”: rising reliability → lower actuarial risk → cheaper coverage → economic pressure on the signature. Formally, σ_I = f(λ, insurance economics) — politically maintained, economically contestable.

The wedge prize. What F2 fights over can be priced — approximately, and the approximation’s structure matters more than its midpoint. The wedge share of I-class value-added is an explicit assumption, 25–60% [A], with the licensing-premium literature as a weak calibration check rather than a derivation input (the re-derivation lives in A.3); at I ≈ 0.20 of GDP:

Roughly 5–12% of GDP is currently priced by permission on tasks that are technically automatable at the frontier — the lower half of the range better supported than the upper. [I, on an [A] wedge share]

(Derivation and its two slices — private-licensed and sovereign-administrative, which erode by different mechanisms: Appendix A.3.) This is the pool F2 contests, the quantity F1’s guild loop defends, the wage bill of §8’s absorption machine (R12), and the S3-upside the I-gate withholds. It is the sharpest available answer to how much is at stake at the I-gate: something between a twentieth and a tenth of the economy.

Leading indicator — the best in the framework: AI E&O/malpractice product launches and their premium ratio to human-equivalent coverage, running years ahead of legal change. Dashboard row ι-1. The first material F2 event is predicted to be preceded by insurance repricing, not legislation (§11, module falsifier M2).

F3: P→D — actuation

Robotics reclassifies embodied tasks as capability-gated. Definitionally: S3 ≡ mass F3 before ~2035. Monitored by robot $/hr against the ~2× median-wage threshold and by units shipped — production data, never demos. Honest note carried from §12: this is the least-instrumented flow in the framework; nothing in the current robotics data materially constrains its timing, and the P-clock is accordingly the widest error bar in the model. [E]

F4: X-manufacture

New positional goods are continuously created: credential tiers, attention formats, luxury variants, access hierarchies. X expands in variety, absorbing expenditure, while — being zero-sum — contributing no productivity and little employment. It is a demand sink, not a production sector.

F4 is also the mechanism by which abundance manufactures its own scarcity: as D-goods approach zero price, differentiation migrates to what cannot be copied. The volume of X-innovation is itself an indicator of D-saturation. [I]

The test asset, restated

Housing observes three flows at once. If F3 arrives, robotic construction deflates the structure component and the residual price isolates land plus entitlement — making zoning politics brutally transparent, because the technical excuse for scarcity is gone. Standing prediction [F]: the construction-cost share of home prices falls over the decade while the land-plus-entitlement share rises in every scenario; only S3 bends the structure component itself. The YIMBY conflict is the X/I frontier made visible, and its intensity should scale with F3 progress.

3.5 Welfare incidence

Setup

Household welfare growth ≈ own-income growth − Σ s_i^h · π_i, with household-specific expenditure shares. Aggregate statistics average over households whose class exposures differ systematically — which is where v1’s “the statistics will gaslight the participants” stops being rhetoric and becomes arithmetic.

Four archetypes

Archetype Income source Basket tilt Modal (S2) outcome
Capital owner concentrated equity X-heavy Income ↑↑; X-inflation bites, but they own the appreciating assets. Unambiguous winner
I-incumbent (licensed senior professional) Baumol-protected wages I + X Wages track I-inflation by construction → hedged. Housing tenure decides the rest
P-trades buildout demand P + D Real gains near-term; partially self-correcting as the apprenticeship pipeline (~4–5y lag) delivers supply response
Displaced D-worker falling wage / transfers D + P cheap; I + X dear The squeeze case

The squeeze case, derived

AI makes cheap things cheaper and leaves dear things dear. The displaced worker’s consumption deflates in exactly the categories that were already affordable — entertainment, generic goods, cognition itself — while the pain concentrates where deflation never arrives: healthcare, shelter, education, credentials. All I and X. Relative-price change under this model is regressive even when headline statistics are generous, because the deflating classes are discretionary and the inflating classes are unavoidable. [I]

Household Leontief

Cross-class complementarity is tightest at the household level: entertainment cannot substitute for housing; a free legal draft cannot substitute for the signature that makes it actionable. As household-level σ → 0, welfare growth → the minimum over necessary classes, not the share-weighted mean.

The lived experience of S2 is Leontief: unlimited free cognition, unaffordable housing, healthcare, and credentials. The aggregate surplus is real. The experienced welfare gain is bottlenecked by the dearest necessary class — which is precisely the class the household cannot exit.

This single result explains the central political puzzle the framework predicts: genuine, large, measurable welfare gains coexisting with genuine, widespread, correctly-perceived immiseration — resolved not by calling either side wrong but by observing that they are indexed to different classes.

The I-entry queue

Expenditure flows into the I-class freely (Baumol drift); labor entry does not (residency slots, partnership tracks, licensing exams, apprenticeship caps). Three consequences: incumbents capture the Baumol rents; displaced D-workers queue at the gates; and the rents partially dissipate through queuing costs rather than clearing through wages.

Signature prediction [F]: applications-per-seat surge while incumbent premiums hold or rise — rising premium and rising queue simultaneously, which naive human-capital models read as a contradiction and this model reads as equilibrium. Dashboard row ι-2; module falsifier M1 (§11).

Distributional summary — who bears S2

  • Gains: capital owners; I-incumbents; P-trades near-term.
  • Losses: the D-displaced; I-aspirants stuck in queues; renters in X-inflating metros.
  • Coalition arithmetic: the losing coalition — young, credentialed, urban, renting — is exactly the Wave B electorate (§5.3). What it will rationally demand is not transfers first but access: entry to I-rents (guild seats, licensing reform from below) and X-goods (housing). This is derivable from the incidence table and is why §8 ranks tuition-paths and housing supply above UBI in the near-term policy sequence.

§4 — Macro-Financial Module

Status: nominal and financial layer. Consumes §3’s class engine and §6’s terminals; feeds §5 (policy vector), §7 (capital implications), §10 (dashboard rows F-1…F-5), §11 (falsifiers). Evidence tags per §0. Results numbering continues from §3 (R1–R3).

§3 built the real side: four classes, four clocks, gates. §4 prices it — and adjudicates Musk’s P7 (“deflation will be the issue, not inflation”), which turns out to be one-fifth right, and right in the fifth that matters least for policy.

4.1 The five deflations and the policy trap

The taxonomy, in class language

“Deflation” conflates five phenomena with different mechanisms, different victims, and different policy handles:

Type What falls Class expression Terminal incidence Policy handle
R — relative automatable output vs. residual π_D ≪ 0 while π_I, π_X > 0; post-F3, π_P joins the deflators All terminals — the only certainty in the framework. Dispersion is maximal in S4a none — this is the technology expressing itself
A — aggregate the price level requires D-deflation to outweigh I/X-inflation and the CB to permit it Low everywhere [I]: under inflation targeting, A-type is a policy choice, not an MV=PY mechanical outcome fully monetary
C — collateral AI-capital and adjacent asset prices D-class capital melts (§4.2); can occur with CPI > 0 S1a/S1b high; S2 moderate (vintage turnover); S3/S4a hit legacy assets instead prudential, not monetary
N — nominal GDP aggregate spending overlay-driven (O1), not terminal-specific S1b brings 2–3 years of it; otherwise cyclical conventional stabilization
W — wedge nothing falls; measurement fails near-free D-goods raise surplus without measured output all terminals; worst where D-abundance is greatest (S3/S4a/S4b) statistical reform

P7 adjudicated. Musk predicts A-type. The model says: R-type is certain, C-type is the near-term financial risk, W-type is why the debate will be confused, and A-type happens only if the central bank lets it — which makes his own position incoherent, since a world of aggressive, precisely calibrated monetization to offset D-deflation is a world of an extremely activist central bank, not of money’s irrelevance. The consistent version of P7 refutes P8. [I]

Illustrative S2 class inflation paths [F]: π_D ≈ −5 to −15%/yr quality-adjusted; π_P ≈ +2–4% (buildout phase); π_I ≈ +4–6% (cost disease plus wedge); π_X ≈ +5–8% (rent sink, R3). Headline CPI plausibly prints a placid 2–3% over a 10–20pp cross-class dispersion. The headline is the least informative number in the economy. Dashboard row F-4 (class-dispersion thermometer) exists because of this line.

R4 — the Baumol policy trap

A central bank in the four-class economy holds one instrument against class-divergent inflation, and the inflating classes are monetary-policy-inert: I-inflation is wedge-driven (politically maintained prices, §3.1) and X-inflation is scarcity- and sink-driven (R3). Neither is credit-sensitive demand. The instrument only reaches the credit-sensitive classes — D-capex, P-buildout, housing construction.

The trap has two jaws:

  • Salience jaw. Households experience the cost-of-living crisis in I and X (shelter, healthcare, tuition, insurance) and demand tightening. Tightening does not touch the wedge or the sink; it lands on D-capex and on the P-side supply response — including the housing construction that would have relieved X-scarcity later. Iatrogenic on both counts.
  • Index jaw. A CB that looks through to the D-dragged headline runs easy — and the liquidity flows to the savings sink (§4.3), inflating X-assets without consumer inflation, worsening the §3.5 incidence and the political backlash that feeds Wave A/B.

Whichever index it targets, the CB amplifies one distortion. [I]

Level vs. trend — the X-side premise, stated precisely (v3). “Policy-inert” was drafted too broadly. X-asset prices are among the most rate-sensitive in the economy — housing through the mortgage channel, income-producing land through cap rates; tightening can and does crash X-price levels. What the instrument cannot reach is X-scarcity and therefore the rent-share trend: rates change the discount factor on X-cash-flows and the leverage available to bid for them, not the supply of position or the R3 sink that directs income growth at it. A tightening cycle that drops land prices 20% while the land share of national income keeps climbing is the premise working, not failing. The trap survives the distinction — the salience jaw is about flows households pay (rents, premiums, tuition), which track the trend, not the level — but M4 must be read against it: M4 tests the I-side; on the X-side, only a rent-share trend reversal under tightening counts against R4, never a price-level drawdown. [I]

Corollary predictions [F]: the defining monetary-policy debate of the 2030s is which index to target (ex-shelter, trimmed-mean, “ex-wedge core” proposals); the first central bank to adopt an explicit class-dispersion adjustment is a regime-change marker. Dashboard row F-5.

The trap × the corridor. The Fisher risk is non-monotonic across terminals, and the trap compounds it at the worst point:

Terminal Fisher exposure Mechanism
S1a/S1b low-moderate recessionary disinflation, but the CB eases and refinancing relief partly offsets; the damage is C-type, not Fisher
S2 chronic, not acute CPI target holds; the grind is fiscal (R6), not deflationary
S3 (the corridor) acute D and P deflating together drags headline hard just as ~120%-of-GDP nominal debt [O] meets a redistribution apparatus that does not yet exist. A CB that spent the late 2020s fighting salient I/X inflation enters the corridor tight — tight money + broad disinflation + nominal debt is the 1930s configuration
S4a inverted rents support nominal flows; the sovereign problem is capture, not deflation (R6 variant)
S4b n/a pre-committed structures only (§6)
S6 bloc-split corridor risk lands only in blocs whose gates opened; holdout blocs import D-deflation through tradables at the EU-profile squeeze (D.3)

Measurement conventions (standing box). Every growth or inflation figure in this document states: geography (default US), metric (measured real GDP per capita, 5-yr CAGR unless flagged), and W-wedge applicability. Corollary for welfare claims: CPI is a bad welfare index in this economy — a D-heavy young household’s true cost of living falls faster than measured; an I/X-heavy household’s rises faster than measured. Political sentiment tracks cohort basket inflation, not CPI (§3.5 restated in measurement terms). Cohort-specific deflators are an Appendix A deliverable.

4.2 The debt machine

Where the risk moved

Phase one of the buildout (2023–25) was funded from hyperscaler operating cash flow — capability risk sat on diversified, daily-marked equity, the most benign possible configuration. [O] That phase is over: record IG issuance (>$100B of hyperscaler bonds in 2025), off-balance-sheet SPVs, and rapidly growing private-credit exposure to data-center debt have moved the marginal dollar onto credit — concentrated in levered vehicles, marked rarely, held by insurers unstressed against correlated GPU-collateral impairment. [E, verified per C.2]

[corrected] Two items in the drafting version overreached the source. Vendor financing appears nowhere in BIS Bulletin 120 or the March 2026 Quarterly Review box; the Bulletin references “circular financing within the AI ecosystem” citing press reporting, which is not the same claim and is not BIS’s own finding. Pension funds are likewise never named or quantified by BIS — insurers are (“these structures strengthen links between hyperscalers and non-bank investors such as private credit vehicles and insurers”), and pensions belong to the unquantified institutional residual. What BIS does establish, and what R5 actually needs: the shift “from cash flows to debt,” AI-related private credit above $200B (~8% of outstanding, from near zero, with $40B originated in 2025 and a projected $300–600B by 2030), SPV structures BIS labels “shadow borrowing” — “obligations that are economically akin to debt but largely reside outside corporate balance sheets” — and the pricing tension that AI loan spreads sit close to non-AI spreads, implying lenders price these as ordinary risk while equity markets price outsized returns. R5’s topology claim survives intact on the verified subset; the two deleted mechanisms were decoration.

The topology inversion: the same capability disappointment that would have been an equity drawdown in 2024 is a credit event in 2027+.

Reflexivity note. The expectation of a sovereign backstop (“too strategic to fail” — the Fannie-Mae-for-compute path flagged in v1) compresses spreads now, which raises leverage now, which raises the stakes of the eventual test. The guarantee expectation is partially self-fulfilling in the wrong direction, and positioning ahead of the implicit guarantee is itself a capital flow. [I]

The treadmill, mix-adjusted

The melting-collateral logic applies to roughly half the stack, not all of it:

Asset layer Economic life Share of stack [E] Collateral quality
Accelerators 3–5y (shorter competitive life) ~40–50% melting
Networking 5–7y ~10% soft
Shells/cooling 15–30y ~20–25% durable
Power infrastructure 30–50y ~20–25% durable, regulated

Replacement capex is therefore a range conditional on mix — but the melting half is the levered half, and by ~2028–29 the sector faces the solvency test: inference revenue versus annualized replacement capex plus debt service. This test is the mechanism behind §6’s Layer-3 conditional (P(systemic | steady × shut) = 0.28): the tree’s number is a probability on this test failing in a correlated way. Dashboard row F-1 (the solvency gauge) is the direct observable.

Collateral migration is already the rational lender response: covenant structures repricing toward power contracts, PPAs, and entitled land — the durable, class-P/X half — and away from silicon. [F] This quietly reinforces §3’s theme: even the creditors are migrating toward the bottleneck classes.

The six channels

# Channel Mechanism Balance sheets hit Worst terminals Leading indicator
1 Cash-flow inference revenue < debt service + replacement capex SPVs, private-credit funds, vendor financiers S1a, S1b F-1 solvency gauge
2 Collateral GPU/DC asset impairment; melting half of the stack secured lenders, lessors S1a, S1b depreciation-schedule revisions (the accounting confession); secondary GPU prices
3 Refinancing maturity walls meet post-bust risk premia or higher-for-longer everything levered S1b; all except S4b DC paper spreads; maturity calendar
4 Fisher (aggregate) broad disinflation raises real burdens sovereigns, households, legacy corporates S3 corridor breakevens collapsing while growth expectations rise — the corridor signature
5 Sovereign-fiscal labor-tax erosion + transfer need + interest burden [O: interest > defense] treasuries, then everyone via rates chronic S2; acute S3; capture-shaped S4a labor share of taxable income; interest/revenue ratio
6 Stranded-asset capability progress obsoletes existing stock — chips by chips, offices by D-automation CRE, old-vintage holders S3, S4a — the good scenarios office/CRE spreads; vintage price curves

The correction from the adjudication is hard-coded in rows 1–2: the plateau scenario is dangerous for credit. S1a is the low-drama terminal for technology and the high-drama terminal for private infrastructure debt.

R5 — the capability straddle

Read the table by column and a structure emerges: channels 1–3 price capability disappointment; channels 4–6 price capability success. The financing system loses if AI underdelivers (cash-flow, collateral, refinancing) and loses if AI overdelivers (Fisher, fiscal, stranding). It is safe only in the orderly middle — S2’s grind, where revenue grows into the capex, disinflation stays relative, and stranding proceeds at vintage pace.

The credit system is short a straddle on capability. Its safe corridor is approximately the modal terminal (S2 ≈ 0.28 plus the benign residual), while both tails — jointly ≈ 0.55–0.60 of tree mass — hurt it from opposite directions. [I]

This is the single most decision-relevant sentence in the module, and it carries directly into §7: the correct credit posture is not “bullish or bearish on AI” but long the corridor, hedged in both tails — and almost no current structure is built that way, because the two tails look like opposite views and are therefore held by different institutions, neither hedging the other.

R6 — the fiscal scissors

  • Blade 1 (early): the labor-income tax base — the majority of federal revenue via individual income and payroll taxes [E] — erodes with D-displacement and wage-share decline, front-loaded per §8’s wave sequencing.
  • Blade 2 (mid-decade): fiscal need rises — Wave B transfer demands, tuition-path subsidies, and an interest burden already exceeding defense [O].

The blades cross in the late 2020s–early 2030s: fiscal capacity and fiscal need move in opposite directions during exactly the window when the political system must fund the transition. This derives v1’s assertion and explains why the χ-component of the policy vector (§5.3) arrives structurally late relative to need — the money isn’t there when the demand first forms.

S4a variant: in the rentier decade the base exists — rents are enormous — but it is concentrated, mobile, and politically defended. The fiscal question mutates from “where did the base go?” (S2/S3) to “who can tax the rents?” (S4a): χ versus Ω becomes the central political economy of the terminal, previewed here and developed in §5.2.

The counter-current (v3): the scissors are ambivalent about the informational state. AI-driven tax administration partially closes Blade 1 from the enforcement side — the tax gap (unreported and under-collected liability) is a base-erosion term that continuous, machine-read compliance directly attacks, and it requires no new legislation, only deployment the state grants itself (the sovereign carve-out, §3.2). This cuts both ways and the framework declines to net it to a sign: computational governance is simultaneously the machinery of the control story (O3, §6.4) and the one visible mechanism that could make χ feasible on the framework’s timeline — the same systems that score citizens can find the base R6 says is eroding. The scissors logic survives (rate-base erosion from wage-share decline is untouched by enforcement), but the darkest scenario and the most hopeful fiscal repair now visibly run through the same technology. Tracked on ι-4. [I]

4.3 Market-implied signals

The contamination lemma

r takes pressure upward from capex demand and growth expectations, and downward from the savings glut that AI-driven concentration itself produces: falling labor share → income concentrates at high-saving households and firms → savings chase X-assets — land, equities, positional stores of value. The X-class is a savings sink, and the sink deepens with AI success.* [I]

Consequence: real rates are non-monotonic in AI outcomes. S4a — extreme capability success with terrible distribution — plausibly prints lower real rates than S2, because concentration out-supplies the capex demand for savings. “The bond market will tell us” fails precisely in the scenarios where the answer matters most. No single market is treated as efficient when convenient (the selective-EMH repair from the review round): cross-market joint patterns are the only admissible signals.

The rates × land discriminator

Pairing real rates with prime-land real appreciation separates what either alone cannot:

Land boom No land boom
Real rates high genuine takeoff repricing → S3/S4b fiscal crowding-out — not an AI signal
Real rates low the sink cell: S2-with-bad-distribution or S4a plateau → S1a

Honest limitation: S2 and S4a share the sink cell. Both are concentration-plus-Baumol worlds from the bond and land markets’ perspective. The tiebreakers are distributional and real-side, not financial: rent share vs. wage share of national income, the magnitude of F-4 class-inflation dispersion (S4a’s is far wider), and the τ-rows (S4a requires recursive capability; S2 doesn’t). Dashboard row F-3 carries the 2×2; the tiebreaker set is listed with it.

The confirmation basket

S3 repricing is indicated only by the joint pattern — all four, sustained:

  1. real 10-year rates +100bp or more; and
  2. exposed-sector TFP accelerating against matched controls; and
  3. equity breadth widening beyond the mega-caps; and
  4. capex growth maintained despite (1).

Any subset short of all four carries a likelihood ratio ≈ 1 (§10 bands) — including, per the lemma, rates alone in either direction.

Credit-side gauges and the current reading

The credit market’s own confessions rank above its prices: F-1 (inference-revenue/annualized-capex — the solvency gauge), F-2 (DC paper spreads and depreciation-schedule revisions — the latter being the accounting system admitting what secondary GPU prices already said), plus PPA terms and collateral-mix drift in new covenants as the slow-motion record of lender class-migration.

As of the document date [E]: real rates ~2%, breadth narrow, capex enormous, prime land firm — the market sits in or near the sink cell with an equity option premium on the tails. That is, the market’s implied posterior approximately matches the tree: mixture-pricing, S3/S4 as paid-for option value, no takeoff certainty. The framework’s disagreement with the market is not about the mixture; it is about the straddle (R5) — the tails are priced in equities and unpriced in credit.

§5 — Structure: Markets and Policy

Status: structural layer. Supplies the two structures §§3–4 kept referencing but did not define: who competes (the determinants of pass-through π) and who decides (the policy vector Λ). Consumes §3’s classes and flows, §4’s R4–R6, §6’s terminals; feeds §§7–8, §10 (rows M-1, M-2, Λ-A, Λ-B), §11. Results continue: R7–R9. Complexity-budget bookkeeping: this section formally retires three v1 scalars — μ (market structure), λ (political throttle), δ (distribution regime) — replacing them with a per-layer concentration vector, the policy vector Λ, and a four-channel outcome decomposition. The scalars were doing hidden vector work; the disaggregation adds no net free parameters, because each component is pinned to named observables.

5.1 Market structure: four layers, one choke point

The stack

“AI market structure” is not one concentration number; it is four, and they move independently:

Layer What it sells Current concentration [E] Contestability
L1 — Frontier development trained model weights few labs contested by open weights; capital moat only as durable as training-cost growth
L2 — Compute/infrastructure FLOPs, power, shells top-3 cloud dominant; power queues worse physically gated (κ); the least contestable layer on any horizon
L3 — Access/orchestration APIs, routing, indemnified deployment moderate, falling prices [O] high — unless trust wrappers re-concentrate it
L4 — Distribution/application the customer relationship, workflow lock-in, default placement inherited from pre-AI platform economics: high low — attention and defaults are positional

The choke-point rule

Consumer-facing deflation requires competition at every layer between model and user. A single concentrated layer converts upstream deflation into margin at that layer:

Deflation stops at the choke point. π for a value chain is set by its least competitive layer, not its average. [I]

This is the §3.3 rent-capture pattern given an address. The D-class can deflate 90% at L1/L3 while retail AI-service prices barely move — the wedge between the API price index and the consumer price is measurable rent capture in real time. Dashboard row M-2 (the pass-through gauge) is exactly that wedge.

The open-weights fallacy

Free weights commoditize L1 only. They do not commoditize:

  • L2 — you still rent the machine, and the machine is power-gated;
  • L4 — the customer, the data, and the default remain owned;
  • the trust wrapper — enterprises do not buy weights; they buy someone to sue. Indemnified, compliance-wrapped deployment is an I-class product (§3.1) emerging inside the stack, and its scarcity is legal, not technical. Note the direct link to ι-1: the insurance economics that gate F2 in the outside economy also decide whether L3 stays competitive or re-concentrates around a few indemnifiers. [I]

R7 — the stack recapitulates the class engine

Map the layers onto §3’s classes: L1 is D (pure cognition — commoditizing), L2 is P (embodied, buildout-gated), the trust wrapper is I (permission-scarce), L4 is X (attention and defaults — positional, zero-sum). Then apply the framework’s own logic to its own industry:

Scarcity migration operates inside the AI stack exactly as it does outside it. As the D-like layer commoditizes, rents migrate to the P-, I-, and X-like layers: compute, indemnification, distribution. The terminal scarcity theorem (R2) predicts the terminal rent pools of the industry itself: power contracts, trust wrappers, and attention. [I]

Investable corollary (→§7): the durable margins in AI are not in models. Vendor-financing circularity (§4.2) reads correctly in this light — it is vertical integration by balance sheet, an attempt by L1/L2 actors to pre-own the layers where the rents will land. Dashboard row M-1 tracks per-layer concentration: open-weight lag to frontier, top-3 cloud share, indemnifier count, app-layer margin distribution.

5.2 Distribution as an outcome

v1 treated distribution (δ) as a regime parameter. It is not a parameter; it is the residual of four channels, each with its own custodian and its own clock:

Channel Mechanism Set by Clock Custodian institution
π — pass-through gains → consumers via prices layer competition (5.1) fast where competitive antitrust (weak, slow)
b — bargaining gains → wages class structure + labor law contract-speed none — the orphan channel
Ω — ownership gains → retained rents, accumulating to equity holders market structure + private-capital plumbing fastest — automatic shareholders themselves
χ — fiscal gains → taxes and transfers legislation slowest — and scissors-bound (R6) treasury/legislature

Identity: for any productivity dividend, incidence ≡ consumer surplus (π) + wage share (b) + retained rents (Ω) + fiscal capture (χ). The size of the dividend says nothing about its address.

Two structural observations:

The orphan channel. Four of the five policy instruments in Λ (below) have institutional custodians; labor bargaining has none — union density is historically low [O], D-labor has no guild (§3.4’s control case), and b is not a component of any ministry’s mandate. b is therefore not one number but a class map: b_I high (the guild wedge is bargaining power, captured by incumbents, not queue entrants — §3.5), b_P temporarily high (shortage-driven, self-compressing on the ~4–5y apprenticeship lag), b_D → 0. Aggregate wage share can fall while incumbent-I wages rise — the barbell, derived from the channel structure.

R8 — default distribution is decided by clock speed

In the early transition, incidence defaults to Ω-dominance — not by anyone’s design, but because Ω is the fastest channel (accumulation is automatic) while χ is the slowest (legislation) and is fiscally scissors-bound (R6) precisely when the dividend arrives. Concentration is the path-of-least-resistance outcome of channel kinetics. [I]

This reframes the political economy: the question is never “will there be a distribution fight?” but “how large does the Ω-accumulated stock grow before the χ-channel activates?” — and R6 says the χ-channel is capacity-constrained through exactly the window when the stock compounds fastest.

Terminal incidence map:

Terminal π b Ω χ Grievance signature
S1a/S1b small dividend — takes the losses (equity, then credit per §4.2) dormant “it was a bubble”
S2 modest (choke points hold) barbell steady accumulation squeezed queues + X-inflation
S3 high (broad deflation) collapsing large racing displacement, behind displacement
S4a low (rents) minimal extreme capture war (R6 variant) rents
S4b — — — pre-commitments only —
S6 bloc-dependent bloc-dependent gate-gradient arbitrage premium (§7.5) per-bloc machinery cross-border comparison — visible foreign abundance against domestic holds

5.3 The policy vector Λ

Components

Λ = {ρ, χ, g, φ, ν} — five instruments, five constituencies, no requirement of coherence:

Component Instruments Feeds (tree object) Current posture [E] Rows
ρ — restriction licensing, liability, sectoral bans, siting/moratoria I-gate (and κ via siting) rising on both waves ι-2, Λ-A
χ — redistribution transfers, capital tax, sovereign stakes incidence; the Wave B answer dormant; scissors-bound Λ-B
g — public investment grid, permitting reform, tuition paths, housing/zoning κ; X-supply; queue relief underpowered relative to need κ-rows
φ — stabilization CB regime, target-index choice R4 trap; the S1b conditional orthodox F-5
ν — security acceleration export controls, subsidies, compute alliances, procurement τ (accelerates capability) strong, bipartisan Λ-treaty tracker

Note what g is: the only component that relieves rather than redistributes or restricts — it attacks κ (Wave A’s actual cause), X-scarcity (the entitlement wedge in housing), and the queue (tuition paths). It is also chronically underpowered, because its benefits are diffuse and lagged while ρ’s are targeted and immediate. [I]

The race asymmetry

ν and ρ rise together, and this is not incoherence — they answer different constituencies (the security establishment vs. domestic professions and ratepayers) and act on different objects (capability vs. deployment). The historical precedent is exact: nuclear — weapons development accelerated without pause while civilian deployment was throttled into a multi-decade freeze. v1 cited nuclear as the cautionary base rate for “political throttle”; the ν/ρ split is the mechanism that base rate was pointing at.

In tree terms: ν pushes Layer 1 toward recursive while ρ pushes the Layer-2 I-gate toward shut. The state is pumping the top of the funnel while pinching its middle.

The two waves

Wave A — power-bill politics (2026–28). Constituency: ratepayers, localities, incumbent industry. Instruments: PUC rate cases, DC-specific tariffs, siting moratoria. Feeds ρ (siting) and, if it matures well, g (grid buildout). Danger: it throttles κ before any TFP is visible to defend the buildout — the O1 overlay’s political expression. Monitor: Λ-A — docket/moratorium/tariff count across states. Promotion test per §6.4 already defined.

Wave B — cohort politics (2028–32). Constituency: §3.5’s losing coalition — young, credentialed, urban, renting; the queued and the displaced. Its demands, in incidence order: access (guild seats, housing) then transfers. Its organized vehicle is the guild loop (F1) — which produces a coalition with a fault line: incumbents want restriction that closes doors (protects rents); the queued want restriction that protects jobs but opens seats. Prediction [F]: Wave B splits per-profession into “close the doors” vs. “open the guilds” factions, and which faction wins is the micro-observable of σ_I. Monitor: Λ-B, bundled with ι-2 per §10’s correlated-indicator rule.

The ADM-scandal sub-current (v3). Wave B carries a second current that is the I-in gate’s political expression, distinct from the guild fight: episodic, scandal-driven backlash against institutional inference on persons. It has already fired three times at full scale, all pre-cutoff [O]: the Netherlands’ SyRI welfare-fraud scoring system, struck down by The Hague District Court (Feb 2020, Art. 8 ECHR grounds); Australia’s Robodebt income-averaging debt scheme, ruled unlawful with a ~A$1.8B settlement and a royal commission (report 2023); and the Dutch toeslagenaffaire, the childcare-benefits algorithmic enforcement scandal over which the Rutte III government resigned (Jan 2021). The pattern is the state-side gate operating ex post — through legitimacy, administrative law, and courts rather than insurance pricing — and it is why I-in’s clock is faster and lumpier than I-out’s (§3.1). The EU AI Act’s asymmetry (prohibiting some citizen-facing scoring while regulating institutional high-risk use) is the same split reaching statute. Not a third wave: it shares Wave B’s cohort energy and rides the same electoral cycles, so it books as a sub-current under Λ-B, monitored on ι-4. Feeds ρ in the I-in direction — and O3’s promotion test (§6.4). [E15–E17, B.2]

R9 — the default policy vector points at S4a

Assemble the module:

  • R6 (fiscal scissors) blocks χ exactly when Wave B peaks → political energy routes into ρ, because restriction is off-budget: its costs are diffuse, unmeasured, and laundered through I-class consumer prices, which is precisely what a fiscally constrained legislature can afford. Licensing is redistribution laundered through prices.
  • The race asymmetry holds ν high regardless → capability stays hot.
  • g stays underpowered → κ and X-supply stay tight.

The political system’s two strongest currents — fiscal constraint at home, security competition abroad — jointly produce accelerate-capability-restrict-deployment (ν↑, ρ↑, χ blocked, g weak). That posture is a machine for manufacturing S4a: recursive-but-bottlenecked, the rentier decade. [I]

And the I-gate becomes a fight, not a switch: F2’s insurance economics push it open (reliability → cheap coverage → wedge collapse) while Wave-B-fed ρ pushes it shut (statute resisting what insurance has already priced). S3 is the fight resolving toward economics; S2 is politics holding; S4a is the fight failing to resolve while capability compounds overhead.

The v3 refinement, from the directional split (§3.1): ρ throttles I-out; I-in accelerates by default under the selection asymmetry (control-enhancing deployment approves itself — the sovereign carve-out). The state is not merely pumping the funnel’s top while pinching its middle; it pinches the middle for the private economy while widening it for itself. R9’s machine therefore manufactures not just S4a but S4a-with-O3-active — the configuration the source correspondence names computational rentierism (§6.4). Falsifier M17 bounds this: scandal-class events (the ADM sub-current above) are the mechanism by which the default fails.

Jurisdictional postures (sketch — full treatment deferred to Appendix D)

Bloc ν ρ χ Implied tilt
US high rising scissors-bound S4a-shaped posture (R9’s home case)
EU low high higher baseline gates shut without a frontier: capability importer, funder of others’ buildout (v1), I-gate holds longest. S2-with-extra-steps regardless of global τ [F]
China maximal discretionary n/a (different δ machinery) the I-gate is administrative: the state can absorb AI liability by fiat — a sovereign insurance-flip, F2’s state-run variant. China’s binding gate is therefore physical (chips, κ), not institutional. If O2-lithography resolves, China opens gates faster than any democracy → cross-border gate divergence (§6’s S6 terminal, promoted in v3) becomes the likely path, and the first mass F2 event may be Chinese [F]

§6 — The Scenario Tree

Status: probability architecture. Consumes §3’s gate-state language; feeds §§7–8 (implications), §10 (updating mechanics), §11 (falsifiers). All layer priors are [A] — subjective, stated to be attacked. All composed masses are [I] — derived, and only as good as the conditionals above them.

6.1 The model vs. the bookkeeping

The causal model is a feedback system, not a tree:

                 ┌────────────────────────────────────────────────┐
                 ▼                                                │
   Capability clock (τ) ──▶ gate pressure ──▶ flows F1–F4 ──▶ class shares s_i
                 ▲                │                               │
                 │                ▼                               ▼
        investment ◀── expected returns          incidence (§3.5): queues, squeeze,
                 ▲                                rents, Wave A/B political energy
                 │                                                │
                 └── policy vector Λ ◀── institutional clock ◀────┘
                     (ρ throttles gates; ν accelerates capability)

Displacement feeds politics; politics sets ρ; ρ gates the flows; the flows determine displacement. The loop is real and cannot be represented in a tree without loss.

The tree is retained anyway, for one reason: probability bookkeeping requires a partition, and loops don’t partition. The resolution is procedural rather than structural — the tree is re-conditioned quarterly on realized dashboard values (§10), so the feedback runs through the analyst rather than through the diagram. This is an approximation, stated as such. Its known failure mode: the tree will lag any feedback that completes inside a quarter (e.g., a fast political cascade following a visible displacement event). [A]

What a terminal is. A terminal is not a world-state; it is an equivalence class over leaves — a decade signature. Two different underlying paths can produce observationally similar decades (S2 arrived at via steady capability, and S2 arrived at via plateau-plus-deployment-overhang, look nearly identical in the national accounts). The dashboard’s job is partly to discriminate within terminals, not just between them; the τ-rows exist precisely because the GDP print cannot tell those two S2s apart (§3.2, R1).

6.2 Layers, conditionals, terminals

Layer 1 — the capability clock (τ)

Branch Prior [A] Definition
Plateau 0.18 Horizon-curve doubling degrades materially; frontier progress stalls at approximately current-trajectory levels
Steady 0.47 Trend continues; no self-accelerating R&D loop demonstrated
Recursive 0.35 AI-R&D automation with measured compounding

Re-anchor log (v3, 2026-08-13; §10.2 discipline). Evidence: τ-1 — the verified E7 correction (TH1.1: post-2024 horizon doubling ~2.9 months, below the drafted 4–7-month floor; >16h horizons saturation-unreliable). Assigned LRs vs. steady: recursive ≈ 1.6 (weak–moderate band — the acceleration sits in the most recent slice, which is where a nascent loop would first show); plateau ≈ 1.05 (the same release’s saturation caveat is weakly plateau-compatible, offsetting). Drafted prior (0.20, 0.55, 0.25) → posterior (0.18, 0.47, 0.35); terminals recomputed mechanically below. Netting caution, logged: E7, R15’s ratchet audit, and the adjudicated feedback’s control asymmetry converge on raising S4a by three routes that plausibly channel one intuition. Only this τ-1 update is applied. The recursive×both-open throttled split is deliberately unchanged at 0.45/0.55 — the control-asymmetry argument that would strengthen it enters structurally (O3, ι-4) and its probability expression waits at O3’s promotion test. S4b’s rise to ≈ 0.09 is the mechanical consequence of that restraint, to be revisited if O3 promotes.

Layer 2 — gate state by ~2034 (conditional on τ)

The two gates from §3: P-gate (embodiment; opened by κ — F3 at scale) and I-gate (permission; opened by insurance economics and law — F2 at scale). X never opens; D needs no gate.

Given τ = … Both shut One open Both open Rationale
Plateau 0.80 0.15 0.05 The previously missing row. Gates can still creak open for already-feasible tasks (insurance can reprice existing capability), but without capability growth the pressure on both gates decays
Steady 0.65 0.28 0.07 Insurance flip plausible within a decade; robot production scale-up mostly isn’t (§3.4, F3 caveat)
Recursive 0.15 0.40 0.45 Reliability soars → insurance flips with high probability; but the P-gate needs factories, and even recursive cognition cannot compress physical production scale-up below some floor [I]

Which gate opens, when one does: under steady capability, either is plausible (split taken 50/50 [A]); under recursive capability, the I-gate is the more likely opener (0.60/0.40 [A]) — reliability is what insurance prices, while robot fleets remain production-constrained.

Coherence conditional (v3, feeding S6). Gate openings are jurisdictional where capability is not: the I-gate is national and state law, administrative in China (D.4), statutory in the EU (D.3), insurance-priced in the US (D.2) — so a “one open” world must also answer whether the opening coheres across blocs. Divergence is likelier under steady capability (0.35 [A]) than recursive (0.25 [A]): at steady, the opening margin is thin enough that jurisdictional machinery decides — China’s fiat-F2 plausibly fires alone; under recursive capability, pressure is overwhelming and actuarial import (D.6.1) couples blocs back toward coherence. Divergent arms route to S6.

Routing note (v3) — steady × one open. All coherent steady×one-open mass routes to S3 regardless of which gate opened. Justification, owed since drafting: a steady-capability I-opening — mass wedge collapse without robots — is economically S3-like even though its physical content differs: a growth surge from rent collapse rather than actuation, a hard labor-share fall in the wedge-priced occupations, Wave-B politics arriving mid-transition. The decade signature (§6.1’s equivalence-class definition) matches; S3’s profile below carries the two variants explicitly. The alternative — routing steady×I-open into S4a — fails S4a’s own definition (recursive capability).

Layer 3 — resolutions (only where they differentiate the decade)

Leaf Resolution Split [A]
plateau × shut Does the capex correction dominate the decade, or does overhang diffusion? bust-dominated 0.60 → S1a; overhang-diffusion 0.40 → S2
steady × shut Does the ~2028–29 solvency test fail systemically (correlated credit event, 2+ year investment recession)? systemic 0.28 → S1b; contained/clears 0.72 → S2
steady × one open Does the opening cohere across blocs? coherent 0.65 → S3 (either gate; routing note above); divergent 0.35 → S6
recursive × one open Cohere across blocs? If coherent, which gate opened? divergent 0.25 → S6; coherent 0.75, of which I-gate 0.60 → S4a, P-gate 0.40 → S3
recursive × both open Is takeoff throttled in time (ν vs ρ race, §5.3)? throttled 0.45 → S4a; unthrottled 0.55 → S4b

Composed terminal masses

Terminal Composition Mass [I]
S1a — capability plateau 0.18 × 0.80 × 0.60 ≈ 0.09
S1b — valuation bust 0.47 × 0.65 × 0.28 ≈ 0.09
S2 — Baumol-institutional (modal) 0.47 × 0.65 × 0.72 + 0.18 × 0.80 × 0.40 ≈ 0.28
S3 — compressed industrial revolution 0.47 × 0.28 × 0.65 + 0.35 × 0.40 × 0.75 × 0.40 ≈ 0.13
S4a — recursive-but-bottlenecked 0.35 × 0.15 + 0.35 × 0.40 × 0.75 × 0.60 + 0.35 × 0.45 × 0.45 ≈ 0.19
S4b — unbound takeoff 0.35 × 0.45 × 0.55 ≈ 0.09
S6 — split world (v3) 0.47 × 0.28 × 0.35 + 0.35 × 0.40 × 0.25 ≈ 0.08
Hybrid/residual plateau × (one + both) + steady × both ≈ 0.07
Σ = 1.00

Terminal profiles

S1a — capability plateau (~0.09). Gate-state: both shut, capability stalled. Signature: capex unwind dominates 2027–29 (investment recession, −1 to −2pp), but the deployment overhang keeps diffusion running for 3–5 years — existing capability is unexploited (the J-curve intangibles are still being built). D-class labor pressure therefore persists through the bust. Discriminator vs S1b: the horizon curve itself, nothing else.

S1b — valuation bust (~0.09). Gate-state: both shut, capability fine; financing broke. The solvency test (inference revenue vs. depreciation treadmill) fails systemically; the credit event lands on the SPV/private-credit structures (§4.2). Signature: capex halt, credit stress, resumption late-decade. Hard-coded: not labor-bullish. Cost pressure in a bust accelerates automation adoption; deployed systems are not un-deployed. Capital drawdown and D-class labor drawdown co-occur. (Propagated to §7.)

S2 — Baumol-institutional (~0.28, modal). Gate-state: I-gate holds; P-gate shut within horizon. The §3 engine at full expression: front-loaded 3–5% growth decaying per R1; relative-price divergence; household Leontief; I-queues; land share rising per R2. Two sub-paths (steady-capability vs. plateau-overhang) indistinguishable in the national accounts — τ-rows only.

S3 — compressed industrial revolution (~0.13). Gate-state: ≥1 gate opens coherently under steady capability, in two variants sharing one decade signature (routing note above). P-variant (the drafted identity): F3 at scale, ≡ mass actuation before ~2035; I-gate partially holds. I-variant (v3): mass F2 — coherent wedge collapse without robots; growth arrives from rent collapse rather than actuation. Both print: high-single-digit growth; labor share falls hard; the corridor debt-deflation risk from §4 is live here specifically — disinflation against nominal debt before redistribution mechanisms exist; Wave-B politics arrives mid-transition, not after it. Discriminator between variants: κ-2 vs ι-1 — which row moved first.

S4a — recursive-but-bottlenecked (~0.19). Gate-state: capability recursive; at least one gate shut, or takeoff throttled. Three inflows: both-gates-shut (pure bottleneck), I-gate-open-P-shut (cognitive abundance, physical scarcity), and throttled takeoff (the bottleneck is political). Signature: the rentier decade — extreme rents on complements (energy, compute, entitled land, permission), measured growth high but concentrated, wage share collapsing without an employment collapse. This is the terminal scarcity theorem (§3.2, R2) made flesh: recursive capability with unopened gates converts productivity into X- and I-rents rather than into broad income. With O3 active (§6.4), this terminal carries the name the source correspondence gave it: computational rentierism — cheap cognition for the population, extraordinary cognition for institutions, formal rights intact while effective power migrates into models, datasets, thresholds, and optimization functions that ordinary political visibility barely reaches; arrived at incrementally, legitimately, administratively, one efficiency improvement at a time.

S4b — unbound takeoff (~0.09). All gates open, unthrottled. Conventional-terms analysis expires; only pre-committed structures matter — ownership breadth, compute governance, constitutional constraints set in advance (§9). The framework’s contribution here is knowing when it stops applying (τ-2 tripwire). Mass note (v3): the rise from 0.06 is the mechanical product of the Layer-1 re-anchor with the throttled split deliberately held at 0.45/0.55 — see the netting caution in the re-anchor log; O3’s promotion would move mass from here into S4a.

S6 — split world (~0.08, promoted v3). Gate-state: divergent across blocs — the same gate open in some jurisdictions and shut in others for two-plus years, under live (steady or recursive) capability. This was “substantially the residual’s content” (D.6.4) while §5.3 called it the likely path; the tension is resolved by naming it. Modal content: China’s administrative F2 fires while insurance-priced and statutory jurisdictions hold (the D.4 route); or a US insurance-flip against an EU statutory lag. Signature: no single world-state to name — per-bloc terminals (US S4a-posture, EU S2-plus, China conditional-fast, India S2-minus, D.2–D.5) with the gate gradient itself becoming the economic object: capital and talent arbitrage it (§7.5 havens and enclaves), actuarial import (D.6.1) leaks across it, and domestic politics is organized by cross-border comparison — visible foreign abundance against domestic holds. The coupling force pulling back toward coherence is actuarial import; the wedge force pushing apart is sovereignty over the I-gate. Discriminator: G-2 — jurisdictional F2 events diverging, sustained. [F]

Hybrid/residual (~0.07). Structurally honest leftovers, slimmer after S6’s promotion: gates opening in 2035–36 (decade straddles two signatures); plateau-with-gate-opening oddities; within-bloc mixtures no single profile fits.

6.3 Sensitivity

If you believe X, the tree says Y

Belief Mechanical consequence
P(recursive) = 0.50, not 0.35 S4a+S4b: 0.27 → ~0.39; S2 falls to ~0.21. Note what this does not do: it does not make S4b modal. Capability optimism flows overwhelmingly into S4a — see below
I-gate fragile (insurance flips cascade, F2 fast) steady-row shifts toward one-open: S2→S3/S6 transfer of ~8–12pp; the modal decade becomes contested
κ relief (robotics scales early) recursive one-open split shifts toward P-gate (S4a→S3); steady both-open rises; S3 becomes strong second mode
Throttle strong (ρ wins the race with ν) S4b 0.09 → ~0.05; S4a absorbs. Throttle cannot reduce S4a — it feeds it. This is also the direction O3’s promotion would move the numbers (re-anchor log, above)
η_D ≫ 1 (Jevons wins) No mass moves. Demand beliefs re-color terminals (D-employment within S2/S3 softens) without moving gates. The tree is a gate model; not all disagreements are tree disagreements
Systemic-bust probability 0.28 → 0.45 S1b ≈ 0.14, S2 ≈ 0.23; the bust becomes a co-modal outcome — this is the single conditional doing the most work in the S1b/S2 boundary
Divergence conditionals doubled (0.35/0.25 → 0.70/0.50) S6 ≈ 0.16; S3 thins to ~0.07. The feedback-vector’s Split World mass (0.13) sits between baseline and this row — a conditionals disagreement, not a structural one, now that S6 exists

The four-coins result, restated on this tree

Grant Musk certainty on capability: P(recursive) = 1. Then P(S4b) = P(both gates | recursive) × P(unthrottled) = 0.45 × 0.55 ≈ 0.25. Total capability certainty buys abundance a one-in-four chance; the other three quarters go to the rentier decade, to S3, and — via the coherence conditional — to the split world. The remaining coins — embodiment, permission, politics — are not his to flip, and no value of τ flips them for him (§3.2, R2: the I-gate is orthogonal to capability by construction).

Robust vs. fragile conclusions

Robust (hold across ≥0.85 of mass): - D-class labor disruption proceeds — every terminal including both busts (the capex cycle and the labor cycle are decoupled) - Relative-price divergence (R-type deflation) — all terminals - Land/X income share rises — all terminals with positive growth (R3) - The insurance/licensing margin is pivotal — it is the S2/S3/S4a/S6 boundary - Energy buildout continues — all except deep S1a

Fragile (flip on one conditional): - All growth point estimates; all dated windows - The identity of the decade’s central political grievance: stagnation (S1), queues (S2), displacement (S3), rents (S4a), or cross-border comparison (S6) — materially different politics, one layer apart

6.4 Overlays

Three standing overlays (O3 added in v3; the count change is logged in §0.1). Overlays modulate timing and politics within terminals; they do not move terminal mass — until promoted.

O1 — Macro-cyclical. Ordinary recessions; financing wobbles short of systemic; and the energy-price squeeze (~2026–29): buildout-driven power costs hitting households before any TFP arrives — stagflationary in feel, temporary in structure, politically potent. Illustrative: P(≥1 year of material power-price political salience by 2029) ≈ 0.35–0.50 [A]. Primary effect: accelerates Wave A, which feeds ρ, which pressures the I-gate shut — the overlay leans on the tree without entering it.

O2 — Geopolitical model-breakers. (i) Taiwan/supply shock: dominates every parameter simultaneously; handled by hedging posture (inventory, geographic diversification), not by mass reallocation — a probability weight on a model-breaker is false comfort. (ii) Chinese lithography breakout: re-conditions κ for the China bloc and — via race dynamics — suppresses domestic throttle (raises the unthrottled split, S4b up slightly). (iii) Compute nonproliferation regimes: a standing institutional layer on ν.

O3 — Computational governance (v3). The directional I-gate (§3.1) as a standing modulation on every terminal: institutions — the state first among them — adopting AI for inference on persons faster than the economy is permitted to adopt it for them, because the sovereign carve-out (§3.2) exempts self-deployment from the min-function and the selection asymmetry favors control-enhancing applications. Within-terminal effects: hardens S2’s queues (continuously recomputed eligibility replaces episodic review); sharpens S4a into computational rentierism (profile annotation above); accelerates R6’s counter-current everywhere; and in S6 the I-in gradient across blocs becomes its own divergence axis. The overlay’s graded prior ladder, registered as [A] from the adjudicated feedback: AI-assisted administration by the mid-2030s extremely likely; continuous computational governance in several major domains likely; cross-domain citizen modeling materially plausible, jurisdiction-dependent; a single universal citizen score much less likely — and unnecessary (federated per-institution models gate different doors without any central system to oppose). Counterweights the overlay must carry honestly: the state-side gate has already fired ex post at full scale (E15–E17 — scandal, administrative law, courts), and the state funnel has longer middle stages than the private one (legacy data, procurement, administrative law, judicial review) — exposure-stage evidence like the OECD 97% figure (E18) is not deployment-stage evidence. Falsifiers M17–M18. Monitored on ι-4. Promotion test: promote into the tree iff statutory I-in/I-out divergence appears in ≥2 major jurisdictions, or state ADM measurably shifts the steady-row gate conditionals (ι-4 read against ι-1/ι-2). Expected promotion direction, pre-registered per the netting caution (§6.2): strengthen the recursive×both throttled split (S4b → S4a) and harden the steady-row I-out holds.

Promotion rule. An overlay is promoted into the tree if and only if it changes the distribution of Layer-2 gate outcomes, not merely their timing. Worked test: if data-center moratoria spread such that annual interconnection additions fall for two consecutive years, the energy squeeze has stopped being timing and started being a κ-shifter → promote into the steady-row conditionals. Until that test fires, it stays an overlay.

Updating mechanics — likelihood-ratio bands, indicator bundling, and the worked propagation example — live in §10; the tree’s job here is to be the thing §10 updates.

§7 — Capital Implications

Status: positive analysis of capital flows and asset-risk structure, not allocation advice. Consumes §3 (classes, flows), §4 (R5 straddle, six channels, treadmill), §5 (R7 stack, R9 posture), §6 (terminals). Feeds §8 (trades/buildout labor demand), §10 (uses existing F/M/κ/G rows; adds G-2), §11. Results continue: R10–R11.

7.1 The collateral hierarchy and the reversal symmetry

The hierarchy

Lenders and owners should price class mix, not “AI exposure”:

Class Canonical assets Collateral quality Dominant risk Worst terminals for holders
D GPUs, model weights, undifferentiated software/content businesses melting — 3–5y economic life, shorter competitive life financing/obsolescence: the next capability vintage S1a/S1b (impairment); S3/S4a (velocity of obsolescence)
P power generation, grid equipment, shells, robots, PPA streams durable — 15–50y supply response: overbuild and stranding when buildout succeeds late S2 (crowding); post-F3 S3 (the robots that opened the gate get overbuilt next)
I licenses, indemnification franchises, audit/fiduciary brands, regulatory moats politically contingent — durable until repriced in one event wedge collapse: F2, one insurance flip away S3, S4a-with-I-open
X entitled land, attention platforms, provenance/originals, positional franchises structural — cannot be manufactured political: taxation, zoning reform, confiscation any terminal where χ finally activates

R10 — the reversal symmetry

Read the risk column against §3.1’s constraint definitions and a symmetry appears:

Each class’s dominant asset risk is the reversal of the mechanism that makes it scarce. D is scarce because capability is new → its risk is newer capability. P is scarce because buildout is slow → its risk is buildout succeeding. I is scarce because permission is withheld → its risk is permission granted. X is scarce by position → its risk is politics, because position is the one scarcity only politics can reach. [I]

Two consequences:

  1. There is no scarcity without embedded reversal risk. “Buy the bottleneck” is always simultaneously “hold the reversal.” The four reversal mechanisms are different in kind — technical, economic, legal, political — so diversification across classes is diversification across reversal mechanisms, which is more robust than diversification across sectors (sectors mix classes per §3.1 and co-move within a reversal).
  2. Corollary — the terminal tax target. R2 says value terminally pools in X + residual-I. R6/R9 say χ activates late but eventually. Put together: when the fiscal channel finally moves, it must aim at X — land, positional wealth — because that is where the model says the value will be sitting. The safest collateral in the hierarchy is also the terminal tax base. Holding X is holding the eventual object of the distribution fight. [I]

Channel 6 (§4.2) slots in here as the P/D cross-term: the good scenarios strand legacy assets — chips by chips, and offices by D-automation. The CRE-short thesis is an S3/S4a thesis, not an S1 thesis; in the busts, offices survive and GPUs don’t.

7.2 Posture: operationalizing the straddle

The two-tail custody problem

R5 said the credit system is short a straddle on capability. §7’s question is who holds which tail — and the answer explains why the straddle persists unhedged:

Tail Exposure Held by Do they know?
Disappointment (S1a/S1b) channels 1–3: cash-flow, collateral, refinancing private credit, insurers, SPV paper (pensions sit in the unquantified institutional residual — not BIS-named, per C.2) mostly not — marked rarely, unstressed against correlated GPU impairment [E]
Success (S3/S4a) channels 4–6: Fisher, fiscal, stranding sovereigns, CRE holders, legacy corporates, sellers of undifferentiated cognition not as AI risk — held as duration, real estate, “stable” services equity

The two tails are held by different institutional populations with mandates that prevent holding the other side; each population is the other’s natural hedge, and no mechanism connects them. The straddle is a coordination failure wearing a market structure. [I]

R11 — the corridor asset paradox

Cross-map the collateral table against tree mass and one profile dominates. P/X-complements — firm power, interconnect-adjacent entitled land, shells, grid equipment, long PPAs:

  • pay in S2 (buildout grinds on; Baumol pricing; R3 land drift) — 0.28;
  • pay maximally in S3/S4a (bottleneck rents; the rentier decade is their decade) — 0.32;
  • and are the senior, recoverable collateral in the S1 workouts — 0.18.

That is a positive-or-protected profile across ~0.78 of tree mass — still the only asset class with one, plus whatever share of S6 resolves toward the holder’s bloc building (the corridor asset is bloc-local by nature; in a split world it pays wherever the buildout runs). The v3 recomposition trimmed the profile from ~0.85 mainly by moving mass into S4b, where nothing conventional is safe. This is why §4.2’s collateral migration is already happening: lenders are discovering the corridor asset empirically.

But apply the standing caveat (scarcity ≠ returns) with full force, because here it has teeth three ways:

  1. Visibility → capitalization. The corridor asset is the legible trade; entry prices can fully capitalize the scarcity before the holder earns it.
  2. Visibility → ρ-magnetism. The same legibility makes it the political target: Wave A aims precisely here (rate cases, DC tariffs, “why does the data center pay less than grandma”). The corridor asset is the most rate-regulation-and-windfall-tax-exposed asset in the economy — R10’s I/X reversal risk arriving early.
  3. Crowding → κ-relief. Capital flooding into P relieves the very scarcity that made P pay — R10’s P-reversal, self-inflicted.

R11. The corridor asset is real, already being discovered, and self-eroding on a 5–10 year horizon through capitalization, regulation, and its own supply response. What survives the erosion is its X-component — land and position — which is R10’s terminal tax target. Every corridor road leads to X, and X is where the fiscal scissors are eventually pointed. [I]

The model’s honest summary of “safe assets in the transition”: there is a corridor, it is crowded-by-construction, and its terminus is the tax base.

What the posture looks like

Positive description, not advice: a corridor-long, tail-aware book holds P/X-complements bought before the migration completes (entry price is everything, per R11.1); treats D-collateral as consumable, not storable (underwrite to melt schedules, not useful life); prices I-moats with an explicit F2 haircut keyed to ι-1 (insurance pricing), not to statute; and buys the cheap tail insurance each institutional population is structurally ignoring — credit-side stress protection for the disappointment tail, duration/stranding protection for the success tail. The last item is where R5’s coordination failure makes hedges systematically underpriced. [I]

S1b hard-coding, capital-side: a bust does not un-deploy systems. Inference demand persists through the workout and migrates to the cheapest surviving capacity — so distressed-compute recovery values are utilization-driven, not replacement-cost-driven, and the workout winners are whoever controls power contracts under the stranded shells. (The labor-side counterpart is hard-coded in §8.)

7.3 Composition rotation, class-conditioned

v1’s four phases, upgraded with tree conditioning:

Phase Window Dominant flow Class signature Tree conditioning
1 — Training land-grab now–2028 chips, shells, power procurement D-capex atop P-complements universal until Layer 3 resolves
2 — Monetization squeeze 2027–30 inference infra + organizational intangibles D + the ι_int shadow the solvency-test window (F-1); S1b ≡ phase 2 failing systemically
3 — Bottleneck arbitrage 2028–33 whatever the shut gates make scarce P/X-complements + trust wrappers the S2/S4a core phase
4 — Actuation wave 2031+ robot fleets, reshored physical production P-gate opening S3-conditional only (~0.13 plus the recursive P-branch); do not finance phase 4 off phase 1 evidence

The intangibles shadow. Measured capex captures hardware (D+P). It misses the expensed organizational redesign — ι_int — that the electricity and IT transitions showed to be larger than the hardware bill and 5–15 years behind it [O, historical]. Two consequences: capex statistics overstate the meltable share of the boom and understate its durable share; and the modal-scenario equity winners are the reorganizers (within-sector productivity dispersion leaders, dashboard ι-int row), not the FLOP-owners. The market prices the visible half of the investment cycle; the invisible half is where S2’s returns actually accrue. [I]

7.4 Stack positioning

R7’s implication, operationalized: models are the commodity; the rent pools are L2 (compute/power), the trust wrapper, and L4 (distribution).

  • Vendor-financing circularity (§4.2) re-read: it is vertical integration by balance sheet — L1/L2 actors pre-buying the layers where R7 says the rents land. Judge those structures as stack-position acquisitions, not as revenue.
  • The indemnifier is the emerging I-class asset inside the stack. Whoever successfully wraps AI output in insured, compliance-grade liability cover owns a wedge (§3.1) at the choke point — and ι-1 prices its birth in real time. Trust assets generally (audit, provenance, brands) appreciate as content approaches free: verification is scarce exactly in proportion to generation being cheap. [I]
  • Sellers of undifferentiated cognition — generic software, content, analysis, advisory — are the D-deflation’s landing zone. Whether their prices or their margins absorb it is the M-2 wedge question (§5.1): in rent-capture worlds their margins survive longer than their valuations deserve; in pass-through worlds neither does. A material share of listed services equity is priced on human-scarcity assumptions the model says are melting [I] — this is the success-tail stranding (channel 6) in equity form.
  • Sovereign backstop front-running exists as an incentive and is already visible in spread behavior (§4.2 reflexivity): paper that becomes strategically guaranteed re-rates. Watch: strategic-asset designations, government compute offtake, DPA-style actions, sovereign co-investment. Noted as an observed flow driver, with the reminder that the expectation itself worsens R5’s stakes.

7.5 Cross-border: capital chases electrons and gates

The 20th-century gradient — capital → cheap labor — was substantially a D-class arbitrage (BPO, IT outsourcing: cheap cognition). That is precisely what melts. The new gradient has two axes:

  1. P-inputs: firm power, cool climate, permissive siting, transmission access;
  2. I-gate openability: jurisdictions whose permission structures can move (§5.3).

Consequences, carried and class-sharpened:

  • US as absorption machine (world savings fund the buildout; current account widens) and EU as funder without a frontier — high-ρ, low-ν, gates shut, capital exported into other blocs’ P-complements [F]. Politically corrosive at home, as v1 noted.
  • Gulf compute recycling: petrodollar → compute-for-security bargains; oil-for-security template with FLOPs as the new barrel.
  • China: parallel, capital-controlled system; binding constraint physical (chips), not institutional; the sovereign F2 option (§5.3) means the first mass I-gate opening may be administrative and Chinese — with different reliability incentives than an insurance-priced opening [F].
  • The EM ladder is kicked away (India IT headcount, Philippine BPO [E, pending verification]) — and the replacement rungs are narrow. Power-rich, institution-poor countries can host compute only as enclaves: sovereign-guaranteed zones with extraction-enclave economics — capital-intensive, employment-light, diffusion-poor. The development content of the new gradient is far lower than the old one’s. [F]
  • I-gate havens [F]: as F2 pressure builds, expect small-jurisdiction liability arbitrage — an “AI sign-off Delaware,” medical-AI domiciles, audit flags of convenience. Same logic as shipping registries: permission is a product, and someone will sell it cheap. Monitor under G-2 (cross-border gate/enclave tracker, added under the complexity budget as the row this section’s falsifiers need).

On r: v1’s +50–150bp drift is retained only as terminal-conditional* — §4.3’s contamination lemma stands, S4a plausibly prints low rates with a land boom, and the F-3 cell logic replaces any single-number forecast.

7.6 Geopolitics triad (carried)

  • Taiwan/supply shock: a P-shock to L2 that dominates every parameter simultaneously; handled as posture (inventory, geographic diversification of compute), never as a scenario weight — a probability on a model-breaker is false comfort (§6.4, O2).
  • Lithography breakout: κ-CN relief; compresses chip rents (bad for incumbent L2 margins); intensifies race dynamics → suppresses domestic throttle via ν (§5.3), nudging the unthrottled split.
  • Compute nonproliferation: a standing institutional layer on ν; if formalized, it makes compute a treaty-controlled commodity — the strongest possible confirmation of L2 as the strategic layer.

§8 — Labor Implications

Status: the framework’s arithmetic becomes people. Consumes §3 (employment corollary, flows, queues, incidence), §5 (b class map, Wave B fault line, g components), §6 (terminals), §7 (phases, S1b capital-side pairing). Feeds §9, §10 (L-rows, formalizing v1 dashboard items 7–8 — no net-new rows beyond re-labeling), §11. Results continue: R12–R14.

§3 classified output; §8 classifies workers. A career is a path through the class structure — the classes are not just output categories but career geography, and a worker’s decade is determined by which class their tasks sit in, which side of a gate they stand on, and where they are in a queue.

8.1 Employment by class: the absorption machine

The class employment table

Class Employment trajectory Mechanism Timing
D falling fast displacement at capability speed; reinstatement only via η_D’s new-category component (§3.3) underway [E: Canaries, with caveats below]
P rising through the buildout, then κ-gated phases 1–3 demand (§7.3); F3 reverses it in S3 now → ~2030+; terminal-dependent
I rising — the absorption machine Baumol drift + static productivity → employment share rises at least as fast as expenditure share (§3.2 corollary) continuous while the gate holds
X thin periphery service layer around positional consumption; low-productivity by construction grows with concentration

The waves were the gates

v1’s wave sequencing survives intact — re-derived, it turns out to have been the gate schedule seen from the labor side:

  • Wave 1 (routine tradable cognition, underway) = D-class, no gate required. Fires in every terminal, including both busts. Capability-only.
  • Wave 2 (structured mid-level cognition, “gated by trust, liability, licensing — not capability”) = wedge-priced D-work behind the I-gate. Fires iff F2. v1’s intuition that Wave 2 was institution-gated is exactly the I-gate, named before the framework existed.
  • Wave 3 (routine physical work) = P-class. Fires iff F3; S3-conditional (~0.13 plus recursive branches).

The tree therefore is the labor forecast: S2 is a one-wave decade, S3 is a three-wave decade, and S4a is a one-to-two-wave decade with the growth accruing elsewhere.

R12 — the absorption result

Where do Wave-1’s displaced go? Follow the arithmetic. I-expenditure share rises (Baumol drift, §3.2); I-productivity is static under protection; therefore I-employment rises — in review, compliance, administration, coordination, queue-servicing, human-sign-off roles. The displaced are absorbed into the protected sector’s inefficiency:

The modal labor outcome is not unemployment but absorption. The I-class is the labor market’s shock absorber, and its absorption capacity is proportional to its inefficiency. Aggregate employment holds; composition, mobility, wage quality, and cohort experience deteriorate. [I]

Status caveat, carried from B.3 and stated here where the result lives: R12 is predicted, not yet observed — the absorption trace (rising I-employment share, L-2) is a forecast awaiting data, and §8’s presentation of absorption as the modal outcome inherits that status.

This derives the “located gaslighting” mechanism: statistics say “strong labor market” because absorption is real employment; the lived experience says “every good door is closed” because the employment created is lower-productivity, lower-wage, queue-gated work. Both are true, and now for a stated reason.

The perverse implication — the module’s uncomfortable finding: cost disease is doing the welfare state’s job. The I-class’s inefficiency is functioning as distributed unemployment insurance, funded through I-class consumer prices rather than taxes (consistent with R9: off-budget redistribution is what a scissors-bound polity can afford). Consequence: F2 efficiency gains would fire the shock absorber. A large-scale wedge collapse doesn’t just strand I-incumbents — it destroys the absorption capacity currently catching D-refugees. This gives even non-guild actors — macro policymakers staring at R6 — a rational reason to tolerate the I-gate staying shut. The gate has more friends than the guilds. [I]

The contingency v3 adds, from the directional split (§3.1): the absorption machine assumes the state does not self-automate its own administrative layer. The absorber’s jobs — review, compliance, queue-servicing, eligibility processing, human sign-off — are exactly what I-in deployment under the sovereign carve-out eliminates, and no F2 event, no insurance flip, no statutory change is required: the permitter is the operator. R12’s scope is therefore conditional on I-in staying slow in the absorptive institutions specifically (health administration, benefits, education bureaucracy), whatever happens elsewhere. The falsifier consequence is M14’s attribution requirement (§11): an I-employment decline with the private gate verifiably shut does not kill R12 — it may be the carve-out working — and a falsifier that fires without attribution teaches the wrong lesson. Tracked on ι-4 against L-2. [I]

8.2 The squeeze zone and the escalator

The squeeze zone is a moving front

Definition: workers whose tasks are (a) D-class by last-constraint (no embodiment, no wedge protection), (b) currently above the capability frontier — not yet displaced, and (c) queued out of I-entry. Mid-level unprotected cognition: analysts, marketing and project management, mid-tier developers, coordination work.

The zone is not a place but a front that travels up the complexity gradient at capability speed — pressed from below by the frontier, blocked from above by shut gates and queues, holding no positional assets. Each capability advance converts the zone’s lower edge into Wave-1 displacement and recruits new members at its upper edge.

R13 — the unfunded tuition result

Junior D-work was always a bundled product: spot output plus option manufacture — the firm bought grunt work and got, as a byproduct, the manufacture of future I-capable seniors (judgment, context, client trust). The bundle was priced as one thing because the spot output covered the bill.

AI unbundles it: the spot value collapses toward inference cost, and the option side — never separately priced, never separately funded — loses its funding vehicle. Firms individually rationally stop buying (Canaries: adjustment through hiring, not wages [E]); collectively, the I-capable cohort of 2032–35 is simply not manufactured.

The escalator failure is an option-market failure: the tuition that junior work used to fund implicitly is now unfunded, and no actor’s individual incentive rebuilds it. [I]

Policy sequence, derived from the structure rather than asserted:

  1. Near-term — re-fund the tuition explicitly (apprenticeships, articling, residency-equivalents for cognitive fields, subsidized junior seats). This is the cheapest intervention in the entire framework: it is a market-completion move, not a transfer, and it operates through g while χ is scissors-bound.
  2. Medium-term — the seats fight: access to I-rents (below, §8.5).
  3. Long-term — ownership: if gates open (S3/S4a), tuition and seats both stop mattering and Ω/χ is everything (§5.2).

Evidence anchor with required caveats [E]: [updated] the Canaries result (~19% relative employment shortfall as of the August 2026 revision, up from the ~13–16% carried at drafting; 22–25-year-olds, exposed occupations) is directionally consistent and mechanically apt (hiring-margin adjustment). It remains the framework’s single most load-bearing contested datapoint, and the verification pass sharpened rather than settled it: the authors have retreated to calling the finding descriptive rather than causal; ADP’s panel is acknowledged to overstate the gap relative to national survey benchmarks; and the identification alternatives are handled within the paper (robustness checks excluding technology firms and computer occupations) rather than defeated. R13 should continue to be read as theory plus a widening trace, not as a measured effect.

8.3 Wage structure: barbell mechanics

From §5.2’s b class map: b_D → 0 (no guild — the control case), b_I high for incumbents only (the wedge is bargaining power; queue entrants get none of it), b_P temporarily high (shortage-driven). The barbell is the b map expressed in wages; the squeeze zone is its hinge, and the hinge moves.

The trades premium — real, dated, twice self-liquidating. Phase 1–3 buildout demand (§7.3) against a thin apprenticeship pipeline (~4–5y lag) produces genuine premium — the cleanest intersection of the capital and labor stories (the marginal data center bids for electricians). But it is the labor analogue of R11’s corridor asset, with the same erosion logic and a weaker profile: it pays across S2/S4a/early-S3 (~0.60 of mass, plus S6’s building blocs), stalls in S1a/S1b (buildout halts — unlike the corridor asset, which retains workout seniority, the corridor job has no seniority in a halted build), and is reversed in S3 by the very actuation it helped construct (R10’s P-reversal, labor edition). Plus the supply response: premium → applications → 4–5y later, compression. v1’s “high confidence through 2030” is formally downgraded to modal-conditional, peaking mid-window. [F]

The credential premium splits along the class boundary. “The college premium falls” is mis-specified. The model’s prediction: credential-as-I-gate-ticket appreciates (licensed tracks — the queue signature is rising demand for gate tickets, ι-2) while credential-as-D-skill-signal depreciates (generic degrees certifying capability-frontier-adjacent cognition). Evidence to date: mixed, sector-specific [E] — the split, not the average, is the observable. Enrollment prediction [F]: licensed-track applications up, generic-track down, simultaneously — which averages to “no clear trend” in aggregate enrollment data and will be misreported as stability.

Within-occupation variance rises even where between-occupation means look stable: superstars get levered by the tools; the competent-undifferentiated get commoditized toward inference cost. Row L-3.

8.4 Countervailing forces: R14 — the mistimed dividend

Demographics run against displacement in the rich world and China: shrinking working-age populations [O] absorb D-displacement, tighten P/I labor further, and make automation partially welcome (Japan as the limiting case). The EM world has the opposite configuration: the youth bulge peaks in South Asia and Africa exactly as the D-arbitrage ladder — BPO, IT services, the cognition-export rung — melts (§7.5).

The demographic dividend and the automation shock are geographically anti-correlated. Aging blocs get their transition cushioned by their own labor scarcity; young blocs arrive at the ladder to find it gone. [I]

Migration is the arbitrage connecting the two — aging-bloc P-class and care shortages against young-bloc surplus — which makes immigration policy a g-component with unusual economic leverage and maximal political cost. The framework notes the lever exists and that R9’s logic (off-budget, restriction-biased politics) predicts it stays mostly unpulled. [F]

Terminal victim profiles (sharpening §6’s grievance table into constituencies):

Terminal Worst-hit labor constituency
S1b entrants — D-displacement continues (cost pressure accelerates automation; deployed systems are not un-deployed) while the trades stall: both growth ends of the barbell close at once. The bust is the entrants’ worst terminal, not their reprieve — hard-coded, labor side
S2 the queued — absorbed, underemployed, paying I/X prices (household Leontief)
S3 mid-career incumbents in newly-opened wedge work — Wave 2 fires into the cohort that thought it was safe
S4a everyone without assets — the distributional worst case; wage share collapses without an employment collapse to organize around
S6 bloc-dependent — holdout blocs get S2’s queued plus visible foreign abundance (the politically combustible mix, D.3); opened blocs get S3’s mid-career incumbents

8.5 Wave B operationalized: doors vs. seats

Wave B’s fault line (§5.3) — incumbents want restriction that closes doors; the queued want protection that opens seats — resolves per-profession. Prediction rule [F]: seats open where (shortage severity × state fiscal interest) exceeds incumbent rent-defense capacity.

Profession Shortage Incumbent defense Prediction
Accounting/audit severe, pipeline already failing [E] weakening (150-hour rule relaxations already visible) seats open — the archetype
Nursing/allied health chronic, state-salient fragmented seats open
Medicine severe + aging demand strong, but residency caps are state-funded chokepoints hybrid: scope expansion (NP/PA) + partial residency opening
Engineering (PE) buildout-driven (Wave A/g) moderate seats open moderately
Therapy/counseling demand boom weak seats open + AI-adjacent tiering
Law none — incumbent oversupply at top strong guild doors close — the archetype

The table is the σ_I micro-forecast: each row is a falsifiable call, and the two archetypes (accounting opens, law closes) are the cleanest tests. Monitored under ι-2/Λ-B (bundled per §10).

What the statistics will show (v1’s table, class-derived):

Indicator Modal expectation Row
Headline unemployment cyclical 4–6%; no AI signature (R12) —
Hires + quits rates the real signal: frozen market — low firing, very low hiring; displacement via non-replacement L-2
22–25 vs 35+ employment ratio, exposed occupations continued divergence L-1
Youth underemployment rising; degree-holders cascade down the ladder, crowding non-degree workers below L-1
I-class employment share rising (the absorption trace) L-2
Labor share −2 to −5pp (S2); −8 to −12 (S3/S4a) [F] δ-1
Median real wage flat-to-modest; cohort-deflator-dependent (§4.1) — the median household’s number and the median entrant’s number diverge F-4
Within-occupation wage variance rising L-3

One carried v1 line, inverted properly: the gardening analogy (“work becomes an artisanal hobby”) describes X-periphery and human-premium I-work as a jobs category, not a hobby — care, hospitality, coached experience. In the modal terminal, “optional work” is a class of employment that the non-optional economy’s winners purchase. Work does not become optional; whose work is optional becomes visible.

§9 — Synthesis: Four Clocks, One Economy

Status: synthesis; deliberately short. Consumes everything; adds R15. The module-by-module findings converge on three sentences, stated here once each.

9.1 The clocks

Each class runs on exactly one clock, fixed by its last unrelieved constraint (§3.1):

Class Clock Period Fast hand
D capability months model releases
P buildout years interconnection queues; apprenticeship and production lags
I institutional decades insurance repricing — the slow clock’s one fast hand (ι-1)
X none — —

The political clocks — Wave A (2026–28) and Wave B (2028–32) — are not a fifth class; they are inputs to the institutional clock, and they mostly wind it backward (ρ up).

Every named pathology in this framework is a clock mismatch. The squeeze zone: the D-clock outrunning the I-clock through a workforce. The solvency test: capital melting on the D-clock financed on the P-clock. The Baumol trap (R4): one monetary instrument facing four clocks. The fiscal scissors (R6): displacement on the D-clock, legislation on the I-clock. Default Ω-dominance (R8): distribution decided by channel speed. The absorption machine (R12, predicted rather than yet observed — L-2): the I-clock refusing to accelerate — which is precisely what lets it absorb the D-clock’s output. The unfunded tuition (R13): careers built on decade-clocks funded by spot work on the month-clock. The mistimed dividend (R14): the demographic clock anti-phased across geographies. The decade’s story is not any clock’s speed; it is their desynchronization.

9.2 The convergence

Three sentences the modules kept arriving at independently:

  1. The transition’s fights are not about AI; they are about who stands where relative to the two gates that don’t open on capability’s schedule. (Capability opens neither the P-gate nor the I-gate; §3.2, R2.) The I-gate alone prices the wedge — roughly 5–12% of GDP, lower half better supported (§3.4, A.3) — so the fight is not abstract.
  2. Capital and labor are converging on the same coordinates from opposite directions: capital migrates toward X and the trust wrapper ahead of the taxman (R10–R11); labor queues at the I-gate while the welfare function hides inside cost disease (R12, §3.5). Both flows are the terminal scarcity theorem expressed in people and money.
  3. The safest asset and the terminal tax base are the same asset (R10 corollary) — which is the framework’s sharpest political-economy prediction: the endgame fight is pre-addressed, and both sides are already moving into position without naming it.

9.3 R15 — the deployment ratchet

The drafting process logged three feedbacks the tree could not natively hold: the central bank as an S1b path (§4, note 1); Λ endogenous to the displacement it governs (§5, note 2); the absorption constituency defending the I-gate beyond the guilds (§8, note 1). Inspect their directions: every one pushes mass toward the bound branches — toward S1b and S2, away from open-gate terminals — while ν keeps capability itself hot.

R15 (restated in v3). Unmodeled feedbacks discovered during construction were, with the exceptions logged below, deployment-suppressing and never capability-suppressing. The framework’s residual bias therefore has a sign — but the sign claim holds for power-dispersing deployment only: the ratchet suppresses deployment that dissolves institutional positions; it can accelerate deployment that strengthens them. Unqualified, R15 likely underestimates bound and rentier outcomes (S2, S4a) and overestimates clean resolution (S3, S4b). Quarterly re-conditioning should carry a standing ratchet audit (§10): if newly logged feedbacks keep pointing bound-ward, revise the steady-row gate conditionals themselves, not just the layer priors. [I]

The counter-column now has two classes, and they differ in kind. Exogenous: actuarial import (D.6.1) — administrative deployment in a fiat-liability jurisdiction generates exactly the loss experience Western insurers lack, so foreign failure data reprices ι-1 and erodes the domestic wedge with no domestic political event required; the ratchet is a property of feedbacks internal to a polity, and this force breaks it from outside. Endogenous (v3, from the adjudicated feedback): institutional self-strengthening deployment — the selection asymmetry of §3.1: control-enhancing applications face stronger adoption incentives than power-dispersing ones, because the approver profits from the former and is threatened by the latter. This class is large, domestic, and structural — R15’s original drafting treated the counter-column as exotic and foreign when its biggest member was sitting inside the sovereign carve-out. The corrected reading: the ratchet points bound-ward for the governed economy while pointing deployment-ward for the governing institutions — which is R9’s S4a machine and O3’s content stated as an audit rule.

R15 is also R9 restated as an epistemic result: the political machine that manufactures S4a has a counterpart inside any model built by observers embedded in that machine.

9.4 The verdict on Musk, scored

Claim Framework adjudication
P1 — capability > summed human intelligence ~2031 A pure τ-claim; the tree prices it (recursive = 0.35, v3) and monitors it (τ-rows); base-rate multiplier says mid-2030s+
P2 — “Stockfish” in all cognitive work Funnel error (§1): benchmark stage-1 evidence for a stage-5 claim; true only of unwedged D-tasks
P3 — robots at scale before 2036 The P-gate; ~S3-conditional (≈0.13 + recursive branches); the framework’s least-instrumented parameter
P4 — quasi-infinite output False by construction in its strong form: requires all gates open, and X never opens (R2). The defensible steelman — abundance in D∪P subsistence — survives in S3/S4b
P5 — constraints are electricity and chips Half right: physical constraints are real and binding (κ); wrong that capital and institutions aren’t constraints — κ is capital-through-institutions, and the I-gate is the deeper one
P6 — “Treasury issues checks” A χ-claim with no mechanism, contradicted by R6 (the base erodes when the need arrives) and R8 (χ is the slowest channel; Ω is the fastest)
P7 — “deflation will be the issue” One-fifth right (§4.1): R-type certain, C-type the actual danger, A-type a policy choice — and the consistent version of P7 requires the activist central bank that refutes P8
P8 — “money won’t matter in 2036” Inverted and now derived: money collapses onto claims over position and permission (R2), i.e., it matters more, and more concentratedly, than at any point in the postwar era

The v1 verdict — reliable observer of physical constraints, poor modeler of social technologies — stands, with two codas.

From the funnel: the error term has a scale. §2.7 located the timeline multiplier in funnel length; §1.2’s Stockfish tell completes it. Chess is the domain with a zero-length funnel, which is why “Stockfish level” is exactly true there and nowhere else; rockets have a short funnel — integration and physics, no permission layer; regulators, insurers, and licensing boards are the long-funnel case. The same forecaster being reliably good at hardware dates and reliably wrong about institutional ones is therefore not two talents but one error term scaled by the number of stages between feasibility and autonomy. The verdict above is that single observation read at two funnel lengths.

From the tree: by the four-coins result, capability optimism mostly purchases the rentier decade, not abundance. If Musk is right about τ, the modal consequence of his being right is the world least like the one he predicts.

§10 — Dashboard

Status: the framework’s error-correction mechanism. The probabilities in §6 are scheduled to be wrong; this is how they improve. Cadence: quarterly re-conditioning; annual conditional-structure review; ratchet audit each pass (R15).

10.1 The row registry

Twenty-one rows in eight clusters (ι-4 added in v3, bundled); correlated rows bundle into ~12 effective evidence units (a bundle counts once per quarter, per the discipline below).

Row Indicator Updates Bundle
τ-1 METR horizon doubling time + frontier benchmark suite + agentic-coding evals Layer 1 one unit (all τ-capability measures are correlated)
τ-2 Demonstrated AI-R&D compounding with measured loop Layer 1 recursive; S4 tripwire standalone
η-1 Jevons Meter B: constant-capability price × paid production volume; saturation check terminal coloring (not gates) standalone
ι-1 AI liability-insurance pricing: products, premium ratio vs human-equivalent I-gate (F2’s fast hand) standalone — the framework’s best leading indicator
ι-2 Queue + doors/seats ledger: applications-per-seat, licensing bills, scope-of-practice, §8.5 profession table I-gate (F1 side) bundled with Λ-B/ι-4
ι-3 Within-sector firm productivity dispersion (ι_int) realized β inside S2 standalone
ι-4 State ADM tracker: government AI procurement, algorithm registries (Amsterdam/Helsinki precedents), explanation-rights litigation, personalized-pricing regulation, ADM-scandal events I-in gate (§3.1); O3 promotion test; M14 attribution; M17 bundled with ι-2/Λ-B (correlated politics)
κ-1 TWh added; interconnection queue duration; transformer/turbine leads; storage costs P-inputs / buildout one unit
κ-2 Robot $/hr vs 2× median wage; units shipped (production, never demos) P-gate (F3) standalone
F-1 Solvency gauge: inference revenue ÷ (annualized replacement capex + debt service) Layer 3 systemic conditional bundled with F-2
F-2 Credit confessions: DC-paper spreads; depreciation-schedule revisions; secondary GPU prices; covenant collateral-mix drift same with F-1
F-3 Rates × land cell + tiebreakers (rent-vs-wage share, F-4 width, τ-rows) terminal discrimination standalone
F-4 Class-inflation dispersion thermometer (π_D/π_P/π_I/π_X); cohort deflators when built R4 trap; S2/S4a tiebreak standalone
F-5 Target-index debate intensity φ regime; proxy for trap operating standalone
M-1 Per-layer concentration: open-weight lag, top-3 cloud share, indemnifier count, app-layer margins R7 migration one unit
M-2 Pass-through wedge: consumer AI-service prices vs API cost index live π estimate standalone
Λ-A Wave A ledger: PUC dockets, DC tariffs, siting moratoria (promotion test attached, §6.4) ρ/κ; O1 politics standalone
Λ-B Wave B ledger: χ bills (transfers, capital tax, sovereign stakes) χ activation; R9 falsifier with ι-2/ι-4
G-1 Lithography milestones; export-control regime; compute treaties O2; ν one unit
G-2 Cross-border gates: jurisdictional F2 events, enclave deals, I-gate-haven registrations S6 divergence conditionals + residual; §7.5 with G-1
L-1 / L-2 / L-3 Cohort ratio (22–25 vs 35+, exposed occs); absorption trace (hires/quits freeze + I-employment share); within-occupation wage variance R12/R13 in progress one unit (common ADP/JOLTS drivers)
X-1 Housing decomposition (structure vs land+entitlement share); land share of national income R3; test-asset prediction standalone

10.2 Updating discipline

LR bands (judgmental, logged, pending calibration): weak ≈ 1.5 · moderate ≈ 3 · strong ≈ 10.

Rules: 1. A bundle contributes one evidence unit per quarter, at the strength of its strongest member — correlated indicators are one observation wearing several hats. 2. Discretion lives only in the band assignment, and every assignment is logged with rationale. Propagation is mechanical: update layer odds by LR, renormalize, recompute every terminal mass through §6’s conditionals. No hand-adjusting terminals. 3. Overlay promotion strictly per §6.4’s test. 4. Ratchet audit (R15): log any newly identified feedback with its direction, in two columns. Bound-ward (the standing bias): three consecutive bound-ward quarters trigger review of the steady-row conditionals themselves. Counter-column (deployment-accelerating), two entries per R15’s v3 restatement — actuarial import (D.6.1), exogenous: foreign fiat-liability deployment repricing domestic ι-1 without a domestic political event; tracked on G-2, read against ι-1. Institutional self-strengthening deployment (§3.1, §9.3), endogenous: control-enhancing state/institutional adoption under the sovereign carve-out; tracked on ι-4, read against ι-1/ι-2 and L-2. Audit rule for the second entry: it counts counter-column for deployment volume but bound-ward for the private I-out gate — one observation, two ledger directions; do not double-count it as good news.

10.3 Worked example

Evidence: τ-1 flattens two consecutive readings. Assigned likelihoods: P(E|plateau)=0.6, P(E|steady)=0.2, P(E|recursive)=0.06 — i.e., moderate against steady, strong against recursive. (Recomputed on the v3 priors and partition; the direction is the mirror of the E7 re-anchor, which is the point — the machinery runs both ways.)

Layer 1: (0.18, 0.47, 0.35) → joint (0.108, 0.094, 0.021) → posterior (0.48, 0.42, 0.09).

Terminals, recomputed mechanically:

Terminal Prior Posterior
S1a 0.09 0.232
S1b 0.09 0.077
S2 0.28 0.352
S3 0.13 0.088
S4a 0.19 0.050
S4b 0.09 0.023
S6 0.08 0.051
residual 0.07 0.126

Readout — and this is why the example earns its space: one moderate-strong observation nearly triples the plateau-bust tail and collapses the recursive tails by ~73%, while the modal terminal moves seven points (0.28→0.35). The tree is tail-responsive and mode-stable — which is the correct epistemic shape for a framework whose robust conclusions (§6.3) were deliberately chosen to be mode-independent. The machinery runs identically in reverse if the curve re-steepens.

10.4 What the dashboard cannot see

Intra-quarter cascades (§6.1’s stated lag); the W-wedge (welfare is not on any row — measurement debt, §12); non-US gates (G-2 is a sketch); and task birth — every labor row measures displacement; none measures reinstatement (§12’s largest blind spot).

§11 — Falsifiers

Status: kill conditions, not surprises. A surprise moves priors through §10; a falsifier retires a named claim. Direction matters: at least one falsifier below is an outcome to hope for.

Gate-state verification (v3). Several conditions below turn on “gates verifiably hold,” which was previously unoperationalized — an immunization risk, since an unfalsifiable gate-state makes K1/K3 and M14 undischargeable. Definitions, on existing rows:

  • I-out gate holds iff all three: (i) ι-1 premium ratio — AI professional-liability coverage priced at ≥ 1.0× the human-equivalent premium in the top licensed lines (no sustained sub-parity pricing for 2+ consecutive quarters); (ii) no statutory removal of mandatory human sign-off in any of the top-10 licensed professions by I-value-added (ι-2 ledger, §8.5’s table as the census); (iii) no mass F2 event, private or jurisdictional, on G-2.
  • P-gate holds iff κ-2 stays below threshold: robot $/hr above 2× median wage, and production shipments (never demos, per E10’s lesson) below the mass-actuation floor.
  • I-in erosion (for M14’s attribution): documented state ADM deployment displacing administrative headcount on ι-4 — procurement records, algorithm-registry entries, agency staffing series — with no corresponding private F2 event.

A falsifier that cites gate-state without these observables has not fired.

11.1 Tier 1 — framework kill conditions

# Observation What dies What survives
K1 D expenditure share rises persistently (2+ yrs) despite D-price collapse σ<1 — the Baumol engine, share drift, R1–R3, the S2 case the funnel critique; the gates as deployment brakes; the straddle (the six channels are engine-independent)
K2 I-class prices fall at scale with no legal or insurance change the I-gate was imaginary; institutional scarcity was fake; R2 halves (X-only); R9/R12 gutted P-gate logic; financial module; X-class results
K3 Sustained measured growth >5% while D-share falls and gates verifiably hold the accounting — class boundaries or measurement mis-specified re-derive Appendix A before trusting any share arithmetic

K2 is the deepest: it is the observable form of “does scarcity migrate again?” — the question the entire framework reduces to.

11.2 Tier 2 — module falsifiers (consolidated from codas)

# Observation Kills Src
M1 Queue signature absent: premiums rise but applications-per-seat don’t §3.5 queue model §3
M2 First mass F2 event not preceded by insurance repricing ι-1’s leading-indicator status §3
M3 Sustained A-type deflation while CB actively targets 2% “A-type is a policy choice” §4
M4 CB tightening materially moves I-class inflation monetary inertness of I (R4’s premise) §4
M5 Systemic credit event with no prior F-1 deterioration the solvency-gauge specification §4
M6 Open-weight parity ≥12 mo with rents still pooling at L1 R7’s migration claim §5
M7 Broad consumer pass-through despite concentrated L4 the choke-point rule §5
M8 Major χ legislation enacted pre-2030 despite the scissors R9’s routing claim — the good-news falsifier §5
M9 Corridor assets broadly underperform in confirmed-S2 R11’s profile (wrong, not early) §7
M10 GPU secondary values hold through demonstrated plateau melting-collateral spec §7
M11 EM D-services exports resume durable growth through 2028 ladder-melting thesis §7
M12 L1 pure-play sustains pricing power 3+ yrs against open-weight parity R7 at the layer that matters §7
M13 Junior hiring rebounds in exposed occs while capability advances R13 (unfunded tuition) §8
M14 I-employment share falls while the I-out gate verifiably holds — attribution required (v3): via private F2 → kills R12 as specified; via state self-automation (I-in erosion, ι-4) → confirms the §3.1 carve-out mechanism and narrows R12 instead. Unattributed, the falsifier teaches the wrong lesson R12 (absorption) §8
M15 No trades premium through a confirmed phase-1–3 buildout barbell mechanics §8
M16 Law opens seats while accounting closes doors §8.5’s prediction rule (inverted) §8
M17 State ADM deployment stalls or reverses after scandal-class events (E15–E17 as the historical calibration) across ≥2 major jurisdictions, despite capability gains the selection asymmetry (§3.1) — kills or bounds the endogenous counter-column (R15), and O3’s default-acceleration premise with it §9
M18 Citizen-side adversarial agents achieve documented parity in contested decisions (appeals, negotiations, eligibility challenges) at scale the asymmetry thesis’s strong form — institutional inference advantage as durable, rather than a transitional data-position gap §5

Already-live tests: M16 (both professions are moving now), M2/ι-1 (products exist to price), η-1’s saturation check, M6 (open-weight lag is measurable today), and M17’s baseline (the state-side gate has already fired ex post three times — E15–E17 — so the bound it tests is calibrated by history, not hypothesis). The framework starts accumulating a track record immediately — by design.

Graceful degradation: the modules were built with distinct load paths. K1 kills the growth story but not the credit story; K2 kills the political economy but not the buildout economics; most Tier-2 losses amputate one named result. The one non-redundant member is §8’s empirical anchor: if Canaries falls to its identification critiques, R13 loses its only direct observation and reverts to theory-plus-traces — logged in §12, not here, because a study failing is a surprise, not a falsifier.

§12 — Open Questions and Residuals

Status: the honesty register. Three parts: questions with closure paths, structural residuals without them, and the deferred-work ledger.

12.1 Open empirical questions

Question Why it matters Closure path Horizon
η_D magnitude (routine vs new-category weights) decides whether D-labor pessimism is front-loaded or permanent η-1 Meter B + saturation check 2026–28
σ_I decay under fiscal pressure R6 vs R12 collide: the scissors-bound state needs the growth an open gate would buy, but the absorption machine is its off-budget welfare system. Does it eventually sell gate-openings for tax base? Genuinely indeterminate — the framework’s deepest unresolved tension watch a fiscally desperate jurisdiction (state-level first) trade licensing reform for growth; G-2 late decade
Capex melt share (true D vs P mix of the installed base) sets treadmill severity and S1b loss-given-default disclosure improvements; F-2 depreciation revisions 2026–27
Cross-border gate asymmetry (does China’s administrative F2 fire first?) would relocate the first mass I-opening — and its reliability precedent — outside insurance-priced law G-2 2027–30
Occupational birth statistics the framework measures task death precisely and task birth not at all; reinstatement R (§3.3) is an unmeasured free parameter — the largest blind spot new-occupation tracking in JOLTS/O*NET revisions; posting taxonomies structural
P-clock instrumentation (robot production data) half of Musk’s model, thinnest data; flagged since v1 and unimproved through this entire exercise κ-2 buildout; supplier disclosures structural
Citizen-side inferential standing (v3) — does a “right to symmetrical intelligence” (explanation rights, audit standing, adversarial-agent standing) emerge as I-in’s counterforce? The constitutional analogy is Gideon v. Wainwright: the state already concedes that facing an institutional apparatus without a professional advocate is a due-process failure decides whether the I-in gate acquires a citizen-side operator — the selection asymmetry (§3.1) is bounded if it does (M18’s test), durable if it doesn’t; the commoditized-cognition anomaly (η_D’s new-category channel) runs in the citizen’s favor, so the binding gap is data position and action rights, not model quality explanation-rights litigation, algorithm-registry adoption, adversarial-agent case law (ι-4); M18 late decade

12.2 Structural residuals (no closure path; carried openly)

  1. US-shape. Class shares, doors/seats, Λ postures — all US-institutional. The EU and China rows are sketches; the framework’s claims should be read with jurisdiction subscripts it doesn’t yet have.
  2. Single-study dependence. R13’s direct evidence is one contested paper. Identification alternatives (§174, tech normalization, ADP coverage) remain live. [updated] Partly relieved and partly worsened by the C.2 pass: the pattern now has independent LinkedIn-based support in the US and UK, but two studies point the other way (Johnston & Makridis find employment increases in more-exposed QCEW cells; Humlum & Vestergaard find near-zero displacement in Danish registers), and the original authors have themselves retreated to descriptive language. R13 is better traced and no better identified than at drafting.
  3. The DAG lags the loop. Three feedbacks logged; R15 gives the bias a sign but not a fix. The quarterly cadence is a patch, not a solution.
  4. The S3/S4a boundary is the taxonomy’s softest joint (§6, note 2); the rent-vs-wage tiebreaker it needs sits on top of national-accounts rent imputation that doesn’t exist yet — measurement debt, alongside the cohort deflators.
  5. Self-crowding. R11 applies to this document: legible analysis of the corridor accelerates its capitalization. The framework is a participant in the flows it describes, and its best trades decay on publication.
  6. Skewed error distribution. [new, from the C.2 pass] Six of the seven drafting errors ran in the direction of a more dramatic analysis; the one running the other way was the one that raises the probability of the terminal the framework least wants. No procedure here corrects for that — R15 gives institutional feedback a sign, but nothing gives the author’s own error distribution a sign. Recorded because the pattern is only visible once an external check has been run, and it will not be visible again until the next one is.
  7. Maintenance burden. [new, v3] The likeliest failure mode of this framework is not falsification but neglect: twenty-one rows re-conditioned quarterly by hand, an annual conditional-structure review, a ratchet audit with two ledger directions — and no custodian beyond the author. A missed quarter degrades silently; the document keeps reading as current while its probabilities go stale, which is a worse epistemic state than being visibly wrong. (The v2→v3 E7 lag is this residual’s first realized instance: a verified correction sat unpropagated in the priors for one full cycle.) No closure path — the cadence either survives contact with the author’s attention or it doesn’t; the honest register is this line.

12.3 Deferred-work ledger

Appendix Contents Status
A Classification memo: NAICS→class mapping, decomposition rules, cohort deflators, rent-imputation construction delivered; §3.2 adopts its share vector (A.2), §3.4 its wedge prize (A.3). The C.2-owed re-derivation of A.3 on the dual premium is closed in v3: wedge share restated as an explicit [A] (25–60%) with the premium demoted to a calibration check; prize now ~5–12% of GDP, lower half better supported, decomposed into private-licensed and sovereign-administrative slices
B Evidence ledger: every [O/E/I/A/F] tag with source and date running; load-bearing subset and dependency table delivered, verdicts attached 2026-08-13
C Verification checklist: all post-cutoff sources (interview, Fed, BIS, headcount data) against primary documents executed 2026-08-13 — six corrected, four updated, five confirmed, the interview alone unresolved. Gate open; sign-off at C.3
D Jurisdictional decompositions and Λ postures (US/EU/China + one EM) delivered at sketch-grade; §12.2’s US-shape residual stands

The register’s closing stance is v1’s, kept deliberately: probabilities are subjective, rounded, and intended for scenario discipline rather than point forecasting; the dashboard is the mechanism by which they improve. One decade-scale addition the framework has earned: the numbers in §6 are the least durable part of this document; the gate structure, the clock mismatches, and the falsifier ledger are the parts built to survive being wrong.

Appendix A — Classification Memo

Status: measurement foundation. Unlocks the §3 arithmetic for citation (§0’s gate), with one logged discrepancy. Everything here is [E]-grade construction — the point is not precision but auditability: every share in §3.2 now has a visible derivation that can be attacked line by line.

A.1 The scoring procedure

Unit of classification: the task, not the job or sector. Procedure:

  1. Take the sector’s task inventory (O*NET-style task lists, weighted by wage-bill share).
  2. Score each task on the four §3.1 dimensions (digitizability, embodiment, verification/liability, positionality), 0–3 each.
  3. Assign class by last unrelieved constraint (§3.1’s rule): positionality > 0 → X; else liability ≥ 2 → I; else embodiment ≥ 2 → P; else D.
  4. Wedge check (the §3.1 refinement): a task scored D on dimensions but priced under license/mandatory sign-off is tagged I-wedge — I-class value-added, D-class technical content. This tag is the F2-collapsible fraction.
  5. Sector class-vector = task scores aggregated by value-added weight; economy shares = sector vectors aggregated by GDP weight.

Vintage rule: class membership drifts — the squeeze-zone front (§8.2) reclassifies tasks from “above frontier” to Wave-1; flows F1–F3 move boundaries. The mapping is stamped vintage 2026 and re-scored annually; the drift itself is dashboard data (it is the empirical trace of the flows).

A.2 Sector decomposition and aggregation

Condensed table (full task-level worksheets are working files, not reproduced). All rows [E]; GDP weights approximate US value-added shares:

Sector (NAICS) GDP wt % D P I X Notes
Agriculture (11) 1 .10 .70 .05 .15 X = farmland rents
Mining/energy (21) 1.5 .05 .60 .05 .30 resource rents → X (boundary case, A.4)
Utilities (22) 1.5 — .70 .20 .10 I = franchise/rate-permission value
Construction (23) 4.5 .05 .85 .10 — entitlement value capitalizes into land (X), not builder VA
Manufacturing (31–33) 10 .25 .60 .10 .05 pharma is an outlier: huge I-wedge (approval assets)
Wholesale (42) 6 .40 .50 .05 .05
Retail (44–45) 6 .30 .40 .05 .25 X = prime locations
Transport/warehousing (48–49) 3.5 .10 .85 .05 —
Information (51) 5.5 .60 .10 .05 .25 X = attention/platform rents (L4)
Finance/insurance (52) 8 .40 .05 .45 .10 I = charters, fiduciary wedges
Real estate (53, incl. imputed rents) 13.5 .05 .30 .05 .60 the X anchor; A.4
Professional/business svcs (54–56) 13 .55 .10 .30 .05 the D/I-wedge battleground
Education (61, private) 1.2 .30 .05 .50 .15 instruction D; credential I; elite seats X
Health (62) 7.5 .20 .45 .30 .05 per §3.1’s worked decomposition
Arts/entertainment (71) 1 .30 .30 — .40 liveness and stardom are X
Accommodation/food (72) 3 .10 .70 — .20
Other services (81) 2 .30 .60 .10 —
Government (92) 12 .30 .20 .50 — permission production

Aggregation: D ≈ 0.28, P ≈ 0.37, I ≈ 0.20, X ≈ 0.15.

Adoption note. This is the vector §3.2 uses. An earlier drafting estimate ran (0.30, 0.35, 0.22, 0.13); the mapping lands X at 0.15–0.16, driven by the imputed-rent land share (A.4), so the earlier X-figure was conservative. R1 sensitivity check: moving between the two vectors shifts the strong-S2 row only from ~5.1% to ~4.9% and leaves the Acemoglu corner at ~1.2% — no result changes sign or interpretation; the deceleration paradox is share-error-robust to at least ±5pp on any class. The arithmetic is unlocked.

A.3 Wedge measurement — and the wedge prize

Operational proxy for the I-wedge fraction: value-added in licensed or sign-off-mandated occupations whose task content scores D on dimensions (stage-2-feasible per current benchmarks). Inputs: [updated] 21.6% of employed US workers hold an occupational license, 24.0% a license or certification (BLS CPS 2025 annual averages — an 11-month average excluding October, the federal shutdown month, and so not strictly comparable with prior years) [O]; licensed-occupation wage premia over observably-equivalent unlicensed work [E]; task-content scoring per A.1.

[corrected] The premium input is softer than drafted, and this is the one verification result that materially weakens a framework quantity. The 10–15% range was carried as though both endpoints were established. Only the upper endpoint is: Kleiner & Krueger’s ~15% comes from demographic controls without detailed occupation fixed effects. Gittleman, Klee & Kleiner (2015), using 3-digit occupation fixed effects on SIPP panel data, obtain 6.5% — and say explicitly that this “falls below the consensus range of licensing wage premia between 10 percent and 15 percent.” Cato-affiliated work finds ~8%, Nunn 5–8%. The literature trends downward as controls improve, which is the expected direction if part of the raw premium is selection into licensed occupations rather than the license itself.

Consequence for the wedge prize, stated rather than buried: the ~8–12% of GDP figure was built on the higher premium. On the 6.5% estimate the wedge narrows materially — plausibly toward the mid-single digits as a share of GDP. This does not touch the sign or the mechanism (F2 still contests a real rent, R2 is derived from the classification rule and not from this number, and the I-gate remains the S2/S3/S4a boundary). The re-derivation the verification pass demanded is executed below (v3), closing the §12.3 item.

The re-derivation (v3). The derivation’s honest structure, restated so it can be attacked at the correct joint:

  1. The wedge share is the primary assumption, not a derived quantity. Share of I-class value-added that is wedge — permission-priced D-feasible work rather than residually-human institutional work: 25–60% [A]. The drafted 40–60% wore an [I] tag it had not earned: nothing in the premium literature ever derived it, and pretending the premium was an input laundered an assumption into an inference. At I ≈ 0.20 of GDP, the assumed range prices the prize at ~5–12% of GDP.
  2. The premium literature is a calibration check, and a weak one — deliberately demoted from derivation input. What Gittleman-type estimates price is the intra-human licensing rent: licensed labor against observably-equivalent unlicensed labor. Run the arithmetic: 21.6% of workers licensed [O] × a ~58% labor share of GDP puts the licensed wage bill near ~12–13% of GDP; the 6.5–15% premium band then implies identified intra-human rents of only ~0.8–1.9% of GDP. The wedge prize is a different estimand (C.2’s lesson, applied to the framework’s own number): the gap between permitted price and technical cost on D-feasible tasks, whose counterfactual is near-zero-marginal-cost inference, not unlicensed human wages. The premium arithmetic therefore disciplines the floor — rents at the I-boundary demonstrably exist, at just under 1% of GDP on the best-controlled estimate — while saying almost nothing about the ceiling, which rests on the stage-2 feasibility judgments of A.1 (dimension iii) and on how much of the billed price actually collapses when the signature does. Hence the standing qualifier everywhere the number is quoted: lower half better supported.
  3. The decomposition (v3, from the directional gate). The wedge has two slices with different erosion mechanics. The private-licensed slice — clinical sign-off, fiduciary duty, audit attestation, legal practice; the F1/F2 battleground — erodes only through I-out: insurance repricing and statute, the slow clock with the ι-1 fast hand. The sovereign-administrative slice — the D-feasible administration inside government’s I-share that A.6(iv) already flagged as overstated durable-I; permission processing as distinct from permission production — erodes through the sovereign carve-out (§3.2): state self-automation attacks it with no insurance event, no statute, and no guild on the other side. On A.2’s weights, government contributes roughly 0.06 of the economy’s 0.20 I-share, a material fraction of it processing; the sovereign slice is plausibly a fifth to a third of the prize. Propagated consequence (R12, M14): the slice that erodes first under O3 is the one whose jobs are the absorption machine’s.

The wedge prize, re-derived: roughly 5–12% of GDP is currently priced by permission on tasks that are technically automatable at the frontier — the lower half of the range better supported than the upper, and a fifth to a third of it sovereign-administrative, exposed to state self-automation rather than to F2. This is the pool F2 fights over, the quantity the guild loop defends, the absorption machine’s wage bill (R12), and the S3-upside that the I-gate withholds. [I, on an [A] wedge share]

§3.4 carries this number in the main text, where it prices what F2 contests. It is the single best quantification of “what is at stake at the I-gate” the framework produces, and the derivation lives here so it can be attacked on its inputs: the licensed-workforce share [O], the wedge-share assumption [A], the stage-2 feasibility judgments of A.1, and the slice weights — each of which moves the range, and only the first of which is settled.

A.4 X-share construction (the rent-imputation debt, §12.2)

National accounts do not separate land from structures. Construction:

  • Land component of housing: residual method — property value − depreciated structure replacement cost = land value; land rents = land value × capitalization rate. Applied to owner-imputed and tenant rents within NAICS 53.
  • Other X: resource rents (partially — deposits are intrinsically scarce; tagged as a boundary case, since their scarcity is geological rather than positional, but the rent behaves X-like: no productivity growth is defined over it); attention/platform rents (Information’s L4 component, proxied by advertising-market economics); positional consumption (luxury, elite admissions, art — CEX/industry data).
  • Published land-share-of-income estimates range 4–8% depending on method [E]; the A.2 X-total (~15%) includes structures-adjacent rents and positional consumption beyond pure land, which is why it exceeds the land-only literature.

Row X-1’s data spec: annual re-estimate of both the land component and the total-X decomposition; the trend matters more than the level (R3 predicts the trend).

A.5 Cohort deflators (spec + illustration)

Construction: CEX expenditure shares by cohort (age × housing tenure) mapped to class price indices (F-4’s four series), yielding cohort-specific inflation. Illustration under §4.1’s S2 paths (π_D=−10, π_P=+3, π_I=+5, π_X=+6.5) [I]:

Cohort D P I X Cohort inflation Effective, after housing hedge
Young renter .15 .30 .15 .40 +2.8% +2.8% (unhedged — bears X as cost)
Older owner .10 .35 .35 .20 +3.1% ~+1.5–2% (X-inflation accrues as wealth)

The measured gap is small; the hedged gap is 1–1.5pp/yr compounding — the §3.5 incidence result in deflator form. The politically operative number is the hedged one, and no statistical agency publishes it. Deliverable flagged for F-4.

A.6 Known failure modes

(i) Sectors that resist task decomposition (management, small business — bundled tasks); (ii) the X-boundary (resource rents, brand equity); (iii) wedge identification depends on stage-2 benchmark judgments that drift with τ; (iv) government’s I-share conflates permission-production with D-feasible administration — probably overstates durable I; v3 promotes this caveat into mechanism (the sovereign carve-out, §3.2, and the sovereign-administrative wedge slice, A.3.3). All are attackable line items, which is the memo’s purpose.

Appendix B — Evidence Ledger

Status: running register. Format plus the load-bearing subset; the full tag census is a working file. The highest-value section is B.3 — the dependency table that operationalizes §11’s graceful degradation.

B.1 Ledger format

Fields: claim · tag · source · date · verification tier (per C.1) · used in · dies/wounded if false.

B.2 Load-bearing evidence (the subset whose failure propagates furthest)

Verification status as of 2026-08-13 (Appendix C executed). Every row was put to five independent primary-source research passes. Verdict records the outcome; now reads gives the corrected value where the drafted one failed.

# Claim as drafted Tag Verdict Now reads If false
E1 China ~10,583 TWh (2025); > US+EU+India [O] confirmed Comparators 9,398 TWh (US 4,520 · EU 2,797 · India 2,082); margin +1,180. 10,583 is generation; Ember’s 10,573 is demand China profile weakens; little else
E2 Canaries: ~13–16% relative decline, 22–25s, exposed occs [E] stale ~19% (Aug 2026 revision, data→Jun 2026). 13/16 are data vintages, not rival specs. Authors: “descriptive… rather than causal” R13 → theory + traces; §8’s only direct observation gone
E3 St. Louis Fed: AI capex 1.3→0.48pp GDP contribution 2025 [E] corrected Quarterly, not H1: 1.30pp Q1 · 1.16pp Q2 · 0.48pp Q3; 0.97pp over three quarters credit list shrinks; nothing structural
E4 BIS: financing shift to debt/SPVs/private credit [E] confirmed, two deletions Debt/SPV/private-credit confirmed (>$200B, ~8%, $40B originated 2025, $300–600B by 2030). Vendor financing and pension exposure are not in the BIS texts straddle reverts to equity-absorbed — R5’s location claim dies, not its logic
E5 US federal interest > defense [O] confirmed FY2025 net interest $970.4B vs national defense $916.6B (Treasury MTS Table 9); CBO’s narrower defense basis $893B widens the gap —
E6 METR RCT: 19% slowdown, own repos [O] confirmed, scope-limited Exact: 16 devs, 246 issues, CI +2%→+39%. Early-2025 tools only; METR’s own follow-up is self-labelled unreliable stage-3 rebuttal loses its citation; funnel logic stands
E7 METR horizon doubling ~4–7 months [E] stale — faster TH1.1 (Jan 2026): 196.5d all-history · 130.8d since 2023 · 88.6d since 2024 (~2.9mo, below the drafted floor). Horizons >16h unreliable (saturation) re-anchor Layer 1 — done in v3 (§6.2 re-anchor log)
E8 ~20–25% of US workforce licensed [E] confirmed 21.6% licensed / 24.0% licensed-or-certified (BLS CPS 2025, 11-month average) wedge prize rescales
E8b Licensed wage premium 10–15% [E] corrected — weakens A.3 Upper endpoint only. 6.5% with 3-digit occupation FE (Gittleman et al.); ~8% Cato; 5–8% Nunn. Wedge prize narrows — see A.3 wedge prize magnitude falls; sign and mechanism unaffected
E9 India top-5 IT: negative net headcount FY26 [E] corrected — overgeneralized Aggregate negative but TCS-driven (−23,460); three of five firms grew (Infosys +5,016 · HCL +3,761 · Wipro +8,810); TechM −1,108; aggregate ≈ −6,981 ladder-melt loses its trace
E10 Humanoid production: tens of thousands (2025) [E] corrected — unsupported No primary global production count exists. IDC ~18,000 shipped; AGIBOT 5,000 audited; Unitree >6,500 built/>5,500 delivered; Tesla discloses none, output in the hundreds, Gen-3 slipped to 2026 P-gate timing shifts (already widest error bar)
E11 Hyperscaler capex ~$600–760B guided 2026 [E] confirmed on a defined basket Four-company CY2026 ≈ $700–725B (Amazon ~$200B · Alphabet $195–205B raised Jul 2026 · Microsoft ~$175–190B · Meta $130–145B). Requires naming constituents and lease treatment treadmill rescales, logic intact
E12 GPU economic life 3–5y [E] corrected Disclosed lives are 5–6y (Amazon 5–6 · Meta 5–5.5 · Alphabet 6 · Microsoft 2–6 · Oracle 6 · CoreWeave 6). The 3y floor is pre-2023 NVIDIA internal policy. Economic life genuinely contested (Burry 2–3y; H100 rentals −64%); no auditable resale curve exists melting-collateral softens (see M10)
E13 US private capital stock ~$70T [O] updated ~$68.0T current-cost net stock of private fixed assets (end-2024). Three of five passes could not close this from BEA primary at all —
E14 Working-age population declining, OECD+China [O] corrected — overbroad China confirmed (declining since ~2010–14). “Across OECD” is many, not all: Australia, Canada, Israel, Mexico still growing on migration and fertility —

What the pass did not touch. Every [I] and [A] object — the funnel, the four-class engine, the gate structure, the tree’s priors, R1–R15 — is framework-internal (T3) and was never in scope. No mechanism was overturned by any of the fourteen rows above.

Two rows moved in the framework’s favour. E10 shows the P-gate more firmly shut than drafted. E7 shows capability compounding faster than the drafted range — which, run through §6.3’s own sensitivity table, raises P(recursive) above the 0.25 prior and flows overwhelmingly into S4a, not S4b. The corrections push probability toward the rentier decade, which is the framework’s least comfortable conclusion rather than its most convenient one. (v2 carried this conclusion in prose without re-conditioning the masses; v3 executed the propagation — §6.2’s re-anchor log, netting caution included.)

v3 additions (feedback integration, 2026-08-13 — not part of the Appendix C pass; T-tiers per C.1). E15–E17 are pre-cutoff, T1-verified [O]; E18–E19 are post-cutoff, T2 unverified pending and quoted here estimand-first per C.2’s logged lesson:

# Claim Tag Tier Used in Dies/wounded if false
E15 SyRI (NL): welfare-fraud risk-scoring system struck down by The Hague District Court, Feb 2020, Art. 8 ECHR grounds [O] T1 §5.3 sub-current; O3; M17 calibration state-side ex-post gate loses a leg; M17’s historical bound weakens
E16 Robodebt (AU): income-averaging debt scheme ruled unlawful; ~A$1.8B settlement; royal commission report 2023 [O] T1 same same
E17 Toeslagenaffaire (NL): childcare-benefits algorithmic enforcement scandal; Rutte III government resigned Jan 2021 [O] T1 same same
E18 OECD 2026 Digital Government Outlook: AI used in ≥1 government area in 97% of OECD countries; oversight lagging. Estimand named first: “used in at least one area” is exposure-stage (funnel stage 1–2) — classically ambiguous in exactly the way C.2 warns about, and not evidence of deployment-stage governance [E] T2 pending O3’s prior ladder (as color, not load-bearing) O3 ladder loses its breadth citation; mechanism unaffected
E19 IMF technical notes (2025): AI in tax administration — compliance, anomaly detection, gap recovery [E] T2 pending; consistent with pre-cutoff knowledge R6 counter-current (§4.2) counter-current reverts to [I] inference from the carve-out alone

EU AI Act asymmetry (prohibits certain citizen-facing scoring/manipulation; regulates most high-risk institutional use): verified accurate, pre-cutoff [O] — cited in §3.1 and §5.3 without a ledger row of its own; JRC algorithmic-management findings: consistent with pre-cutoff knowledge [E].

B.3 Results-dependency table

R Depends on Dies / wounded if
R1 A.2 shares [E] + σ<1 [A] + gate structure [I] K1 kills; A-share errors only dent (A.6, §A.2 sensitivity)
R2 classification rule [A] + X-definition K2 halves it (X-only version survives)
R3 X fixed supply [O] + top income-elasticity [A] M-falsifier: land share flat through measured growth
R4 R1–R3 dispersion [I] + single-instrument CB [O] + level/trend distinction on X (v3, §4.1) M4 kills the I-side inertness premise; X-side, only a rent-share trend reversal under tightening counts
R5 six channels [I] + E4 E4 false → topology claim dies; straddle logic survives at equity
R6 US fiscal composition [O] + displacement timing [E2]; ambivalent counter-current logged (AI tax enforcement, E19/ι-4 — v3) timing wounds; composition is verifiable; a large realized enforcement dividend would blunt Blade 1 without killing the rate-base blade
R7 layer mapping [I] + concentration [E] M6/M12
R8 channel-speed ordering [I] near-definitional; M8 tests the consequence
R9 R6 + race asymmetry [O] M8 (the good-news falsifier)
R10 classification symmetry [I] falls with K2 for the I-row; X-row robust
R11 R10 + tree masses [A] M9
R12 Baumol drift [I] + I-out gate holds [A] + state does not self-automate the absorptive layer [A, v3 — the I-in contingency, §8.1] + absorption trace: predicted, not yet observed M14 with attribution (v3): private F2 kills; state self-automation narrows scope and confirms the carve-out instead — L-2 read against ι-4
R13 unbundling logic [I] + E2 (single study) M13; E2’s fall demotes to theory
R14 E14 [O] + E9 [E] demographic half is safe; ladder half rests on E9
R15 internal meta-observation [I]; sign claim scoped to power-dispersing deployment (v3, §9.3) self-auditing via §10’s ratchet audit (two counter-column entries); M17 bounds the endogenous counter-class

B.4 Demotion log

v1→v2 demotions, recorded for calibration: “real rates are the cleanest test” — retired (§2.2); Canaries [O]→[E]; “5% of GDP” → upper bound; trades premium “high confidence” → modal-conditional; “I-class capability-immune” → politically maintained, economically contestable. Pattern note: four of five demotions ran against the framework’s early confidence — consistent with R15’s sign.

Verification-pass demotions (2026-08-13). E3 H1→quarterly; E9 “all five firms”→TCS-driven aggregate; E10 “tens of thousands produced”→single-digit thousands built; E12 “3–5y”→5–6y disclosed, 2–3y contested; E14 “across OECD”→many not all; E4 loses vendor financing and pension exposure; E8b’s premium loses its lower endpoint and gains a 6.5% rival. Calibration note, and the uncomfortable one: the drafting errors were not random. Six of seven ran in the direction of making the analysis more dramatic — faster robots, bigger GDP contribution, broader financial contagion, a wider wedge, a more uniform employment collapse. The single error running the other way (E7, capability compounding faster than drafted) was the one place the framework was too conservative, and it is the row that most raises the probability of the outcome the framework likes least. A framework whose error distribution is skewed toward its own thesis should treat that as a finding about itself, not about the world; §12 gains a residual accordingly.

Appendix C — Verification Checklist

Status: executed 2026-08-13. Circulation was gated on execution, not on this appendix’s existence; the gate is now open. Sign-off at C.3.

C.1 Protocol

  • T1 — pre-cutoff verifiable: cite primary source in B; no gate.
  • T2 — post-cutoff claimed: verify against primary document before circulation; failing items force re-tagging and softening of every dependent claim (B.3 gives the blast radius).
  • T3 — framework-internal: no external verification possible; audit the derivation chain instead.

Rule: no circulation while any T2 item is unresolved. Partial circulation with failed items requires inline strikethrough-and-caveat, not silent deletion.

C.2 The T2 checklist

Item Verify against Blast radius if failed (via B.3) Outcome
The interview itself: existence, quotes, P1–P8 fidelity recording/transcript §1 rebuilt; framework survives — the position analyzed is held far beyond one speaker, and §§3–12 analyze the position, not the person not located — the one open item
E1 China TWh Ember/OWID published dataset minor confirmed
E2 Canaries (version: 13 vs 16%; replication; §174 critiques) paper + responses R13 demoted; §8 anchor lost stale → 19%; §174 answered in-paper
E3 St. Louis Fed contributions the Jan 2026 note §2.1 credit corrected — quarterly, not H1
E4 BIS Bulletin 120 / QR Mar 2026 BIS publications R5 location claim confirmed less vendor financing, pensions
E7 METR horizon curve METR site Layer-1 re-anchor stale — faster than drafted
E9 India IT headcount company filings R14 trace, M11 baseline corrected — aggregate, not universal
E10 humanoid production industry data κ-2 baseline corrected — no production series exists
E11 capex guidance earnings documents treadmill scale confirmed on a named basket
Interest>defense, capital stock, licensing share, demographics CBO/Treasury, BEA, BLS, UN T1 — verify-and-close interest and licensing confirmed; capital stock →$68T; demographics narrowed

Protocol lesson, logged for the next execution. The rule “verify against a primary source” underdetermined the answer in exactly the cases that mattered. Where one institution publishes one canonical number, all five passes agreed. Where the estimand was ambiguous — units produced vs shipped vs ordered vs capacity, accounting vs economic life, per-firm vs aggregate, which data vintage — they diverged sharply, and every one of the six corrections lives in that second category. The next iteration of C.1 should require the estimand to be specified before the source is: naming the measure is the verification step; finding the document is the easy part. That the checklist’s failures cluster exactly where measurement is ambiguous is the funnel thesis reappearing one level up, in the machinery built to police it.

C.3 Sign-off

Executed-by / date / items-failed / re-tags-applied — the blank table was the gate. It is now filled.

Field Entry
Executed 2026-08-13
Method All eleven T2 items put independently to five research systems under a primary-sources-only protocol; verdicts adopted on cross-system agreement, not on any single pass
Items corrected (wrong as stated) E3 (H1→quarterly) · E4 (vendor financing, pensions deleted) · E9 (not all five firms) · E10 (production unsupported) · E12 (3–5y→5–6y disclosed) · E14 (not all OECD)
Items updated (stale, directionally right) E2 (→19%) · E7 (→2.9mo post-2024) · E8b (premium 6.5–15%) · E13 (→$68T)
Items confirmed E1 · E5 · E6 (scope-limited) · E8 · E11 (on a defined basket)
Unresolved The interview itself. No pass closed it against a recording or transcript. §1’s reconstruction rests on quotes as given; §§3–12 do not depend on it
Re-tags applied E13 [O] stands at a corrected value; E8 [E]→[O]; E4 narrowed to the verified subset; E6 scope-limited to early-2025 tools; E2 carries its authors’ own descriptive-not-causal qualification
Blast radius realized None of B.3’s “dies if false” conditions fired. R5 keeps its topology on the verified subset; R13 keeps theory-plus-trace; A.3’s wedge magnitude narrows without changing sign — the single substantive casualty
Standing caution Verification quality varied sharply across passes. One system returned near-uniform confirmations, including a verdict contradicted by the anchor it had itself quoted, and per-firm figures irreconcilable with three others. Agreement across passes — not the confidence of any one — is what these verdicts rest on

Appendix D — Jurisdictional Decompositions

Status: promoted from §5.3’s sketch; still the framework’s thinnest module (§12.2 stands). Four profiles + the interaction layer, which contains the appendix’s one new mechanism.

D.1 Method and caveat

Each profile: approximate class-share deltas vs the US baseline [E, coarse], gate structure (who operates the I-gate; whether F2 is insurance-priced or administrative), Λ posture, clock notes, terminal tilt. These are posture sketches for scenario discipline, not country models.

D.2 United States (baseline)

Shares: A.2. I-gate: state-level licensing + liability law + private insurance — F2 is insurance-priced (ι-1 is a genuine leading indicator here). Λ = {ρ↑ on both waves, χ scissors-blocked, g weak, φ orthodox, ν high}. Terminal tilt: S2/S4a per R9 — the posture is the S4a machine. Distinctive exposure: the absorption machine (R12) is large (health, education, government, compliance ≈ the biggest I-employment complex anywhere), so the S2 cushion is thick — and the F2 stakes (A.3’s wedge prize) correspondingly maximal.

D.3 European Union

Shares vs US: I higher (co-determination, professional codes, public-sector weight), D lower (thin frontier), X comparable. I-gate: statutory and EU-regulation-driven more than insurance-priced — F2 requires legislation, not just actuarial repricing, so the gate is slower in both directions. Λ = {ρ high and codified, χ baseline-high, g medium, φ single-mandate, ν low}. Terminal tilt: S2-with-extra-steps regardless of global τ [F]. Mechanism, owed since §5.3: the EU imports D-deflation through tradables without earning D-production income — consumer of abundance, producer of little — while its intact I/X core inflates domestically. Better-cushioned distribution (high χ baseline), worse growth, and capital exported into other blocs’ buildouts (§7.5). The politically explosive combination: visible foreign abundance, domestic cost disease, and pension-funded financing of both.

D.4 China

Shares vs US: P far higher (manufacturing, construction), I present but state-operated, X concentrated in land-finance. Binding gate: physical (chips), not institutional — power is abundant (E1). The I-gate is administrative: the state can absorb AI liability by fiat — F2 as sovereign act rather than insurance event (generalized in v3 into the sovereign carve-out, §3.2: every state has this power over its own administrative deployments; China merely has it over the whole economy). Λ = {ρ discretionary, χ different machinery, g maximal, φ managed, ν maximal}. Demographic clock: aging cushions displacement (R14’s favorable side). Terminal tilt: bound-S2-like while chip-constrained; if lithography resolves (O2), gates open faster than in any insurance-priced jurisdiction — the first mass F2 event is plausibly Chinese [F].

Reliability caveat: administrative self-insurance removes independent actuarial pricing — deployment accelerates but correlated failure risk goes unpriced. The demonstration runs faster and the tail is fatter.

D.5 India (the EM profile)

Shares vs US: D-export sector large relative to formal economy (the ladder, §7.5), P constrained by grid and logistics, I strong in form (Anglo-model professions) but the formal I-sector is small relative to the workforce — which matters enormously: the absorption machine is thin. R12’s cushion barely exists; displaced D-workers have no protected-sector sponge to absorb them. Demographic clock: the youth bulge peaks as the ladder melts — R14’s epicenter. Λ = {ρ rising, χ thin, g ambitious but constrained, φ orthodox-ish, ν modest}. Terminal tilt: S2-without-absorption — the same terminal as the US with the shock absorber removed, implying earlier, larger, more open political contest [F]. The framework’s grimmest jurisdictional cell.

D.6 Interactions — the gate-coupling layer (G-2’s content)

  1. Actuarial import. Chinese administrative deployment generates at scale exactly the operational failure data Western insurers currently lack. Loss experience travels through reinsurers and technical literature even across political blocs → China’s fiat F2 feeds Western ι-1 pricing. The Western I-gate is therefore partially exogenous to Western politics: a guild can win every domestic fight and still watch its wedge repriced by foreign actuarial evidence. This materially strengthens the F2 side of the §5.3 gate-fight — and it is the one identified mechanism running against R15’s deployment-suppressing ratchet. [I]
  2. Havens. As wedge pressure builds, small-jurisdiction liability arbitrage (§7.5): sign-off domiciles, audit flags of convenience. Watch first in maritime-law-style registries and medical tourism corridors.
  3. Enclaves. Power-rich, institution-poor hosts take compute as extraction-style enclaves — capital-intensive, employment-light (§7.5); D.5’s thin-absorption logic says enclaves do not rescue the EM labor story.
  4. Divergence as S6’s content (v3). What this item carried as “substantially the residual’s content” is now the named terminal S6 (~0.08, §6.2): the world resolving into different terminals per bloc — US S4a-posture, EU S2-plus, China conditional-fast, India S2-minus — with no single world-state to name. This appendix supplies S6’s per-bloc profiles; G-2 feeds its divergence conditionals; the residual keeps only the timing-straddles and oddities.

D.7 Jurisdiction subscripts for the results registry

US-shaped and non-portable: R6, R9 (fiscal structure and posture). Portable wherever inflation-targeting meets big I/X: R4. Strong where formal protected sectors are large, weak in D.5-type economies: R12. Universal where white-collar careers exist: R13. Scope-limited to insurance-priced jurisdictions: M2 (the falsifier is mis-specified for China by construction). Portable everywhere by construction: R2, R10. The directional I-gate / sovereign carve-out (v3, §3.1–3.2): most portable where administrative states are strong and courts weak — the inverse of M2’s scope limit; in D.3-type jurisdictions the scandal/administrative-law brake (E15–E17) binds hardest, and in D.4 the carve-out is the whole gate.