Experimental — Dyno-Sim is under active development. These runs are research-grade: reduced-form models with disclosed dials (not fitted to your data), reporting relative, structural results — the mechanism and the comparison between runs, not any single level, year, or outcome. Interrogate them as decision-support. Illustrative / decision-support, not investment advice. Figures below come from Comuvia's Dyno-Sim using its shipped, stylized sample scenarios — they illustrate the mechanism and that the tool works, not a forecast of any specific portfolio. A client engagement calibrates the model to your data.
Updated 2026-07-26. Added: a 5,000-replica parameter-robustness check on the divergence verdict (with its event definition stated inline), the model's own WID-calibrated 2000–2024 hindcast with a held-out 2016–2024 fit, the measured 2007–2011 crisis asymmetry, and the final-ratio row that the peak-only table previously left ambiguous. Every original number re-verified against the committed scenarios before this update; none changed.
Updated 2026-08-07. Added four bright house-theme simulation-chart figures — the ×47 divergence, the three-lever comparison, the 5,000-replica robustness histogram, and the WID hold-out fit — drawn from the same runs as the companion explainer video. No numbers changed.
Updated 2026-08-13. Editorial-only: added a one-line note distinguishing this ×47 wealth ratio from the 47% verdict-flip statistic used in our Minsky knife-edge pieces (different quantities, same digits), and added trend arrows to the cycle and intervention metric tables. No numbers changed; every figure re-confirmed against its committed run.
If you steward private capital — a family office principal weighing resilience, a chief of staff filtering research for a decision-maker, an investor stress-testing a thesis — most commentary on the "K-shaped economy" hands you a chart and an opinion. Neither helps you decide anything. For a book that carries the tail, the useful questions are mechanical: how fast does concentration compound, when does leverage turn dangerous, and which interventions actually reduce the modeled tail — rather than just feeling prudent?
Averages can't answer that. A mechanism you can interrogate, stress, and reproduce can. So Comuvia's system doesn't write an essay about fragility: it runs a simulator Comuvia owns, and it shows the work below. This page was produced by the same AI-managed system that ran the simulations — we run the stack we advise on.
Two forces, one tail
Concentration risk in 2026 is the product of two reinforcing dynamics:
- Wealth concentration (the "K"). Automation lifts the capital share of income without lifting payroll; capital income accrues in proportion to wealth already held; the upper cohort saves and compounds while the lower cohort spends. Left alone, ownership concentrates — and a narrow, highly-levered system is a more fragile one, not a safer one. The last crisis measured it: bottom-half US household wealth fell ~80% peak-to-trough 2007–2011 while the top decile's fell ~11% (computed from committed WID series, corroborated by the Fed's DFA bottom-50 share) — leverage concentrates the damage exactly where the buffers aren't.
- Private-debt leverage (Minsky-Keen). Calm breeds risk-taking; financing migrates from hedge → speculative → Ponzi; the change in private debt — not its level — is the crisis signal. High debt that persists while credit growth weakens is the dangerous configuration.
The tail is what happens when a concentrated, levered system meets a credit turn.
Diagram: two reinforcing loops — wealth concentration and private-debt leverage — converging into tail risk when the credit turn arrives
Two reinforcing dynamics, one failure mode. Both forces run inside a single stock-flow-consistent model, so their interaction is computed, not asserted.
What the simulator says (reproducible)
Comuvia's system ran two of Dyno-Sim's stock-flow-consistent models on their shipped sample scenarios.
1. Concentration compounds — fast
The two-cohort K-shaped model, run with no policy intervention:
| Metric (stylized sample run) | Result |
|---|---|
| Peak upper-cohort wealth share | 97.9% |
| Peak wealth ratio (upper : lower) | 47.1× |
| Final inequality (Gini proxy) | 0.96 |
| Lower-cohort share of capital ownership, final | 2.1% |
The lower arm doesn't just earn less — it is squeezed out of ownership (down to ~2% of the capital stock), which removes its only path onto the compounding side of the trade. That hollowing-out is what makes the macro system brittle.
The upper-to-lower wealth ratio climbing from about ×9 in 2026 to ×47 by 2056 under no intervention — with nothing done, the top arm ends 47× wealthier than the bottom
Concentration compounds. Left alone, the gap between the two arms of the K runs from roughly ×9 today to ×47 over thirty years — the same no-policy trajectory the table above summarizes.
(A note on a numerical coincidence: this ×47 wealth ratio — one cohort's wealth as a multiple of the other's — is a different quantity from the 47% verdict-flip figure in our Minsky knife-edge and quantum pieces, which is a probability that a boom-bust replay reverses an employment call. Same digits, unrelated measures.)
2. Leverage cycles, and the crisis signal is the credit turn
The Minsky-Keen cycle model on its sample scenario:
| Metric (stylized sample run) | Result | Trend |
|---|---|---|
| Private debt-to-output | starts 1.04, peaks 1.47, then deleverages to 1.26 | ↑↑ peak → ↓ |
| Employment rate | cycles from 0.94 down to 0.92, recovers to 0.96 | ↓ → ↑ |
| Distributional swing (wage share) | up to 0.93, with profit share compressed to 0.035 | wage ↑ / profit ↓↓ |
The model reproduces the Keen geometry: leverage builds during the calm, the wage/profit split swings, and the system turns when credit growth rolls over — exactly the "high debt + weakening credit" pre-crisis configuration that levels alone would miss. For when that turn becomes visible in market data, the companion piece now carries a four-signal tripwire graded against the 2000/2008/2020 drawdowns, with a sealed, dated prediction attached — see the US fragility check.
3. The one lever that measurably bends the curve
The reason to own a simulator is to test interventions, not just describe doom. Re-running the K-shaped model with an ownership-broadening lever (structurally routing a slice of new capital ownership to the lower cohort — worker equity / citizen-dividend style) instead of income transfers:
| Metric | No intervention | Ownership-broadening | Change |
|---|---|---|---|
| Peak wealth ratio | 47.1× | 15.0× (during the transition) | −68% ↓ |
| Final wealth ratio (year 30) | 47.1× | 2.2× | −95% ↓↓ |
| Final Gini proxy | 0.96 | 0.37 | −61% ↓ |
| Lower-cohort capital ownership | 2.1% | 31.3% | ~15× ↑↑ |
Moving ownership (a stock) rather than income (a flow) is the structurally durable fix — the model makes the difference quantitative, not rhetorical. (The intervention's ratio peaks at 15× mid-transition, then keeps bending: by year 30 the same economy ends at 2.2×.)
Three policy paths from a shared year-8 switch point: doing nothing ends at ×47, a 30% income hand-out only bends it to ×38, while broadening who owns the capital collapses the gap to ×2.2 by 2056
The lever, three ways. Income transfers (the dashed path) leak back out as spending and barely move the ratio; broadening ownership (the green path) puts the lower cohort inside the compounding loop and bends ×47 down to ×2.2.
4. Does the divergence verdict survive parameter luck?
A single run is a story, so we re-ran the no-intervention baseline 5,000 times with run-level parameter draws — automation rate ±1.5pp and upper-cohort consumption propensity ±8pp per replica (disclosed illustrative sigmas, not estimates). Across those plausible parameter worlds, the final wealth ratio lands between 38× and 56× (10th–90th percentile, median 47×), and the worst draw still ends at 28×:
| Robustness question (5,000 replicas) | Result |
|---|---|
| Replicas whose final ratio falls below the intervention's peak (15×) | 0 of 5,000 (95% CI 0.0–0.1%) |
| Final wealth ratio, 10th–90th percentile | 38× – 56× |
Reading it honestly: parameter uncertainty moves the magnitude of the divergence, never
its direction — no plausible parameter draw makes doing-nothing end even as well as the
ownership lever's worst moment. The verdict is not an artifact of one lucky parameterization.
(Surface sample-k-shaped-economy-tail-risk-20260726; the definition is stated in the table
because a robustness number without its event definition is decoration.)
Histogram of the final wealth ratio across 5,000 parameter draws: the bulk lands between ×38 and ×56, and even the ownership lever's worst moment (×15) sits to the left of every do-nothing run
Parameter luck moves the magnitude, never the direction. Across 5,000 replicas the do-nothing gap stays wide (80% land ×38–×56); the green line marks the ownership lever's worst moment (×15) — still below every one of the 5,000 do-nothing outcomes.
Calibrated against the measured world (WID 2000–2024)
The stylized runs above illustrate the mechanism; a calibrated sibling scenario points the same model at the measured world. Fitted by grid search to the WID world top-10% wealth share — trained on 2000–2015, judged on a held-out 2016–2024 window it never saw — the calibrated model tracks the holdout at RMSE 0.0037 on the share (train 0.0021), roughly 22× better than default parameters. Two honest notes travel with that: the measured world top-10% share sits near 81–84% over that window — the stylized baseline's 97.9% peak is a runaway endpoint of the no-policy mechanism, not a forecast — and the capsule's real-data anchors (WID top-1% income share, BLS labour share) are quoted on the model's own pages with their vintages. A registered next version (the wealth-curve adoption plan) replaces the two arms with a full distribution curve, gated on beating this two-arm hindcast out-of-sample.
The calibrated model's fitted curve tracking WID.world top-10% wealth-share data points, tuned only on 2000–2015 and holding out 2016–2024 — on the shaded held-out years the model tracks reality to 0.0037 error
The honest test. The model is tuned only on 2000–2015; the shaded region (2016–2024) is data it never saw. On those held-out years it tracks the measured WID top-10% wealth share to RMSE 0.0037 — the mechanism, not just the story, survives contact with the real world.
Diagram: the compounding loop between capital stock and household income — an income transfer enters as a flow and leaks out as spending, while the ownership-broadening lever moves a stock directly into the loop
Why the lever works. A transfer enters the loop as income and leaks back out as spending; the ownership-broadening lever moves a stock — it puts the lower cohort inside the compounding loop itself.
Dyno-Sim Atlas dashboard: K-shaped economy, baseline vs ownership intervention — stat tiles and the two wealth-ratio trajectories diverging at the year-8 policy switch
The two policy paths in Dyno-Sim's Atlas dashboard (screenshot of the live operator view). The baseline (solid) runs away toward a 47.1× wealth ratio; the ownership-broadening intervention (dashed) switches on at year 8 and bends the same economy back to a 2.2× final ratio at year 30.
Why this is different from a slide
-
It conserves. Every step the model checks its income and wealth accounting identities. In these runs the maximum identity error was on the order of 1×10⁻¹⁴ — machine precision. This is not a curve fit; it's a conservation-consistent system. Numbers that don't balance get caught.
-
It fits real history. The same simulator family is calibrated to real FRED data across the GFC and COVID windows — see the US fragility check, where a documented recalibration matched the real ~10% GFC unemployment and the 2009 zero lower bound. So this isn't only a stylized illustration; the engine reproduces real macro history when pointed at it.
-
It's owned and on-prem. No vendor model deciding what it will or won't say about inequality or instability; a first-principles mechanism with explicit, auditable assumptions. (The model proposes; the verifier — here, the accounting identity — decides.)
-
It's reproducible. Same scenario in, same numbers out. Each result above is one command against a shipped sample scenario, for example:
dyno-sim run --scenario scenarios/samples/sample-k-shaped-economy.json
The baseline run's page in the Dyno-Sim operator dashboard: decision question, fidelity label, and run overview
Every number above traces to a run id. The baseline run (artifact-sample-k-shaped-economy,
120 quarterly steps, 29 metric columns) as it appears in the operator dashboard — including
the model's own fidelity label: "Illustrative — not statistically calibrated." Naming what a
run is NOT is part of the discipline.
What this looks like for your portfolio
The runs above use stylized sample scenarios; the engagement version points the same machinery at your exposure. A Simulation Decision Pack calibrates these mechanisms to your data and answers questions a benchmark can't:
- Where is your book on the concentration/leverage map, and how close is the credit turn?
- Which hedges or allocation shifts actually reduce the modeled tail — and which only feel safe?
- For a fiduciary or insurer: what does the pre-crisis configuration look like in your data, with enough lead time to act?
And when the modeled scenarios feed a larger choice — an allocation shift, a hedge program, a thesis you're about to fund — a Decision Map Workshop turns them into a decision map with a recommended path. If you carry the tail rather than the upside, start with simulation-backed decision support.
Comuvia builds decision support on owned, accounting-consistent simulators. This piece uses stylized sample scenarios to show the method; engagements calibrate to your data. Not investment advice.