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 a forecast. 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 the US economy or any portfolio. A client engagement calibrates the model to your data and questions. This is decision support, not investment advice.
Updated 2026-07-25. The robustness section now runs on the composed fiscal-monetary model itself (the June version quoted the companion debt-cycle surface — retained below, correctly attributed); every reported event now carries its definition inline; a lead-time section, two computed figures, and a sealed, dated prediction were added. Nothing from the June run was re-fitted; one June headline number was removed for the reason stated where it stood.
Updated 2026-07-26. Added the 2030 fragility outlook (2,000 bootstrap driver paths + named stress replays, sealed as
pred-2026-07-25-015) and the measured private-balance-sheet decomposition (seven Z.1/FRED series through 2026Q1) to "Whose balance sheet, exactly?".Updated 2026-08-01. Added "Is the AI-capex boom a Minsky cycle?" — the timely reframe of the same debt-cycle mechanism onto hyperscaler/AI-capital-expenditure leverage. Mechanism-level only: no fitted AI-capex calibration has been run, so no AI-bubble probability is quoted; that scenario is registered future work. No existing number changed.
Updated 2026-08-07. Added three bright house-theme figures drawn from the same runs as the companion explainer video — the Minsky hedge → speculative → Ponzi financing ladder, the 47% knife-edge balance, and the "no bubble number" discipline card. No numbers changed.
Updated 2026-08-13. The fiscal-monetary numbers, the fan figure, and the "What the simulator says" section were refreshed at a converged integration step (dt 0.0625, 20,000 replicas); every figure traces to a committed converged run. The 47% knife-edge — a separate, already-converged runtime — is unchanged.
If you steward private capital — a family office balance sheet, a fund's risk budget, a PE portfolio — you already have macro views. What you don't have is a way to test them. The house-view memo tells you what will happen; it cannot tell you how robust that call is when the shock actually arrives — and you, not the author, carry the downside when it's wrong. What a risk committee needs is a mechanism it can interrogate, stress, and reproduce, with the uncertainty stated rather than hidden.
So instead of another conviction essay, Comuvia's system ran the question through a fiscal-monetary simulator it owns — and shows its work below: the run, the calibration against real history, and how often the answer flips across 20,000 replays. This is the AI-managed organization we advise enterprises to build, operating visibly: the simulation, the calibration, and this publication came off the same stack.
The question a benchmark can't answer
Take the 2025 US configuration the committed baseline encodes: a policy rate of 4.3%, US private (non-financial) debt near 1.5× GDP, elevated public debt. Now hit it with an ordinary recession — not a tail event, just unemployment rising toward ~10% — and let the Fed and Treasury respond the way their post-2020 reaction functions say they will. (By this update the Fed had already eased to ~3.6% — EFFR monthly mean, April 2026. The run keeps the higher 2025 starting rate: the conservative direction for this question, since the private debt-service burden scales with the rate.)
Does the policy stack hold the system out of crisis? And — the question that actually matters for positioning — how robust is that answer to the luck of the draw?
What the simulator says (reproducible)
Comuvia's system ran a US-2025 baseline through Dyno-Sim's composed fiscal-monetary model — a Taylor rule, a zero-lower-bound, a financial-stability override, and a post-2020 Treasury reaction function. A recession shock at quarter six pushes employment from 0.96 to 0.90.
The recession bites — but the central case doesn't spiral
| Metric (stylized sample run, converged dt) | Result |
|---|---|
| Policy rate response | cut to the zero lower bound (0.25%), QE deployed |
| Public debt-to-GDP | ≈1.20 → ~1.00 — stabilized through nominal growth |
| Private debt-to-GDP | 1.50 → peaks ≈1.57 → mean-reverts to ≈1.42 (no spiral) |
| Private debt service (interest share of output) | ≈8.9% → ≈8.6% (peaks ≈9.4%) |
| Crisis-stress index | climbs to ≈0.67, then eases to ≈0.52 |
| Employment | troughs at ≈0.888 (~11% unemployment); no recovery in-horizon |
| Inflation | dips to ≈−0.9% (mild deflation), then recovers to ≈+0.7% |
The monetary-fiscal machinery does its job — and in the central case it does it on both ledgers: the rate cut, asset purchases, and fiscal support stabilize public debt through nominal growth, and the private-debt ratio rises only modestly after the shock (to ≈1.57×) before mean-reverting to ≈1.42×. The recession still bites the labour market — employment troughs near 11% unemployment and does not recover in-horizon — but the private balance sheet does not spiral. That is the point that sets up everything below: a single run can look like it holds. The fragility only appears when you re-run it across plausible worlds.
Diagram: one recession shock hits the policy stack — the Fed cuts to the zero lower bound and deploys QE, Treasury adds fiscal support — and in the central case both ledgers stabilize (public debt eases, the private-debt ratio peaks near 1.57x then mean-reverts to 1.42x); the private balance sheet is where the tail risk concentrates, with about one in four plausible worlds tipping it into a spiral
One shock, two balance sheets. In the central case the policy response stabilizes public debt (top) and the private path mean-reverts (bottom) — but the private balance sheet is where the tail risk lives: about a quarter of plausible worlds tip it into a spiral.
The decision-relevant finding isn't the central path — it's what a single run hides. Re-run this same configuration across plausible worlds (next section) and the private balance sheet is where the fragility concentrates: about a quarter of those worlds tip into a spiral even though the central case mean-reverts, while public debt stays stabilized across the board. The location of the tail risk — private, not public — is the finding; a strategy that watches public debt and the Fed's response can be looking at the wrong balance sheet entirely.
How robust is that call? (the part most models skip)
A single run is a story. To know whether to act on it, you need the distribution — and that is the step most macro commentary skips entirely. The system re-ran the same composed fiscal-monetary scenario 20,000 times, each replica drawing a plausible productivity-growth and private-rate offset once per path (a parameter-uncertainty ensemble: "across plausible worlds, does the call survive?").
Fan chart: private debt-to-GDP across 20,000 replicas of the composed fiscal-monetary model. The 10–90% and 25–75% bands widen after the quarter-6 recession shock; ~26.6% of replicas end above the 2.5x debt-spiral line; the median path ends near 1.5x
The distribution behind the story: the median path stays under the spiral line — and a fat cohort
of plausible worlds does not. The p95 path ends near ≈5.7× GDP. Computed from the committed
scenario at build time; surface sample-fiscal-monetary-us-2025-tail-risk-dt0p0625.
| Decision-grade metric (20,000 replicas, composed model) | Result | What the event means (implemented rule) |
|---|---|---|
| Debt-spiral rate | 26.6% (95% CI 26.0–27.2%) | final private debt-to-GDP ends above 2.5× |
| Severe-crisis rate | 35.6% (95% CI 35.0–36.3%) | crisis-stress index peaks above 0.80 at any point |
| Severe labour-market floor | ~0.04% (95% CI 0.02–0.07%) | employment touches 0.85 (~15% unemployment) at any step |
| Recommendation-flip probability | 18.0% (95% CI 17.5–18.6%, Wilson) | a replica flips when its end-of-run employment verdict (above/below the no-shock baseline's final value) differs from the median replica's verdict |
The thresholds are documented and informative, not magic constants — they mark where the model's own crisis language starts to bind, and every probability ships with its confidence interval, not as a bare point estimate.
Reading it honestly: the median world muddles through under the spiral line — and roughly about a quarter of plausible worlds don't, with about a third entering severe-crisis territory at some point. That is exactly the kind of statement a benchmark or a house-view memo cannot make, and the kind a risk committee can actually budget against.
The companion cross-check (the number the June version of this article led with): a
Minsky-family debt-cycle model calibrated to real US unemployment (1990–2020) and hit with
20,000 per-step TFP shock paths — a different uncertainty type, path luck rather than parameter
doubt — shows a 47.0% flip rate (95% CI 46.3–47.7%). There the median replica's final
employment (0.9070) sits a hair below the no-shock calibrated baseline (0.9081): a genuine
knife-edge, and 47% of shock paths land on the other side of it. (The June version also quoted a
"99.6% debt-collapse rate" from this surface. We removed it: inspection showed that event measured
the model's debt ratio ending near its zero floor — a fully-deleveraged endpoint, not a debt
crisis — and it has been renamed final_debt_near_zero in the artifacts. A number that misleads
readers doesn't get to stay because it's dramatic.)
Balance bar of 20,000 sealed shock paths split at the median employment verdict: 10,594 paths agree with the median call and 9,406 flip it, a fulcrum under the decision line — 47% land on the other side
A genuine knife-edge. The median replica's final employment sits a hair from the no-shock baseline, so 47% of the 20,000 shock paths flip the verdict — noise alone, not signal, decides which side nearly half of them land on.
Is the AI-capex boom a Minsky cycle?
The 2026 version of this question isn't recession-in-the-abstract — it's the AI-capital-expenditure
boom: hyperscaler and model-lab spending running at a pace that increasingly leans on off-balance-sheet
and debt financing. The reflexive market read is sentiment — greed now, fear later. The Minsky-Keen
lens the companion model encodes says something more testable: what signals a turn is the change in
private debt and whether financing is migrating hedge → speculative → Ponzi — not mood. A boom
that levers up while the incremental output per borrowed dollar falls is the pre-crisis configuration
whether the collateral is houses or GPUs. The same machinery that produces the 47% knife-edge above is
the honest instrument for the AI-capex question, because it fires on leverage, not on fear — the
companion crowd-dynamics runtime (market_crowd_abm_v1) is built on exactly that distinction.
The Minsky financing ladder as three linked stages — HEDGE (income covers interest + principal) to SPECULATIVE (income covers interest only) to PONZI (needs asset prices to keep rising) — under a banner reading the tripwire is the change in private debt, not its level
The lens fires on leverage, not fear. As a boom runs, financing migrates hedge → speculative → Ponzi; the crisis signal is the change in private debt (Δd) — the flow of new credit, not the stock — which is testable whether the collateral is houses or GPUs.
The honest limit, stated up front: Comuvia has not run a fitted hyperscaler-capex calibration, so this piece quotes no AI-bubble probability. That scenario — pointing this debt-cycle machinery at the AI-capex leverage picture — is registered future work, and until the run is committed with its inputs on the table, there is no number to publish. What the lens already rules out is the lazy version of the debate: "it's just a vibe." The mechanism is what transfers today; the figure waits for the run. (The model proposes; the verifier decides — and it has not yet been asked this exact question.)
Two cards: "No bubble number" — Comuvia has not run a fitted AI-capex calibration, so it quotes no bubble probability, that scenario is registered not answered — beside "What it rules out" — the lazy version, "it's just a vibe", because the lens fires on leverage not fear
The discipline, stated plainly. There is no fitted AI-capex run, so this piece prints no bubble probability — the scenario is registered future work. What the Minsky lens already rules out is the lazy read: "it's just a vibe." The mechanism transfers; the figure waits for the run.
When would you see it coming?
The scenario above assumes the recession arrives. The standing question for anyone who carries the tail is different: would you get any warning? Comuvia's portfolio includes a market-based fragility score built for exactly that — four public signals (the 10y–2y yield curve with an inversion memory, Baa credit spreads, the VIX, and Shiller CAPE valuation), graded against the three modern drawdowns.
Line chart: the smoothed fragility score 1997–2026 against the dot-com, GFC and COVID drawdown windows, with watch/elevated/fragile thresholds. The score is elevated from 2022Q3 to 2026Q1 and eases to watch (0.34) in 2026Q2
The tripwire, computed from committed FRED/Shiller packs (1997Q1–2026Q2). It crossed "elevated" in 1999Q1 — a year ahead of the dot-com peak — reached "watch" by 2007Q1 into the GFC, and stayed quiet through the long benign stretches: no false alarms 2003–2007 or 2010–2019.
What it says now: after sitting elevated from 2022Q3 through 2026Q1 (inversion memory plus stretched valuations), the score eased to "watch" (0.34) in 2026Q2 — the yield curve has re-steepened long enough for the inversion memory to fade, credit spreads are calm, and the one leg still pinned near its ceiling is valuation (CAPE ≈ 40, 0.98 of the leg's max).
Three honesty notes travel with that reading. The score warns for credit-cycle build-ups, not exogenous shocks — its own backtest shows it reacting weakly to COVID, which is disclosed as a feature of the instrument, not hidden. The 2024+ valuation inputs come from a public mirror of the Shiller dataset (the primary file wasn't fetchable at retrieval; graded accordingly in the data manifest). And a reading of "watch" is not "safe" — it means the leading legs are quiet while valuations alone are stretched.
And we put a date on it. Sealed prediction pred-2026-07-25-014 (registry, immutable,
sha-sealed): the fragility score stays below the "elevated" threshold (0.40) through 2026Q4,
because the inversion-memory leg — the one that held it elevated for four years — fully expires in
2026H2 while credit stays calm. Registered against the regime-persistence baseline (the four-year
elevated regime continues), due 2027-02-28, scored publicly either way. A resolved miss drives
recalibration; that is the point of sealing it.
And through 2030?
Quiet now is not quiet for long. We ran 2,000 forward driver paths to 2030Q4 — each a block-bootstrap draw from the four drivers' own 29-year joint history, appended after the full committed record so the score's memory state is exact at the 2026Q2 handoff — plus three named stress replays (the GFC's, dot-com's, and 2022-tightening's actual driver deltas, re-based onto today's levels):
Fan chart: the fragility score's committed history through 2026Q2, then 2,000 bootstrap driver paths to 2030 with 10–90% and 25–75% bands, three dashed stress-replay paths, and the watch/elevated/fragile thresholds. 77% of paths spend at least one quarter at elevated during 2027–2030
The outlook, with its conditionality in the title: IF the drivers keep behaving like their own 1997–2026 climate, 77% of paths spend at least one quarter at "elevated" during 2027–2030 (44% touch "fragile"; median peak score 0.54, 80% band 0.30–0.69). Two structural reasons the odds are that high: the valuation leg starts at 0.98 of its ceiling, and half the sampled history contains build-up regimes. The honest stress finding: replaying the GFC's own deltas from today's steep-curve/calm-credit start peaks at only 0.50 — 2007-style warnings come from multi-year build-ups, not single shocks.
This too is sealed: pred-2026-07-25-015 — the peak quarterly score over 2027–2030 lands in
[0.30, 0.69] (central 0.54), against the persistence baseline that the tripwire never
re-enters elevated at all — due 2031-02-28. The near-term and long-term predictions bracket the
claim: quiet now, most likely not quiet for four years. What the bands cannot contain, by
construction: a driver regime the 1997–2026 sample never saw, or a shock markets don't price in
advance — the same COVID-shaped blindness disclosed above.
Whose balance sheet, exactly?
"Private debt spirals" is an aggregate statement. The named audiences of this piece — family offices, insurers, PE — sit in different seats inside that aggregate. First, the measured seats — a seven-series Z.1/FRED pack (committed 2026-07-25, all through 2026Q1) decomposes who owes the debt and who intermediates it outside banks, as shares of GDP:
| Stock (share of GDP) | 2007Q4 | 2026Q1 | Reading |
|---|---|---|---|
| Household debt | 96.8% | 66.1% | the great deleveraging — 2008's epicenter is structurally lighter |
| Nonfinancial-corporate debt | 45.1% | 45.4% | unchanged — the model's single private-debt dial now has measured arms |
| Broker-dealers (assets) | 41.9% | 21.0% | half the pre-Lehman share |
| ABS issuers (assets) | 31.5% | 5.9% | the private-label securitization machine never restarted |
| Money-market funds (assets) | 21.0% | 26.0% | the one stock at a record share — run-prone by construction |
| Finance companies (assets) | 16.1% | 9.2% | the closest Z.1 line to direct lending — and it under-captures it |
| Household debt service (of disposable income) | ~15.9% (peak) | 11.2% | mid-range: above the 2021 trough (9.1%), far below the GFC peak |
Today's shadow-bank map is genuinely different from 2008's: the old leverage chains are smaller, households are lighter, and the risk concentration moved — toward money funds (record share) and toward direct-lending vehicles that Z.1 aggregates only partially see (a disclosed, open data gap; the FSB's monitoring reports are the qualitative reference). One boundary note so the numbers reconcile: Z.1 household+corporate sums to 1.11× GDP, while the article's ~1.5× anchor is the broader BIS total (which adds noncorporate business) — both are right about different boundaries, and we say which one we're quoting.
Second, the seats as mechanisms, in Comuvia's owned runtimes:
- Households. A cost-of-living squeeze runtime tracks essential-spending distress (housing, food, energy, insurance, medical) under exactly the debt-deflation conditions the run above produces. In the last measured crisis of this type, the damage was radically asymmetric: bottom-half US household wealth fell ~80% peak-to-trough 2007–2011 while the top decile's fell ~11% — the private-debt spiral is a distributional event before it is a macro one.
- Banks vs insurers. A securitization-crisis runtime replays how pooled-loan structures (2008-style MBS→banks, CLO→insurers chains) turn one aggregate spiral into two very different loss experiences: mark-to-market institutions take the drawdown immediately; book-value holders take it slowly, then all at once.
- The public seat. The run above already shows it: the sovereign balance sheet stabilizes. The question your seat determines is which side of the private ledger you're funding.
A K-shaped concentration-risk companion piece works the distributional seat in detail. An engagement points these runtimes at your seat.
Validated against real history (GFC + COVID)
A stylized scenario shows the mechanism; fitting real data shows the model isn't a toy. The same fiscal-monetary engine was calibrated to real FRED data — effective fed funds rate and unemployment — across the 2007–2012 GFC and 2019–2022 COVID windows. COVID employment fit out of the box. The GFC fit initially overshot — a Great-Depression-scale collapse — and a documented recalibration closed it:
| GFC 2007–2012 fit (model vs real FRED) | Before | After calibration |
|---|---|---|
| Employment trough (real ~0.90, i.e. ~10% unemployment) | 0.72 (≈28% — too deep) | 0.91 (matches) |
| Policy rate at the 2009 floor | missed the ZLB | hits the ELB |
| Combined employment + rate error (RMSE) | 0.088 | 0.037 |
The point isn't the exact figures; it's that the model that produced the fragility call above reproduces real history when pointed at it, and the calibration that got it there — which lever moved, by how much, and the residual that's still off (~1pp on the pre-crisis rate peak) — is written down, not hand-waved. That is the difference between a simulator you can stake a decision on and a chart you can't.
Why this is different from a slide
-
It's traceable to a run. Every figure above maps to a scenario file, an equation, and an artifact: the ensemble surfaces carry the generating scenario's sha256 and timestamp, the two charts are computed from committed inputs at build time, and the prediction is sealed in a registry that CI verifies. Nothing rests on "our model says so."
-
It conserves and it's owned. The stock-flow-consistent core checks its accounting identities every step; no vendor model is deciding what it will or won't say about instability. The model proposes; the verifier — the accounting identity and the out-of-sample calibration — decides.
-
It's reproducible. Same scenario in, same numbers out:
dyno-sim run --scenario scenarios/samples/sample-fiscal-monetary-us-2025-baseline.json # tail-risk surface at the converged step size (dt=0.0625) the numbers above cite: python scripts/run-tail-risk-surface.py \ --scenario scenarios/samples/converged/sample-fiscal-monetary-us-2025-tail-risk-dt0p0625.json --runs 20000 -
It's honest about its limits. This is a single short rate (no term structure), a canonical policy reaction function (not a fitted DSGE), and an illustrative scenario rather than a point forecast. One private-debt aggregate (the household/corporate split is registered future work), interest-only debt service, and a valuation input that currently leans on a mirror source — all named. Naming what it doesn't do is how you know the rest is real.
The fiscal-monetary run's page in the Dyno-Sim operator dashboard: decision question, cross-check badge against real FRED data (GFC-2008 and COVID-2020), and the run overview
Every number above traces to a run id. The central fiscal-monetary run (run-20260624T062013Z,
24 quarterly steps, 120 metric columns) as it appears in the operator dashboard — including its
cross-check badge against real FRED data (fed funds rate + unemployment, over the GFC-2008 and
COVID-2020 episodes) and the five validation profiles that target it.
What this looks like pointed at your portfolio
The run above answers a public question with public data. A Simulation Decision Pack points the same machinery at yours — bounded scope, one target question, calibrated to your exposure:
- Where is your book on the leverage/credit map, and how close is the turn that the public-side policy response won't catch?
- Which hedges or allocation shifts measurably 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 the tripwire above standing watch — and how robust is the read across the ensemble?
For stewards who want this standing rather than one-off — the same owned simulators plus senior judgment on retainer, as questions arrive — that's the fractional-advisor relationship. Either way the engagement starts from a run you can audit, not an opinion you have to trust.
Bottom line. Achieved: an owned, accounting-consistent simulator — calibrated to real GFC + COVID FRED data — locates US fragility in the private balance sheet and quantifies how robust that call is on the composed model itself: a 26.6% debt-spiral rate and an 18.0% recommendation-flip, every event defined inline, every probability with its interval — plus a sealed, dated prediction (
pred-2026-07-25-014) the reader can score against us in early 2027. Business value: a decision-grade, auditable read that states its own uncertainty — what a risk committee can budget against and a benchmark or house-view memo cannot deliver.
Comuvia builds decision support on owned, accounting-consistent simulators, and works as a fractional advisor to the people who carry the tail. This piece uses stylized sample scenarios to show the method; engagements calibrate to your data. These runs cost roughly $0 of inference (local compute, free public data). Not investment advice.