Conditions
29
Across 6 frameworks
A framework without break conditions is a religion, and a break-condition page nobody scores is decoration. So we score it. Each condition below is quoted in condensed form from the published note that set it, dated to the day it was set, and carried at one of six states, defined in the vocabulary below. Each condition also carries a stable ID, the date it was last reviewed, the next disclosure or event we expect to score it against, and, where it shares an underlying observation with other conditions, a cluster tag. A condition never disappears from this page; it changes state in public, and when its text is ever revised, the original stays printed above the revision, struck but legible.
Board revision, September 6-7, 2026
This board now carries six states in place of the old holding-or-triggered binary, because "no trigger observed" is not the same thing as confirmed healthy, and we will not color a condition green when the evidence it needs does not exist. Three conditions, TW-1, NV-2, and TT-1, received an arithmetic revision: the originals compared growth rates where the correct test is the product of price and volume changes, and each original remains printed above its revision, struck but legible. The board also now discloses which conditions share an underlying observation, so 29 conditions are not mistaken for 29 independent pieces of support. Every change here strengthens what it would take to falsify these frameworks; no threshold was softened.
This board now carries six states in place of the old holding-or-triggered binary, because "no trigger observed" is not the same thing as confirmed healthy, and we will not color a condition green when the evidence it needs does not exist. Three conditions, TW-1, NV-2, and TT-1, received an arithmetic revision: the originals compared growth rates where the correct test is the product of price and volume changes, and each original remains printed above its revision, struck but legible. The board also now discloses which conditions share an underlying observation, so 29 conditions are not mistaken for 29 independent pieces of support. Every change here strengthens what it would take to falsify these frameworks; no threshold was softened.
The board
Unchanged
18
As of Sep 6, 2026
Improving
4
As of Sep 6, 2026
Deteriorating
2
As of Sep 6, 2026
Insufficient evidence
5
As of Sep 6, 2026
Breached
0
As of Sep 6, 2026
A board with no breaches deserves a caveat rather than a victory lap: every framework below was published within the past six weeks, and several conditions were set only days before the latest review. Five conditions are marked insufficient evidence precisely because the observable they need is not yet published; we say so instead of counting them as support. The proof accumulates one re-underwriting at a time, on this page.
We also grade our old calls
In February 2026 we argued the winter selloff was rotation, not destruction. Six months on, we graded all eight legs: five right, one early, one wrong, one still open.
In February 2026 we argued the winter selloff was rotation, not destruction. Six months on, we graded all eight legs: five right, one early, one wrong, one still open.
Status vocabulary
UnchangedReviewed on the printed date; no new evidence moved the condition either way.
ImprovingNew dated evidence moved the condition further from triggering.
New dated evidence moved the condition closer to triggering, but the threshold in its own text has not been met.
BreachedThe threshold was met per the condition's own text. We state what the evidence was, and what changed in our view because of it.
Insufficient evidenceThe observable the condition needs is not currently published or measurable. We say so instead of coloring it green.
SupersededReplaced by a revised condition. The original stays rendered on its card, struck but legible, above the text that replaced it.
One rule governs all six: "no trigger observed" is never silently rendered as confirmed healthy. A condition earns an improving mark only from dated evidence, and a condition whose evidence cannot currently be observed says exactly that.
The Three Waves framework
The Three Waves of AI
Source: The Three Waves of AI and the $250 Billion Question, July 27, 2026. Four conditions, monitored quarterly per the note.
Improving
The efficiency ratio turns. Total token consumption growth falls below token price deflation for four consecutive quarters, breaking the volume flywheel under waves one and two. Set Jul 27, 2026
Set Jul 27, 2026
Revised September 6, 2026: arithmetic correction. The original compared growth rates, but revenue is the product of price and volume changes, so the original test could stay silent through a revenue contraction.
The efficiency ratio turns. Realized AI compute revenue contracts for four consecutive quarters: the product of token price change and token volume change (revenue ratio = price ratio x volume ratio for comparable units, or appropriately weighted realized revenue where model and workload mix shifts) falls below 1, breaking the volume flywheel under waves one and two. Volume growth merely exceeding the percentage price decline is not sufficient: price down 50 percent with volume up 60 percent still shrinks revenue 20 percent.
Unchanged
Set Jul 27, 2026
The profitable buyers blink. Alphabet, Microsoft, Amazon, and Meta begin cutting AI capital spending while only speculative, credit-dependent buyers keep building.
Unchanged
Set Jul 27, 2026
The financing spreads. Guarantees extend to the third, fourth, and fifth customers, meaning the demand curve as a whole cannot fund itself: the telecom-era pattern.
The lock-in fails anyway. OpenAI proceeds with large-scale custom silicon despite the financing relationship, leaving the credit risk without the strategic benefit.
The constructive NVIDIA view
NVIDIA After the Q2 FY2027 Print
Source: NVIDIA's $96 Billion Quarter: What the Print Settled, and What It Didn't, August 30, 2026. Five conditions, tracked quarterly per the note.
Unchanged
Set Aug 30, 2026
Guarantee contagion. Customer guarantees spread materially beyond the single disclosed OpenAI project toward the broader customer list: the Lucent signature, which would change the view regardless of reported growth.
Improving
The demand ratio turning. Industry token volume growth slows below token price deflation on a sustained basis, shrinking AI compute revenue while capex is still landing. Set Aug 30, 2026
Set Aug 30, 2026
Revised September 6, 2026: arithmetic correction. The original compared growth rates, but revenue is the product of price and volume changes, so the original test could stay silent through a revenue contraction.
The demand ratio turning. Realized AI compute revenue shrinks on a sustained basis while capex is still landing: price ratio times volume ratio below 1 across comparable units, or weighted realized revenue declining where mix shifts. Supplier-reported AI revenue is an indirect proxy for this and is flagged as such when used; end-customer monetization is tracked separately.
Unchanged
Set Aug 30, 2026
Profitable-buyer capex rolling over. Two consecutive quarters of hyperscaler capex guide-downs, signaling the buildout's foundation cracking before any NVIDIA print would show it.
Unchanged
Set Aug 30, 2026
The margin floor breaking. A gross-margin print below the guided 71 to 72 percent trough, or a push-out of the recovery, signaling memory costs are structural rather than cyclical.
Unchanged
Set Aug 30, 2026
A cut to the fiscal 2028 outlook. Any reduction of the roughly 70 percent growth guide is a material negative event, whatever the explanation.
The Token Tollbooth framework
The Token Tollbooth
Source: The Token Tollbooth: Where AI's Value Settles When the Waves End, August 30, 2026. Four conditions, revisited quarterly per the note.
Improving
The Jevons ratio turning. Token demand growth falls below token price deflation for four consecutive quarters, meaning the industrialization phase is not funding itself and the core of the framework breaks. Set Aug 30, 2026
Set Aug 30, 2026
Revised September 6, 2026: arithmetic correction. The original compared growth rates, but revenue is the product of price and volume changes, so the original test could stay silent through a revenue contraction.
The Jevons ratio turning. Realized token revenue (price ratio times volume ratio) contracts for four consecutive quarters, meaning the industrialization phase is not funding itself and the core of the framework breaks. Where realized revenue is not directly observable across the industry, we state what cannot be observed rather than substituting a growth-rate comparison that can miss revenue declines.
Unchanged
Set Aug 30, 2026
Distribution without a landlord. A frontier lab demonstrably serves planetary-scale distribution without a hyperscaler partner: its own datacenters, power, and enterprise and consumer reach.
Unchanged
Set Aug 30, 2026
Structural separation. Antitrust or sovereignty remedies force a durable separation of model, cloud, and distribution, dismantling the integrated-stack advantage by rule rather than by competition.
Improving
Set Aug 30, 2026
The capex throttle failing. AI revenue growth decelerates materially while capital spending guidance keeps rising for several consecutive quarters, ending the self-correcting flywheel argument.
The Energy + Nuclear + AI thesis
Energy + Nuclear + AI
Source: Powering Intelligence: The AI Buildout's Next Bottleneck Is Electricity, August 30, 2026. Five conditions.
Unchanged
Set Aug 30, 2026
Repeatable fusion net gain. A demonstrated, repeatable, system-level net-energy result from any credible developer, or a fusion PPA priced at or below firm fission power with delivery guarantees.
Unchanged
Set Aug 30, 2026
First-power slippage past the procurement window. Leading SMR programs slip materially past their 2027 to 2028 first-power targets while hyperscaler procurement is filled by gas, restarts, and renewables; new marquee offtake stops flowing to SMR developers for several consecutive quarters.
Unchanged
Set Aug 30, 2026
A safety incident anywhere in the class. Any significant safety event at any advanced reactor, in any country, re-rating the complex and likely reversing the policy tailwind. Reassessed that day, not the following quarter.
Unchanged
Set Aug 30, 2026
Regulatory reversal. A change in administration posture, a slowdown in NRC modernization, or a rollback of the DOE pathway, restoring the decade-long licensing cycle the thesis assumes is gone.
Unchanged
Set Aug 30, 2026
The demand side softening. Two consecutive quarters of hyperscaler capex guide-downs, or grid interconnection debottlenecking materially faster than expected, removing the scarcity that makes time-to-power valuable.
The Fourth Leg framework
The Fourth Leg
Source: The Fourth Leg: Where the AI Trade Goes After Memory, August 31, 2026. Five conditions; framework first developed June 10, 2026.
Insufficient evidence
Set Aug 31, 2026
Agent revenue decelerating hard. Disclosed agent revenue lines at the large software platforms decelerate sharply and persistently from their 2026 growth rates, breaking the monetization inflection on its own evidence.
Insufficient evidence
Set Aug 31, 2026
Seat loss showing up in retention. Net revenue retention at the anchor platforms begins visibly reflecting seat compression across multiple quarters, vindicating the market's first-quarter 2026 verdict.
Unchanged
Set Aug 31, 2026
Monetization skipping the application layer. AI revenue growth concentrates overwhelmingly in hyperscalers and model providers while application-layer agent lines stagnate: the fourth leg exists, but the expression was wrong.
Unchanged
Set Aug 31, 2026
800VDC failing to leave the reference design. Major hyperscalers decline to adopt 800-volt DC distribution outside NVIDIA's own rack architectures, or the next rack generations slip materially, flattening the power-silicon content slope.
The macro overwhelming everything. A sustained rate-hiking cycle or an AI-financing credit event de-rates every leg together regardless of fundamentals, making the framework unactionable for as long as the regime lasts.
The Displacement Cascade framework
The Displacement Cascade
Source: The Displacement Cascade: What Happens If AI Empties the Office, August 31, 2026. Six conditions; framework first developed February 2026.
Unchanged
Set Aug 31, 2026
The unemployment crossover reversing. White-collar unemployment falls back below blue-collar and stays there across several quarters of BLS data, making the framework's labor-market signature cyclical noise.
Insufficient evidence
Set Aug 31, 2026
AI attribution staying trivial. As attribution methods improve, explicitly AI-linked job cuts remain a low single-digit share of total announced cuts through 2027.
Unchanged
Set Aug 31, 2026
Household credit stabilizing. Serious delinquency rates in the New York Fed data plateau or decline from early-2026 levels while savings rates recover, unloading the cascade's transmission mechanism.
Insufficient evidence
Set Aug 31, 2026
Seat counts holding at the platforms. Disclosed retention and seat metrics at the large per-seat software platforms show no compression through fiscal 2027, breaking the third order's software leg in both its original and refined forms.
Unchanged
Set Aug 31, 2026
Payer mix holding. Commercial covered lives and insurer benefit-expense ratios remain stable through 2027, whether because displacement stalled or because policy absorbed it.
Insufficient evidence
Set Aug 31, 2026
New-role absorption showing up. Job creation in AI-adjacent categories visibly absorbs displaced workers at scale in BLS category data, proving the WEF projection right and ours wrong.
Independence: what actually moves together
Twenty-nine conditions do not mean twenty-nine independent pieces of support. Several conditions across frameworks read the same underlying observation, so one data point moves all of them at once. Accounting for the clusters below, the 29 conditions rest on roughly 22 independent observations, and we print the clusters so readers do not double-count support.
Cluster A, token demand vs price: TW-1, NV-2, TT-1, and the revenue side of TT-4 share the same underlying observation. One data point moves all four.
Cluster B, hyperscaler capex: TW-2, NV-3, EN-5, and the capex side of TT-4 share the October and January guide cycle.
Cluster C, vendor financing spread: TW-3 and NV-1 are the same observable read at two thresholds.
Cluster D, per-seat software retention: FL-2 and DC-4 are the same disclosure read for two frameworks.
"Frameworks that cannot be falsified are marketing, not research. This page is where ours stand trial."
How this page works
Checked, not assumed. Every state on this page reflects a review performed on the printed date, most recently September 6, 2026. States are re-evaluated whenever a framework is re-underwritten: on the quarterly cadence each source note commits to, and immediately when material evidence arrives between checks. An improving or deteriorating mark cites dated evidence; it is never a mood.
A breach gets announced, not buried. If a condition's threshold is met, the state changes here, the change is dated, and the note that explains what the evidence was and what we did about it gets published and linked from the row. Quietly deleting a breached condition would defeat the only purpose this page has.
Original text is immutable; revisions append. When a condition's text is revised, the revision carries a date and a reason, the original stays printed on the card, struck but legible, and the set-date never changes. The first three revisions on this board (TW-1, NV-2, TT-1) are arithmetic corrections dated September 6, 2026, disclosed in the note at the top of this page rather than left for readers to discover by diffing.
The wording is condensed, never changed. Row text is shortened from the source notes for scanning; the full condition, with its context and caveats, lives at the linked note, which is the authoritative version. If a condensation ever alters a meaning, that is an error and we will correct it in place, the way The Record handles corrections.
Machine-readable. This board is also published as conditions.json, free to read and cite with attribution, so research agents can consume the same states, revision history, and clusters that human readers see here.
Members receive the notes where these conditions are set, argued, and re-underwritten, as they publish, for $99.95 per year. The founding rate closes September 29, 2026.
No payment today; founding rate locked. First member mailing: mid-September 2026. Membership details
How to read this page
Each condition is a condensed restatement of a break condition published in the linked source note, which remains the authoritative text for the original wording; revised text on this page is authoritative for the revision and is dated and reasoned where it appears. States reflect our own judgment as of the review date printed above, based on the evidence available to us at that time; they are not audited, and an UNCHANGED or IMPROVING state is a snapshot, not a prediction. Break conditions are opinions about our own frameworks: statements of what evidence would change our views, not forecasts that such evidence will or will not arrive. This page is updated periodically and was last reviewed September 6, 2026.
Not investment advice
This page is published by Three Douglas, LLC for informational and educational purposes only. It is not investment advice, an offer, or a solicitation to buy or sell any security. Three Douglas, LLC and affiliated persons hold positions in securities discussed in the linked notes and may transact in them at any time without notice. Investing involves risk, including possible loss of the entire investment. Past performance is not indicative of future results.