Four separate studies sat in four separate files: a factor presence matrix of 56 factors scored against sixteen independent sources, a stack coverage matrix of seven AI layers against five diagnostic pillars, a diagnostic work order, and a data plan of eight variable categories with their measurement problems. Each answered a different question and none of them talked to the others. Since merging, the panel has taken on company filings, four research houses, one payments index, and three feeds of its own.
The governing text is The Architecture of Bubbles, revised edition, September 2026. The four studies supply the raw material, the attestation counts and the instrument status. It decides what belongs here, how it is organised, and which observables are allowed to count as instruments. It supersedes Book 2 and leaves it structurally intact: the same four Parts in the same order, the same fourteen books in the bibliography, and Part IV, which governs the whole Distress phase, unchanged in substance. What it adds is one new section in Part III, Geopolitical Amplifier: A Global Economy More Expensive to Finance, and with it Pape, Mearsheimer, Macgregor and Sachs, who enter the governing text for the first time. Its Part I supplies the second axis in 01b, its Part II the liquidity section in 02b, its Part III the AI capital cycle, and its Part IV the whole of the Distress phase and its convergence test. The scoping text that opened this project is superseded and is kept for provenance rather than authority.
This panel puts them on one spine. The spine is the five-phase sequence, cut into nineteen named signals. Every signal carries three things and only three: how many independent sources attest it historically, what instrument reads it today, and what that instrument currently says.
Every signal is now read. Eleven run entirely on fetched series, eight on a mix of fetched and hand-collected readings, and none is left unread. Where a figure cannot be fetched the panel says which one and why, and where a factor cannot be instrumented at all it says that too: two of them are historical by construction, and their observables are 1897 asset growth and VOC share-trading records.
The sequence itself is the best-attested object in the corpus: eight of the sixteen sources describe the same displacement to revulsion path independently. What follows is that path cut into nineteen signals. Three came from the Distress chapter and a fourth, official price support, from the pattern of state intervention in prices. Attestation counts how many independent sources name the signal. Instrument is what the project can actually read: a wired series, a dated reading collected by hand, or nothing. The two are not correlated, and the gap between them is the point of the board.
The instrument column runs on four states. Pulled live means a series measures the signal and a current value was fetched. Wired, not pulled means the series exists in the project but no value was taken. Read by hand means a dated reading from a named source with no series behind it. No instrument means nothing. A factor is promoted to live only where a pulled series measures the thing the factor claims, and adjacency does not count: the global liquidity cycle stays read by hand because the index itself is not wired, however much M2 and the Fed balance sheet sit beside it.
The state is decided by coverage rather than by the best carrier. A signal reads pulled live only when a majority of its factors are measured by a fetched series; one live factor among five earns part pulled, not more. This matters: official price support would otherwise have read as fully live on the strength of a breakevens series that measures a different condition entirely.
Signals repeat across phases where the sources repeat them. Credit appears at 8, again at 10, again at 12, because credit growth, credit acceleration and leverage are three different readings of one mechanism at three different pressures, and the sources treat them separately.
The signal board above is arranged on one axis: the five-phase sequence, displacement through revulsion. That axis answers where in the sequence are we, which is a timing question the corpus explicitly declines to answer. There is no crash clock in the framework, and high valuation does not specify when valuation will fall.
This is the second axis, and it asks a different question: which mechanisms are present, and are they feeding one another? Five load-bearing conditions run through all five phases at once. Their importance lies less in their presence than in the connections among them, and the book is explicit that convergence matters more than the presence of any single symptom. Each of the historical episodes became consequential because mechanisms that could have stayed individually manageable began to reinforce one another.
How the tagging was done, and its limit. A factor is tagged to a condition only where the source text itself names that condition for that mechanism, and every tag carries the quotation it was made on, shown on the factor. Where the text is silent the factor stays unassigned and is counted as unassigned below. Tagging all fifty-six would be authoring a mapping and then presenting it as though the corpus had supplied it, which is the one thing this panel exists not to do.
Phase 4 was the hole in this panel. It now carries three signals of its own, and an instrument that the other four phases do not have: a convergence test. Distress cannot be declared by any single variable, because every individual observation is ambiguous. High capex can accompany excellent opportunities. Wider spreads can come from macro conditions unrelated to AI. Insider sales can reflect diversification. The diagnosis strengthens only when independent channels begin to reinforce one another.
Signal 16 is the newest and works the same way. No single act of official price support is diagnostic. Governments buy assets, release reserves and talk to markets in ordinary times as well as fraught ones. What the architecture asks is whether the interventions converge: equities, bonds, currency and oil defended at once, by the same actors, in the same window, and with diminishing effect. That conjunction is what separates policy from support.
Read the scorecard below for rates of change and divergences, not for levels. The absolute high-yield spread matters less than whether it is widening while equity valuations stay elevated. The level of capex matters less than whether capex accelerates while cash conversion weakens. Every figure below carries its own pull date, and the masthead stamp says when the rail was last refreshed. One reading in this section is an exception and is labelled where it appears: the SEC filing set spans fiscal years that do not align. The single-name CDS levels are no longer read by hand. ICE Clear Credit publishes end-of-day settlement prices for five-year single-name CDS free and without a key, so eight of the nine names are now fetched; Micron has no cleared contract in that file and is carried as absent. What ICE publishes is a price, so the spread is a stated conversion rather than a quoted level.
This panel already pulled the monetary rail and, until now, read it as one thing. The argument this section carries is that money, credit, funding liquidity and market liquidity are not synonyms. They move together for long periods, which encourages the habit of speaking of them as one substance, and in a disturbance they separate abruptly. The divergence between them is the signal; the level of any one of them is not.
Money is the asset in which obligations are finally settled. Credit is a contractual claim that lets expenditure happen before income is earned. Funding liquidity is whether an institution can raise cash by issuing liabilities. Market liquidity is whether an asset can be sold quickly and in size without moving its price. Three of the four are now instrumented on this board. The fourth is the honest hole, and it is stated as one rather than approximated with something adjacent.
Eight data categories, each with its variables, its sources and the measurement problem that will not go away. These map almost one to one onto the five diagnostic pillars of the attached panel, which is the join between the historical framework and the live case. Each card lists which signals it feeds.
Four factors sit outside the phase sequence. They are not conditions of a bubble, they are attempts to measure one. Two of them currently contradict the framework, and both contradictions are informative rather than fatal.
The framework above is a claim about all bubbles. This is the live case, and it is deliberately kept separate, because a scale reading is not a verdict. The move that makes it tractable is decomposition: "is AI a bubble" is unanswerable as a binary claim about all AI assets, because risk can concentrate in one layer while another is genuinely supported by cash flows. Score seven layers against five pillars, then label a segment only where several independent method families agree.
Two inputs build this panel: the stack coverage matrix, which is the honest inventory of what can and cannot be measured per layer, and the geopolitical case study, which supplies the shock that turns a stretched market into a broken one. The second used to be the least instrumented well-attested factor in the corpus. It was wired on this pass, and the title passed to regulatory complacency and moral suasion failure, attested by seven sources and still carrying nothing.
Seven layers down, five pillars across. Owned means an instrument exists and points at this layer. Ready means working code exists in the project but is aimed elsewhere. Proxy means an index-level approximation only, which is not layer resolution. None means nothing, and the empty rows are left empty rather than filled with an index stand-in. Open a row for what the project currently stands on and the cheapest next instrument.
One feature of this cycle deserves separate treatment, because it is the mechanism through which capex is manufactured rather than demanded. A hardware supplier or cloud provider funds its own customers in a closed loop. It is not illegal, and it is an echo chamber: cash out logged as long-term investment, cash in logged as immediate revenue. Faculty at INSEAD, among them Boris Vallée, draw the near-exact parallel to telecom vendor financing in 1999 and 2000.
The condition that has to hold for the build-out to be an investment rather than a transfer. Revenue must cover depreciation, operating cost and the cost of the capital tied up. Everything turns on L, the useful life of the asset, which nobody discloses and which the whole argument between the bulls and the bears is really about. Move the sliders.
The thesis this section now carries is not that a war will break the market. It is that the accumulation of simultaneous conflicts makes the global economy more expensive to finance, and that the transmission into AI equity valuations runs through the discount rate rather than through any single military event. Higher defence spending, larger deficits, costlier energy, duplicated supply chains and heavier sovereign issuance each reduce one strategic vulnerability while raising the demand for capital. The cumulative effect is a higher required return, and long-duration assets are arithmetically the most sensitive thing to it.
That reframes what a geopolitical section is for. The question is not whether AI will generate extraordinary productivity gains. It is whether those gains arrive in a financial environment capable of supporting the valuations already attached to them. A technology can transform the economy while its securities deliver poor returns, which is the same proposition signal 01 reads from the NIPA record and Chancellor's capital cycle reads from the railways.
This section used to be the widest gap on the board. War finance and sovereign events as displacement is attested by six independent historical sources and carried no instrument at all. It now carries one: the financing channel is readable, and the variables are below with their dates. What is still missing is the event chronology to read them against.
Where each historical signal lands in the live stack. A signal with a strong historical attestation and a single exposed layer is a concentrated risk. A signal that lands everywhere is a systemic one.
Every empty cell on the coverage matrix, with the instrument that would fill it, sorted so the cheap and public work comes first. This is the section that turns the panel from a diagnosis into a work order. Open a layer row above for the same content in context.
Four cells resist filling for reasons that are not about effort, and they are listed separately underneath. Naming them is part of the answer, not a failure to answer.
Six sources on two dates. Three added on 26 August 2026: two are forecasts from named houses and are labelled as forecasts throughout, never as observations, and the third is a supplied chart that was reconciled against the filing. Three supplied charts added on 6 September 2026 follow underneath: an off-balance-sheet tally from Morgan Stanley, an issuance ratio from J.P. Morgan Asset Management, and a spend-concentration series from Ramp. Where a figure could not be reconciled, or could only be read by eye, that is stated on the card.
The reason this material matters more than its volume suggests: UBS supplies revenue forecasts split by layer, which is the denominator this panel has been missing since it was built. The capex-payback calculator in section 07 has been running on a placeholder because no disclosed AI revenue series exists at layer resolution. A named forecast is not a disclosure, but it is a testable proposition, and it can be put against the capex.
Both free, both keyless, both added on this pass. Token pricing is the only input in the project that cannot be built retroactively, so it starts accumulating from today. Filing mentions measure how fast the language is spreading, which is not the same as how fast the technology is being deployed, and the gap between the two is the finding.
Everything fetched, with its identifier, its observation date and its value. Two sources, both free and needing no API key: FRED's CSV endpoint, and the SEC's XBRL company-concept API. Re-run refresh.py in this folder to re-stamp it.
Two caveats carried here rather than buried. The Wilshire market-cap series FRED used to publish is discontinued, so the Buffett indicator below is computed from Z.1 corporate equities over nominal GDP, a different definition that does not reproduce the 237% hand-collected reading. And margin debt is not on FRED at all: it stays a FINRA hand-pull, and is labelled as such wherever it appears.
Nine names, split between the firms doing the spending and the firms booking it as revenue. Latest annual figures, and the fiscal years do not align, so each row carries its own period end. Comparing Nvidia's January year against Microsoft's June year is a limitation of the source, not a choice.
Twenty voice codes, every one resolving to a named source. Fourteen carry factor attestation and six do not. The attestation counts elsewhere on this board are stated against sixteen, and that is deliberate: the factor matrix was scored against sixteen sources and no later arrival has been backfilled into counts it did not earn. Pepper and Shin entered from the bibliography, ZHS was resolved from the source register, and Pape, Mearsheimer, Macgregor and Sachs are the geopolitical contributors. Their standing changed with the revised edition. All four were absent from Book 2, Sachs but for a single mention, and the September 2026 revision argues them at length in the geopolitical amplifier section of Part III, so they are cited to the governing text rather than to a reading the panel cannot check. No document of their own is held in this room, so the primary sources remain absent, and none of the four carries factor attestation: the matrix was scored against sixteen sources before any of them arrived. The count after each name is how many of the nineteen signals that source attests.