Permissioned Abundance
Pre-registered forecasts on modern mercantilism × AI
By 2030 intelligence is abundant and cheap, and the scarce factor is permission, the energized megawatt, the licensed chip, the approved supply chain, the chartered settlement rail; these fourteen forecasts price where that binds and who keeps the rent. Read the full research paper →
The Risk Museum: I test ideas for trading strategies the way a quant desk would, on an engine I built, and I publish the ones that fail my gates. New here? Start at the entrance →
- What this is. 23 yes/no predictions on the 2026–2030 macro cycle: tariffs, AI, and where they collide. Each one has a hard, objective resolution rule (spelled out on its card below) and a probability I committed to in advance.
- Why it counts. Every price was locked in and dated (2026-07-21), each revision is logged, and the slate freezes at submission, so nothing gets graded with hindsight. The same discipline the museum applies to its trading strategies, pointed at the macro cycle.
The fine print: revisions, conviction tiers, and the base-rate band
Priced 2026-07-21; dated revisions 2026-07-23 (log: 23a igv 29->35, btos 45->58; 23b adversarial pass wedge 33->34, nvda 39->34, igv 35->30, usg 45->52, sec301 72->83). AMENDMENT 2026-07-24: the revision window, originally closing 2026-07-25, is extended to 2026-07-30, the week's evidence flow (chokepoint stress, a second open-model shock, major-bank rate views) warranted more updating time. Every revision remains dated and logged; CORRECTION 2026-07-28: F1 base_rate_current tier counts aligned to USTR's final-action fact sheet, 17 economies at the flat 10% (ban / ART commitment / partial regime), 43 at 12.5%, five partly net of MFN; evidence field only, probability and criteria unchanged. CORRECTION 2026-07-28b: F9 base_rate_current adds CY2023 (1,037t, WGC) so the consecutive-year count is checkable on the page; evidence field only. the slate freezes permanently when it is submitted to an external forecasting contest on 2026-07-30, after that, no edits are possible. The base-rate band shown on each forecast is our reference-class estimate, how often events like this have happened historically, computed before we committed a probability, so you can see where our price agrees or disagrees. REVISION 2026-07-26: the original F1 (Sec 301 duties in effect on >=20 of 60 economies at 2026-12-31, priced 83) is RETIRED, not repriced. Its resolving event occurred on 2026-07-24 when duties took legal effect on all 60 economies, and a decided event does not belong in a forecast slate. It is replaced by a forward test of whether the wall survives deals and litigation, >=40 of 60 still in force at 2027-12-31, priced 70. The retirement is logged rather than hidden. BLOCK O AND G PRICED 2026-10-07: the owner accepted the anchored proposals in docs/research/soros_bet/07_pricing_anchors.md (O1 8, O2 30, O3 75, O4 70, O5 20, O6 5, G1 35, G2 30, G3 18); claim text, criteria, sources and horizons are unchanged, verified by recomputing the registration hash 2472c493 with the three pricing fields reset; the priced records hash to 7820b79bc06cc227. ERRATUM 2026-10-07 on O1 (oracle-ig-exit-2027) base_rate_current, frozen and unedited: the 249bp protection print is the DTCC daily median of 1 and 2 Oct 2026, read on 5 Oct 2026, not a 5 Oct print. Priced so far: 23/23.
Each forecast carries a declared conviction tier: CONVICTION (defended under pressure), AGAINST THE CROWD (a priced disagreement with consensus), DECLARED UNCERTAINTY (honestly near the coin, and saying so), or WORKING BOOK. Tiers follow the probabilities; if a price is revised before the freeze, its tier moves with it.
The revision log: every change, dated, in the open
The museum's macro wing: 13 binary forecasts on Modern Mercantilism, AI, and their intersection, the same pre-registration discipline the museum applies to trading strategies, pointed at the next five years. Objective resolution criteria; status tracked by the nightly pipeline. REVISION 2026-07-28c: slate narrowed 14 to 13 before the freeze. Withdrawn: china-chip-selfsuff-50 (third-party estimate stack too soft to headline), wgc-cb-gold-750t-2026 (WGC's own projection brackets the threshold, low information per Brier point), usg-equity-stakes-5-more-2028 (cross-loads the F3/F4 chokepoint driver). Added: xlu-igv-permission-rent-2029 (the rent-capture test the book lacked) and eu-gpai-enforcement-2027 (the non-US permission regime). All changes pre-submission; nothing had begun resolving. btos repriced 58 to 48, into its own band, shading survey-curve flattening below trend extrapolation. CORRECTION 2026-07-28d: F5 and F6 evidence fields add Apollo mid-year-outlook anchors (Slok, Jun-2026); evidence only, probabilities unchanged. CORRECTION 2026-07-28e (pre-freeze criteria repair, external red-team): F1 adds a 50% import-value floor so token residual duties cannot resolve YES; F10's vintage rule made symmetric (archived monthly vintages govern both YES and NO); F5 and F12 claim-descriptions rewritten from 'direct test' to the proxy/blunt-instrument descriptions they always were. Probabilities unchanged. REVISION 2026-07-28f (pre-freeze, merged from a second independent red-team build): slate widened 13 to 14 with seat-saas-economics-2030, a frozen 12-name cohort resolving on GAAP revenue growth and gross margin, the direct test of the seat thesis; F4 statement made identical to its dual resolver; F6 capex line defined (excludes acquisitions, finance leases; Microsoft calendarized); F7 result deadline extended to 2027-12-31 for postponement robustness; F11 pinned to the exact BTOS question with a successor rule; the EU forecast raised to a meaningful-action bar (Art. 92/93/101, an RFI alone does not count) and repriced 65 to 48 accordingly. All pre-submission; nothing had begun resolving. CORRECTION 2026-07-28g: F5 and F6 evidence fields add citations to Bridgewater's published AI-capex paper (Jan-2026); evidence only, probabilities unchanged. CORRECTION 2026-07-28h: punctuation normalized across meta and text fields (em-dashes to commas), house style; no substantive content changed. CORRECTION 2026-07-30i (the held-open relock, executed as promised): F5, F6, F10 and F3 evidence fields updated on the Q2 GDP advance, the Jul-29 hyperscaler prints, the Jul-29 FOMC, and H1 customs data. All probabilities reviewed against the new evidence and held unchanged. The slate is now FROZEN for submission. REGISTRATION 2026-10-05: nine forecasts added as a second family, backlog-credit-2026 (Block O, six claims on Oracle's investment-grade commitment and its allies, registered before the first Bloomberg rebalance in their window on 2026-10-30; Block G, three claims on PJM's capacity price collar and its 2030/31 sunset). The fourteen forecasts of the July slate are unchanged. The new nine carry probabilities of null until priced by hand, and the page shows 'pricing' until then; they are graded FORECAST ONLY and are not gate results. DISCLOSURE ON F7 (pjm-fourth-cap, frozen 2026-07-30, untouched): the 2029/30 price cap in force is $337.83/MW-day UCAP ($256.75 ICAP over a 0.76 reference factor), so F7's YES zone is the cap plus the sliver of the curve between $325.00 and $337.83; and PJM's 2026-08-13 filing ER26-3515 would exclude uncovered new large load from auction demand beginning with 2029/30, a rule change filed two weeks after F7 was frozen, with FERC's statutory date 2026-10-12. F7's wording and probability stand as frozen; the new claim pjm-2029-below-cap is scored jointly with it so that two YES results count as one event.
Before pricing these macro bets, the same engine made 620 predictions in a Jump Trading forecasting competition and was graded by a score, not a story. It came out calibrated (when it said 70%, things happened about 70% of the time) and finished in the top 6% of entrants.
The curve is the proof at a glance. Each dot is a group of forecasts, and the closer it sits to the dashed line, the better the calibration. My own misses are left in: a little underconfident in the middle, overconfident on a few near-certain calls.
One honest caveat: that competition was on soccer. It proves the method is calibrated and honest, not that the macro calls below are right. Those get their own score when they resolve (2026–2030). how the scoring works →
The Observatory
The live book · pre-registered bets, five themes, resolution order within each
| Forecast | Tier | P | Band | Expressed via | Crowding | Resolves |
|---|---|---|---|---|---|---|
| Trade & the tariff rebuild2 of 23 | ||||||
| The tariff wall holds, or the deals dismantle… | conviction | 70% | 60–80% | Macro, no clean sleeve | partly crowded | 2027-12-31 |
| Statutory vs collected, the wedge | contrarian | 34% | 25–45% | Macro, no clean sleeve | uncrowded | 2028-03-31 |
| China chokepoints & chips2 of 23 | ||||||
| Magnet flows, the chokepoint in a customs line | contrarian | 21% | 15–35% | Rare-earth miners (REMX) | uncrowded | 2028-02-28 |
| The double-veto chip trade | book | 34% | 25–45% | Semis (NVDA, SOXX) | uncrowded | 2028-03-31 |
| The AI buildout, capex & power4 of 23 | ||||||
| The +140bp claim, tested in the national acco… | conviction | 66% | 55–70% | AI-capex complex | partly crowded | 2027-03-31 |
| Is scarcity pricing structural? | conviction | 72% | 55–75% | Power / IPPs (CEG, VST) | partly crowded | 2027-12-31 |
| The first measurable signature of a bust | book | 62% | 50–65% | Hyperscalers + semis | partly crowded | 2028-03-15 |
| Does the permission holder keep the rent? | book | 40% | 28–45% | Macro, no clean sleeve | low | 2030-03-31 |
| The monetary order2 of 23 | ||||||
| Where the collision prices | contrarian | 55% | 55–75% | Long-end rates (steepener) | partly crowded | 2028-06-30 |
| Level vs share, the contrarian read | book | 60% | 50–65% | USTs / rates (TLT) | uncrowded | 2028-08-31 |
| The state & the real economy4 of 23 | ||||||
| Does the second permission regime act? | book | 48% | 40–60% | Macro, no clean sleeve | low | 2028-01-31 |
| Buildout vs adoption, measured | book | 48% | 35–55% | Broad AI adoption, diffuse | uncrowded | 2029-01-31 |
| The 82% survival bet | book | 30% | 15–30% | Software (IGV) short | partly crowded | 2029-12-31 |
| Does the SaaS tail preserve its public-equity… | book | 58% | 55–80% | Macro, no clean sleeve | low | 2030-06-30 |
| Elsewhere9 of 23 | ||||||
| Does the ally keep paying? | uncertain | 75% | n/a | Macro, no clean sleeve | partly crowded | 2027-01-31 |
| Does a referee start a clock? | uncertain | 30% | 5–25% | Macro, no clean sleeve | partly crowded | 2027-03-31 |
| Does the top of the loop keep raising? | contrarian | 70% | 60–85% | Macro, no clean sleeve | partly crowded | 2027-03-31 |
| Does the last notch hold? | contrarian | 8% | 2–20% | Macro, no clean sleeve | partly crowded | 2027-09-30 |
| Is the mechanism doing work? | book | 20% | 0–27% | Macro, no clean sleeve | partly crowded | 2027-09-30 |
| Delivered, and broke: the fiber pattern? | book | 5% | 0–8% | Macro, no clean sleeve | partly crowded | 2027-09-30 |
| Does the redefinition print a price? | uncertain | 35% | 24–40% | Macro, no clean sleeve | partly crowded | 2027-12-31 |
| Does the collar lapse? | contrarian | 30% | 15–50% | Macro, no clean sleeve | partly crowded | 2027-12-31 |
| If the collar lapses, does the price break? | book | 18% | 5–30% | Macro, no clean sleeve | partly crowded | 2027-12-31 |
Effective breadth, computed: these 23 forecasts carry the precision of about six to seven independent ones, so the standard error on a mean Brier over the book runs roughly 2.1 times what counting to 23 would suggest. The tightest link needs no judgement call: two of them settle on the same instrument, IGV total return, so one number moves both. 9 of the 23 sit downstream of another forecast in the declared causal graph. Nothing here is estimated from outcomes, because none have resolved; the dependency structure is declared and the arithmetic on top of it is deterministic. Skill compounds with the root of independent bets, and I will not pretend these are 23 of them. The scoreboard is the Brier score, reported when each forecast resolves.
Each forecast in full
Reading each gauge below: the shaded bar is the range of base rates (how often things like this have happened historically, set before I priced it); the marker is my committed probability, and it turns blue when it sits outside that range (I’m taking a view); Brier if YES / if NO is the score the call earns under each outcome, where 0 is perfect and lower is better.
Trade & the tariff rebuild
The tariff wall holds, or the deals dismantle it convictionP = 70%
On 2027-12-31, Section 301 forced-labor duties of at least 10% ad valorem remain in legal effect on imports from at least 40 of the 60 economies covered by USTR's 2026 final action.
Resolution criteria (binding)
Resolves YES if, as of 11:59pm ET on 2027-12-31, an additional ad valorem duty of at least 10% imposed under USTR's Section 301 forced-labor proceeding (dockets USTR-2026-0265/-0266; affirmative determinations 2026-06-02; final action 2026-07-23; effective 12:01am ET 2026-07-24) is then in legal effect on imports from at least 40 distinct economies, AND the economies still covered accounted for at least 50% of the original 60-economy group's calendar-2025 US goods imports (US Census/USITC country import values, first-published annual figures). Both conditions must hold; the import-value floor prevents token residual coverage from counting as a standing wall. Counting rules: (a) economies as USTR enumerates them, EU counts as one; (b) an economy counts only if the duty under THIS proceeding is then in force, meaning it has not been terminated, has not been suspended for any reason including an executive agreement, trade deal, or compliance determination, has not been reduced below 10%, and is not vacated, stayed, or universally enjoined by a court order then in effect; plaintiff-specific relief does not disqualify an economy; (c) the 10% threshold is this proceeding's rate alone, never stacked with Section 122/232/338, IEEPA or successor, or other 301 dockets; (d) partial product coverage qualifies; (e) suspended duties do NOT count, whether or not the suspension is later reversed. Fewer than 40 qualifying economies at that moment resolves NO. Resolve from the Federal Register, USTR releases, and CIT/CAFC/SCOTUS dockets.
Base rate / current: SCOTUS voided the IEEPA tariffs 6-3 on 2026-02-20, demonstrating live judicial risk to tariff authority. USTR made affirmative determinations in all 60 forced-labor investigations on 2026-06-02, took final action 2026-07-23, and duties took legal effect 2026-07-24 as the Section 122 stopgap lapsed. Seventeen economies carry the flat 10% rate, reached through an existing forced-labor import ban, an Agreement on Reciprocal Trade commitment, or a partial regime (USTR final-action fact sheet); the other 43 carry 12.5%, five of them partly net of MFN. Affected governments began filing challenges immediately. The 2018-19 China Section 301 lists survived their court challenges.
What it tests: Modern Mercantilism is durable or negotiable. The wall was rebuilt on statutory authority within a day of the Section 122 lapse; the open question is whether deals and courts dismantle it faster than it was built.
Statutory vs collected, the wedge against the crowdP = 34%
US calculated duties as a share of customs value of goods imports for consumption exceed 9.0% for calendar year 2027.
Resolution criteria (binding)
Total calculated duties divided by total customs value, imports for consumption, all countries and commodities, CY2027, from USITC DataWeb (Census source data). Vintage: the annual figure as retrievable after the December-2027 statistical month first publishes (~early Feb 2028) and on or before 2028-03-31; later revisions ignored. Unrounded ratio; YES iff strictly >9.0% (exactly 9.0% = NO). This is the calculated-duties measure (PWBM tracker basis), NOT Treasury/CBP cash collections, refunds do not enter. If DataWeb is unavailable: official Census imports-for-consumption data with calculated-duty and customs-value fields (USA Trade Online); never collections data. If December 2027 is unpublished by 2028-03-31: the same ratio over cumulative CY2027 year-to-date months available.
Base rate / current: Yale statutory rate ~11.8% pre-substitution vs 7.2% collected (May 2026, PWBM), a 3-4.6pp wedge from exemptions, substitution, and transshipment. Pre-2025 baseline was 2.3%.
What it tests: Tariffs as a durable revenue and rebalancing instrument, tested at the cash register, not the announcement.
China chokepoints & chips
Magnet flows, the chokepoint in a customs line against the crowdP = 21%
China's CY2027 exports of metal permanent magnets (HS 8505.11, containing sintered NdFeB) to the US exceed its 2022-2024 annual average tonnage in GACC data.
Resolution criteria (binding)
Primary: China GACC (stats.customs.gov.cn), exports to the US under HS 8505.11 (sum of all Chinese 8-digit children incl. 8505.1110), kg converted to tonnes. YES if the CY2027 total (first GACC release covering full-year 2027) strictly exceeds the mean of CY2022/2023/2024 annual totals as displayed on the same platform at resolution; equal or lower = NO. First-retrieved figures govern. Fallback: if GACC has not published CY2027 country-level 8505.11 by 2028-02-21, or has discontinued/reclassified the line for 2022-2027, use US Census general imports from China, HS 8505.11 (USA Trade Online, customs quantity), identical comparison within the Census series. Resolution date 2028-02-28.
Base rate / current: H1-2026 flows still ~20% below the 2022-24 average despite the truce; extraterritorial controls postponed only one year; the reciprocal-tariff suspension expires 2026-11-10; US demand substitution is ramping. US-bound China magnet shipments kept falling in 2026 (Jan-Feb 994t, -22.5% y/y) even as global exports set records.
What it tests: Weaponized interdependence, chokepoints as standing instruments of statecraft. The purest Mercantilism x AI collision visible in a single customs line.
The double-veto chip trade working bookP = 34%
In any covered quarter from FY2027 Q2 through FY2028 Q4, Nvidia reports either at least $2.5 billion of China and Hong Kong geographic revenue or at least $2.0 billion of explicitly identified China data-center revenue.
Resolution criteria (binding)
Covered quarters end after 2026-07-21 and on or before 2028-01-31. PRIMARY: China including Hong Kong geographic revenue in a Form 10-Q or 10-K is at least $2.5B for one covered quarter. SECONDARY: Nvidia explicitly states China data-center revenue of at least $2.0B for one covered quarter in CFO commentary or prepared remarks. Qualitative language does not count. A restatement published by 2028-03-31 governs symmetrically. If neither condition is met, NO.
Base rate / current: $4.6B/quarter pre-ban went to zero; two US approvals (H20 with 15% revenue share, H200 with 25% surcharge) produced zero deliveries. Beijing buy-local directives block purchases; B30A unapproved; guidance assumes no China DC revenue.
What it tests: Is US-China tech trade a negotiable toll regime or a structural mutual embargo? Both capitals now hold a veto, the uncrowded read.
The AI buildout, capex & power
The +140bp claim, tested in the national accounts convictionP = 66%
The combined contribution of Information Processing Equipment plus Software to US real GDP growth for CY2026 (annual-average basis, BEA NIPA Table 1.5.2) is >=1.00 percentage points.
Resolution criteria (binding)
Sum the published annual-2026 contribution lines for 'Information processing equipment' (parent line under Equipment) and 'Software' (under Intellectual Property Products) in BEA NIPA Table 1.5.2. Annual-frequency figure (2026 average over 2025 average), not Q4/Q4. Governing vintage: the Q4-2026 third estimate (late March 2027) if published by 2027-03-31; else the most recent BEA release with annual-2026 data by that date; else the first release ever containing it. Later revisions ignored. Published two-decimal values summed, no re-rounding; exactly 1.00 = YES, <=0.99 = NO. Renamed lines resolve to successor lines; a renumbered table resolves to the successor expanded-detail contributions table.
Base rate / current: A widely-circulated institutional estimate (Jan 2026) puts AI capex contribution at ~+140bp to 2026 growth. The observable equipment+software lines contributed ~0.8-1.0pp in 2025 vs a ~0.3-0.5pp pre-2023 decade norm. >=1.00pp is dot-com-peak territory. Apollo's mid-year outlook (Slok, Jun-2026) independently puts AI spending's GDP lift near 1pp, a second named house on the same line the accounts must confirm. Bridgewater's own published estimate (Jensen, DeBois & Zimbler, The Macro Implications of the AI Capex Boom, Jan-2026): ~140bp of 2026 US growth and ~150bp of 2027, with their caveat that a material part is chipmaker profit unlikely to be recycled into the economy. Q2-2026 advance (Jul-30): headline +1.5% with equipment gains led by information-processing and software up; Q1-2026 contributions computed at 0.59pp equipment-IT plus 0.52pp software, 1.11pp combined, above threshold pace; Chair Warsh (Jul-29) cites nearly 20% four-quarter growth in the category.
What it tests: A deliberately broad national-accounts proxy for the AI investment cycle, the closest public series to the most-cited institutional claim of the cycle. A YES supports the capex-boom thesis; it does not isolate AI's own contribution.
Is scarcity pricing structural? convictionP = 72%
PJM's 2029/2030 Base Residual Auction clears RTO-wide at or above $325.00/MW-day, a fourth consecutive scarcity print.
Resolution criteria (binding)
Resolves YES if PJM's official 2029/30 Base Residual Auction results (report or news release on pjm.com) show an RTO-wide clearing price of $325.00/MW-day or higher, nominal, UCAP, as reported; exactly $325.00 counts. The at-cap language used elsewhere is descriptive only; this threshold governs. The auction is scheduled to close 2026-12-15. If PJM has not published official 2029/30 BRA results by 2027-12-31, whether through postponement, redesign or suspension, resolves NO. Latest official PJM figures as of 2027-03-31 govern.
Base rate / current: Cap hit three-for-three (2026/27 $329.17; 2027/28 $333.44; 2028/29 $325.00). Supply cannot respond by 2029/30: interconnection request-to-COD >4 years, turbine lead times 6-8 years, major turbine vendors sold out through 2030; datacenter demand growing ~5%/yr in commercial sales. Cap provenance: PA Gov. Shapiro's Dec-2024 FERC complaint led to the Jan-2025 PJM settlement (cap $325/MW-day, floor $175), FERC-approved Apr-2025 and extended Apr-2026 through the next two auctions, so the 2029/30 BRA prices under the same authored ceiling; PA claims $18.2B cumulative consumer savings (pa.gov, Jul-15-2026).
What it tests: AI capex as a durable macro force, tested at its most binding physical constraint (power), not at the spending line.
The first measurable signature of a bust working bookP = 62%
Combined CY2027 capital expenditure of Microsoft, Alphabet, Amazon, and Meta exceeds combined CY2026 capex by >=20%.
Resolution criteria (binding)
For each company, use the consolidated cash-flow line reporting cash paid for additions or purchases of property and equipment, or its direct successor line. Exclude finance-lease principal, acquisitions, and noncash additions. Calendarize Microsoft by summing the four calendar quarters; use calendar-year filings for Alphabet, Amazon, and Meta. YES if the CY2027 sum is at least 1.20 times CY2026. Restatements filed by 2028-03-15 govern symmetrically; if a company ceases filing, use the latest official filings for both years and apply the same treatment to numerator and denominator.
Base rate / current: CY2025 combined ~$380-400B; CY2026 guidance ~$500B+; the series has never declined year-over-year since 2015; the telecom analog flipped to -40% in year five (2001). The 2025 tax act's 100% immediate expensing subsidizes the marginal data-center dollar into 2027, a fiscal tailwind under the capex line (Apollo mid-year outlook, Jun-2026). Component scarcity corroborates persistence: SK Hynix has stated it is sold out of all memory capacity until 2027 (cited in the same Bridgewater paper). Q2 prints (Jul-29): Microsoft quarterly capex $35.8B, 2.1x year-ago, CY2026 plans unchanged; Meta raised its FY2026 floor to $130-145B.
What it tests: Compute +60%/yr and capex-boom continuation; a miss below +20% is the first falsifiable signature of the bust chapter the consensus macro view warns about.
Does the permission holder keep the rent? working bookP = 40%
Over 2027-2029, XLU (utilities) delivers a cumulative NAV total return at least 5.00 percentage points greater than IGV (software), the book's own thesis priced as one market test.
Resolution criteria (binding)
Window: 2026-12-31 close to 2029-12-31 close, dividends reinvested, net of expenses. PRIMARY RESOLVER: each sponsor's standardized 3-year average annual NAV total return as of the 2029-12-31 quarter-end (sectorspdrs.com/ssga.com for XLU, ishares.com for IGV), converted cum=(1+r)^3-1; resolves YES if XLU_cum - IGV_cum >= +5.00pp. First published vintage governs. Fallbacks in order: successor fund of an acquiring sponsor; S&P DJI gross index total returns for the underlying indices; Bloomberg adjusted closes. Hard deadline 2030-03-31; if no qualifying data by then, resolves NO.
Base rate / current: Computed, not asserted: 265 rolling monthly 3-year windows Jul-2004 to Jul-2026, XLU beat IGV by >=5.00pp in 27.9%; 22 calendar Dec-Dec windows, 7 of 22 = 31.8%; effective independent sample ~7 windows, standard error roughly +/-15pp. Five of the seven historical YES windows start inside the dot-com unwind. The trade has already run: XLU over IGV +25.4pp trailing 3y through Jul-2026. Both legs fee-inclusive.
What it tests: The thesis's own market test: if permission is the scarce asset, its holders should out-earn its renters.
The monetary order
Level vs share, the contrarian read working bookP = 60%
Total foreign holdings of US Treasuries (TIC Major Foreign Holders 'Grand Total', first print) reach >=$10 trillion for any data month from July 2026 through June 2028.
Resolution criteria (binding)
YES if the Grand Total of the Treasury's mfh.txt table is >= $10,000.0B as published to one decimal for any data month July 2026 through June 2028, using the FIRST publication of that month (a new month appears in the release ~mid month+2). Revisions and benchmark restatements ignored. mfh.txt is overwritten in place, first prints are verified contemporaneously or via a release-date archive (e.g. Wayback); the archived first print governs. A renamed table resolves to the successor total-foreign-UST line. The June-2028 print is scheduled ~2028-08-17. NO if no qualifying first print by 2028-08-31; publication delays do not extend the window; later-published figures do not count.
Base rate / current: All TIC figures here are first prints as published by the US Treasury, vintage stated for each. ~$9.2T (Dec-2025 data month), +$1.5T over four years (~+$375B/yr) even as China exits (~-$0.7T) and the foreign share of federal debt falls; ~$9,370B (May-2026 data month, first print); record $9,490B (Feb-2026 data month). Trend crosses $10T ~mid-2028, window-edge by design. No figure here is older than Dec-2025.
What it tests: The flow side of 'foreign demand for US debt is structurally weakening', tested in the level series the de-dollarization debate ignores in favor of share data.
The state & the real economy
Does the second permission regime act? working bookP = 48%
By 2027-12-31, the European Commission publicly begins an Article 92 GPAI model evaluation, adopts an Article 93 measure, or imposes an Article 101 fine against a named GPAI provider.
Resolution criteria (binding)
A named provider must be the subject of a Commission publication announcing an Article 92 model evaluation, an Article 93 corrective/interim/market-restriction measure, or an Article 101 fine under the AI Act GPAI regime. An Article 91 request for information alone, guidelines, speeches, national-authority action, or DSA/DMA/competition action does not count. First publication controls; publication grace through 2028-01-31 applies only to an action dated on or before 2027-12-31.
Base rate / current: Commission GPAI enforcement powers switch on 2026-08-02 (fines to 3% of worldwide turnover or EUR 15m). Analogues: DMA gatekeeper deadline 2024-03-07 to named proceedings in 18 days; DSA VLOP obligations Aug-2023 to formal proceedings against X in ~4 months; first DSA non-compliance fine EUR 120m, 2025-12-05. Against: no official AI Act case register, the Art. 52(6) systemic-risk list remains unpublished as of 2026-07-27, and the 2027-08-02 grace period shrinks the enforceable population.
What it tests: Regulatory mercantilism at the model layer, held to a meaningful-action bar: an information request alone does not count.
Buildout vs adoption, measured working bookP = 48%
The firm-weighted share of US businesses reporting current AI use in the Census Bureau's BTOS survey reaches >=25.0% in any biweekly period ending by 2028-12-31.
Resolution criteria (binding)
YES if any national firm-weighted BTOS estimate for the question labeled business currently using AI in producing goods or services is at least 25.0% for a period ending on or before 2028-12-31 and published by 2029-01-31. The employment-weighted series and future-use question do not count. A successor item counts only if Census explicitly identifies it in release notes or metadata as the continuation of that current-use series; otherwise the original series controls and discontinuation without a qualifying print resolves NO. Exactly 25.0% is YES.
Base rate / current: ~3.9% (fall 2023) → ~9% (2025, old wording); the Nov-2025 rewording lifted the level to ~18% at end-2025; 19.8% at the period ending 2026-05-03, the gap to 25% is ~5pp on the governing series.
What it tests: The buildout-vs-adoption pricing gap: YES = adoption arriving fast enough to justify the capex; NO = the gap persists into 2029, the bubble-scenario precondition.
The 82% survival bet working bookP = 30%
The iShares software ETF (IGV) underperforms SPY by >=25 percentage points in cumulative NAV total return from 2026-12-31 to 2029-12-31.
Resolution criteria (binding)
YES if IGV cumulative NAV total return minus SPY cumulative NAV total return over the window is <= -25.0pp (exactly -25.0 = YES). Primary: each issuer's published 3-year average annual NAV total return as of 2029-12-31 (ishares.com / ssga.com), cumulative = ((1+r)^3 - 1) x 100, difference unrounded. If either figure is not published by 2030-01-31: chain issuer daily NAV with distributions reinvested on ex-date (Morningstar NAV TR is an acceptable equivalent). First figures published after 2029-12-31 govern. Fund closure/merger/delisting → chain-link the realized fund return with the total-return benchmark index (S&P North American Expanded Technology Software TR / S&P 500 TR or successors) through 2029-12-31.
Base rate / current: A -25pp three-year sector shortfall vs the S&P is roughly a 1-in-5 event given 10-15%/yr tracking error; SaaS multiples already de-rated through 2025-26.
What it tests: A deliberately blunt index-level expression of the seat-repricing thesis. IGV contains the platforms that own the disruption, so this is not a direct test of the seat-priced tail; the tail argument lives in Section II.5 and resolves on economics, not this spread.
Does the SaaS tail preserve its public-equity economics? working bookP = 58%
Four or fewer companies in a frozen 12-name seat-priced application-software cohort remain independently public and report both at least 10% year-over-year revenue growth and at least 70% GAAP gross margin in their final qualifying fiscal year.
Resolution criteria (binding)
Frozen cohort at submission: Asana, Atlassian, Box, DocuSign, Dropbox, Freshworks, GitLab, monday.com, PagerDuty, RingCentral, Sprout Social, and Zoom Communications. An issuer passes only if (a) substantially the same operating company remains independently listed on a US exchange at 2029-12-31, and (b) its latest fiscal year ending from 2029-01-01 through 2030-03-31, filed by 2030-06-30, reports GAAP revenue growth of at least 10.0% year over year and GAAP gross margin of at least 70.0% (gross profit divided by revenue if not stated). Acquisition, take-private, liquidation, or delisting fails the independence gate; a renamed public successor counts. YES if four or fewer of 12 pass all gates. This is one joint cohort outcome; no marginal probabilities are multiplied.
Base rate / current: All 12 names independently listed as of 2026-07-28, chosen for seat-priced exposure before outcomes were knowable. The II.5 gate rates imply 2.6-4.2 of 12 passing all gates; all-software median net retention has fallen from 123 (early 2022) toward 105, and the public-software universe has been shrinking at roughly a tenth a year through take-privates and acquisitions.
What it tests: The direct test the index spread cannot give: a frozen, pre-registered 12-name cohort resolving on GAAP lines, one joint outcome, no multiplied marginals.
Elsewhere
Does the ally keep paying? declared uncertaintyP = 75%
Oracle's Form 10-Q for the quarter ending 30 Nov 2026 reports customer prepayments with a significant financing component for the second quarter of fiscal 2027 (the six-month figure minus $11,363m) below $11,363m.
Resolution criteria (binding)
Resolves YES if the cash-flow statement line 'Increase in deferred revenues from customer prepayments with significant financing component' in Oracle's first-filed 10-Q for the quarter ending 2026-11-30 gives a six-month figure such that the second-quarter amount (six-month figure minus the FQ1 figure of $11,363m) is below $11,363m. If the line is renamed, its stated successor governs; if it is removed, void. First-filed figures govern. Void if no 10-Q for that quarter is filed by 2027-01-31. Otherwise NO.
Base rate / current: FQ1 FY27 (quarter to 31 Aug 2026): $11,363m of such prepayments, none a year earlier, 2.47 times all of fiscal 2026 (Oracle 10-Q 0001193125-26-389274). These are funds from the buyers whose credit is the risk. A second quarter at or above FQ1 means the ally is still paying; a fall is the first filed sign that it is not.
What it tests: The reflexive loop refills its weakest member through customer prepayments; this reads whether the refill continues.
Does a referee start a clock? declared uncertaintyP = 30%
By 31 Mar 2027 at least one of: Moody's or Fitch lowers Oracle's senior unsecured rating; S&P places Oracle on CreditWatch with negative implications or lowers it; Moody's places Oracle on review for downgrade; Fitch places Oracle on Rating Watch Negative.
Resolution criteria (binding)
Resolves YES if any listed action is published by the agency on or before 2027-03-31: a downgrade of Oracle's senior unsecured rating by Moody's or Fitch; S&P CreditWatch negative or a downgrade; Moody's review for downgrade; Fitch Rating Watch Negative. A Fitch outlook change alone does not count. Source order: the agency's own release read at source and saved with its date; if it cannot be opened, two independent major news outlets naming the agency, the action and its date. Otherwise NO.
Base rate / current: At registration no agency has Oracle on watch or review. S&P's criteria describe a stable outlook as a rating not likely to change, a negative outlook as at least a one-in-three chance of a change over up to two years for investment grade, and CreditWatch as at least one-in-two within about 90 days. Oracle's revolver and term-loan margins reprice with the rating, so a watch is also visible in its filings. No scheduled event forces an agency to act: the agencies own this clock, which is why the claim exists.
What it tests: Whether the referee tightens before the market relaxes: the 1992 shape needs a clock the defender does not control.
Does the top of the loop keep raising? against the crowdP = 70%
By 31 Mar 2027, OpenAI announces the closing of a primary equity funding round of at least $30bn at a post-money valuation of at least $852bn.
Resolution criteria (binding)
Resolves YES if, on or before 2027-03-31, OpenAI announces (its own announcement; if it cannot be read, two independent major outlets naming size and valuation) the closing of a primary equity round of at least $30bn at a post-money valuation of at least $852bn. If the announcement, or both outlets, give one valuation without saying pre- or post-money, treat it as post-money. A round closed in tranches counts once the closed amount announced by the deadline reaches $30bn. Otherwise NO.
Base rate / current: At registration the last reported closed round was $122bn at an $852bn post-money valuation (SECONDARY), and a further round of at least $30bn at about $1.4tn was reported as in progress (REPORTED, unverified). Anthropic reported $65bn at $965bn. The ally's willingness to keep funding the buyers is the element that held the French franc in 1992 and would hold this label.
What it tests: Whether the anchor ally refuses: the loop breaks only when the upstream raise stops.
Does the last notch hold? against the crowdP = 8%
On at least one Bloomberg US Corporate Index rebalance date (the last business day of each month) from 30 Oct 2026 to 30 Sep 2027 inclusive, the middle of Oracle Corporation's senior unsecured ratings from Moody's, S&P and Fitch is Ba1/BB+ or lower.
Resolution criteria (binding)
Resolves YES if, at the close of any Bloomberg US Corporate Index rebalance date (last business day of each month) from 2026-10-30 to 2027-09-30 inclusive, the middle of Oracle's three senior unsecured ratings from Moody's, S&P Global Ratings and Fitch Ratings is Ba1/BB+ or lower; if only two agencies rate Oracle on that date, the lower of the two governs. Ratings in effect at the close of each rebalance date govern. Source order: each agency's own release or rating page, read at source and saved with its date; if a release cannot be opened, an Oracle SEC filing that states the rating; failing that, two independent major news outlets naming the agency, the action, the resulting senior unsecured rating and its date. Otherwise resolves NO on 2027-09-30.
Base rate / current: At registration: S&P BBB- stable (since 9 Jul 2026), Moody's Baa2 negative, Fitch BBB stable (agency releases not opened at source; SECONDARY). Middle rating two notches above the line, so exit needs a second agency to cut twice. Oracle's stated commitment: the 1 Feb 2026 financing plan to keep an investment-grade balance sheet. Trailing free cash flow -$28.7bn; leases signed but not commenced $288bn; first $20bn equity programme fully used at an average $141.13 a share; customer prepayments with a financing component $11,363m in FQ1 FY27 (WP44 and docs/research/ai_backlog_index). Protection reportedly about 249bp on 5 Oct 2026 against a BB index spread near 190bp, so the label change is largely priced; this claim is about the label, not the price.
What it tests: The AI build-out's weakest investment-grade seller cannot keep the grade, the build its backlog promises, and funding that does not come from the buyers whose credit is the risk, all at once.
Is the mechanism doing work? working bookP = 20%
oracle-ig-exit-2027 resolves YES, scored only if openai-round-852bn-2027q1 resolves NO; void if that claim resolves YES.
Resolution criteria (binding)
A conditional claim. Scored on the outcome of oracle-ig-exit-2027 only if openai-round-852bn-2027q1 resolves NO; void (not scored) if openai-round-852bn-2027q1 resolves YES. Same sources and deadline as oracle-ig-exit-2027.
Base rate / current: A coherence test on the forecaster, not a new fact: if the probability here is close to the unconditional probability of the exit, the reflexive story (the ally's refusal breaks the label) is doing no work in the forecast.
What it tests: Separates a credit view from the reflexive thesis: how much of the exit probability runs through the ally.
Delivered, and broke: the fiber pattern? working bookP = 5%
Both: WP44's registered test P3 resolves YES (Oracle's first-filed revenue for the four quarters to 31 Aug 2027 reaches $83.0bn) and oracle-ig-exit-2027 resolves YES.
Resolution criteria (binding)
Resolves YES only if both hold: (a) WP44 claim P3, as registered on 2026-10-03 and unedited (Oracle's first-filed revenue for the fiscal quarters ending 30 Nov 2026, 28 Feb 2027, 31 May 2027 and 31 Aug 2027 sums to at least $83.0bn, read from its 10-Qs), and (b) oracle-ig-exit-2027 resolves YES. If the 10-Q for the quarter ending 31 Aug 2027 is not filed by 2027-09-30, resolves on its filing. Coherence: this probability cannot exceed the probability of oracle-ig-exit-2027.
Base rate / current: The trilemma cell. In the 2000-02 telecom and fiber cycle, builders delivered the network and the credit failed on the schedule of the debt rather than the schedule of the technology (docs/mse435_enterprise_ai_lenses.md). WP44 P3 needs $20.75bn a quarter against $19.3bn in Oracle's latest quarter; meeting it takes the capex that drains the balance sheet. A forecaster who holds the 1992 thesis and registered P3 must put real weight here, or the thesis is an ordinary credit view.
What it tests: Delivery and financing can diverge: the promises are kept and the grade is lost in the same year.
Does the redefinition print a price? declared uncertaintyP = 35%
PJM's 2029/30 Base Residual Auction RTO clearing price is below $337.83/MW-day UCAP, the cap in force.
Resolution criteria (binding)
Resolves YES if PJM's official 2029/30 BRA results show an RTO-wide clearing price below $337.83/MW-day UCAP, nominal, as reported. Resolves NO if the RTO price is $337.83 or if PJM has not published official 2029/30 results by 2027-12-31. The latest official PJM figures as of 2027-03-31 govern. Coherence with pjm-fourth-cap (frozen at 72): the two claims split the outcome into four cells (at the cap: this NO, F7 YES; $325.00 up to but not including $337.83: both YES; below $325.00: this YES, F7 NO; no results: both NO), so P(this) + P(F7) - 1 is at most the probability of the middle cell, and two YES results are one event, not two hits.
Base rate / current: At registration the 2029/30 cap is $337.83/MW-day UCAP ($256.75 ICAP over a 0.76 reference factor), not $325; the last three auctions cleared at the cap ($329.17, $333.44, $325.00). PJM's 13 Aug 2026 filing (ER26-3515) would exclude uncovered new large load from auction demand beginning with 2029/30; FERC's statutory date is 12 Oct 2026. If the exclusion removes more megawatts than the 6,831 MW shortfall, the price can come off the cap without any price decision. The auction window is 9 to 15 Dec 2026, results about 22 Dec 2026.
What it tests: The defender holds the price by redefining demand: a cleared price below the cap after the exclusion is the quantity break showing up in the price.
Does the collar lapse? against the crowdP = 30%
PJM's 2030/31 planning parameters in force when the auction window opens (scheduled 12 May 2027) contain no FERC-accepted RTO price cap below the VRR curve's own point (a).
Resolution criteria (binding)
Resolves YES if the PJM 2030/31 BRA planning parameters in force when the auction window opens contain no FERC-accepted RTO-wide price cap below the Variable Resource Requirement curve's point (a); any accepted extension or replacement collar for 2030/31 resolves NO. Resolves NO if the 2030/31 BRA window has not opened by 2027-12-31. Sources: PJM 2030/31 planning parameters (due 1 Feb 2027, and revisions) and FERC eLibrary.
Base rate / current: The collar (cap about $325 ICAP-equivalent, floor $175) came from Pennsylvania's Dec 2024 complaint, a Jan 2025 settlement, FERC approval in Apr 2025 and an extension on 28 Apr 2026 through 2029/30, which FERC called time-limited. At registration no party has filed to extend it to 2030/31; the notice-of-intent deadline for the 2030/31 auction is 1 Dec 2026. The defender's will strengthens with scarcity, which favours a re-peg; the filed record shows nobody asking for one yet.
What it tests: The sunset of an administered price: whether the defender renews by decree or lets the ceiling go.
If the collar lapses, does the price break? working bookP = 18%
PJM's 2030/31 Base Residual Auction RTO clearing price exceeds $256.75 divided by the Reference Resource AUCAP factor in PJM's 2030/31 planning parameters.
Resolution criteria (binding)
Resolves YES if PJM's official 2030/31 BRA results show an RTO-wide clearing price, nominal UCAP, above $256.75 divided by the Reference Resource AUCAP (accredited UCAP) factor published in the 2030/31 planning parameters (that is, above the old cap re-expressed at the new factor). Resolves NO otherwise, or if no official 2030/31 results are published by 2027-12-31. Results scheduled about 25 May 2027.
Base rate / current: PJM's own simulation without the cap put the 2028/29 price near $555/MW-day for the region and $777 for ComEd (pre-exclusion). Against that: the demand exclusion, possible load-forecast cuts, the removal of the floor with the cap, and the Reliability Backstop Procurement ($555 maximum willingness to pay, accepted by FERC on 29 Sep 2026 subject to a suspension to 28 Feb 2027). The 2028/29 auction cleared 6,831 MW short at a 14.4% reserve margin against a 20% target.
What it tests: The price leg of the sunset: an uncapped auction clearing above the old ceiling is the only outcome with the 1992 payoff shape, and the defender has two ways to prevent it.