BCR Macro Dashboard

Fires on: Bank of Japan raises its policy rate

Across 6 BOJ hikes (2006-2025), nothing was significant vs random windows: yen strengthened in only 3/6 at 21 days (median USD/JPY -1.24%, CI crosses zero, p=0.27), Nikkei median -1.79% (p=0.28), S&P 500 -0.16% (p=0.33), VIX +1.0pt (p=0.35). The August 2024 'template' was violent but recovered within a month, and the June 2025 hike produced the OPPOSITE (yen weakened, equities rallied). The unwind mechanism that actually bit in Aug 2024 was crowded positioning plus surprise: check speculative yen positioning before invoking that analogy.

Small sample (6 hikes with a 17-year gap 2007-2024).

Fires on: ECB raises the Main Refinancing Rate while euro-area composite PMI < 50 and the inflation impulse is supply-shock-driven

The three stagflation-conditions ECB hikes (Jul 2008, Apr 2011, Jul 2011) were all reversed within 210 days: median 119 days to the first cut: versus a median 686 days for non-stagflation hikes, and the median post-hike equity drawdown was 29% (CI [23%, 47%]) versus 3%. The PMI<50 gate is the load-bearing condition separating reversed hikes from durable ones. Closest historical match to the June-2026 setup by similarity scoring: July 2011 (reversed in 119 days, 23% drawdown).

n=3 vs 3; permutation p=0.20 (small sample limits power). The 2008 episode is Lehman-confounded and drawdowns are peak-to-trough.

Fires on: ISM Manufacturing PMI prints below 50 for 2+ consecutive months (escalate below 48). ISM PMI is not in the pipeline's series set: manual check required.

At the 4 US recession onsets since 1986, the BAA-10Y spread widened in 4/4 cases at the 6-month horizon: median +47bp (CI [+34, +63]bp, p=0.026 vs random entry points): and starting 3 months BEFORE onset the median was +83bp. Widening is a 3-6 month phenomenon; at 12 months results were mixed as spreads compress into recovery. Every US recession since 1948 included consecutive sub-50 ISM readings.

The model's evidence base is NBER recession ONSETS: sub-50 PMI stretches without recession (2015-16, 2019, 2022-24) were never tested, so frame the precedent as 'at recession onsets' when PMI fires without one confirmed.

Fires on: A new coordinated SPR/IEA strategic release is announced (or US SPR stocks fall >5M bbl over 4 weeks)

Across 6 modern strategic releases (1991-2022), WTI's median forward return was -11% / -13% / -16% at 3/6/12 months (12-month bootstrap CI [-39%, -1%] excludes zero; random-window median +4%); only 1 of 6 episodes traded above the pre-release price a year later. The 1-month response is a coin flip (median +0.5%, p=0.49): 'a release breaks price' is mostly myth. The live counter-case: the 2021 coordinated release against a persistent supply squeeze FAILED, with WTI +48% six months later.

Releases are usually a response to disruption, so the lower path partly reflects shocks resolving. Verify which analogue (typical episode vs 2021) the current disruption resembles.

Fires on: A major oil supply-shock event is flagged (operator-set) and XLE-minus-S&P cumulative relative return over the 30 trading days since onset is below the historical +5.2pt median

Across 7 major supply shocks (1973-2022), the 30-trading-day median energy-minus-market spread was +5.2pt (persistent shocks ran higher: 1973 +12.9pt, 2022 +11.5pt at 30 days). Early energy UNDERperformance in a genuine supply shock is rare and historically associated with fast-resolution pricing. The edge fades with horizon: +5.2 / +2.0 / -0.4 / -0.1pt at 30/60/90/180 days: a 30-60 day phenomenon.

The event flag is discretionary: no data series identifies a 'supply shock', and the measured spread is sensitive to the chosen onset date. Known reconciliation issue: the briefing sector dashboard's 1M XLE-vs-SPY figure has previously failed to reconcile with raw data: compute the spread from raw XLE/^GSPC closes.

Fires on: A new China-Taiwan / maritime gray-zone escalation event (major drill series, ADIZ surge, coast guard operation, quarantine rehearsal)

Across 8 gray-zone events (2012-2024), market moves were indistinguishable from noise and event windows were actually CALMER than random ones: median TAIEX +0.21% (1d) / +1.30% (1w), TSMC +0.29% / +1.13%, VIX FELL -1.75pt at 1 week, and yen/gold never bid (all permutation p>0.66). The tested asymmetry: 7 of 8 episodes saw a further equal-or-higher escalation rung within ~12 months: an 88% continuation base rate (CI [62%, 100%]). Fade the market reaction; respect the ratchet.

n=8; escalation coding is judgment over the public record. Absence of market reaction to drills says nothing about a kinetic/quarantine tail scenario.

Fires on: A new China-Taiwan gray-zone escalation event (same class as above)

Asia-Pacific defense primes (6-name JP/KR basket) show no statistically detectable post-event excess return: pooled 3-month post-event median +5.2pp vs +1.4pp for random windows, permutation p=0.22 (n=48 name-events across 8 events). The basket's +411pp excess since the Dec-2022 Japan rearmament pivot is secular rearmament beta that predates and outlasts individual events: and YTD-2026 the basket is -36pp vs home indices (in relative drawdown).

Basket is concentrated (sensitive to Hanwha's parabola) and name-events cluster in time, so effective n < 48. The companion Western model found defense outperformed in only 1 of 4 historical episodes.

Fires on: A new Taiwan gray-zone escalation, evaluated against the 1-week change in EM High Grade OAS (BAMLEMIBHGCRPIOAS)

EM IG credit has never priced a gray-zone event: pooled 1-month event median -4bp vs -2bp normal (p=0.83), and the tested noise band puts the median 1-week move at 3bp with the 75th percentile at 5bp. Read any post-event spread quote against that band: below 5bp/week is noise consistent with precedent; a sustained move above it would be the binary repricing that has not yet happened in the sample.

Proxy is global EM High Grade (no keyless Taiwan/Asia IG series), n=2 in-window events, and the 3-year window contains no genuine credit-stress episode.

Fires on: Japan MOF's monthly release reports yen-buying (USD-selling) intervention, or USD/JPY drops >2% intraday on suspected intervention

All 4 yen-buying intervention quarters since 2022 showed Japanese TIC selling of US Treasuries: -$92.6B, -$43.9B, -$61.3B, -$22.7B (median -$52.6B vs +$8.1B in normal quarters; permutation p=0.002; bootstrap CI entirely in selling territory). Confirmation arrives with the TIC data's 2-month lag. Live episode: ¥11.73T (~$73.6B) deployed 28 Apr - 27 May 2026; Q2-2026 TIC is the forward test.

n=4 intervention quarters; dose-response is descriptive only (r=0.35).

Fires on: TIC Major Foreign Holders shows Japan's UST holdings falling more than $10B over a quarter (TIC reports with a 2-month lag; the pipeline tracks only aggregate foreign holdings, so the Japan line is a manual check)

OLS elasticity: $100B of Japanese net selling maps to +17bp of 10Y term premium (95% CI [2, 32]bp: the memo's 10-15bp falls inside; r=-0.255, p=0.017, n=88 quarters since 2003). Japanese-selling quarters saw the term premium rise 64% of the time (median +4.9bp) vs 46% for buying quarters. Use it as a CONFIRMATORY attribution aid only: the effect appears at lag 0 only, and Japan's flow explains ~6.5% of quarterly variation.

R²=0.065; only lag 0 is significant: Japanese flows do not PREDICT future term-premium moves. Japan holdings are down ~$302B from the Dec-2021 peak.

Fires on: A major de-dollarization announcement hits the tape (non-USD crude pricing deal, yuan-settlement agreement, 'petrodollar collapse' headline cycle)

Across 5 measurable de-dollarization events (2006-present), the broad dollar index strengthened in 60% of cases at 30 days and 80% at 180 days; median +0.51% at 30 days (CI includes zero, p=0.39 vs random windows) and +6.46% at 180 days. No detectable negative dollar impact; the largest single-event weakness (Iran Oil Bourse 2008) was GFC-confounded. Removing any single event does not flip the finding.

Event classification is judgmental; DTWEXBGS starts 2006 so earlier episodes are unmeasurable.

Fires on: A 'buy defense, rearmament cycle' thesis is circulating, or US defense spending growth accelerates sharply (FDEFX YoY > ~10%) on a named geopolitical trigger

Defense industrials outperformed in only 1 of 4 measurable US rearmament episodes (Vietnam -24.6pp excess, Reagan -49.2pp, post-9/11 -25.8pp, Ukraine 2022-present +27.3pp: the lone, still-open exception); monthly surge-period excess returns are indistinguishable from non-surge (p=0.50). On inflation: CPI rose during 3 of 5 episodes (Korean War +10pp was panic-buying-amplified), but the lead-lag is NEGATIVE (r=-0.33 at -4 quarters): surges historically began when inflation was already elevated and tightening followed.

Episode boundaries were judgmental (no tested numeric threshold); equity benchmark is the equal-weighted Fama-French industry average; only 4-5 episodes.

Fires on: A BoJ policy normalization step or a Japanese defense supplementary-budget announcement, with the JGB 10Y near cycle highs

JGB 10Y is +207bp since the Dec-2022 rearmament pivot while the defense budget rose 63%, but the relationship is statistically underpowered (r=+0.38, p=0.20, n=13; OLS slope CI [-7.7, +6.5]bp straddles zero) and the BoJ regime dominates: the entire +174bp breakout lands exactly when YCC ended. Supports DIRECTIONAL duration caution on JGBs only; the slope CI straddles zero, so no quantified rearmament-to-yield elasticity survives testing.

Annual data, descriptive not causal; the monetary regime sets the JGB term structure; the defense budget shows no independent statistical effect.

Fires on: US crude days-of-cover (EIA WCESTUS1/WCRRIUS2) detrended Z vs trailing 5 years falls below -1.0 (weekly EIA data: outside the pipeline's series set, manual/EIA check)

Across 2,176 weeks (1984-2026), tight inventory weeks (days-of-cover Z < -1) preceded LOWER forward crude: 3-month WTI median +0.1% vs +3.3% for non-tight weeks (p=0.000), with the decile staircase monotonic the 'wrong' way (most-ample decile +8.6% forward, tightest -0.7%). The inventory-cliff-equals-price-spike narrative fails its backtest. Volatility separates only at true extremes: at Z < -2.0, forward 13-week realized vol runs 37% vs 30% baseline.

Last reading at model build (2026-05-29): days-of-cover Z -0.92, stocks at the 71st percentile: well away from the tight threshold. Result stable across tightness cutoffs.

Fires on: China tightens or relaxes rare-earth/magnet export controls (a new licensing tier, an extraterritorial rule, or a truce-driven suspension). Compare a rare-earth producer basket (MP, Lynas, REMX, Energy Fuels) vs a magnet-dependent downstream basket (autos/defense) on cumulative abnormal returns in the [-1,+5] trading-day window.

Market-model event study around the Apr 4 2025 (Tier-1 controls) and Oct 9 2025 (0.1% extraterritorial rule) dates. In the immediate [-1,+5] window the predicted sign held at BOTH dates: the producer basket beat the downstream auto/defense basket by +10pp (April: producers +7.0%, downstream -3.1%) and +10pp (October: +6.6% vs -3.5%); the immediate producer rally was individually abnormal vs random windows for Energy Fuels (p=0.04 / 0.003) and MP (p=0.02 in Oct). Longer horizon diverged and is mechanism-consistent: after April producers kept re-rating (+57pp spread by +120d, structural buildout + DoD deal); after October producers spiked then reversed hard (-66pp by +120d) as the extraterritorial rule was suspended ~1yr under the US-China truce: producer equity tracks the policy lever, not a one-way bet.

N=2 events: descriptive, not a powered base rate. Equity proxies are one step removed from the physical NdPr/Dy/Tb prices the claim names (trade-press/paywalled: DATA-GATED); the physical-price and ex-China-vs-China spread-persistence test could not be run. Longer-horizon CARs are confounded (truce suspension, DoD/Pentagon deals, the late-2025 producer-equity bubble).

Fires on: A single-day gold decline of 5% or more (PM-fix basis)

Across 30 distinct 5%+ single-day gold crashes since 1971 (LBMA PM fix), the median 12-month forward return was -1.6% (bootstrap CI [-4.7, +19.5]) with a 47% up-rate (CI 30-63%): indistinguishable from random windows (p=0.71) and below gold's unconditional drift. The oft-cited positive averages are 1970s skew (Nov-1973 +106%, Oct-1979 +68%). Post-2000 sub-sample: median -2.9%, up-rate 44%. Robust across -4% to -7% thresholds.

PM-fix basis understates intraday extremes; 30 episodes give wide CIs; the equity (Dow) version of the claim was not tested here.

Fires on: US tariff regime change that creates or closes an origin-specific wedge on electronics (e.g. a successor Section 301 action naming China or India)

On 99 months of US phone-import data (HS 851712/13, Comtrade/US Census, 2018-01 to 2026-03): India's flip from 4.9% to 65.2% of US phone imports (China 91.3% to 16.1%) sits entirely inside the 2024-26 wedge regime; all ten of the largest India trend breaks in eight years land between 2024-10 and 2026-01 (zero in the prior 81 months); the Dec-2019 placebo is clean for India in all six specs (p>=0.38). Point-event alignment FAILS at exact duty dates because flows move 2-3 months ahead (anticipation), consistent with the chokepoint-regimes-not-events synthesis. Post-wedge stickiness: India held 56.3% in Mar-2026, above the 40% reassessment threshold, with Apr-Jun 2026 data pending.

E3 (Section 122 wedge closure) has only one post-month of data - the stickiness verdict is unresolved until Comtrade publishes Apr-Jun 2026; level shift at E2 is significant (+23.8pp, p=0.03) but slope tests at exact anchors are not.

Fires on: A discrete USMCA or Mexico-tariff action: review round, withdrawal/tariff threat, deal, exemption, or court ruling

Across 14 USMCA/tariff events (2017-2026), peso event windows were CALMER than random windows (mean |move| 0.70-0.96x random, all p>=0.44 pooled) and EWW abnormal returns cleared nothing at 1-5 days. Even the May 2019 tariff threat, the biggest event move (+3.6% USD/MXN in a week), was only borderline vs peso noise (p=0.07) and round-tripped to +0.4% within a month; the 2025 IEEPA orders moved the peso under 0.6%. Meanwhile 100% of Mexico's decade-long equity de-rating (-58.5% EWW vs SPY beta-adjusted, cum log) accrued OUTSIDE event windows (+2.6% inside, -61.2% outside). Third instance of the chokepoint-regimes-not-events pattern (Taiwan gray-zone, rare-earth controls, now USMCA). Do not pitch the event trade; re-check the regime arithmetic instead.

Small per-category samples (4-5). Daily bars hide intraday round-trips (the Feb 3 2025 threat-then-pause session nets to -0.6%). EWW month-long windows show 1.5x excess dispersion (p=0.04) but with inconsistent signs, partly alpha-estimation noise. Drift decomposition is descriptive, not causal.

Fires on: Uranium SPOT price crosses ABOVE the long-term contract price (inversion)

Term-above-spot is uranium's DEFAULT state (88% of months since 1996, including the entire 2011-2020 bear) and carries ZERO forward edge (12m forward +4.4% vs +4.5%, p=1.00) - the memo's 'term premium precedes appreciation' framing is rejected. The tradeable signal is the INVERSION: all 5 spot-above-term episodes in 30 years occurred inside the two manias (2004-07, 2021-24), bracketing the two secular tops ($136 Jun 2007, $100.25 Jan 2024), with a lag (median +50% over the next 12m before 3/5 episodes hit >=35% drawdowns within 36m). As of Jun 2026: NO inversion (term $95.50 - a record print - vs spot $85.00); the warning light is off.

Term series reconstructed from Cameco's public UxC/TradeTech month-end table + 7 Wayback vintages (1,693 overlapping obs, zero restatements); only ~2 independent contracting cycles exist.

Fires on: Major hyperscaler nuclear PPA or US nuclear policy announcement

Nuclear announcement pops are real and abnormal (+1d median CAR: fleet owners +1.5% p=0.02, announcement-stage names +4.3% p=0.003) but the premium regime CHANGED: 2024 events kept re-rating (+53.6% median CAR at 60d for announcement-stage names) while every event since May 2025 pops then fades (-12.5% at 60d; era difference +66pp, permutation p=0.008). The Jan-2026 Meta 6.6 GW trio decayed hardest (OKLO -68.6% at 60d) and the Jun-2026 DOE $17.5B loans produced NO pop at all. The fleet-vs-speculative split does NOT hold (p=0.84). The market has stopped paying for nuclear announcements - treat any future PPA headline as a fade candidate, not a chase.

Six events, two clustered window pairs; DOE event lacks +20/+60d data until ~Sep 2026; era split is an ex-post 2-vs-4 partition; null drawn inside the secular nuclear re-rating.

Fires on: LME copper registered stocks fall below -1.0 detrended Z vs their trailing 5 years (weekly LME data: outside the pipeline's series set, manual/LME check), or a 'copper stocks at multi-year lows' narrative wave

Replicating the oil-inventory-cliff design on copper (967 weeks, 2008-2026, LME registered stocks + cash settlement): tight weeks (stock Z < -1 vs trailing 5y) preceded LOWER forward copper at every horizon: 13-week median -1.4% vs +1.9% for non-tight weeks (diff -3.3pp, p=0.000, matching oil's -3.3pp), 26-week diff -5.5pp (p=0.000). No decile staircase exists (Spearman +0.07); the direction holds on COMEX prices (p≤0.003) and on stock-draw signals. The one honest exception: true squeezes (Z < -2.0, n=12 weeks, 2021-type) ran +13.9pp (p=0.001): an unforecastable positioning event, not a tradeable signal (Z < -1.5 gives -0.0pp, p=0.97). At model build the current reading REFUTES the crowd narrative outright: LME stocks 306,500 t = Z +1.61, AMPLE vs trend, while peer videos lead with 'inventories collapsing'.

LME registered stocks only (SHFE/COMEX/bonded excluded: metal migrates between reporting regimes, especially during the 2025-26 tariff relocation); 18.5y window vs oil's 42y.

Fires on: A Section 232 copper action (investigation step, announcement, proclamation, modification, or the pending refined-copper phased-duty decision)

Across the five Section 232 copper events (EO 14220 2025-02-25; the 50% announcement 2025-07-08; the proclamation 2025-07-30 that deferred refined copper; effective date 2025-08-01; the 2026-04-06 modification): the COMEX-LME incidence spread moved violently and abnormally (E1 +1,147 USD/t p=0.008; proclamation -2,464 USD/t p=0.002: a -2,679 USD/t ONE-DAY collapse when refined cathode was excluded, the -22.3% COMEX day), then round-tripped. The GLOBAL complex did not: LME cash 21d moves ran +4.1/-3.0/+0.7/+1.7/+9.3% (every p≥0.16), median |move| 3.0% vs 3.8% in random windows: events CALMER than random, the Taiwan-gray-zone signature on a tariff calendar. Miners (COPX/FCX market-model CARs) individually insignificant at E1-E4; the lone p<0.05 reading (E5) is confounded (metal +9.3% while miners fell: a broad April-2026 equity-stress artifact, wrong-signed for a tariff shock). Forward read: with the 15%/30% refined-copper phased duty still pending, the arb sits near ~300 USD/t (~2%): the market prices very little of it.

n=5 events, one policy arc, 14 months; per-event nulls + medians only, no pooled significance. Out-of-sample test: if the refined-copper duty IS imposed, the arb should converge toward duty magnitude.

Fires on: OPEC+ announces a production pause or cut (policy flip back to price defense in the live 2025-26 volume episode)

Both completed Saudi volume episodes flipped back to price defense within ~2 years: 1985-86 in 15 months (netback Sep-1985 -> $18 fixed-price accord Dec-1986), 2014-16 in 24 months (Nov-2014 no-cut -> Nov-2016 Vienna cut; Algiers first-signal at month 22). Buffers absorbed the difference: reserves -33% and -27% ($740bn -> $537bn in 2014-16), cumulative deficits SAR -358bn (1983-90) and SAR -1,039bn (2014-17). Price troughed -59%/-60% from episode start in both. The live episode (month 14) diverges on channel, not strategy: reserves are UP ~11% because the war premium holds prices and H1-2026 was financed entirely by borrowing: the buffer burning this time is debt capacity (+SAR 148bn in Q1 alone). 24m real price change from episode starts: median -38% vs -2% random-window null (p=0.052, n=3).

Base rate of two completed episodes: precedent, not probability. The live episode carries a war shock the precedents lacked (memo Counter-Risk 1 at 40%). 1980s reserve drawdown is a floor (IMF definition vs SAMA total foreign assets).

Fires on: Saudi MoF publishes the Q3 2026 quarterly budget report (~Nov 2026): first print containing co-belligerency costs; re-run the model

On H1 2026 data the strong form fails: Q1's net war effect was -17.8 SAR bn (military +26% y/y, no windfall yet) but Q2 flipped to +33.3 SAR bn as war-priced oil revenue (+22% y/y) arrived against a flat military line: the windfall paid the defense bill roughly twice over. Decisive caveat: Saudi co-belligerency began July 29, 2026, AFTER both tested quarters. The historical lens supports the structural concern: in the 1984-88 tanker war, oil revenue fell -85% peak-to-trough while military expenditure held at $13-27bn: the defense bill is a ratchet that does not flex down when the premium fades. War-year milex share 30.7% vs 26.9% non-war (+3.8pp, p=0.129, not significant).

Two usable quarters, both pre-co-belligerency; four of six military figures are computed residuals from cumulative MoF disclosures; subsidy and risk-premium channels excluded (bias: understates war cost). Exploratory by design: not a Video Core dependency.

Fires on: A capped benchmark reaches a scheduled rebalance date (MSCI 25/50: close of last business day of Feb/May/Aug/Nov; Select Sector: after the close of the third Friday of Mar/Jun/Sep/Dec), or an index provider announces a concentration-triggered special rebalance

Across 54 MSCI Korea 25/50 rebalances (2013-2026), Samsung traded 1.40x its 60-day median volume on the rebalance close (CI 1.29-1.54x) vs 1.19x on matched non-rebalance month-ends (p=0.032) and 1.00x on ordinary days (p<0.0001); 81% of rebalances were above baseline and the spike lasts exactly one session. The cap-specific half fails: ten non-capped MSCI Korea large caps traded 1.39x on the same dates, Samsung's excess over them was +3.0% (p=0.323) and SK hynix's was -17.1% (p=0.011, wrong direction), the dose-response on cap excess is negative (-0.049 per pp, CI -0.152 to -0.026), and the record 11.2pp breach of 2026-05-29 produced Samsung at 1.30x against a 2.24x cohort. No price footprint: day-0 CAR vs KOSPI -0.04% (Samsung) and -0.20% (SK hynix), indistinguishable from month-ends (p=0.86, p=0.97), with no +1..+5 reversal. US analogue CONFIRMED to exist: the Nasdaq-100 Special Rebalance of July 2023 is a live US concentration tripwire (8 trimmed names at 2.28x, the highest cohort reading of any 2023 session), and XLK's RIC caps forced roughly $11bn of Apple to be sold at the June 2024 rebalance, when Apple traded 4.60x normal volume, the 96th percentile of all 54 Select Sector rebalance dates. Frame the mandate as real and datable; do not claim the cap trim is separable from ordinary index-review flow.

Volume is unsigned, so the 'selling' direction is only probed indirectly via returns. The 25/50 cap rebalance and the parent MSCI Korea review fall on the same date by construction and cannot be separated in time. AUM tracking 25/50 specifically is a small slice of the Korea passive complex, so a genuine cap increment of a few tens of bp of daily volume would be invisible; NOT DETECTED is not proof of absence. Uncapped weights are reconstructed by price drift from the 2026-05-29 MSCI factsheet anchor (validated to 0.57pp at 2026-03-31, but far looser pre-2020) and model only the Samsung common line, while the 25/50 cap reads on issuer including the preferred. The US legs rest on n=1 clean events; spglobal.com/spdji returns HTTP 403 so the exact Select Sector single-company cap could not be pinned and a US dose-response is a documented gap.

Fires on: A major oil supply disruption (>=1 mb/d) reaches a datable resolution: ceasefire, reopening, or restored output

Across the 5 major oil disruptions with a datable resolution (1990-2019), the risk premium was ALREADY GONE by the resolution date: median peak premium +16.6% vs pre-disruption baseline, median premium remaining AT resolution -1.7%, 3 of 5 episodes BELOW baseline on the day, peak a median 62 days BEFORE resolution. Forward real WTI from resolution is indistinguishable from random at every horizon (all p>=0.43, n=5). Resolution is an anticipated event and anticipated events are not tradeable. REFILL LEG (revised 2026-08-02 after operator challenge; v1 understated it 3-8x by using US SPR data alone): the OECD drawdown is globally coordinated -- OECD commercial stocks fell 255M bbl from the 2025-09 peak to 2026-06, of which 66.3% was NON-US. The 2026 US programme is an EXCHANGE (not a sale) repaid at an 18-22% premium: DOE has AWARDED >133M bbl of the 172M target (77.3% subscribed, ~39M unplaced) to BP, Gunvor, Marathon, Shell Trading, Energy Transfer, Mercuria, Trafigura and Vitol, with one 53.3M tranche clearing at ~28%. On awarded volume that is 157-162M bbl returned over Nov 2026-Sep 2028 = 225-232 kb/d = 0.22-0.23% of world demand, 2.44x the historical discretionary median of 92 kb/d (290-300 kb/d if the full target is placed) -- and it is PRICE-INELASTIC, a contractual obligation on private counterparties buying at market, so the 'governments buy when cheap' endogeneity does not apply. EIA's own STEO projects a +418M bbl OECD rebuild by Dec 2027 = 763 kb/d = 0.74% of world demand (FORECAST, not data).

PREMIUM LEG: n=5 and every CI is wide -- 'p>=0.43 everywhere' means this sample cannot distinguish these outcomes from noise, NOT that they are equal. All five episodes are PRODUCTION disruptions; no transport-disruption resolution exists in the price record, so the set contains no true 2026-Hormuz analogue. REFILL LEG IS NOT TESTED, only sized: there is no historical analogue of an obligated exchange repayment at this scale into a globally-drawn reserve system, so the historical base rate cannot speak to it, and the p=0.30 refill-period price test was measured on DISCRETIONARY US refills -- the wrong population. Every remaining data gap biases the refill bid DOWN not up: China's SPR has no public series and is excluded, IEA member government stocks need a paid subscription and are excluded, and the STEO OECD series is COMMERCIAL not strategic. Read 0.28-0.74% as a LOWER BOUND. The OECD window is 2022+ only, so 'unprecedented coordination' cannot be tested against a long history. The EIA rebuild figure is a projection in which stock change is partly a balance residual. AWARD FIGURES ARE AS OF JUNE 2026 from DOE releases, not a maintained series -- re-verify before subscriber-facing use. The ~39M bbl gap is RESOLVED and is NOT pacing: the June 10 2026 RFP offered 40M bbl and awarded 0.5M (~98.75% failure). Cause was the FORWARD CURVE, not a demand view -- the exchange only pencils in backwardation (borrow expensive now, buy cheaper back later, carry must exceed the 18-28% barrel premium); at launch Aug ~$90 vs mid-2027 ~$78 (~$5/bbl net), by the Jun 15 deadline the FRONT had collapsed to ~$80 while the back held ~$78. Uptake is therefore a function of CURVE SHAPE, not price direction, and the back end barely moved -- this is NOT evidence that counterparties expect cheap oil later. Consistent with this model's anticipation finding: the front is what deflated.

10Y yields rose in 3 of 5 US rearmament episodes (Vietnam +184bp; Ukraine 2022-present +219bp with +55bp of 2s10s steepening); the 2 exceptions had countervailing monetary policy (Volcker, Greenspan conundrum). Defense spending growth LEADS yield changes by ~8 quarters (peak r=0.20, p=0.0007), so the pressure window extends ~2 years beyond a spending-growth peak. The model classifies 2022-present as a live episode.

Only 5 episodes; tested on US yields though the live thesis is partly about EU/German issuance.

The US debt/GDP trajectory aligns with Japan's at a ~23-year lag (level-crossing median 23y, CI [22, 23]; beats 97% of candidate lags, permutation p=0.028), versus the memo's claimed 15. The lag BREAKS on the interest-burden channel: Japan's net interest stayed ~1.4-1.5% of GDP even as debt/GDP rose 207%->251% (ZIRP/QQE mask), while US net interest hit 3.3% of GDP in FY26 (a post-WWII record) rising +0.36pp/yr with no mask.

Descriptive lag between two trending series: a resemblance with no causal force; debt-definition sensitive.

Oil fell from 14.3% of global trade (1980) to 3.8% (2024) while the USD's FX-turnover share ROSE to 89.2%: correlation indistinguishable from zero in every sub-period. On reserves: the USD's -8.6pp COFER share decline since 2016 was absorbed ~6x more by secondary fiat (AUD/CAD/other, +5.4pp) than by RMB (+0.9pp, and DECLINING since 2022 to 1.95%); central-bank gold buying step-changed 2.1x to ~1,007t/yr from 2022. At RMB's realized pace, a 10% reserve share is 60+ years away.

COFER is quarterly with a lag; refresh the decomposition on each release rather than treating these numbers as static.

Over the Korean War (Jun 1950 - Jul 1953): S&P 500 total return +39.8%, defense sector +15.1pp excess, inflation -0.4% -> 9.4% peak -> 0.4% (partly panic-hoarding), 10Y +49bp. The gold/dollar comparison is void: both were fixed under Bretton Woods, so the modern reserve-diversification channel literally could not operate. Single-event reference material for mobilization narratives; it carries no statistical weight.

n=1 historical episode; descriptive only.

After the 2010 Senkaku embargo Japan cut its China rare-earth dependence from ~90% (2010) to ~58% (2017): -32pp over 7 years, ~-4.6pp/yr with full state backing: then PLATEAUED near 58% into the 2020s ('never to zero'). Applying that best-case pace to China's ~91% magnet-REE refining share, the 60% level (G7 Évian 2030 target) is reached only ~2032 (bootstrap 95% CI 2031-2034); P(hitting 60% by 2030 at Japan's speed) ~= 0.1%. The 50% 'asap' goal sits below Japan's own ~58% floor: no single country has diversified that far. The base rate FLATTERS the Western task because Japan only diverted its own demand, whereas the G7 must move China's GLOBAL supply share.

Cross-metric overlay (one country's import dependence vs China's global supply share vs US net import reliance): a difficulty benchmark, not a point forecast. Two sourced anchors (90%, 58%) only; no robust intermediate path. n=1 historical analogue.

Working in share space (China 91 of 100 refined magnet-REE units, ex-China 9), the displacement math is the wall: with China flat, ex-China refining must grow 6.7x to pull China to 60% and 10.1x to reach 50%. A Monte Carlo (20,000 draws) of the announced project pipeline (CRMA, G7 195-project/EUR64B list, MP, Lynas, Energy Fuels) discounted by ~8-year lead-time slippage delivers only ~2.1x ex-China growth by 2030, leaving China's refining share ~84% in 2030 (95% CI 71-91%) and ~80% in 2035 (95% CI 67-89%): bracketing the IEA's own 86%->82% top-3 decline (calibrated, not alarmist). P(China < 60% by 2030) = 0.0%. The G7 target requires diversifying ~17x faster than the IEA models (-6.2pp/yr required vs -0.36pp/yr modeled).

Share-space abstraction, not a project-level construction schedule; pipeline-potential and delivery distributions are judgment-calibrated to the memo's qualitative scale and IEA lead times, not a bottom-up capacity database (the upgrade path). Demand growth omitted (would make diversification harder).

Tested on US core-goods CPI (CUSR0000SACL1E) vs import prices (IR) and an effective-tariff proxy. The effective tariff rose from a 1.41% mean (1995-2017) to 8.0% latest (Jan 2026); the import channel's contribution to core-goods inflation flipped from +0.03pp (1995-2017) to +1.05pp (2018-2026). Import-price pass-through to core-goods CPI steepened from slope +0.024 (r=0.16, n=276) pre-2018 to +0.558 (r=0.79, n=101) post-2018; the +0.534 jump has a 95% block-bootstrap CI [-0.05,+0.74] (p=0.068, just shy of the bar). China import-price slope jump is even larger (+0.095 to +1.247).

Partially supported: the pass-through steepening collapses to -0.072 if 2021-2022 is excluded, so the statistical slope result leans on the COVID spike; the tariff-regime flip and contribution-sign flip are the robust parts. Effective-tariff proxy (customs duties / IMPGS) understates the goods-only rate because IMPGS includes services, so the level reads ~8% vs the memo's ~11% goods figure; the trend is the point.

Tested the Driver 2 mechanism directly. The ECI wage-on-(vacancy/unemployment) Phillips slope FLATTENED from +3.16 (2002-2014) to +2.21 (2015-2025); the -1.03 difference has P(diff<=0)=0.95 (flatter, opposite the claim) and is robust to dropping 2021-2022. The demographic term (working-age-population YoY growth) entered with a +0.43 coefficient (boot CI [+0.08,+0.73]): the WRONG sign vs the predicted negative. Long-history AHETPI test (1965-2026) was uninformative (r~0.09).

NOT SUPPORTED only in the narrow falsifiable form. Wages DID rise and tightness reached historic extremes; the defensible framing is 'tightness went to extremes,' not 'the wage-tightness relationship steepened.' Short modern sample, multicollinearity, and 2021-2022 leverage all limit power. The memo's own confidence table already flagged Driver 2 as its weakest claim.

Jorda local projections of CPI YoY response to a +1pp funds-rate shock, 1965-1990 vs 2015-2026. The 1970s impulse response is POSITIVE at every horizon (trough +0.06pp, +0.53pp by 24m): the textbook 'price puzzle' from the Fed hiking reactively into demand inflation. The modern response is cleanly negative (trough -2.74pp at 24m). The difference (then minus now) runs OPPOSITE the memo's claim and its 95% block-bootstrap CI [-1.09,+7.08] includes zero. Robust across horizon caps and shock definitions; the modern negative response survives dropping 2021-2022 (-2.83).

NOT SUPPORTED as literally framed, but informative: the 1970s positive IRF is endogeneity, itself evidence of the demand-driven, Fed-chasing dynamic the memo describes; and the modern -2.74pp dent is small against a ~3.4% core floor: hikes bend the floor, they don't erase it. Modern window has ~4x less rate-hike variation (std 0.18 vs 0.75pp), so its CI is wide.

Federal net interest is already 18.5% of receipts in FY2025 ($970B), matching the early-1990s record (18.4%, 1991) with zero headroom: 98th percentile of its own history. A debt-rollover scenario on the ~$30T public stock: the average coupon that pushes interest past the 18% peak is ~3.1% today vs ~9.8% at 1981's 32% debt/GDP; past a destabilizing 30% of revenue it is ~5.2% today vs ~16.2% in 1981: a ~3x loss of hiking capacity. The same 12% Volcker-implied coupon costs ~69% of revenue today vs ~22% at 1981 debt/GDP (bootstrap mean ~57%, every draw clears the 18% peak). Today's ~123% debt/GDP sits at the 98th percentile of 1966-2026 (permutation p=0.016).

Scenario arithmetic (steady-state rollover), not a forecast; assumes average maturity 5-7y and 0.8-1.0 pass-through. The empirical anchors (FY2025 interest/revenue at 18.5%, the debt/GDP outlier) are hard facts. Uses net interest / receipts (Treasury FYOINT/FYFR), not BEA gross interest, to match the memo's ~18% early-1990s peak and ~$970B FY2025 figure.

TIPS-era horse race (2003-2025, monthly): DFII10 beats DGS10-minus-CPI on every leg: level correlation with forward 12m real gold returns +0.654 vs +0.366 (r-difference CI excludes zero), non-overlapping annual r +0.638 (p=0.001) vs +0.319 (p=0.14), out-of-sample R2 +0.45 vs +0.10, 12m-change correlation -0.446 vs -0.21. The LEVEL relationship flipped positive in this era (post-2022 gold rose through 2%+ real yields on the official bid), so the sign-stable channel is CHANGES in real rates, not levels. The TIPS-vs-CPI gap itself (currently ~+1.8pp, 88th percentile) carries no forward information (top vs bottom gap quintile median fwd12 +6.1% vs +6.2%, p=1.00).

TIPS era only (2003+); the 2022-2025 official-bid regime dominates the positive level correlation. Qualifies (does not delete) the gold-real-rate-regime trigger: that finding's slope comes from the 54-year annual model and remains valid long-run; its <1% CPI-deflated LEVEL condition is the era-fragile part.

Across 420 months with gold in a 15%+ drawdown (1971-2026): median forward 24m gold return +8.6% (CI +4.4,+15.6) when the real 10Y FELL over the following 24 months vs -1.3% (CI -3.9,+3.7) when it ROSE: a +9.9pp spread, permutation p=0.0038, stable at -20%/-25% drawdown thresholds. The three completed >25% episodes land exactly on the fork: 1975 (rates flat) -> +15.9% and recovery in 43 months; 1981 (rates +5.4pp) -> -16.5% and 319 months under water; 2013 (+1.6pp) -> -1.8% and 107 months under water.

Conditions on the REALIZED future real-rate path: a scenario map, not a timing signal. Only 3 completed >25% episodes; the statistics come from the overlapping month-level layer.

Price-side test around the verified 2022 regime break: median rolling-63-day drawdown post-2022 is -3.12% vs -3.99% (2010-2022) and -4.25% (full pre-2022 era), permutation p<0.001: typical dips are shallower inside the regime, and 10% dips recovered in ~199 days vs the 2011-2020 bear's 3,227. But downside semivol is HIGHER post-2022 (13.5%/yr vs 11.9%) and the largest one-day crash since 1983 (Jan 30, 2026) happened inside the regime: the bid cushions ordinary drawdowns and does not prevent tail events.

DATA-GATED: the direct quarterly purchases-vs-drawdown panel needs the registration-gated WGC statistics download; price-side evidence cannot isolate the official bid from other demand. Post-2022 window is ~4.4 years.

Across the China (2009-18) and India (2020-26) episodes, assembly share always migrates faster than value-added share (H0 slopes-equal rejected p<0.0001 in all 7 specifications). India: assembly +4.98pp/yr [4.44, 5.56] vs gross DVA +0.50pp/yr [-0.20, +1.20] - a 10.0x ratio [4.1, inf), and the DVA CI includes ZERO (P(no value progress) = 10%). China's teardown-value climb was +2.42pp/yr (3.6% in 2009 to 25.4% by 2018, Xing series) - a ~2.1x ratio against saturated assembly. India's value slope runs at ~1/5 of China's historical pace. On GTRI's stricter net-accrual measure India's value share is FALLING (-2.30pp/yr).

Sparse anchor points (2-4 per series) with parametric within-band bootstrap; China ratio is a cross-era construction (assembly saturated ~100% in 2009-18); OECD TiVA upgrade blocked by SDMX API rejection - retry path in MODEL.md.

The iPhone assembly migration to India runs at -4.60pp/yr [CI -5.34, -3.88] - statistically indistinguishable from Japan's post-2010 state-backed rare-earth best case (-4.6pp/yr; difference CI [-1.22, +1.27]). Even at that record pace, P(China <50% of iPhone assembly by 2030) = 0.1-0.5%; the median 50% crossing is ~2032, and a Japan-analog floor (~48%) leaves 2035 at 72%. China's GLOBAL smartphone assembly share is falling ~5x SLOWER than the Japan pace (-0.95pp/yr [CI -2.50, +0.62]; P(<50% by 2030) = 0.8%). 20,000-draw Monte Carlo, 2026 anchors (iPhone 75%, global ~63.5% Counterpoint-derived).

Global-share anchors diverge by source (Counterpoint ~63.5% vs MIIT-style unit counts ~82%) - the decisive robustness axis; the observed iPhone pace was driven by a tariff wedge that closed 2026-02-24, so the forward pace may be slower.

Top-quintile India/EM relative-valuation months (5y trailing percentile, USD) saw median forward-12m India-minus-EM returns of +3.07% vs +11.94% otherwise (difference -8.87pp [CI -15.15, -1.73], permutation p=0.0036, monotonic across quintiles). The window then FIRED: India lagged EM by 40.5% in the 12 months to Jun-2026 and the ratio now reads the 3.3rd percentile of its 5y range (z=-2.31). Per the never-say registry this is a vulnerability-window finding, never a dated timing signal - and it now cuts AGAINST fresh premium-compression claims, since the compression is behind, not ahead, on this proxy.

Price-ratio-vs-trailing-window proxy measures re-rating vs its own range, NOT the forward-P/E premium LEVEL - India can read cheap here while still carrying a P/E premium; significance fades in some robustness specs (3y window, 6m horizon, INDA pairing); effective sample is ~4 episodes.

On IMF monthly bilateral flows (2010-2026), the correlation of Vietnam's import growth from China with its export growth to the US flipped from -0.09 pre-2018 to +0.34 after the Sec. 301 waves (difference +0.43, rotation-permutation p=0.011, robust to excluding COVID 2020-21: +0.40). Mexico's two pipes co-move in both eras (pre +0.58, post +0.41) but did NOT tighten after the tariffs (p=1.00) - the aggregate coupling predates the trade war. The conduit mechanism shows up for Mexico instead as sequencing: post-2018, China-to-Mexico growth leads Mexico-to-US growth by ~3 months (r=+0.58 vs +0.41 contemporaneous). Co-movement at the Apr-2025 escalation was near post-era highs (Vietnam +0.49, Mexico +0.38) but only ~10 months of post-wave data exist.

Aggregate bilateral flows, not the memo's product-level claim - sector-concentrated rerouting can be masked; Vietnam result is significant in the opposite-reporter pairing but flat in the Vietnam-reported mirror; values not volumes; 1 month of data after the Feb-2026 collapse - re-run when IMTS extends past mid-2026.

["Across 8 China-specific tariff escalations (Section 301 lists 2018-19; IEEPA 2025), China's share of US goods imports fell a mean -2.51pp within 6 months on a seasonally-adjusted basis (p<0.001 vs random windows), while Mexico+Vietnam's combined share rose +2.25pp (p<0.001) - the price-driven rerouting signature. The REBOUND after the wedge collapsed (2026-02-24 flat 10% Section 122) got its first US-side confirming print on 2026-07-09: May 2026 share 7.54%, +0.95pp off the April trough (6.59%), with the YoY share change turning POSITIVE (+0.13pp) for the first time in the sample - exactly the tripwire the April-vintage run named as earliest confirmation. Post-Section-122 mean (7.13%) is still below the pre-reset 8.32%: a turn, not yet a recovery. US-measured imports ran 67% of same-month 2024 in May vs ~90% on the China-customs basis - the two bases still diverge and are never interchangeable."]

One confirming month (n=4 post-reset); the June 2026 FT-900 (~Aug 5) confirms or reverts the rebound leg. 2018-19 event windows overlap; 2025 front-running distorts both directions; China-customs and US Census import data diverge materially in 2025-26.

OECD TiVA 2025 (input-output data through 2022): Chinese value-added rose from 5.1% to 6.7% of Mexican gross exports over 2016-2022 (+0.32 pp/yr, 95% CI [+0.22, +0.42], p=0.005) while China's direct share of US goods imports fell from 21.2% to 16.6% (-0.86 pp/yr, p=0.007) and kept falling to 9.0% by 2025. Divergence +1.18 pp/yr, P(<=0)<0.0001, robust across all windows, denominators, and sectors (electronics +0.39, autos +0.55 pp/yr). The widely-cited 'one-third' figure is NOT Chinese VA as a share of Mexican exports (that is ~7%) - it is China's share of the FOREIGN content of Mexico's electronics exports (~32% in 2022, 33% in 2020).

Data-vintage limit: latest measurable input-output year is 2022 (TiVA 2025 edition); the 2023-2026 tariff escalation post-dates the content data. The Brookings 'one-third' reconciliation is an inference about their unstated methodology. TiVA vs INEGI disagree on Mexican electronics DVA (46% vs ~20%). Re-run on the next ICIO/TiVA vintage (~2027). OECD TiVA fetch route: sdmx.oecd.org/sti-public/rest/ (documented in fetch_data.py - also unblocks the India model's flagged TiVA upgrade).

BLS locality-of-origin import price indexes (2003-2026) show NO Mexico-specific conduit premium: since Dec-2017 Mexico's index is +20.0% vs the EU's +20.9% (Japan +8.9%, ASEAN +5.8%), and diff-in-diff vs EU+Japan controls is insignificant at every tariff wave (2018-19: -1.6%, placebo p=0.87; 2025: +0.75%, p=0.97; 2021+ era: +2.4%, p=0.77). The raw '+68% above pre-2018 trend' stat is an artifact - non-conduit Canada reads +167% on the same metric because the 2014-16 oil crash sits in the fit window. What IS in the data: China's index fell to -1.3% below its Dec-2017 level and -19.2 log points vs controls over the reallocation era (placebo p=0.029), a +12.7pp Mexico-minus-China divergence driven by China's decline - consistent with Alfaro-Chor's ~71% tariff pass-through (Chinese exporters did not cut pre-tariff prices in 2018-19; the later slide tracks China's 2023-25 PPI deflation).

BLS locality indexes EXCLUDE import duties - this tests pre-tariff goods prices only. No Vietnam-specific index exists (ASEAN aggregate used). Fixed-basket indexes cannot see a premium expressed as new higher-cost supply chains entering the basket (composition channel) - Alfaro-Chor's product-level unit values could still be right at that level. Mexico's index carries petroleum weight and USD denomination effects.

Tested against IMF BOP quarterly FDI components (2000Q1-2026Q1) and the Caldara-Iacoviello TPU index: Mexico's new-investment (equity ex reinvestment) FDI ran $12.0bn in 2017 and $11.3bn in 2018, at or above the 2014-16 average of $10.1bn, through the entire NAFTA renegotiation. The apparent freeze vs a 2012-16 trend (-12.7%) flips to +14.0% on a 2013-16 baseline (the 2013 Modelo deal distorts any trend), and the placebo permutation test returns p=0.54 (total FDI p=0.19). Local projections of FDI growth on trade-policy uncertainty are null at every horizon (min p=0.15). In the current episode, TPU peaked at 1,151 (4.3x the 2018 peak of 266) while new-investment FDI rose from $4.1bn (2024) to $7.3bn (2025), +36% vs trend in 2026Q1; the soft component is reinvestment of earnings (-21.5%). Do not cite 2017-18 as an established freeze-and-thaw base rate for total or new FDI.

The Chinese-FDI-specific chill (Rhodium announcement data: $272M Q1 2025, $57M Q2 2025; BYD withdrawal) is NOT testable publicly and is not contradicted; the tested object is total/new FDI, all nationalities. BOP equity ex reinvestment proxies Banxico's narrower 'new investments' category (Banxico SIE needs a token). Post-2022 new investment is structurally depressed ($4-7bn/yr vs $11-14bn 2014-2019), predating the 2025-26 episode. Power against a genuine sub-10% chill is low given FDI lumpiness.

Of 177 US power reactors that began construction 1954-2013, 76.3% were completed (95% CI 70-83%); completion fell 97.6% (pre-1968 starts, median 4.7y builds) to 70.5% (1968-78, median 9.7y) to 50% for the modern large-reactor sample (Vogtle 3/4 completed ~7.5y late at 10.4y each; V.C. Summer 2/3 abandoned after 4.4y and ~$9B). Monte Carlo on EO 14302's '10 large reactors under construction by 2030': base case P = 0.09% (median outcome 2 units); even a never-observed optimistic scenario reaches only 20.2%, and P(all 10 operating by 2035) stays ~0.2%. Restarts and uprates are the only fast lane; new large-reactor GW before the mid-2030s is a tail outcome.

Reactor dates transcribed from Wikipedia (secondary to IAEA PRIS; spot checks passed). Monte Carlo site-conversion and lead-time parameters judgment-anchored to sparse modern data; varied loudly across scenarios.

Across AEO vintages 1994-2025, 81% (17/21) overshot realized 10-year US electricity demand growth (mean error +0.57pp/yr, bootstrap CI [+0.25,+0.87]). Every vintage that projected >=2.0%/yr for a decade (n=5, all 2001-2005) overshot by a mean +1.37pp/yr against a realized median of 0.68%/yr; the 2025 NERC LTRA's implied ~2.17%/yr (+224 GW/+24%) sits exactly in that class. The honest counter is quantified: 2020-2023 vintages UNDERSHOT by -0.41pp/yr as data-center load outran the models. Treat +24%/10y as an at-risk ceiling with a floor higher than the flat-decade reflex assumes.

AEO SALES (TWh) used as proxy for NERC PEAK (GW) forecasts (LTRA vintage archive is PDF-bound). Overlapping 10y windows overstate effective n; the >=2%/yr class (n=5) is descriptive.

Difference-in-differences on EIA-861M retail prices (6 data-center states vs 14 region-matched controls): +0.60 log points, CI [-3.75,+4.59], placebo p=0.80 - no divergence through April 2026, and Virginia (the epicenter) LAGS its South Atlantic peers (+18.2 vs +26.2 log pts since 2019). The one leading-edge signal is residential-only: +5.30 log pts, p=0.082 (suggestive), consistent with cost-shifting onto households while data-center load sits on special tariffs. The PJM capacity-auction spike ($333.44/MW-day cap, $16.4B) clears into 2025-2027 delivery-year bills - the ratepayer impact is in front of us, not behind us. Do not claim present-tense data-center bill inflation on camera.

Treatment is a proxy classification (no keyless state-level data-center MW series); retail prices lag wholesale by rate-case timing; PJM cost socialization biases toward null. 2025-26 EIA data preliminary.

Top-decile delinquency LEVELS preceded a recession 0 of 2 times at 4 and 8 quarters (1991-2026), and the 2026 record episode (NY Fed card 90+ stock 13.12%) was a third failure out of sample. A sustained 3-quarter rising streak in the card rate preceded recession within 4 quarters in 50% of episodes (3/6, CI 17-83%) vs a 9.6% base rate (p=0.015; ex-COVID 40% vs 7.0%, p=0.041), but both post-2010 streaks (2015, 2022) were false positives. The lead/lag structure is coincident-to-lagging: delinquency change correlates best with GDP five quarters EARLIER and with the recession indicator at +1 quarter.

One fragile significant cell among ~80 grid tests; only 3-4 recessions in sample. Aggregate bank-book series, not income-split (distributional) data: that half of the claim is DATA-GATED. Treat a sustained delinquency upturn as a weak vulnerability marker, never a timing signal; a record LEVEL has never timed anything.

2025 payroll first prints were revised down a mean 67K/month first-to-current (bootstrap CI [-100, -34], permutation p=0.018 vs 2000-2024; p=0.008 ex-COVID), with 81% of months revised down vs 49% historically and a cumulative -1.07M markdown; the bias survives benchmark-free horizons (3rd-vs-1st -40K, p=0.029). Qualifications: it is a 2025 phenomenon (2026 prints so far unbiased, n=5, p=0.96), revisions are procyclical near turning points (-39K near recessions vs +15K in expansions, p<0.0001) but the sign was NOT predictable in real time (persistence 50.2% vs 42.9% base). None of BCR's five graded trigger verdicts (Dec 2025-Jun 2026) flipped between first print and current vintage, though 3 of 5 months crossed at least one verdict-family line.

BLS CES revision table 1979-2026 via same-week Wayback capture (direct fetch blocked); 5 ALFRED spot-check vintages match exactly. 2026 sample is 5 months; re-run at year-end.

Monte Carlo (20,000 draws) anchored to published S&P Global / USGS figures and deliberately supply-generous: existing mines decline toward 22 Mt by 2040 (S&P no-new-investment path), the sanctioned wave delivers by 2030 (validates near S&P's 33 Mt total-production peak), recycling grows to 7-10 Mt, demand runs to 42 Mt (S&P Jan 2026). Result: median 2040 gap +8.5 Mt (95% CI 3.8-13.2, bracketing S&P's published ~10 Mt: calibrated), P(gap>0) = 100%, P(gap>5) = 93%, P(gap>10) = 26%. The lead-time wall does the work: at a ~17-year copper mine lead time, projects FID'd 2026+ contribute a median of just 0.6 Mt by 2040. Even the everything-delivers case leaves +3.0 Mt. 2035 is NOT yet the crisis (P(gap>10) = 0% in 2035): this is a later-decade wall.

Works from published aggregate anchors (S&P project database is proprietary); no price-feedback (demand destruction / substitution is the memo's 35% counter-risk, handled there); recycling band 7-10 Mt is the widest wedge. Re-run on S&P/ICSG revisions.

In share space (USGS 2025e: China 14.0 Mt of 29.0 Mt world refined output = 48.3%, up from 44.9% in 2024), the displacement math requires ex-China refining to grow 1.40x for China to fall below 40% (1.89x for 33%) with China FLAT: while China adds capacity at ~5x lower build cost (Columbia CGEP) and holds 12 of the 20 largest smelters. A diversification-generous Monte Carlo puts China's share at median 48.8% in 2030 and 47.6% in 2035, with P(<40% by 2035) = 0.2% and a 41% probability the share is HIGHER in 2035 than today. The observed 2024→2025 move ran +3.4pp toward China: ~4pp/yr against the diversification direction. Honest cap: ~48% is concentration risk, not the ~91% REE-style monopoly; this is the copper thesis's supporting leg, not its spine.

Refinery-production share (output), not capacity share (S&P: ~40%); no dynamic TC/RC/closure loop: CGEP smelter economics inform the bands qualitatively. UNCTAD page-level cite for '>45%' still open in the memo.

Scenario decomposition of S&P's 42 Mt 2040 demand (AI/DC and defense each ~triple, adding a combined ~4 Mt) propagated through the supply-gap model's 20,000 supply draws: the 2040 median gap runs +10.0 Mt under an AI boom (P>10 = 50%), +8.5 base, +7.5 with a 50% AI-growth haircut, and +6.5 Mt with a FULL AI capex reversal (AI/DC frozen at today's level): P(gap>5) = 74% even then, P(gap>0) = 100% in all scenarios. The AI leg swings ~2.0 Mt (23% of the base gap; 1.2-2.8 Mt across split assumptions): the largest marginal swing factor, and the difference between a manageable deficit and a 10-Mt-plus one: but it does not create the gap: EV/grid/reshoring demand does the lifting. TECHNOLOGICAL-SUBSTITUTION CHANNEL (added 2026-07-24): fibre and co-packaged optics displacing copper interconnect remove a median 0.47 Mt from 2040 demand (90% range 0.22-0.87), moving the gap from +8.5 to +8.0 Mt and P(>5 Mt) from 93% to 90%; the compound worst case (full AI capex reversal AND optics on what remains) still leaves +6.4 Mt median, P(>5) = 71%, P(>0) = 100%. Optics remove ~5% of the median gap. The bound is physical: fibre carries data, not power, so transformers, switchgear, busbars and power cabling stay copper, and CRU expects that layer to keep growing with the buildout. Parameters were set generous to the counter-argument and the offsetting power-density/grid copper growth was not credited.

The 50/50 AI-vs-defense split of the current ~2 Mt combined is an assumption (swept 30/70-70/30); data-center t/MW intensity deliberately not load-bearing (memo carries it as [DATA NEEDED]); inherits the supply-gap model's limitations. The interconnect share of data-centre copper (modeled 15-40%) is unverified. Separately, NO t/MW intensity figure is load-bearing anywhere in this model: NVIDIA's widely circulated 'half a million tons per GW' was a unit-conversion typo (they meant pounds, 2,200x smaller), quietly corrected to ~200 kg/MW for rack architectures, so published intensity spans two orders of magnitude on scope alone.

IPEDS panel 2019-2023 (4,496 Title IV degree-granting institutions): exit from the active directory is strongly predicted by log-enrollment and 2015-19 enrollment trend alone (private nonprofit 4yr AUC 0.785, both predictors p<0.001, permutation p<0.0001). The gradient is an order of magnitude: top fitted-risk decile (median 121 students, -34% trend) exited at 22.5% in 4 years vs 0.9% for 5,000+ institutions; small-and-shrinking (<1,000 students, trend <-20%) exited at 19.3%. Scoring the 1,566 surviving private nonprofit 4-years on 2019-2023 data implies ~86 exits over the next 4 years (~21/yr, vs ~16/yr tracked closure announcements; exits include mergers), rising to ~111 (~28/yr) under a further -10% enrollment shock. The Philadelphia Fed's 80-closures figure is the abrupt -15%-overnight worst case, not a baseline.

Exit = left IPEDS actives (includes mergers, runs above pure closure counts by construction); no finance covariates this pass; training window spans COVID; never name at-risk institutions in content (aggregate counts only).

NY Fed college-labor-market series (1990-2026, 435 months): 88 total inversion months in seven episodes. The current episode (Jan 2021 - ongoing) is 63 months, 5.7x the longest precedent (11 months, 2019-20), and twice as deep (mean gap +0.81pp, CI [+0.70,+0.92], vs +0.41pp across prior inversion months; permutation p=0.0003). Block-bootstrap of the pre-2021 regime puts P(a 63-month run) at ~0.000. Base-rate warning for the normalization case: of six completed episodes, four closed by the GRAD rate falling, but every true normalization came from episodes of 5 months or less; the other two closed by the all-workers rate rising to meet it (2001, COVID), which ends the inversion without helping graduates. Current gap +1.40pp (5.63% vs 4.23%, Mar 2026).

Establishes persistence and abnormality, not cause (AI vs tech-hiring shift vs supply glut unresolved); NY Fed all-workers comparator (16-65 excl. enrolled students) is not headline U-3; 3-month-MA source data.

NY Fed early-career wage data (1990-2025, 2025 dollars): the median recent-grad premium over young HS workers was $19,785 in 2010 and $20,000 in 2025 (+1.1% real in 15 years; post-2010 slope CI [-133,+338] straddles zero). The statistically detectable structural break is in the tail, not the median: the 25th-percentile grad premium flipped from +$99/yr pre-2010 (CI excludes zero) to -$116/yr after (Chow F=4.59, p=0.018), sitting at $5,000/yr in 2025. Within-graduate dispersion (p75-p25) widened ~16% (avg $31.7K 1990-2009 vs $36.8K 2010-2025, ~$40K in 2025). Median real recent-grad wage has been ~flat since 2000 ($56.1K then, $58.3K 2010-25 mean). Never say 'the premium collapsed': the median held; the tail eroded and the spread widened.

Early-career (22-27) only; premium measured before debt service (the ROI layer is Abel/Deitz); 36 annual points limit test power; 2025 values provisional in the source.

WICHE 11th-edition projections (HS grads peak 3.86M in 2025, -5.5% by 2030, -12.5% to 3.37M by 2041) crossed with the NCES immediate-enrollment rate (69.8% in 2016 -> 62.0% in 2022, drift -0.51pp/yr since 2010): 20,000-draw Monte Carlo puts P(traditional freshman enrollment below the 2025 peak) at 91% in both 2030 and 2035 (median index 90.6 and 86.4, peak=100; 75.9 by 2041). Holding freshman enrollment flat requires the participation rate to reach 65.6% by 2030 and 70.9% by 2041, ABOVE the series' all-time high, while its actual trend is down. The decline is two-legged (drop to 2030, plateau, second leg after 2035), so 'cliff' overstates the near-term slope. Aggregate enrollment growth (NSC spring 2026 +1.3%) is adults/certificates around the shrinking traditional core.

Indexes the traditional (recent HS completer) pipeline only; adult/transfer/international channels excluded by design; WICHE error at 15y+ horizons wider than the calibrated band; NCES rate is CPS-sampled (SE ~2pp), latest 2022.

India's real manufacturing value added grew 6.15%/yr over 2015-25 ($328B to $595B constant-2015 USD, ~+82%) and 8.22%/yr in the 2020-25 PLI window - the flat/falling GDP share (15.6% to 13.5%) is purely the denominator: the manufacturing-minus-GDP real growth differential is +0.39pp/yr [CI -2.90, +3.47], p=0.83, indistinguishable from zero in every window tested. The two-clock contrast survives in sharper form: genuine share-gaining industrializations ran POSITIVE differentials (China 1995-2010 industry proxy +1.03pp/yr, p<0.001; Vietnam 2010-25 +2.44pp/yr, p=0.0005). Correct framing: 'manufacturing is growing but not outgrowing the economy'; WRONG framing: 'manufacturing went backward'.

Nominal share vs real differential mixes deflator effects (manufacturing relative prices fell); China benchmark uses industry incl. construction (true manufacturing differential likely higher, widening the contrast); WDI 2025 values provisional. This model corrects METRIC wording, not thesis direction.

Across 8 completed >=15% drawdowns from running ATHs (LBMA daily 1971-2026): median depth -29.3%, median 13.6 months peak-to-trough, median 12.2 months trough back to the prior ATH, median mechanical ride-back +41.5%: but the medians hide 1980 (19.5 years to trough, 28 years peak-to-recovery) and 2011 (4.3 years down, 8.9 total). Maturation is NOT SUPPORTED (depth vs era Spearman r=-0.02, p=0.96). The depth-entry test INVERTS the dip-buying intuition: 24-month median forward return was +47.0% buying within 10% of the ATH vs +3.4% buying >30% below it (deep vs shallow p<0.0001); within drawdowns, the real-rate fork dominates depth in every cell. The 2026 episode sits at -26% (June 25 trough-so-far), in the one bucket (-20/-30%) with decent conditional history (+27.1% median 24m, 73% up-rate): an era-confounded average, not a regime-neutral expectation.

Overlapping daily windows make bootstrap CIs optimistic; the deeper-than-30% bucket is mostly the 1980-1999 bear (the confound is the finding); 8 completed episodes only.

Tradable backtest (monthly, 1-month lag, T-bill cash leg, 1972-2026): the combined rule returns 7.21% CAGR vs 8.92% buy-and-hold (circular-shift null p=0.495 - timing luck on returns). The direction-only variant (12m real-rate change < 0) is the honest winner: 8.32% CAGR, Sharpe +0.60 vs +0.53, max drawdown -24.8% vs -61.8%, invested only 53% of months - near-B&H returns with 40% of the drawdown. But the edge lives pre-2000 (rule 8.89% vs B&H 7.18%, dodging the post-1980 bear) and DIED post-2022: the TIPS-basis rule since 2003 returns 3.13% vs 10.88% B&H (null p=1.00) because it exits during the official-bid rally that ran through high real yields. Current signal state is split: CPI basis HOLD (12m chg -1.74pp), TIPS basis EXIT (+0.24pp, level +2.25%). PART 2 (operator iteration): the 12m window IS too slow - direction-only short windows (1-3m) beat it on return and Sharpe (w=2m: 10.41% CAGR, Sharpe +0.76, above buy-and-hold while invested 51% of months; treat the exact w=2 cell as parameter luck since neighbors print ~8.6%), and at 6 of 8 completed troughs the 2-3m signal was ON within 0-3 months, capturing 68-100% of the trough-to-prior-ATH ride. The exception both times is post-2022: every window was months-to-years late at the 2022 trough because the rally ran while real rates rose. Long windows keep the drawdown crown (-24.8% at w=12m vs -32% at w=2m).

No transaction costs; the <1% level threshold is in-sample (the direction leg is the transferable part); verdict is regime-conditional - if the sovereign-bid regime fades, the pre-2000 rule behavior may reassert.

MoF oil revenue moves 10.2 SAR bn per $1/bbl of Brent (bootstrap CI [9.0, 11.5], R2=0.96 on 2010-2025; 8.4-10.5 across robustness variants; ~$2.3-2.8bn per $1). Quarterly revenue follows Brent with a one-quarter lag (R2=0.77), and the held-out 2026 war quarters landed within 2-4% of Brent-implied. Model-implied fiscal breakeven: $96 at plan spending / $101 with a typical 4% overrun: independently inside the memo's $94-111 Bloomberg/IMF band. 2027 scenarios at 2025 run-rates: $60 Brent -> SAR -365 to -421bn; $70 -> -263 to -318; $80 -> -160 to -216; $90 -> -58 to -114. Even at $90 war prices the budget does not balance at current spending.

Scenario grid holds spending/non-oil revenue at 2025 actuals (a sensitivity, not a forecast); GFSM-2014 break at 2017 bounds the slope (8.4 on the 2017+ subsample); above ~$100 the 80% royalty tier makes revenue convex: the linear fit understates windfalls.

As literally worded the claim is falsified: but only by carve-outs: Norway sat at 8-32% oil share while the world's #3 crude exporter (1993-2003) because it NEVER had the dependence to cut, and Russia's sub-40 prints (2016-25, while #2) are price crashes and sanctions shrinking the numerator, not diversification. The version that matters for Saudi Arabia is 0-for-6: no petrostate that launched a diversification program while oil-dependent (>50% of revenue) has reached sub-40 while remaining a top-5 exporter. The genuine diversifiers: Indonesia (60%->6%) and Malaysia: exited the top exporter ranks doing it (Indonesia rank 9 -> 26). Sub-40 is significantly rarer among top-5 exporters (odds ratio 0.48, permutation p=0.027). The Saudi clock: revenue share 67% at the 2016 launch -> 60% (2024), post-launch pace 0.55pp/yr with CI [-3.7, +4.7]: statistically indistinguishable from no decline; ~2060 at own pace, ~11 years even at Indonesia's pace.

Panel gaps documented (Saudi 2000-10, UAE 2016-17); cross-country basis differences (central vs general govt, investment-income denominators) cap precision at the which-side-of-40 level; if post-sanctions Russia stays sub-40, it becomes a genuine counterexample: watch-item.

Across KOSPI, the S&P 500 and TAIEX, 2015-2026 (8,419 market-days), the 60-day correlation between an index and its cross-sectional median constituent falls as the top two names take over the market cap: pooled slope -0.484 with market fixed effects (block-bootstrap 95% CI [-0.616, -0.316], Newey-West t=-5.95, circular-shift permutation p=0.0003), which is -0.048 of correlation per +10pp of top-two share. The loss is a top-decile effect rather than a smooth slope: deciles 1-9 are flat, and only the top decile drops (KOSPI 0.79 vs 0.89, S&P 500 0.78 vs 0.97, TAIEX 0.83 vs 0.92, all p<0.0002). On 2026-07-31 the S&P 500 reading is +0.25, the lowest in the whole 2015-2026 sample, and SPY vs RSP is at the 0.6th percentile of its own history; KOSPI is at +0.60 (3rd percentile). Sign holds in 19 of 19 specifications including non-overlapping first differences and a real SPY-vs-RSP cross-check.

The CROSS-MARKET half of the claim is NOT supported: TAIEX averages more top-two concentration than KOSPI (35.3% vs 28.8%) yet tracks its median stock better (0.901 vs 0.856), and on 2026-07-31 Taiwan is at the 57th percentile while Korea is at the 3rd. With n=3 markets there is no cross-market inference. Korea alone has no significant LINEAR slope (-0.269, CI [-0.413, +0.224]); its loss shows only in the top decile. Part of the negative sign is arithmetic, since a cap-weighted index dominated by two names must move with them, so quote the threshold and magnitude rather than the sign. Panels are current-membership with fixed share counts and large-cap medians (196 of ~946 KOSPI names), all of which bias toward understating divergence. The broad EM leg of the memo's four-market test was dropped for lack of constituent data. Contemporaneous association only, with no forward-return content: this is not a timing signal.

Tested on primary KOFIA daily data (margin-credit balance and forced liquidation, 2010-2026, fetched from the FreeSIS JSON endpoint) against KOSPI, Samsung and SK hynix returns. As the memo words it, the claim holds: with a same-day chip control, five lags of margin change add 0.21pp of R-squared to an 83.7% baseline and fail to reject at p=0.132 (HAC) and p=0.133 (10,000-resample block bootstrap); a 1-sd margin change maps to +1.3bp of KOSPI over five sessions, CI [-12.3,+12.5]. The null is well powered: the test would have caught any effect above 0.34pp of R-squared, and on the 4,066-day sample the floor falls to 0.12pp with p=0.344. It is not exactly true: KOFIA records on a settlement-date basis with an empirically identified 3-session reporting lag (corr 0.447 at k=+3, ~0 elsewhere), and realigning to trade date produces a small, non-tradeable residual effect (p=6.5e-06, 1.09pp of R-squared) that survives nonlinear own-return controls and flow-innovation orthogonalisation. The magnitude ranking is not close: forecast-error variance of the KOSPI at one day is 83.0% Samsung+SK hynix, 16.9% own shocks, 0.11% margin shocks; the reverse direction (index predicts leverage) runs p=3.6e-21 with 24.8pp of incremental R-squared, 23x the strongest margin result and 118x the published-alignment one, with an elasticity of 0.34% margin balance per +1% index day. Primary data also corrects the record: the 2026 single-day forced-liquidation peak was 169.8bn won on 2026-06-09, a day the KOSPI ROSE 7.87% after an -8.65% session, not the 142.2bn of 2026-07-09 that the press called the peak. Every memo leverage figure reproduced exactly from KOFIA (38.6328trn peak on 2026-06-24; 32.7492trn on 07-23; -5.88trn unwind = 0.32% of the 1,833trn cap decline; 425.8bn cumulative forced liquidation 07-01/07-10).

The same-day chip control absorbs any margin-driven move that itself passed through Samsung or SK hynix, so the design is biased toward the memo's conclusion and that bias is unquantified. The 3-session reporting lag is inferred from a correlation peak, not from KOFIA documentation. KOFIA margin data ends 2026-07-30, so the 2026-07-31 rebound (+17.91%, record retail net selling of 8.2737trn won) is outside the sample; re-run when that print lands and again at end-August 2026 after the FSC leverage-rule changes (both rules accelerated to take effect 2026-07-31 per the 2026-07-24 FSC decision, with a 20% per-investor portfolio cap announced 2026-07-29). Single-stock leveraged ETFs (~13.02trn won) are a separate channel not in the KOFIA margin series and are not tested. Daily frequency only, so intraday amplification is invisible. Sub-samples are unstable (2024 p=0.031, 2025 p=0.618, 2026 YTD p=0.023); do not quote single years.

Rolling 250-day R-squared of the S&P 500 daily return on the cap-weighted Magnificent 7 return rose from a 2015-2019 mean of 0.638 (CI [0.631, 0.645]) to a 2023-2026 mean of 0.785 (CI [0.781, 0.790]) as the cohort went from 10.6% to 33.4% of index market cap. Regressing R-squared on contemporaneous weight gives 0.0085 of R-squared per 1pp of weight (HAC lag-250 t=3.11, p=0.0019; block bootstrap CI [0.0014, 0.0147] per pp). The drawdown leg is the strongest result: across 11 peak-to-trough S&P declines over 5% since 2015, the Mag 7 share of the fall has a median of 14.8% pre-2021 (n=7) versus 48.5% from 2021 (n=4), reaching 60.1% in the 2026-01-27 to 2026-03-30 episode. Cohort dose response (2015-16 vs 2025-26 mean R-squared): top 2 0.33 to 0.64, top 3 0.38 to 0.75, top 5 0.55 to 0.79, top 7 0.65 to 0.82, top 10 0.72 to 0.86 - today's top three explain as much of the index as the top ten did a decade ago. Weight series rebuilt from a 628-ticker constituent market-cap panel and validated to 32.84% against SPY's published 32.61% Mag 7 weight on 2026-07-30.

About 76% of the R-squared rise is reproduced by the weight alone with the 2015-2016 correlation structure frozen, so this is largely the arithmetic of a cap-weighted index rather than a new co-movement regime; the correlation channel added only 22%. The relationship has already saturated: the 2021-2026 subsample slope is 0.143 (p=0.854) because R-squared plateaued in the 0.75-0.88 band while weight kept climbing. In monthly first differences the relationship is absent (slope -0.153, p=0.456) and the block-permutation null is 0.0796 two-sided, so treat this as a description of current structure, not a predictive relationship. Weights are full market cap rather than float-adjusted, membership comes from Wikipedia rather than the index provider, and 129 of 757 ever-members lack usable price history.

Across 98 US years (1927-2024, Kenneth French NYSE BE/ME breakpoints), the share of firms trading below book FALLS when the market advances: slope -0.480 per 1.0 of annual return (Newey-West t=-9.21, p=4.3e-26, R2=0.689, block-bootstrap CI [-0.598, -0.381]). In up years the below-book share rose only 28.4% of the time [18.9, 39.2] with a median change of -4.27pp; in years above +20% it rose only 9.8% of the time [2.4, 19.5] with a median -8.82pp. Concentration IS a real moderator: the return x concentration interaction is +1.608 (t=+2.80, p=0.0062, bootstrap CI [+0.254, +2.973], permutation p=0.0129), and the slope flattens monotonically from -0.648 in the least concentrated tercile to -0.390 in the most. But the sign never flips: the implied slope is -0.279 even at the December 2025 all-time-record top-decile cap share of 78.2%, and would reach zero only at 95.6%, 17pp beyond anything ever recorded. Korea's 2026 print is therefore an outlier against US history, not the general rule: from January to June 2026 the KOSPI below-book share went 66-63-67-70% (net +4pp) while the index rose 62% with trailing 12m returns of +107% to +214%, a move matched or beaten in only 2 of 41 US big-up years. On camera, say concentration WEAKENS the normal relation and Korea is the extreme case; do not say concentration reverses it.

Major scope reduction: the memo's monthly four-market panel could not be run. The Korean series is five press-reported points from one Seoul Economic Daily article and carries NO inferential statistics; Japan and Taiwan have no obtainable constituent book-value panel; the workspace Nasdaq Data Link key resolves to the Sharadar free sample (calendar 2018 only). The US leg is annual, not monthly, and concentration is proxied by the top-decile share of total US market cap (French size deciles) because no free series carries a US top-two weight back to 1926, so the specific two-name structure Korea exhibits cannot be tested. The base relation is close to mechanical (prices move, book is stale), so the informative result is the interaction, which rests on 33 high-concentration years in one country. Korea's 70%-to-75% July step coincides with a 33% index fall and is the ORDINARY response, not evidence for the claim: quote only the January-to-June segment.

Measured as a share of world supply at the time, 2026 Hormuz is the largest disruption in a 12-episode taxonomy spanning 1973-2026: 13.6% on a crude-only basis (14 mb/d of 103) = 2.09x the 1990-91 Gulf War (6.5%) and 1.79x the largest prior event on record (1973 Arab embargo, 7.6%); 19.4% on a crude+products basis (20 mb/d) = 2.98x the Gulf War. Median prior disruption: 3.4%. It is also the least re-routable: bypass capacity (Petroline + Habshan-Fujairah, 2.6-4.2 mb/d) covers only 24.3% of at-risk crude flow (17.0% on the products basis), leaving ~10.6 mb/d with no physical bypass. STRUCTURAL: ~2.7 of ~3.0 mb/d of world spare capacity sits BEHIND Hormuz: in past disruptions the substitute supply was outside the affected envelope and could sail.

ALWAYS STATE THE BASIS: 'several times larger than 1991' is defensible on the crude+products basis (~3x) and overstated on crude-only (~2x). The 2026 figure is AT-RISK throughput, not measured outage; ship-to-ship transfers continue. The episode table is hand-assembled and volumes carry ~+/-20% analyst disagreement (the 1973 absolute is share-anchored because the memo's 4.5 mb/d did not survive fact-check). Spare-capacity split is a hardcoded assumption. NO price-path-by-kind test exists: there are zero resolved transport disruptions with price data, and none should be inferred.

The oil-inventory-cliff finding reproduces strongly on 2004-2026 weekly data: tight accessible inventory (days-cover Z < -1, n=70) preceded a median -8.69% forward 13-week WTI change vs +3.83% when loose (n=1073), difference -12.5pp, p=0.000. The proposed mechanism: a stranded surplus discounting to clear: FAILS the decisive test: among tight-inventory weeks, forward crude was -8.27% with a wide location discount (n=42) and -9.61% with a narrow one (n=28), difference +1.34pp, p=0.784. Tight inventory predicts weak forward crude whether or not a stranded-barrel discount is present. Step 1 survives only weakly: Cushing share vs Brent-WTI r=0.36 [0.31,0.41] full-sample, but r=0.09 excluding the 2011-2014 landlocked episode that carries it.

The proxy measures the WRONG STRAND: Cushing is a US midcontinent hub, while the live 2026 question is a Gulf-producer strand no available series measures, so a null here is weak evidence about the Gulf. Brent-WTI is a dirty location spread (quality, freight, the 2015 export-ban lift). n=70 tight weeks splitting 42/28: p=0.784 is a genuine failure to detect a difference, not proof of none. Panel ends 2026-05-29, so the live July-August 2026 episode is not in sample. oil-inventory-cliff therefore remains a finding WITHOUT a mechanism.

Across 11 real oil price shocks (crude +40%/6m, 1987-2026), US petroleum demand was below its 5-year pre-shock trend in only 40% of episodes at 12 months, median -336 kb/d (i.e. demand ran ABOVE trend), bootstrap CI [-895, +1042]. A high oil price on its own has NOT reliably destroyed demand. The split that matters is recession: episodes followed by an NBER recession destroyed a median +1,042 kb/d at 12m vs -389 kb/d (above trend) with no recession; at 6 months the difference is +1,203 vs -380 kb/d, p=0.036. World-scaled, the recession branch is +5,369 kb/d (~24x the 225 kb/d SPR exchange refill bid) and the no-recession branch is -2,002 kb/d. P(recession within 12m of an oil shock) = 27.3% (3 of 11). The oil-buffer memo's '45% demand destruction' leading counter-risk is therefore a DISGUISED RECESSION CALL, not an independent price-rationing mechanism, and oil-shock-transmission-lag's '20% destruction without recession' looks generous.

PRE-1986 BLIND SPOT that cuts AGAINST this conclusion: WTISPLC is a posted/regulated price before 1986, so the 1973-74 embargo and 1979-80 Iranian revolution -- the best-documented demand-destruction episodes -- cannot be detected and are absent from every parameterisation, biasing the sample against finding destruction. Both were also recessions, so including them would likely STRENGTHEN the recession conditionality, but that is an expectation not a result. n=11 with only 3 recession episodes: 12m/24m splits are directionally clear but insignificant (p=0.22/0.19); only 6m clears p<0.05. Recession is ENDOGENOUS to the shock in several episodes -- this is a conditional split, NOT a causal decomposition, and must never be presented as 'oil shocks don't matter unless a recession happens anyway.' US demand proxies world demand by linear scaling when OECD/non-OECD elasticities differ. Linear trend counterfactual. A 24-month de-clustering gap in the first build silently dropped the 1990 Gulf War and 2022 Russia-Ukraine; fixed to 12 months and swept.

Market Cap / GDP at 191%: well above the 180% extreme zone. Equity Risk Premium at -0.49%. Shiller CAPE in the 99.1th percentile of its 155-year history. The Buffett Indicator is the simplest valuation framework: is the stock market bigger than the economy that supports it? At 191%, the answer is emphatically yes. Previous readings above 180% preceded the dot-com crash and the 2021-2022 correction. This doesn't tell you WHEN to sell: but it tells you the margin of safety is gone. Any negative catalyst (earnings miss, rate surprise, geopolitical shock) hits a market that has zero cushion. One honest caveat: the indicator remains above the 180% extreme zone while mean-reverting from its peak (Z=-0.8, 89th percentile): it has been easing for two quarters. Extended valuation is the backdrop; the direction of travel is currently working the extreme off.

Watch for: This is a slow-burn indicator: it doesn't trigger corrections, but it determines how deep they go. The trigger will be something else (earnings, rates, event); the Buffett Indicator tells you the magnitude risk.

Positioning crowding score at 2.8 across 5 futures markets (one most-extreme trader-category reading per contract, |Z| beyond 1.5σ on 3-year positioning levels): CFTC TFF: Euro FX Leveraged Funds Net % (short, Z=-2.6), CFTC TFF: 10Y Treasury Note Asset Manager Net % (long, Z=+2.2), CFTC DCOT: Silver Producer-Merchant Net % (long, Z=+1.9), CFTC TFF: British Pound Asset Manager Net % (short, Z=-1.9). Crowded trades unwind fast. One data surprise or policy shift and the deleveraging becomes self-reinforcing. The most crowded trades have the worst risk/reward because everyone's exit is the same door.

Watch for: Any data print that contradicts the consensus direction of these positions triggers the unwind. Watch for the first sign of capitulation in the weekly CFTC data: a one-week Z-change of 1.5+ in the crowded contracts is the tell.

The Equity Risk Premium has turned NEGATIVE at -0.49% (earnings yield 4.26% minus 10Y yield 4.75%). Stocks are no longer compensating investors for the extra risk of owning equities over risk-free Treasuries. This is the most expensive the market has been relative to bonds since the dot-com era. When the ERP goes negative, either earnings need to surge or bond yields need to drop: otherwise stocks reprice lower. However, the Real ERP (adjusting for inflation expectations) remains at +1.78%, suggesting some of the nominal compression is driven by inflation priced into bonds. If inflation cools, the nominal ERP could improve without stocks needing to move.

Watch for: Watch for bond fund inflows to accelerate. When retail realizes they can earn more in risk-free Treasuries than in stock earnings yield, the rotation is sudden.

SOFR volatility spiking (Z=+1.9). Equities have been mixed (SPY -3.0% 1W, -0.1% 1M, drawdown -1.4%). This is the modern TED-spread equivalent: the spread between risky (unsecured bank) and risk-free (overnight secured) dollar funding is widening, meaning banks and dealers are hoarding liquidity. Historically, SOFR-FFER spreads above 8bp precede broader funding dislocations (Sept 2019 repo spike, March 2020 COVID, March 2023 SVB). When dollar funding costs rise faster than policy rates, (1) levered trades across credit, rates, and FX start to unwind; (2) dealer balance sheets shrink, reducing market making in Treasuries and corporates; (3) short-term funding costs for corporates rise, squeezing issuance. This is the leading edge of any broader liquidity event.

Watch for: Watch SOFR vs IORB (Interest on Reserve Balances): if SOFR trades above IORB for 3+ consecutive sessions, the Fed's Standing Repo Facility (SRF) usage will spike. SRF take-up above $50B is a red flag that dealers can't self-fund. Secondary confirmation: CP 30-day yield rising faster than T-bills, money market fund outflows from prime funds to government funds.

Jul 06, 2018 (similarity: 75%): Reflation at 49%, DTWEXBGS depressed. During this regime (1w regime): SPY +0.0%, TLT +0.0%, GLD +0.0%, HYG +0.0%, UUP +0.0%

Mar 09, 2018 (similarity: 73%): Reflation at 44%, no extreme indicators. During this regime (4w regime): SPY -7.3%, TLT +3.6%, GLD +1.4%, HYG -0.9%, UUP +0.2%

Jun 17, 2005 (similarity: 73%): Reflation at 35%, no extreme indicators. During this regime (8w regime): SPY +2.0%, TLT +6.2%, GLD +3.2%

Jul 13, 1973 (similarity: 72%): Reflation at 54%, FEDFUNDS elevated, CPIAUCSL depressed. During this regime (4w regime): outcomes unavailable

Dec 08, 1978 (similarity: 72%): Reflation at 49%, no extreme indicators. During this regime (3w regime): outcomes unavailable