Executive Summary
The October 23 packet describes one machine with two engines pulling against each other: an economy running hot enough that prediction markets now price a December Federal Reserve hike near 78 percent, and an AI capital complex whose marginal buyer has quietly become the credit market, with Morgan Stanley pegging 2026 AI-linked debt issuance near $570 billion. Energy is the fulcrum: Brent topped $113, the highest since May, on Gulf shipping attacks and post-midterm strike planning. An oil shock inside a hiking cycle is the scenario equity AI valuations have not priced.
U.S. data keep the hawks armed. September CPI ran 3.7 percent year over year, led by gasoline up 9.1 percent, while jobless claims fell to 194,000, the lowest since July. Friday's flash PMI is the last clean activity print before the October 28 FOMC. The labor market is not deteriorating, which removes the excuse to look past energy-driven inflation.
Abroad, the story is divergence under strain. Eurozone composite PMI slipped to 51.9 against a 52.7 forecast, and the IMF still pegs 2026 global growth at 3.1 percent with 4.7 percent inflation. China's Q3 GDP beat at 4.5 percent on an export surge reduces Beijing's urgency even as the mix worsens, and India is burning reserves — $60 billion in six weeks — defending the rupee.
Corporate America delivered the week's most consequential reversal: Meta raised 2026 capex guidance to $135–145 billion and printed negative free cash flow of $2.1 billion, while Tesla cut its capex plan to roughly $20 billion — the first megacap AI-cycle capex cut. The market repriced the electricity leg of AI immediately: Talen Energy fell 8.3 percent and Eaton 7 percent.
Artificial intelligence itself is commoditizing at the frontier while failing in deployment. Meta's Muse Wonder entered the frontier at $4/$20 pricing, and an 816-billion-parameter open-weight model landed the same week. Yet 69 percent of U.S. businesses report AI failures while 92 percent plan more spending. Spending continues; returns do not. That gap is the widest in the packet.
Supply chains are taxing everything downstream of AI. Commodity DRAM contract prices rose 93–98 percent in a single quarter as HBM eats wafer capacity, and HBM4 pricing is set to rise 121 percent into 2027 with all three suppliers sold out. The quiet capex escalator is memory, and it lands in 2027 budgets next year.
This edition previews twelve sections: the U.S. macro-to-micro transmission; the global macro signal; the India and China lenses; the micro earnings implications; a prediction; a positioning move; the AI disruption lens; AI supply chain risk; a blind spot radar of signals outside the brief's normal beats; a trend predictor with current, historical, and projected paths; and the dead canaries. Each ends in forward implication. The single highest-conviction call: the next Fed move is up, not down, and the AI complex's funding stack — not its earnings — is the pressure point to watch into December.
U.S. MACRO-TO-MICRO
The transmission chain this month runs energy to headline CPI to the Fed's reaction function to the AI funding stack. September's 0.6 percent monthly CPI print, with energy up 5.2 percent and gasoline up 9.1 percent, is not a demand story — core added 0.3 percent — but the Fed cannot credibly cut into an energy shock, and Warsh has rejected forward guidance, making the October 28 statement the volatility event. Markets price a hold at 3.75–4.00 percent with roughly 15 percent odds of a surprise hike.
The labor side removes the dovish hedge. Claims at 194,000 for the week ended October 17 are the lowest since July — there is no labor-market deterioration to justify looking past energy inflation. The last clean read before FOMC is Friday's flash PMI: September's composite hit 58.4, a five-year high implying near 5 percent annualized growth, and October consensus sees only mild moderation. An upside print cements December-hike odds above the current 77 percent.
Analysis, labeled: Rates have already begun moving. The 10-year yield fell 17–20 basis points to roughly 4.95 percent in a risk-off bid, with the 30-year at 5.29 percent — note the direction of the move: yields fell on AI-capex doubts, not on growth fears. When the flight-to-quality bid is about AI skepticism, credit spreads in data-center-adjacent sectors are the variable to watch, not the 10-year level.
Follow the Money
The micro consequence lands in funding-sensitive corners first. A Fed that may hike in December against a 5.29 percent 30-year reprices every leveraged AI financing done in the assumption of easing — the $570 billion AI-linked debt stack, neocloud vendor paper, and project finance for power. Flash PMI Friday and the FOMC statement Wednesday week are the two dates that decide whether this repricing accelerates. Follow the spread between AI-sector corporate bonds and treasuries, the forward calendar into FOMC, and who is still issuing: hyperscalers with negative free cash flow are the natural sellers of new paper into a widening market. Money is also moving from duration risk into cash — the same rotation Michael Burry signaled by closing Nvidia and Palantir puts rather than rolling them. When bears close shorts and issuers lean on financial engineering in the same fortnight, the funding leg is where the stress shows first.
MACRO SIGNAL — GLOBAL
Europe is softening while the U.S. runs hot, and that divergence is the global macro signal. The eurozone flash composite fell to 51.9 versus 52.7 expected, with Germany at 52.8 and France softening — still expansion, but the miss argues against a near-term ECB move even as the IMF expects one more hike this year. The IMF sees at least one more 25-basis-point ECB hike in 2026 after June and September moves; the October PMI miss versus the oil premium is exactly the debate.
Japan is the other side of the carry trade. 83 percent of BOJ watchers expect the next hike, off 1.25 percent, at the December meeting with October 30 a hold. A December BOJ hike stacked on a December Fed hike is the configuration that historically unwinds yen-funded carry — and AI equities have been a major destination of that carry. The two central banks are scheduled to tighten into the same risk complex in the same month.
Analysis, labeled: The IMF's frame is divergence: 3.1 percent global growth in 2026 with 4.7 percent inflation, where energy-importing economies outside the tech value chain face the hardest outlook. That is a split world — AI-value-chain participants grow, commodity importers stagnate — and it explains why capital keeps concentrating into a narrower set of assets even as their valuations stretch.
Follow the Money
China held policy still. The PBoC kept the one- and five-year LPR at 3.0 and 3.5 percent for a seventeenth month, with Governor Pan Gongshen shifting the policy benchmark toward the overnight repo. No fresh cuts means the stimulus burden stays fiscal, and a 4.5 percent GDP print reduces urgency. The global implication: none of the three major central banks is adding stimulus into an energy shock, which removes the traditional policy offset for any downturn. Follow reserve flows and currency interventions, not just policy rates. India burned $60 billion defending the rupee in six weeks; the ECB faces a stagflationary oil premium; the BOJ must choose between yen stability and market function. When multiple central banks intervene or tighten simultaneously, dollar liquidity tightens globally — and the highest-leverage assets, AI infrastructure included, feel it first through their funding costs.
INDIA LENS — MACRO-TO-MICRO
India is the packet's largest reserve-burning operation, and the arithmetic is now public. Reserves fell to $726.0 billion from a $785.7 billion peak as the RBI sold dollars against high crude, rising U.S. yields, and $26.8 billion of FPI outflows. Reserve cover is shrinking fast enough to test the rupee's ~95 floor — a floor defended with visible, depleting ammunition.
Bankers' estimates say half the war chest is gone. Roughly $75–80 billion of the June–August mobilization is already deployed through spot sales and swaps. If the remaining capital halves again at the current pace, the choice narrows to a stepwise rupee adjustment or tighter domestic liquidity — both inflation-relevant, both equity-negative.
Analysis, labeled: Strategists already treat depreciation as structural. CNBC's survey sees the rupee weakening toward 100 per dollar by mid-2027 despite administrative curbs, with forward-book hedging costs and a $200 billion forward-liability overhang keeping pressure on. The forward overhang matters most: every dollar the RBI spends defending spot today enlarges tomorrow's delivery obligation.
Follow the Money
The micro consequence lands on importers and dollar borrowers first. Crude above $113 raises the import bill exactly when the defense drains reserves; companies that hedged cheaply in the forward book earlier in 2026 face renewal at materially worse rates. Expect either a stepwise adjustment or liquidity tightening once the war chest halves again — and position for rupee volatility, not rupee stability, through mid-2027. Follow the forward book and FPI flows, not the spot fix. $26.8 billion of FPI outflows is the pressure behind the intervention, and swap-based defense is borrowing from future supply. Money is also moving into dollar assets and out of emerging-market equities at the margin — watch whether the RBI pivots from spot defense to capital controls, which historically marks the late stage of a defense rather than its success.
CHINA LENS — MACRO-TO-MICRO
China's beat is real but unbalanced. Q3 GDP grew 4.5 percent year over year against a 4.4 percent forecast, on an export surge while property and consumption drag. The beat cuts both ways: it reduces Beijing's urgency for aggressive stimulus even as the growth mix becomes more dependent on the one component — exports — most exposed to Western rebalancing pressure.
Policy stayed still. The PBoC held the LPR at record lows for a seventeenth month while signaling the overnight repo as the new policy benchmark — a framework modernization, not a stimulus. With the growth print beating, expect fiscal measures targeted at property completion and consumption only if the export engine stalls, not before.
Analysis, labeled: The composition is the risk. Tech-export strength is offsetting a deepening property slump, and the question is whether export momentum survives EU and U.S. rebalancing pressure as the surplus compounds through Q4. A 4.5 percent economy with a shrinking domestic demand core is a supply shock generator: more goods chasing constrained foreign demand means more trade friction, not less.
Follow the Money
For the AI complex, China's stillness is a wildcard. A holding PBoC and a GDP beat mean China is not competing for capital or commodities this quarter, which marginally relieves global inflation pressure — but an export surge into Western protectionism is the likeliest trigger for the next tariff escalation. Watch whether Q4 export momentum holds; a stall would force the fiscal response markets have been waiting for. Follow the trade surplus and where it is reinvested. A compounding surplus with no domestic absorption flows into treasuries, gold, and strategic commodities — and the property slump means the traditional sink is gone. Stimulus is rolling out at the margin, but the fiscal burden stays until consumption revives. The money to watch is state-directed capex into semiconductors and equipment, which competes directly with the Western AI buildout for the same supply chain.
MICRO IMPLICATION
The earnings week restructured the AI capex narrative in two prints. Meta's revenue surged to $64.6 billion, but capex guidance rose to $135–145 billion and Q3 free cash flow printed negative at $2.1 billion, with a new non-voting C share class created. Meta is now funding AI at the balance-sheet level, not from earnings — the funding template for the entire hyperscaler complex, and a template that requires benign credit markets.
Tesla went the other way. 2026 capex was cut from above $25 billion to roughly $20 billion, with Q3 spend 41 percent below Street estimates at $4.7 billion and the tens of billions deferred to 2028. The first megacap capex cut of the AI cycle is a demand-side admission dressed as prudence, and the October 22 selloff in power infrastructure — Talen Energy down 8.3 percent, Eaton down 7 percent — traces partly to this print.
Analysis, labeled: Old-economy cash flows are quietly funding the new economy. IBM lifted its outlook while CSX grew free cash flow 80 percent and Kinder Morgan beat its EBITDA budget — industrials raising guidance is the breadth tell that the economy itself is fine even as AI names wobble. The bifurcation: cash-generative old economy funding, and capex-hungry new economy burning.
Follow the Money
The foundry sits in the middle. TSMC beat and raised with AI demand "far exceeding supply," holding 2026 capex at $56–64 billion — deliberate under-investment relative to demand is the mechanism that keeps pricing power, and choke-point leverage, at TSMC through 2027. Every hyperscaler's buildout plan runs through a supplier deliberately rationing capacity. Follow the funding structure, not the capex number: Meta's negative FCF plus a new share class means dilution and debt are next, with the C-class structure already in place. Follow also who cuts: Tesla's deferral shifts tens of billions out of 2026–27 demand for power and equipment, which is why the electricity trade unwound the same week. Capex guidance is now a leading indicator of AI-infrastructure revenues with a two-to-four quarter lag.
THE PREDICTION
Prediction: the Federal Reserve raises rates at the December meeting, and the AI-infrastructure credit complex — not equities — is where the damage concentrates. The evidence stack: 3.7 percent CPI with energy driving the upside, claims at cycle-low 194,000, and prediction markets pricing the October hold near 85 percent but the December move as a hike at roughly 74–78 percent — an inversion of the entire 2024–25 easing narrative. Real money has already switched sides.
Second prediction: Brent holds above $100 through the U.S. midterm window, with strike risk keeping a fat tail under the price. Prediction markets swung the "no US strike on Iran" contract from 82 to 29 to 64 percent inside 30 minutes on October 19 — a post-midterm strike is now priced at roughly one-in-three, and WTI is up 20 percent over eight sessions with 18 attacks on Gulf shipping in two weeks. Energy-driven CPI is what forces the Fed's hand; this is the mechanism connecting geopolitics to December.
Analysis, labeled: Third prediction: AI-linked credit spreads widen measurably before year end. Roughly $1 trillion was wiped from AI and semiconductor names in the week to October 16, and Morgan Stanley pegs 2026 AI-linked debt issuance near $570 billion — when the marginal buyer of a buildout is the credit market and rates rise, spreads do the adjusting. CoreWeave's 25 percent weekly drop and bond-yield spike is the template, with management insisting no Nvidia backstop is needed — historically the sentence that precedes needing one.
Follow the Money
The falsifiers: a dovish October 28 statement or press conference, an October flash PMI below 55, or a rapid de-escalation in the Gulf would each break the chain. Friday's flash PMI is the last clean activity read before FOMC — treat it and the statement as the two decision points that either confirm or kill this prediction. Follow the pricing of the December meeting, not October: Polymarket's economy board carries a persistent ~16 percent October hike probability on $32 million of volume, while December carries the real hike pricing. Money positioning for a December hike will front-run the FOMC itself — watch leveraged AI credit, yen funding costs, and the oil-curve back end, because that is where a December hike is discounted first and repriced hardest.
THE MOVE
The positioning move for this environment: reduce duration in AI-infrastructure credit and buy energy-curve convexity. The core exposure to cut is anything financed on the assumption of Fed easing — the neocloud financing stack and AI-linked debt that has reached roughly $570 billion of issuance reprices first in a December-hike world. CoreWeave's spiral shows the shape: a 25 percent weekly loss and a bond-yield spike framed by consensus as idiosyncratic. It is not idiosyncratic; it is the sector's financing model in miniature.
The hedge is energy. With WTI up 20 percent in eight sessions and Aramco HQ evacuating staff, the oil curve's back end remains the cheapest insurance in the packet — a one-in-three priced strike probability means the fat tail is still underpriced in options, not in futures. Energy exposure also hedges the CPI mechanism that drives the Fed decision itself.
Analysis, labeled: For enterprise operators rather than traders, the move is contracting discipline. With 69 percent of businesses reporting AI failures and 92 percent still planning more spend, the expected-value play is to gate incremental AI capex on measured deployment outcomes, and to keep portability — data, evaluations, workflow evidence — contractually portable across vendors. Four frontier labs within five index points means today's vendor is not guaranteed next year's price.
Follow the Money
For supply-chain exposed businesses, lock memory now. DRAM contract prices rose 93–98 percent in a quarter and IDC sees 2026 PC shipments down 11 percent — anyone buying devices, servers, or memory-bearing equipment in 2027 should contract in 2026, because HBM4 pricing up 121 percent with all suppliers sold out is a confirmed escalator, not a forecast. The money in this move is in the funding gap between the two engine types: cash-rich old economy (CSX free cash flow up 80 percent, IBM raising outlook) versus FCF-negative AI builders (Meta at negative $2.1 billion). Rotating exposure from the funded-by-debt complex toward the funded-by-cash complex — while keeping energy convexity — is the single trade consistent with every section of this brief.
AI DISRUPTION LENS
The frontier commoditized this fortnight. Meta's Muse Wonder — its first frontier model — entered at $4/$20 pricing with an Artificial Analysis index of 56 versus Opus 5.5's 58. Four frontier labs within roughly five index points means pricing discipline is fragile: with near-parity quality, the rational lab strategy is price competition, and the rational buyer strategy is multi-vendor with portability.
The open-weight flank arrived simultaneously. Muse Spark Open is an 816-billion-parameter open-weight MoE released frontier-class in the same window as the closed launches. A frontier-class open model commoditizes the mid-tier API market within quarters — closed vendors' moats must move up the stack to agents, verticals, and enterprise integration or be competed away on price.
Analysis, labeled: Deployment, not capability, is the binding constraint. Roughly nine in ten executives report no measurable productivity or employment effect despite 69 percent adoption, with MIT pegging 95 percent of genAI pilots as failing. The gap between $790 billion of 2026 AI capex and realized enterprise revenue is the widest spread in the technology economy — and it is a demand-side problem, not a model problem.
Follow the Money
The disruption to expect next: verticalization and failure costs. OpenAI's Rosalind Academic Plan pushes a dedicated life-sciences workbench model into academia — the next adoption wave is domain-locked tools, not general chat. Meanwhile BSI finds 69 percent reporting failures including security breaches while 75 percent expect higher AI costs — a margin headwind for the enterprise software layer that consensus is not modeling. Follow where model economics force revenue: with frontier quality converging, $4/$20 pricing against a 56 index compresses inference margins, so labs pivot to enterprise contracts, verticals, and financial engineering (see Meta's new C share class funding the compute bill). The money to watch is enterprise AI procurement — the $790B capex figure needs a revenue story, and verticalized models with measurable outcomes are where that story gets written or dies.
AI SUPPLY CHAIN RISK
Memory is the squeeze. HBM consumes roughly three times the wafers of DDR5 per bit, so commodity DRAM contract prices rose 93–98 percent in Q1 and phone memory packages nearly doubled in a quarter, with IDC projecting 2026 PC shipments down 11 percent. AI demand is now a direct tax on consumer electronics — the cost of the AI buildout is being distributed to every device buyer.
The escalator continues into 2027. HBM4 stack pricing is set to rise 121 percent year over year, from roughly $600 to $1,300, with all three suppliers sold out and TSMC's CoWoS advanced packaging allocated through mid-2027. Accelerator bill-of-materials inflation is the quiet capex escalator inside every hyperscaler's 2027 budget — the budgets announced this quarter already understate it.
Analysis, labeled: The foundry is absorbing the node transition. TSMC's Q3 beat-and-raise on AI and HPC demand comes with N2 ramp costs trimming gross margin even as revenue hits records. The strategic question for 2027: whether A14 pricing power restores margin or the margin peak is already in — the answer determines whether the choke-point supplier can keep under-investing, which is what sustains its pricing power.
Follow the Money
Consumer hardware is the release valve — and the victim. When HBM crowds out commodity DRAM, either consumer prices rise or shipments fall; IDC's -11 percent PC forecast says both. For procurement teams, the forward implication is explicit: memory-bearing purchases made in 2027 will cost multiples of 2025 levels, and supply allocated to hyperscalers will not be reclaimed by spot buyers. Follow the wafer allocation, not the press releases: TSMC's $56–64 billion capex against "far exceeding supply" demand is a deliberate rationing choice, and sold-out HBM4 through 2027 means allocation, not price, is the currency. Money will flow to whoever holds 2027 memory and packaging slots — the scarcity rents are migrating from chip designers to memory makers and packagers.
BLIND SPOT RADAR
Signal outside the brief's beats: housing is transmitting the rate scare weeks before it hits AI multiples. Pending home sales fell 7.3 percent month over month in September against a -0.4 percent forecast — the worst miss of the cycle — buried in a market wrap next to a Dow up 1 percent day. Low confidence in the read itself: the print is one month, but its scale is not noise.
Nasdaq internals are near-bear-market while the index is flat. On October 20 the Nasdaq logged 361 new 52-week lows against 31 highs on a day the indexes rose. Consensus is anchored on index calm; the internals say rotation into cash. Low-confidence flag: breadth is a timing tool, not a direction tool.
Analysis, labeled: Prediction markets themselves are now systemic. Kalshi printed a record $4.5 billion day on October 18 and a $22.4 billion week — driven mostly by sports parlay volume — big enough that regulatory backlash (a Massachusetts action and a pending Supreme Court case) is the sector's own canary. A sports-contract crackdown would drain liquidity from the same markets now pricing Fed policy.
Follow the Money
Quiet regulatory and adjacent-tech items: AI failure rates now include security breaches as the headline category — expect cyber-insurance repricing for AI deployments within quarters. And the 30-minute 82→29→64 percent swing in Iran-strike odds shows event markets are thin enough to be moved by single flows — a leverage and manipulation surface nobody is watching. The blind-spot money is in repricing laggards: cyber-insurance and D&O premiums for AI-heavy balance sheets, municipal and housing-adjacent credit as the pending-sales miss transmits through realtors, lenders, and title chains, and the $22.4 billion weekly event-market float whose regulatory shock would hit the fintechs intermediating it. None of these are in the AI conversation; all of them reprice on the same December decision.
TREND PREDICTOR
Fed funds path — current: 3.75–4.00 percent with an October hold near 85 percent priced; historical: a 2024–25 easing cycle now fully inverted; projected: a December hike priced at 74–78 percent on prediction markets, the first hike of a new cycle. What the data illustrates: real-money conviction moved before the FOMC; the market is not waiting for the statement.
10-year Treasury — current: ~4.95 percent after a 17–20 basis point risk-off rally; historical: above 5 percent as recently as October 21; projected: resumption above 5 percent into a December hike, with the 30-year near 5.29 percent anchoring long-end term premium. The yield-curve-style read: long-end premium plus a hiking Fed is a tightening double-barrel for duration-sensitive AI credit.
Analysis, labeled: Oil — current: Brent above $113, WTI up 20 percent in eight sessions; historical: May's ~$90 base before the Gulf escalation; projected: a $100+ floor through the midterm window with a strike-probability fat tail, given 18 shipping attacks in two weeks. This is the input that keeps the hawk case alive regardless of core inflation.
Follow the Money
AI complex — current: ~$1 trillion of equity value shed in the week to October 16 and ~$570 billion of 2026 AI-linked debt outstanding; historical: capex guidance cuts were unthinkable at mid-year; projected: memory and packaging inflation of 100+ percent into 2027 plus credit-spread widening compresses 2027 buildout economics — capex plans will be quietly rebased, as Tesla already has. The trend to watch is guidance language shifting from "raising" to "phasing." The predictor's money metric is the spread: AI-sector bond spreads and their weekly change. Current readings are near cycle tights; historical spreads widened 200+ basis points in comparable funding shocks; projected widening into December as negative-FCF issuers come to market and December hike odds hold above 70 percent. When the spread turns, equity AI multiples follow with a lag — the bond market is the leading indicator now.
DEAD CANARIES
Canary one: Bitcoin as the leveraged liquidation valve. BTC slid 3.4 percent below $81,500 on October 22, down sharply from 2026 highs and decoupling downward from equities. Crypto is behaving like the leverage-release mechanism of the AI-risk complex — a classic late-cycle flow signal that consensus dismisses as crypto being crypto.
Canary two: CoreWeave as the neocloud template. A 25 percent weekly loss, a bond-yield spike, and management stating no Nvidia backstop is needed — consensus treats it as idiosyncratic, but it is the canary for the entire GPU-collateralized financing stack. When the template issuer's paper gaps wider, every similar issuer's cost of capital is set by that print.
Analysis, labeled: Canary three: index calm over broken breadth. 361 new 52-week lows against 31 highs on the Nasdaq on a flat tape is the same internal structure as late 2021. The "everything fine" headline is what consensus anchors on; the internals already moved.
Follow the Money
Canary four: the housing transmission. A -7.3 percent pending-sales month against a -0.4 percent forecast, receiving one paragraph of coverage, is rate-hike fear hitting the real economy through the most rate-sensitive channel — weeks before it touches AI equity multiples. The canary's message: the tightening is real and already flying; the debate is only about the landing. The canaries all route through one place: funding. Crypto liquidations, neocloud paper, breadth-driven margin calls, and mortgage-rate lockout are four valves of the same pressure. Watch which valve opens next after December FOMC — the first systemic funding accident of this cycle will announce itself there, not in an earnings call.
Section Addendum — Methodology, Source Boundary & Monitoring Notes
This edition is built from a frozen packet of 39 verified URLs (editions/2026-10-23/sources.md), captured October 23, 2026, 04:30–05:30 EDT. Every citation was independently re-verified by curl with a browser user agent before insertion; 401/403/406 responses are recorded as bot walls with search-snippet confirmation, not dead links. Subagent verification claims were not trusted; all verification was repeated at the parent.
Monitoring cadence for recurring signals: prediction-market Fed pricing (daily), AI-sector credit spreads (weekly), Gulf shipping incident count (daily), RBI reserve releases (bi-weekly), DRAM/HBM contract pricing (quarterly with monthly color), Nasdaq breadth internals (daily), CoreWeave and neocloud bond spreads (daily until resolved). Any source that dies (404/000 on re-check) should be swapped for a live same-article copy before the next edition.
Word-count and audio calibration: the listen script (Executive Summary through Dead Canaries) targets 15–20 minutes at the calibrated synthesis rate; the gate enforces a 4,300-word floor and measured wpm from a probe slice. The prediction in THE PREDICTION carries explicit falsifiers and should be scored in the December 26–28 editions.