Every question this desk holds itself to, from docs/questions-the-desk-must-answer.md. An answered one carries a value, the basis that produced it, the date it was taken and the observation that would overturn it. An unanswered one carries the kind of gap it is, because a refusal with no verdict is not a research position.
Is vol-of-vol elevated — is the VIX itself unstable, not just realized vol?
72.4 annualised %
annualised 20d close-to-close realized vol of VIX/USD daily candles; NOT VVIX, which this source does not carry
Is skew steepening or is call skew bid — which tail is being paid for?
38.61 vol pts
25-delta put IV minus 25-delta call IV, delta-matched, single expiry 30d; positive = puts bid
Is the AI complex crowded / consensus-long, on 13F, flows, and prime brokerage data?
29 names lit on more than one style
tools/confluence.py cross-style crowding over 165 names; short interest from data/projection/positioning.json (153 of 239 names carry si_pct_float). edgar_13f.json is corporate equity books, not institutional ownership, so it does not bear on this; prime brokerage is excluded by the desk's own joint-selection rule and by a recorded kill test
also answers Q-1F-2
Is dispersion low and correlation high — is everything trading as one factor?
0.1926 mean pairwise correlation
data/projection/returns_matrix.json, 241 names, min_overlap 250 sessions; stressed_rho is the same panel over the stress window declared in that file
also answers Q-1F-3
Is realized above or below implied — is vol cheap or rich versus what is actually happening?
0.221 percentile of trailing 1y SPY 20d vol
tools/market_regime.py, rule-based and untuned, appended daily to data/history/measurements.csv as spy_vol_pctl and vix
also answers Q-1B-2
Is the regime vol-suppressed or vol-expansionary?
0.221 percentile of trailing 1y SPY 20d vol
tools/market_regime.py, rule-based and untuned, appended daily to data/history/measurements.csv as spy_vol_pctl and vix
also answers Q-1B-1
Is it gap-prone — do moves happen between sessions rather than within them?
13.9 % of daily variance arriving overnight
var(log(open_t / close_t-1)) over var(log(close_t / close_t-1)), 4,999 session pairs from LSE daily candles for NVDA. The desk's own equity_closes.csv cannot produce this: it is close-only, so the session boundary is not in it
Multiple-driven or earnings-driven?
109 % of price move that is multiple
data/series.json weekly trailing P/E and price, 86 names with both legs over 52 weeks. Equal-weighted mean of LOG ratios, which add exactly, after dropping whole rows whose either leg exceeds a 10x move — trimming rows rather than legs is what keeps the identity closing
Macro-dominated or idiosyncratic — are single names moving on their own news or on the 10-year?
14 strongly-correlated pairs with no graph path
data/projection/returns_matrix.json rho and explained panels; explained is a 0/1 flag for a <=2-hop directed path, so unexplained means the graph offers no mechanism, not that none exists
How duration-sensitive is the complex?
0.0411 % index move per bp of 10y, conditional on the index
tools/macro_inbound.py fitted partial, index-conditional, n=651; the arithmetic duration table in tools/consistency.py and the UNDERPOWERED verdict from tools/duration_backtest.py are the two disagreeing readings and travel with this answer rather than being reconciled away
also answers Q-3E-3
Reserve-scarce or ample-reserve?
-1 bp (SOFR minus IORB)
pair:sofr-iorb:SOFR minus pair:sofr-iorb:IORB in basis points, both legs read bounded to a common anchor reference date by macro_bridge.py's PAIRS and recorded as a matched set, so the two legs cannot be as-of different days; the underlying series were adopted into the credit-duration channel on 2026-08-27 after sitting as unowned probe artifacts
also answers Q-3E-4
Are spreads compressed — credit spreads near multi-decade tights, meaning no cushion and full pass-through from Treasury yields to corporate borrowing costs?
267 bp (HY OAS)
FRED BAMLC0A0CM and BAMLH0A0HYM2, option-adjusted spreads in basis points — NOT the LSE .IBLUS credit indices, which are total-return levels with no reachable history endpoint
What is the risk-on/risk-off correlation regime — do stocks and bonds hedge each other or not?
-0.561 correlation of daily SPY return with daily 10y yield change
data/history/dgs10.csv against SPY in data/history/equity_closes.csv, 120 overlapping sessions ending 2026-08-21. The bond leg is a YIELD CHANGE in points, not a bond return; the sign convention is therefore inverted relative to a stock/bond return correlation and is stated rather than assumed
Is this a late-cycle capex boom?
79 % of multi-style names that dissolve on cycle adjustment
tools/confluence.py, which describes itself as a late-cycle phase measurement; marked = multi-style lit, dissolved = fails after the cycle adjustment
Shortage-priced or capacity-normalized?
3 scored layers still tight
scored layer states in data/regime.json; a layer is scored only where the ontology carries constraint coverage for it, and 9 of 14 are not
What bullwhip phase is the chain in — where does pricing power currently sit?
3 layers reading tight
layer states from data/regime.json (tools/regime.py); the sibling structure from tools/layers.py LAYER_DAG. regime.py records its own lag_q ordering as a STRUCTURAL PRIOR, untested, and says do not size on it
Are we pre-supply-response?
27 names at or above their own 80th-percentile capex intensity
supply_response.own_history_pct_rank in entity frontmatter, 100 companies, capex/revenue over matched annual windows from SEC filings ranked against each company's own history. Fiscal periods span 2021-2026 and are NOT aligned: 9 of 100 names report a period two or more years behind the newest close, so this is a cross-section over mixed vintages rather than a single-date reading
What financing stage is the buildout at — internal cash → debt → equity → pension/insurance?
13 companies outspending their own cash flow
tools/export_financing.py, capex_to_ocf on trailing-year figures from the same filing; 97 filers refused for missing ocf/capex and 2 stale filers excluded from the count, because a count of who outspends their cash flow reads as a statement about now. The $56.5bn shortfall sums ONLY the 10 rows the artifact marks `aggregatable` — converted at the period-average rate with both legs dated together — and excludes 0 that fail it rather than summing them silently. It is also SCOPED to the AI complex by each entity's own `exposed_to`, which drops american-airlines, boeing, southwest-airlines: cash-negative, but not this question's cohort. The price ladder is Note 10 of CoreWeave's 10-Q for the quarter ended 2026-06-30, read off one table.
Is breadth narrow — is there an advance-decline divergence?
45.6 % of desk universe above its own 50-day mean
regime:breadth in data/history/measurements.csv, appended by tools/regime_observables.py from data/history/equity_closes.csv; names with fewer than 50 closes are skipped rather than counted as below, because a short series is not a weak one. 691 prior reading(s), so no percentile is quoted yet.
Is the market factor-crowded — when value and small-cap indices both hold semis, has style diversification stopped working?
29 names lit on multiple styles
tools/confluence.py over 165 names on value, growth, quality, momentum, income, threshold 60% — a cross-style crowding gauge, not a flow measure
also answers Q-1A-6
Is this a single-factor market?
0.1926 mean pairwise correlation
data/projection/returns_matrix.json, 241 names, min_overlap 250 sessions; stressed_rho is the same panel over the stress window declared in that file
also answers Q-1A-7
Does the founder's one-sentence characterization hold, clause by clause, each against its own observable?
3/10 clauses holding
entities/regimes/ai-compute-market-regime.md, whose clause table names an observable and a reading per clause; the composite's own state is PARTIALLY_SUPPORTED
Is zHBM real, and how big is the claim?
HBM5 the baseline the headline claim lacks
tools/export_feasibility.py gate scoring; the desk reached this independently of the corpus, which records the same shrinkage — 8x an imaginary HBM5 at FMS became 2.3x a shipping HBM4E at Hot Chips three weeks later (data/projection/feasibility.json carries no `generated` key; this is the FILE MTIME, i.e. when it was last built, not the period its facts describe)
also answers Q-2C-1
Is hybrid bonding ready for 20-Hi?
0.676 20-high stack yield at 99% die / 99% bond
stack yield = Y_die^N * Y_bond^(N-1) from tools/fab_physics.py (Plummer, Deal & Griffin Ch 4); the readiness claim is SK hynix first-party as recorded in docs/questions-the-desk-must-answer.md Q-2A-2. TWO HALVES WITH DIFFERENT DATES: the arithmetic has none — it is an identity and does not age — while the readiness claim is dated by when the first-party statement entered the corpus. The as_of below is the CLAIM's, because that is the half that can go stale, and a fixture caught this returning no date at all
Where does the stack's thermal bottleneck sit?
0.55 lowest system_thermal gate probability
system_thermal gate across 4 scored technologies in tools/export_feasibility.py, defined as 'does the heat get out once it is in a real package'. data/projection/feasibility.json carries no `generated` key; this is the FILE MTIME, i.e. when it was last built, not the period its facts describe
Does zHBM survive the thermal physics?
HBM5 the baseline the headline claim lacks
tools/export_feasibility.py gate scoring; the desk reached this independently of the corpus, which records the same shrinkage — 8x an imaginary HBM5 at FMS became 2.3x a shipping HBM4E at Hot Chips three weeks later (data/projection/feasibility.json carries no `generated` key; this is the FILE MTIME, i.e. when it was last built, not the period its facts describe)
also answers Q-2A-1
What does any of it cost?
1 technologies whose binding gate IS cost
tools/export_feasibility.py gate scoring over 4 technologies; the yield_economics gate is defined as 'can you make N at a price, not can you make one'. data/projection/feasibility.json carries no `generated` key; this is the FILE MTIME, i.e. when it was last built, not the period its facts describe
AI capex is heading toward $4T. Can the capital markets keep up?
13 companies outspending their own cash flow
tools/export_financing.py, capex_to_ocf on trailing-year figures from the same filing; 97 filers refused for missing ocf/capex and 2 stale filers excluded from the count, because a count of who outspends their cash flow reads as a statement about now. The $56.5bn shortfall sums ONLY the 10 rows the artifact marks `aggregatable` — converted at the period-average rate with both legs dated together — and excludes 0 that fail it rather than summing them silently. It is also SCOPED to the AI complex by each entity's own `exposed_to`, which drops american-airlines, boeing, southwest-airlines: cash-negative, but not this question's cohort. The price ladder is Note 10 of CoreWeave's 10-Q for the quarter ended 2026-06-30, read off one table.
Why is Nvidia pulling pension and insurance money into AI, and what does that say about `Q-1E-5`'s financing stage?
13 companies outspending their own cash flow
tools/export_financing.py, capex_to_ocf on trailing-year figures from the same filing; 97 filers refused for missing ocf/capex and 2 stale filers excluded from the count, because a count of who outspends their cash flow reads as a statement about now. The $56.5bn shortfall sums ONLY the 10 rows the artifact marks `aggregatable` — converted at the period-average rate with both legs dated together — and excludes 0 that fail it rather than summing them silently. It is also SCOPED to the AI complex by each entity's own `exposed_to`, which drops american-airlines, boeing, southwest-airlines: cash-negative, but not this question's cohort. The price ladder is Note 10 of CoreWeave's 10-Q for the quarter ended 2026-06-30, read off one table.
If AI cracks, does the US economy crack with it?
5.84 x — share of market-cap gains divided by share of weight, same window
tools/gain_concentration.py over data/history/equity_closes.csv and the market caps on entities.jsonl, window 2026-05. Caps are BACK-CAST from today at constant share count, and the weight is taken AT THE START of the window, never today — read off today's caps the pair is 4.7% of the universe and the ratio is 4.3x, because the very move being measured is what lifted the weight. 234 of 241 priced names carried both a cap and a price. The denominator is this desk's universe and NOT MSCI ACWI. The household-equity share is FRED BOGZ1FL153064486Q as-of 2026-01-01 via macro_bridge.py, whose channel reports INCOMPLETE.
Ben Thompson's case: can AI keep winning even if the bubble bursts?
11 nodes at loop_gain >= 0.5
tools/capital_formation.py loop_gain over 15 nodes; the asymmetry is stated in the artifact's own doubly_exposed block (data/capital_formation.json carries no `generated` key; this is the FILE MTIME, i.e. when it was last built, not the period its facts describe). D1 applies: this is descriptive of a mechanism and is not a dated forecast
Why do rising Treasury yields hit AI tech stocks?
0.0411 % index move per bp of 10y, conditional on the index
tools/macro_inbound.py fitted partial, index-conditional, n=651; the arithmetic duration table in tools/consistency.py and the UNDERPOWERED verdict from tools/duration_backtest.py are the two disagreeing readings and travel with this answer rather than being reconciled away
also answers Q-1C-4
Is RRP exhaustion the reserve-scarcity trigger?
-1 bp (SOFR minus IORB)
pair:sofr-iorb:SOFR minus pair:sofr-iorb:IORB in basis points, both legs read bounded to a common anchor reference date by macro_bridge.py's PAIRS and recorded as a matched set, so the two legs cannot be as-of different days; the underlying series were adopted into the credit-duration channel on 2026-08-27 after sitting as unowned probe artifacts
also answers Q-1D-1
every one of these 9 carries the same basis, verbatim: the coverage audit resolved this against premises the model already holds; no resolver is written for it in REGISTRY yet
Is it fragile — low realized vol with high tail risk?
Discount-rate-driven or growth-driven?
Is this a rate-of-change regime — does the market trade the second derivative rather than the level?
Is it duration-supply-heavy — how much long-end paper must the market absorb?
What holds until then?
Why can Samsung attempt memory-on-logic?
Which product gets hybrid bonding first?
What are SK hynix's HBM5 and custom-HBM plans?
Can the serving stacks schedule against a pooled CXL tier with predictable tail latency?
What does the EMIB-T roadmap change about who can build 2.5D at scale?
What are the real HBM4 packaging challenges, and which of them are shared across all three memory vendors rather than vendor-specific?
Is microfluidic cooling on a product path or a conference path?
Where do photonic interconnects actually bind — bandwidth, energy per bit, or reach — and at what date?
Does custom HBM change the memory vendors' margin structure, or only their customer concentration?
Does the grid support the buildout, or does 40GW+ of behind-the-meter datacenter by 2028 become the base case — 50%+ of new DCs per year?
Behind-the-meter is a workaround, not a relief. What does the model's relocates it?
Is CXMT set to challenge the DRAM incumbents — and on what axis: node deficit, wafer adds, China HBM, or the LTA structure?
What does the CXMT IPO disclose that was previously unobservable?
Does China's property bust into manufacturing superpower — the "second China Shock" reaching the frontier rather than the low end — change the indigenization thesis?
Inference is becoming heterogeneous while the dominant abstraction is still tokens-in/tokens-out, one model, one hardware backend. A single multimodal request may hit a…
If the token is the wrong unit, what is the right one, and what would the desk have to measure to price it?
every one of these 12 carries the same basis, verbatim: audited: the model could hold this premise and does not. Nothing external is required — an entity, edge or signal is.
Is hedging charm-heavy — decay-driven, concentrated into expiries and quarter-ends?
Narrative-driven or flow-driven?
Is "good news is bad news" live — does strong data raise hike odds and hurt equities?
Is the market collateral-constrained?
When does CXL computational memory ship?
Does PIM need system redesign?
Can the three thermal claims be put on a common metric?
Does the claim-shrinkage prior generalise?
Vendors present the axis they are winning on. Samsung talks base die because it has a logic node; SK hynix talks packaging because it has MR-MUF and 48% share. Does this…
Is JEDEC thickness revision the leading indicator for when hybrid bonding becomes mandatory?
Does the drop-in-socket test predict adoption better than performance does?
Can buybacks actually stop long-term yields from rising, or only improve liquidity in the securities bought?
Is dealer positioning short or long vanna — how does dealer delta respond to a change in implied vol rather than spot?
option open interest by strike plus a dealer sign convention. OCC daily Volume & Open Interest is free and public and unwired; the IV and greek half already arrives from lse_macro.chain(), and vanna is closed-form off those fields.
Is spot above or below the gamma flip level?
aggregate per-strike open interest on a reference underlying. lse_macro.gamma_flow() ships a volume-weighted FLOW proxy and explicitly refuses the gamma-flip level rather than faking it.
How passive-dominated is the float — what share of shares are price-insensitive, such that marginal fundamental news moves price less than it should?
index-provider constituent weights or an ETF-holdings feed. The denominator is held — float_shares on 237 of 239 names — so 1 minus float/outstanding gives the CLOSELY-HELD share, which is a different quantity and must be labelled as one.
Does the projected share change hold — SK hynix 48% of HBM bit shipments this year, then Samsung 41% vs SK hynix 39% next year?
supplier-level HBM bit-share series. TrendForce and Omdia are both paid; the free TrendForce press centre was checked back through April 2026 and the split is not there. HBM share of DRAM wafer input and of DRAM bit supply ARE sourced and are different quantities.
How short will Bessent go on $32T of Treasuries?
Treasury FiscalData maturity composition and buyback operation results. Public and keyless, and unwired — which makes it unacquired data rather than a modelling gap. The forecast half of the question ('how short will he go') is barred by the reflexive-layer rule and must be a scenario tree against what is priced.
generated by tools/export_answers_page.py from site/data/answers.json · built 2026-08-28 · nothing on this page is hand-written