One hundred and eighty-seven ideas tested.
What each one failed at.

An earlier test (R-126) put Foundry at a fifth of the account beside ULTRA and found about the same money with a worst fall about two points smaller, after charging back an estimate of what rebuilding young funds from their holdings today flatters. On that result it was offered, beside ULTRA only. This test (R-136, written down before it ran) replayed the shipped package day by day from July 2007: whole shares, commissions, cash that has to settle before it is spent, splits, dividends and the basket’s own pauses. Most of the funds are under two years old, so each traded its real prices from its first day and the rebuilt history before that, charged the same estimate, with synthetic splits keeping each price between $12 and $60 so whole shares cost what they would on a real fund. The pass mark had two parts, both needed at both sizes in every stretch: a smaller worst fall than the engine alone, and more growth than simply keeping a fifth in cash at the same worst fall. It beat cash in most stretches, by +0.49% a year over the whole stretch from $10,000, but it never made the worst fall smaller, so it fails. A harsher charge for the rebuild changed nothing that matters. What it teaches: a result measured on a rebuilt history and a simplified account can vanish when the real machinery trades it, the same lesson Rotation taught a day earlier; the replay is now the gate for anything offered.

An outside reviewer proposed this basket, and we froze it before any data was loaded (R-133, commit 838e5a61): SOXX, VGT, XLI and XLV at 25% each, no exit, at 10% of the account beside ULTRA, run as a buyer’s own basket by the real package, day by day from July 2007, with whole shares, commissions, cash that has to settle before it is spent, splits and dividends. One declared change: the reviewer asked for a quarterly rebalance, and the product checks monthly and trades only when a holding drifts more than a quarter from its target, so it was tested the way a buyer would run it. It was compared with the engine alone, the engine on 90% with 10% in cash earning the Treasury bill rate, and the engine on 90% with 10% in two total US market funds. The pass mark: beat both 90% versions at the same worst fall, from 2007 to 2026 and in each half, from $10,000 and from $25,000. It passed every stretch from $25,000 and failed from $10,000 in 2007 to 2012. Reported without letting it decide: the engine itself makes slightly different calls at different account sizes, because whole shares and its rebalancing thresholds land differently, and in 2015 and 2016 that alone moved the worst fall by about 1.6 points between two books with the same rule; the basket’s edges were of the same size. Two assumptions, stated: dividends were paid on their documented dates where known and 45 days after the ex-date otherwise, and cash earned the bill rate at no cost, which a small real account does not. This retires this basket under these rules; it does not show that every sector basket must fail.

We wrote the rules and the pass mark down before any data was loaded, so nothing could be tuned afterwards (R-112, commit 2907b5ce). One basket was fixed in advance, with no search: SMH 40, XLV 20, XLE 20, XLP 20, rebalanced monthly, each holding stepping aside while its month-end is at or below its 10-month average, exactly as the Foundry preset runs. Only the real funds were used. The health, energy and staples funds have prices from January 1999; the chip fund lists in June 2000, so the test starts in July 2000. The pass mark, the same as the Foundry’s own history test (grave 177): the engine with 30% in the basket had to earn more a year than the engine turned down to the same worst fall, the rest in Treasury bills, from mid-2000 to 2026 and in each half, 2000 to 2012 and 2013 to 2026. It won over the whole stretch (13.78% a year against 13.18% for 78% of the engine) and from 2013 (20.18% against 17.83%), and lost from 2000 to 2012 (6.73% against 8.18% for 83% of the engine). One lost half is a fail. We said before the run that we expected a fail: a sector all-weather book with health, staples, utilities and energy had already failed beside this engine (grave 156). Reported without letting them decide: in the tech bust, July 2000 to October 2002, the engine alone lost 1.7%, the engine with Four Corners 13.9%, and Four Corners on its own 38.0%. In the 2008 crash the mix lost 14.9% against 15.9% for the engine alone, and in 2022 14.0% against 19.5%. Equal quarters in each fund also lost from 2000 to 2012, by 0.92% a year; with the exits turned off the basket was behind over the whole stretch as well, by 1.32% a year.

We wrote the rules and the pass mark down before any data was loaded, so nothing could be tuned afterwards (R-110, commit 6b735ec1). The basket is the Foundry exactly as set: SMH 35, GRID 25, GOOGL 15, ANET 10, URA 5, SPCX 4, QTUM 3, BOTZ 3, rebalanced monthly, each holding stepping aside when it closes a month at or below its 10-month average. Until each real fund has a price, an older one stands in: the Philadelphia chip index for SMH (price only, so its dividends are missing), half utilities and half industrials for GRID, the Nasdaq-100 for Alphabet before 2004 and for the space, quantum and robotics funds, Cisco for Arista before 2014, and Cameco for URA before 2010. The stand-ins were picked by someone who knew how 2000 to 2002 went; we say so rather than hide it. The pass mark: the engine with 30% in the Foundry had to earn more a year than the engine turned down to the same worst fall (89% of the engine, the rest in Treasury bills) over 2000 to 2026 and in each half, 2000 to 2012 and 2013 to 2026. It won over the whole stretch (14.49% a year against 14.19%) and from 2013 (21.05% against 20.01%), and lost from 2000 to 2012 (7.49% against 7.86%). One lost half is a fail. The engine alone beat both: 15.55% a year with a worst fall of −27.6%, against the mix’s −24.9%. Reported without letting them decide: the Foundry’s own exits are what kept it survivable. In the tech bust the Foundry alone lost 29.8% with its exits and 61.2% without them. Without the exits it would have made more over the whole stretch (13.34% a year against 11.30%), but fallen 72.0% at worst against 37.3%. In the 2008 crash the mix lost 15.4% against 15.9% for the engine alone, and in 2022 18.7% against 19.5%, so it cushioned those two falls a little and deepened the tech bust. Smaller slices were measured too and are listed below; choosing a size after seeing them would be tuning, so none of them counts. This test changes nothing live: the Foundry’s sealed forward watch from 2026-09-24 keeps running.

We wrote the rules and the pass mark down before the test ran, so nothing could be tuned afterwards (R-104, commit f21233f4; amended before any return was computed, 602e971b, to price KRUZ from Tiingo, because Yahoo has no history for it). Both funds listed in February 2023, so the window is short: March 2023 to August 2026, in two stretches, to the end of 2024 and since 2025. Each fund was held with no exit, 20% beside the engine, re-sized monthly. To pass, a seat had to earn more a year than the engine turned down to the same worst fall (the rest in Treasury bills) over the whole window and both stretches, and the fund alone had to beat the S&P 500 over the whole window. A pass would only have earned a forward watch. NANC beat the turned-down engine over the whole window (32.69% a year against 32.40% for 93% of the engine) and to the end of 2024 (0.20% ahead), then trailed since 2025, 21.56% against 21.77% for 99% of the engine. KRUZ’s last close was March 20, 2025, so its record and the mix’s stop there, and any figure for them that runs past it would compare a shortened record with the engine’s full one. We quote them only to the end of 2024. There, KRUZ beside the engine made 40.74% a year against 41.10% for 86% of the engine, and the mix 42.36% against 42.44%. On its own KRUZ made 15.48% a year in that stretch, while the Nasdaq-100 made 36.80%.

We wrote the rules and the pass mark down before the test ran, so nothing could be tuned afterwards (R-103, commit 2172c607). The engine is the same full-history harness as every other seat test; each coin is held only while its close on the first session of the month is above its 200-day average, otherwise the seat earns the Treasury-bill rate; the book is re-sized monthly and each switch costs 0.10%. The window starts in November 2018, the first month with a full year of ether prices for the average, and runs to August 2026, in two stretches: to December 2021, and since 2022. To pass, ether had to do two things in the whole window and in both stretches: earn more a year than the engine turned down to the same worst fall (the rest in Treasury bills), and beat the bitcoin seat on return divided by worst fall. In bitcoin’s place it did both in the first stretch (1.759 against 1.398), then failed both since 2022: 0.619 against 0.779, and 15.57% a year against 16.92% for 91% of the engine alone. Over the whole window it came to 0.970 against bitcoin’s 0.979. Sharing the seat 10/10 tied bitcoin over the whole window (0.979 against 0.979), and a tie is not a pass; since 2022 it trailed, 0.702 against 0.779. The bitcoin seat itself is unchanged by this test.

Two tests, each with its rules and pass mark written down before it ran, so nothing could be tuned afterwards (R-099, commit dd0a0915; R-101, commit ff399e80). The first used the 503 companies in the S&P 500 today and their reports from 2003: buy at the close of the first session that could react to the report, sell 60 sessions later, 0.10% in costs. It passed every registered condition: +2.39% a trade over 8,172 trades, a t of 6.2 against a bar of 3 (t measures how far a result stands above its own noise; each quarter counts once so overlapping trades cannot inflate it), +3.15% in the first stretch (2003–2014) and +1.60% in the second (2015–2026), and +1.74% since 2021. It said at the time that a list of today’s survivors flatters any buy signal, and one number showed it: even the fifth of reports with the worst surprises beat the S&P 500, by 0.91%. The second test ran two registered checks on the same trades from 2010. Counting only companies that were members on the day (the list rebuilt from the index’s changes table) cut the gain to +0.46% (t 1.3). Judging each stock against its own average quarter, which removes the lift a survivor gets every quarter, cut it to +0.58% (t 1.5). With both, the pass mark was a t above 2.5 and a gain in each stretch; it made +0.04% (t 0.1), +0.37% in the first stretch (2010–2017) and −0.27% in the second (2018–2026). The registration said a fail here ends the idea, so the planned test of a real ten-stock sleeve beside the engine was not run. Data: Tiingo, which prices companies that have left the index, has earnings data only for the 30 Dow stocks, so a full rebuild from the companies’ own SEC filings was left for a later test; we did not run it, because the checks we could run already brought the result to zero.

We wrote the rules and the pass mark down before the test ran, so nothing could be tuned afterwards (R-100, commit a1263e16). The basket: 25% Nasdaq-100, 15% US small value, 20% developed markets, 10% international small caps, 20% emerging markets, 10% momentum, rebalanced monthly, each holding stepping aside when it closes a month below its 10-month average. The funds we would actually buy are young, so the test used older funds that track the same things (QQQ, IJS, EFA, SCZ, EEM, MTUM), from May 2003, when the emerging-markets fund has its first full month. Pass mark: the engine with 30% in the Globe had to earn more a year than the engine turned down to the same worst fall, over the whole window and in each stretch of it, 2003–2014 and 2015–2026. It trailed by 0.35% a year in the first stretch (against 90% engine) and 0.03% in the second (against 81%). Two things we reported without letting them decide. On its own since July 2008 the Globe was not a more aggressive world fund but a calmer one: 7.8% a year with a worst fall of −21.0%, against 9.0% and −49.8% for the whole-world fund VT. The exits bought the calm and cost the return. And the real funds, which exist only since October 2019, were too young for their 10-month exits to work in the 2020 crash; they rode it down, a worst fall of −32.4% against −17.3% for the stand-ins. A basket of stocks falls with the engine in a crash, so its only chance was a stretch when markets abroad lead; the first stretch (2003 to 2014) holds such years, and over that whole stretch it still trailed. Same verdict as the chips-and-energy pair beside the engine (grave 169).

We wrote the rules and the pass mark down before the test ran, so nothing could be tuned afterwards (R-098, commit bb2e6f36). It grew out of an earlier test that is not a stone of its own. That one (R-097, commit e2f57a7a) asked whether new members keep rising after the day they join. It was void by its own rule: we could price only 79.3% of the 328 additions against a bar of 80%. What it did show was the wrong way round: −1.16% against the S&P 500 twenty trading days after joining. Its rough five sessions before the joining day showed +1.07%, and this test asked whether an ordinary account can catch that. The announcement date is the one printed in S&P’s own press release; announcements come after the close, so the buy is the next close. 208 of 328 additions (63%, bar 60%) had a usable date, prices and at least one session between buy and sell. Costs: 0.10% a round trip, and a harsh 0.30%. Pass mark: an average gain after costs across all trades, strong enough that luck is an unlikely explanation, plus a gain in each of the two stretches of the window and at the harsh cost. The first stretch, additions from 2010–2017, lost 1.27% a trade, and so consistently that chance cannot explain it. The second stretch, 2018–2026, made +0.12%, too little to tell from nothing. An older stone (grave 46) found the classic index-inclusion pop has shrunk from its textbook size; this is the part of it a daily account could still reach, and it is gone.

We wrote the rules and the pass mark down before the test ran, so nothing could be tuned afterwards (R-096, commit fa55a3a6). The fund is QLD, the two-times Nasdaq fund the engine holds, on daily prices from 2007 to August 2026. The exit is the engine’s, simplified: be in only in months when the previous month-end Nasdaq close was above its 10-month average. Costs: 0.01% of slippage on each side of each trade at the closing and opening auctions, plus the broker’s per-share commission, at a $25,000 account. Pass mark: return divided by the worst fall had to beat simply holding the fund under the same exit, over the whole window and in each ten-year stretch of it, 2007–2016 and 2017–2026. It passed in the first stretch only, where it made 7.2% a year against 11.2% but fell at worst −30.0% against −51.6%. In the second stretch it made 17.1% against 29.5% with the same worst fall. The overnight effect itself is real: without the exit, overnight-only cut the fund’s worst fall from −83.1% to −52.3%. The exit does that job already (−51.7%) and keeps the daytime return, so the two cannot be stacked. Taxes were not counted; in a taxable account every overnight gain would be short-term. We buried the same idea on index futures earlier (grave 104); on the fund it fails for the same reason.

Registered before each run (R-087, commit 25cb0bae; R-088, commit 8709270d). R-087 held each fund as the entire $32,000 Roth, bought once, payouts reinvested, and asked for either more money than its index or a clearly calmer ride: a better return for the risk and a worst fall at least 5 percentage points shallower (say −30% instead of −35%), across the whole window and both halves of it. No fund met either test in any window. 5 were too young to judge and were reported without a verdict: JEPI, SPYI, and the new memory and neocloud funds DRAM, DISK and NCLD. R-088 answered the objection that young funds cannot be tested: it tested the pattern instead of the fund, the same monthly trend rule the engine uses, on 21 themes, winners and flops, each against QQQ bought on the same day. A second version stepped in only when the theme was also beating QQQ; it did worse, a median 13% a year behind. The result held in the older themes, the younger ones, and with the three best removed. The same rule applied to QQQ itself is what works, and it is what the engine already does. Looking through to single companies, the rule did beat holding Micron and the chip index; that is one lead, picked by us, and it is being watched forward, not claimed.

Registered before the run (R-086, commit 23b82616). The watchlist came from 336 combinations tried on one year, September 2025 to September 2026. CHPY has no other history, so only a forward test can judge it; that test is running. Its plain versions have fifteen to twenty-six years, and this used them. The shape is the watchlist’s own: 5% and 5%, carved from ULTRA’s stock side, held all the time, at $25,000. To pass, a pair had to beat the same engine with that 10% in bills and with it in gold, by a margin fixed in advance, with no deeper worst fall than the gold version, in both halves of the window, and with 2008, 2020 and 2022 each left out in turn. None came close on the first rule. The two nearest, chips with utilities and chips with oil majors, added about a point and a quarter a year, and matched it with extra swing. Switching the pair in and out on its own trend, the way every earlier sleeve was tried, did worse. Using SOXX in place of SMH before 2011 did not change a verdict. At $5,000 the pair cannot even be built: one share of SMH costs more than the whole 10% sleeve. It joins the chips (16, 121), energy (154) and the sector book (156). What is left of the income idea is one forward question, whether the option wrapper beats its own plain twin, and history says it will not.

Registered twice, and the first registration was the mistake. R-070 asked a borrowed-money arm to beat a capital-funded one with no deeper worst fall. Borrowing adds exposure without removing any, so it was deeper in all twenty-four cells and no result could have passed; it also compared the two on raw return, where leverage wins by definition. That gate was written before the run and its verdict stands, but a gate nothing can pass is not a test, so the question was registered again properly as R-071 before it was run again. The second time every arm was laddered across eight dial settings so the comparison is one whole risk-return curve against another, and the places they are compared are the quarter, half and three-quarter points of the range where the two curves overlap — fixed by the data, not chosen after the fact. The overlay itself is gold notional equal to a tenth or a fifth of the account, switched on and off by the same monthly trend rule the shipped gold seat uses, financed at the three-month Treasury rate plus a third of a point, carried in its own ledger so it counts toward the account’s value and its sizing but never as cash the engine can spend on shares. Borrowed gold beats bought gold at the deep end of the range in eighteen of twenty-four cells, by two thirds of a point a year on average; in the middle it is a wash, sixteen of twenty-four and eight hundredths of a point; at the shallow end it loses in nineteen of twenty-four, by four tenths. Doubling the borrowing cost to three quarters of a point does not rescue any cell. One group of cells is honestly inconsistent and is recorded as noise rather than averaged in quietly: on ULTRA since 2009 the middle point is negative in all four cells while the shallow point is positive in all four, the opposite of every other dial, and ULTRA’s worst fall is already known to move about four points on a two per cent change in position size. Neither study ran a bitcoin version. The gold arms failed, and reaching for a second asset after a failure is how a result gets manufactured. What the record now says: the funding trick is real and it is not free, and for this book the interest is the larger bill everywhere except the hottest settings. The gold seat stays bought with capital.

Registered as R-066 with the pass mark written first: at least a quarter point a year on both the modelled full window and the real funds, at both sizes, with no deeper worst fall (half a point of tolerance). The first run passed on ULTRA (+1.47 / +1.52 modelled, +0.84 / +1.05 real, falls -28.3% to -26.7%) and failed on SELECT and STEADY; R-068 confirmed the S&P leg gains nothing worth the fall either. R-067 attacked it: positive at equal fall in all three eras (+0.3, +2.1, +2.8), 18 of 27 years, +1.4 after tax, one to four points deeper in a crash session, +0.9 at $3,000. The Lab replayed it independently, 56 simulations, and reproduced every number. Then the Lab asked for the account as it is run, not as the harness idealises it. With the executor’s 98% sizing reserve the real-fund edge at $25,000 fell to +0.70 with the fall 0.9 points deeper, and at $5,000 to +0.07: whole shares and a 2% smaller slice move both arms enough that the difference is noise at small sizes. The Lab then found that the simulator pays parked cash a flat 3% whatever it is told, and rebuilt that line to accrue the historical three-month bill rate on settled and parked balances (eight accounting tests, sixteen exact legacy comparisons). Under that model the edge at $25,000 is +0.47 points (23.3% against 23.8%) with the fall 0.88 points deeper (-26.5% to -27.4%), and at $5,000 +0.13 with 0.78 deeper. The return clears the quarter point at $25,000 and not at $5,000; the fall clears the tolerance at neither. What the record shows: the three-times fund’s advantage is mostly the freed cash, and the freed cash is worth what cash earns; price it at the historical rate, size it as the executor sizes it, and hold it in whole shares at the sizes people actually run, and about a quarter to half a point a year is left, bought with a point of extra fall and a fund that a single 33% session would end. Two things worth more than the idea came out of it: the published record’s cash accrual is a flat 3% that the site now needs to reconcile in a controlled update, and ULTRA’s worst fall moves by up to four points on a 2% sizing change, a path sensitivity the record should state. The levered leg stays QLD on every dial.

R-063 ran the gold seat’s plateau check on the bitcoin seat and found the short-and-fast corner of the grid beating the shipped rule on both windows at both sizes: the 100-day average read every session, +0.06 to +0.15 Sharpe beside the engine. Three objections were written down before anything moved: it was the corner of the grid, it was picked and judged on the same data, and the seat stream charged no trading cost. R-065 answered all three, registered before the run: lengths 50 to 200 days, three cadences, the seat traded by the harness itself (whole units, a $1 commission per fill, next-close fills, T+1 settled cash, the engine leg untouched), the candidate chosen on 2015 to 2019 by its mean lead across the three dials and judged on 2020 to 2026 at $25,000 and $250,000, pass mark +0.03 at both sizes on ULTRA and SELECT with the worst fall no deeper by more than two points. Training chose the same cell R-063 had, 100/D, by a wide margin: +0.252 on ULTRA, +0.235 on SELECT, +0.259 on STEADY; every daily-read cell led by 0.15 or more, every weekly one by about 0.10, so the training years were unanimous. The test years were not: ULTRA -0.028 at $25,000 and +0.004 at $250,000, SELECT +0.030 and +0.012, STEADY +0.017 and +0.007; the book earned 28.8% a year against 30.6% on ULTRA with a worst fall 2.1 points shallower, and the seat traded 12 times a year against 5. What the training years had that the test years did not: 2018, when bitcoin fell 80% over eleven months and a faster read left two months earlier than a monthly one. 2022 was the same kind of fall and the faster read gained a little there too, but 2020 to 2026 also held March 2020 and a run of whipsaws in 2021 and 2024 where a daily read stepped out and back in at a cost, and the two effects cancel. The shallower worst fall is real and small, the extra return is not, and the extra trades are certain. The monthly rule stays, the bitcoin gate stays on its clock, and the corner cell of R-063 is recorded as what it was: a fit to one crash.

Mel asked whether the VIX playbook had been tested. The option hedge (graves 84, 118), every VIX signal (43, 85, 106, 129, 130, 131) and the short side of the volatility funds (102) already had stones; the long side as a small permanent seat did not. Registered as R-061: VIXY (short-term futures), VIXM (mid-term) and VXX, from 2011 and since 2018, four ways in the seat: the fund as the whole seat, a quarter of the seat (5% of the book, the size a hedge would be), that 5% only while the engine is in, and that 5% only while the fund is above its own 200-day; against cash, gold and gold-while-out; pass = +0.03 Sharpe at both sizes with no deeper fall. The expectation was written first: the short-term fund’s roll cost makes it a certain loser, the mid-term fund is the one cell with a chance. Alone: VIXY -48.6% a year and a -100% worst fall, VIXM -17.6% and -96%; correlation with the Nasdaq fund -0.70; inside the 2022-23 stretch the engine sat out, the funds lost -49% and -23% while gold made +7%: a slow bear is not a volatility event. Full window, best construction per dial against gold at $25,000, VIXM: ULTRA -0.004, SELECT -0.040, STEADY -0.011; VIXY worse on every line. Since 2018 the mid-term fund as the whole seat passes SELECT and STEADY at both sizes (book 15.2% a year, -16% worst fall on SELECT) and the 5% short-term seat held while the engine is in passes STEADY. Why that is the trap, not the find: the validation window must contain the failure mode, and a long-volatility seat’s failure mode is a long calm bull market with nothing to insure. 2011 to 2017 is exactly that, and the mid-term fund lost -92% across it, about -32% a year of the seat, before 2018 gave it February, December and then the 2020 crash (+76%). The since-2018 pass is one payoff with its premium cut off. The engine’s exit insures the same event for the cost of being a month late, and gold pays in the slow bears where volatility funds bleed.

Mel asked for reasoning rather than another fund, and the reasoning gave a mechanism: emerging-market earnings, funding and flows are priced in dollars, so their good decades (2003 to 2007, and 2017) sit inside dollar downtrends and their bad ones (2011 to 2015, 2018, and 2021 to 2022) inside dollar uptrends. Grave 150 had already killed emerging markets held outright or on their own trend; this study switched the seat on the dollar’s trend instead: the dollar index’s last close against its 200-day average, read at the month turn, hold EM while the dollar is below, cash while above. Registered first, with a mechanism check written in: if the split by regime is not there, the seat result is moot. The split is there. EEM 11.1% a year and Sharpe 0.58 below the average against 4.5% and 0.30 above; VWO the same. But so is everything else: the S&P 500 12.8% against 8.7%, the Nasdaq-100 19.3% against 10.2%, at higher Sharpe in both regimes. The dollar falls when the world is buying risk, and the engine’s own root is the highest-return risk asset in the set, so the switch adds a slower, lower-return copy of a bet the book already holds. The seat streams alone: EM always 7.8% a year with a -66% worst fall; dollar-gated 5.8% and -42%; both gates 5.4% and -31%; the gate cuts the fall and the return together. Beside the engine the best construction sits -0.099 (ULTRA), -0.093 (SELECT), -0.086 (STEADY) Sharpe below the gold sleeve at $25,000 on the full window, and no better at $250,000 or since 2009; the only cells above cash are the while-out ones since 2009, which is the gold-while-out seat with a worse asset. Uranium and the nuclear theme in the post that prompted this were not retested: URA is inside grave 150, and a supply-timing narrative with an 80% drawdown in its own history (2011) is not a rule the engine can read.

Registered as R-059 part B before the run: the gold seat’s gate and cap as shipped (last close against the 200-day average at the month turn, vol-sized, capped at 1.0) with the holding swapped for GDX (from 2006-05-22), SIL (2010-04-20), GDXJ (2009-11-11) and UGL (2008-12-03), read on gold’s signal and read on the vehicle’s own closes, v2 sleeve machinery, three dials, $25,000 and $250,000, each from its listing, against the shipped gold seat over the same window; pass = +0.03 Sharpe at both sizes with no deeper fall. Alone: GDX 5.6% a year with a -81% worst fall, SIL 2.9% / -83%, GDXJ 1.8% / -89%, UGL 10.4% / -76%; correlation with gold +0.76 to +1.00. In the seat the best read per vehicle sits below the shipped seat on ULTRA by -0.074 (GDX), -0.016 (SIL), -0.035 (GDXJ) and -0.028 (UGL) Sharpe at $25,000, with SELECT and STEADY worse for the miners. Why the analogy with QLD fails: the two-times Nasdaq fund is the index with financing; miners are companies with costs, debt and an equity beta, so they fall with stocks in the stretches the seat exists to cover, and the gate, which reads gold, cannot see that. The two-times gold fund is the honest analogue and it lost anyway: the seat is capped at 1.0 and vol-sized, so doubling the holding’s volatility halves its size and adds the fund’s reset drag and expense for nothing. The gold seat stays SGOL.

The two-times fund carries its own financing: fed funds plus about 0.35% of swap spread plus a 0.95% expense ratio, 1.3% over fed funds per borrowed dollar. R-033 (grave 141) showed a margin loan at the broker’s 1.5% spread costs more, not less. A box spread on cash-settled index options is a synthetic loan priced by arbitrage; the published record puts its rate about 0.35% over the matching Treasury rate on average. Holding the plain fund on box financing therefore costs fed funds plus 0.35% plus twice the plain fund’s 0.20% expense: 0.75% over fed funds, a saving of 55 basis points per borrowed dollar. We registered the test as R-033’s harness with the cost model swapped, three spreads (0.20%, 0.35%, 0.80%), SELECT and ULTRA at $25,000 and $250,000, 1999 to 2026, and the gate: at least a quarter point a year on both dials at both sizes at the baseline spread, surviving the 0.80% stress at one size. Result at 0.35%: SELECT +0.03 / +0.02, ULTRA +0.51 / +0.27 points; at 0.80% ULTRA keeps +0.27 at $25,000 and +0.05 at $250,000; worst falls unchanged. The mechanism is exactly what the cost gap predicts: the saving scales with how much is borrowed, ULTRA (cap 2.0) borrows about twice what SELECT does, and fed funds averaged 2.14% over the window so the whole financing line was small. Real but small, and the registered gate wanted both dials. Not modelled, as in R-033: the path difference between a daily-reset fund and a loan (the band keeps the two close), one-box granularity ($10,000 of borrowing per SPX box, a fifth of a $25,000 ULTRA account’s exposure), quarterly rolls, and the options permission and margin account a box needs. The practical verdict: the futures engine from $25,000 already borrows at futures rates, about 0.2% over fed funds, with one contract and no options; a box spread is a harder way to the same place.

The Donchian quant book (“DON55/20”) ran as a report-only shadow on our box and had already halted itself by its own floor (51 closed trades, -0.16R). Its rule, rebuilt from the published definition: signal when the close is at or above the prior 55-day high, fill at the next open, stop at two ATR below entry, exit on the first close at or below the prior 20-day low once twenty bars have been held, at the stop, or at the close of the thirtieth bar; five positions at most, strongest breakout first; 1.5 bp a side. Registered with the allocator in R-058 before the run. As a five-slot seat from 2005: 7.4% a year, worst fall -24%, Sharpe 0.69, 50 trades a year, 41.7% wins, correlation +0.35 with the Nasdaq fund; -2.8% in 2008, -11.3% in 2022. Inside the engine’s out-stretches it made +7.0% in 2008-09 and lost -5.6% in 2022-23 while gold made +7.1%. Beside the engine the best construction per dial sits -0.070 (ULTRA), -0.080 (SELECT), -0.074 (STEADY) below the gold sleeve at $25,000 and no better at $250,000; it beats cash on SELECT by a hair and nowhere else. The mechanism is the one the graveyard already knows from the exit vote and the shorter averages (graves 142 to 145, R-057): a second, faster trend rule on the same market is the engine’s own signal with more noise. Where it holds something the engine does not (bonds, gold, oil, foreign funds) it is the R-053 sector book again, and those pieces did not pay in the stretch that mattered.

The ETF cross-sectional mean-reversion allocator ran as a report-only shadow on our box from July 2026 (29 forward trades, +0.24R, 72% wins) after a July study of its own found it failed four of five validation gates and lost money at a $5,600 account on the broker’s fixed commission. Mel asked whether anything in the retired shadows could be a seat, then said test all. We wrote the rule and the pass mark down first. The rule was rebuilt from its published definition (not the retired code): z-scores across the universe of the 3- and 5-day returns and the distance to the 5- and 10-day averages in volatility units, plus a two-day RSI term; top two above a threshold, one new position a day, five at most, next-open fills, a 1.5 ATR stop, exit on the first close above the 5-day average or the fifth day; 1.5 bp a side. The rebuild put the shadow’s own pick in its top three on 27 of 29 forward days (the exact pick on 17, the rest rank swaps at near-equal scores and the shadow’s empty July book). As a five-slot seat with idle slots in cash it earned 9.2% a year from 2005 with a -19% worst fall and a Sharpe of 0.96, +18.9% in 2008 and -5.7% in 2022, correlation +0.57 with the Nasdaq fund. Inside the engine’s out-stretches: +25.6% in 2008-09 against gold -1.6%, +4.4% in 2022-23 against gold +7.1%. Beside the engine it beats cash on SELECT (+0.066 Sharpe held always) and loses to the gold sleeve everywhere: the best construction per dial sits -0.040 (ULTRA), -0.039 (SELECT), -0.037 (STEADY) below gold at $25,000, no better at $250,000. Two reasons it cannot be a seat even where it edges cash: it is long the same dips the engine already owns through the two-times fund, so it adds exposure to the engine’s own bad days rather than the opposite; and it needs same-day cash on a settled-cash account that the engine keeps liquid for its own adds. The commission arithmetic is the third: 222 round trips a year on a $5,000 seat is about $450 at the fixed minimum, nine points of a nine-point return.

The pitch was a product rung, not a return: if rounding to whole shares is what makes a $2,000 account drift from the published record, fractional orders, which the broker fills on liquid funds, would open the engine below its floor for free. We registered the two conditions a rung had to meet before running: rounding must cost at least 0.30 of a point a year, and the fractional account must then land within 0.30 of the $25,000 record. Neither held anywhere. Rounding cost between -0.53 and +0.38 of a point (a negative number means whole shares did better, because rounding down holds a little less than the target and the target is a volatility guess, not a truth). The small-account gap is the commission. STEADY places the most orders (three legs), so at $1,000 its $1 fills cost 2.4 points a year and the account earned 10.7% against 12.6% at $25,000; ULTRA, one leg, lost 0.47 at $1,000 and nothing above $5,000. With commissions at zero every size sits within a tenth of a point of the $25,000 line. So the rung question is a broker question: a commission-free account runs the engine at $1,000 as well as at $25,000 in this model, and a $1-minimum account should not run STEADY below about $5,000. The product keeps whole shares; the buyer guide gets the commission sentence instead.

Cross-sectional momentum, buying what rose most over the past year, is the best-documented anomaly in the academic record and has a hundred years of data. Since 1999 the top momentum tenth returned 12.6% a year with a -56% worst fall against the Nasdaq fund’s 10.7% and -83%, and it was flat from 2000 to 2002 while the Nasdaq halved. That is the case for it as the engine’s root. We registered the test first: the frozen rules on the momentum series with a modelled two-times leg (no such fund exists), and on the plain series capped at one times exposure (what a cash account could hold in MTUM), against the shipped dials, at $25,000 and $250,000, 1999 to 2026; then the same on MTUM itself, the real fund, since its 2013 listing. Every arm lost on every dial. ULTRA: 12.6% / -38% against the shipped 16.7% / -28%; SELECT 10.7% against 15.2%; STEADY 11.0% against 12.5%. On the real fund the gap is wider: ULTRA 17.4% and 13.1% against 22.2%. Why the raw advantage vanishes under the rules: the exit and the dial are built for a series that falls hard and recovers hard. Momentum stocks fall with the market but recover slowly, because the portfolio is rebuilt from last year’s winners just as the losers rip; in 2009 the Nasdaq made +54.7% and momentum +7.1%, so the engine came back in for +34.2% on the Nasdaq and +15.7% on momentum. The volatility dial then sizes momentum’s calm stretches up and its crashes arrive from calm. The academic series is also an upper bound: monthly-formed, no costs, microcaps in; the real fund earned less on every line. Noted against ourselves: a two-times momentum fund does not exist, so the only buyable arm was the plain one, and it lost by more.

Managed-futures funds ride trends in many markets at once, long and short, and are sold as crisis insurance: they made money in 2008 and in 2022 while stocks fell. Both facts hold in the raw funds. The question was whether that pays beside an engine that already steps out of the market on its own trend rule. We wrote the rules and the pass mark down first, then ran RYMFX (a 2007 fund on the S&P trend index), AQMIX (AQR, 2010), DBMF (2019) and KMLM (2020) in the 20% seat on ULTRA, SELECT and STEADY at $25,000 and $250,000, each fund four ways, against cash, against the gold sleeve as shipped, and against gold held only while the engine is out. Alone over their windows the funds earned 1.6% to 9.6% a year with worst falls of -20% to -36%, and moved with the Nasdaq fund at a correlation of -0.16 to +0.18: the diversification is real, the return is not. Beside the engine, none of the sixteen constructions met the pass mark on any dial. The timing is the autopsy: the trend funds’ 2022 was made in the first half of the year, before the exit turned off in June; inside the out-stretch that followed (2022-06-01 to 2023-04-03) they gave back -12.8% to -0.0% while gold made +7.1% and cash +2.7%. In the 2008 stretch the one fund old enough returned +1.1%, cash +1.4%. The gold sleeve pays because gold tends to rise in the stretches the engine sits out; trend funds tend to have already risen by then. Same lesson as the energy seat (grave 154): the thing that worked in 2022 worked before the exit knew.

Mel brought a “sector all-weather” book from another chat: growth sectors and semiconductors for good times, health care, staples and utilities for bad ones, energy and gold for inflation, long Treasuries for the opposite, with a five-year backtest that beat the S&P 500 and a bond-heavy 60/40 on every line. The five years were 2021 to 2026, one regime, and its two headline claims flipped on the full cycle: energy’s correlation with the market is 0.27 in that window and 0.60 since 2005, long Treasuries’ is 0.56 in the window and below zero since 2005. We wrote the rules and the pass mark down before the test ran. The book, without the memory-chip fund that listed in April, was run from 2005, which holds 2008, 2020 and 2022, three ways: alone against the three dials; as a fifth of the book beside each dial, held outright and on its own trend gate, at $25,000 and $250,000, against the gold sleeve; and as one of five candidate blends of the sector shelf, with the blend chosen on 2005 to 2015 and judged on 2016 to 2026 so nothing could be picked on the answer. Alone: 11.3% a year, worst fall -30%, against STEADY’s 13.6% and -18%; it lost 2008 by more than twice as much and moves with the engine (correlation 0.65 with ULTRA, 0.85 with the Nasdaq fund). As a sleeve: below the gold sleeve on every dial at both sizes, the best construction by -0.05 Sharpe, and every construction deepened the worst fall. The blend search: the only candidate with a positive training score was the four funds least like the engine, bonds, gold and utilities in equal weight, and on the test window it sat -0.07 below gold on ULTRA and failed every dial. The lesson is the one graves 15, 17, 150 and 151 already carry: sectors are the market in pieces, the pieces move together when it matters, and the one piece that does not, bonds, lost money in the stretch the engine was out.

The futures version of the engine holds its Nasdaq exposure as micro futures contracts with the leftover in the plain and two-times funds; it was scoped in early September and is not sold. Its one known weak era is 2009 to 2016, a calm, steady bull, where a two-times fund that resets every day compounds better than futures that carry a fixed amount, while in choppy years the reverse holds. The engine already measures how violent the market has been over the last twenty sessions, so the natural idea was a switch: futures when the market is violent, the two-times fund when it is calm. We wrote the rules and the pass mark down before the test ran, so nothing could be tuned afterwards. First the model itself was moved into the repository with its inputs frozen and made to reproduce its September figures exactly; then two honesty checks were run on it before any tuning: real day-by-day interest rates on the cash it holds instead of a flat two percent, which left the futures version’s edge unchanged within a tenth of a percent, and the broker’s own continuous futures prices for the two years it serves, which matched the model’s assumed cost of carry within a tenth of a percent a year. Then the switch, at four settings and two account sizes, judged against the plain futures version after tax over the whole history and in each of three eras. Every setting earned less over the full history, the best by -0.07% a year at $40,000, and none won all three eras: the two settings that helped in 2009 to 2016 lost in the other two, and at $40,000 they held no contracts at all on average, which makes them the plain-fund version under another name. The pass mark had three parts and this failed the first, so the other two were not needed. What stands: the plain futures version as scoped, unchanged, with its honesty checks now on the record. What is still outside the model, stated: margin calls, contract months before 2019, and an order path for futures in the sold software, which does not exist yet.

Energy was the one thing that rose in 2022 while the Nasdaq fell, so the natural question was whether the engine should hold energy in the months its Nasdaq exit has it out of the market. We wrote the rules and the pass mark down before the test ran, so nothing could be tuned afterwards. The out-state is the engine’s own monthly reading of the Nasdaq against its 200-day average, which gives three stretches out: 2000-11-01 to 2003-05-01, 2008-03-03 to 2009-06-01, and 2022-06-01 to 2023-04-03; STEADY’s S&P leg has its own exit, so on STEADY this is the Nasdaq leg’s out-state, not the whole book’s. Six energy funds were held in the 20% seat only during those stretches, plain and with their own trend gate, on every dial at $25,000 and $250,000, against the same seat holding cash, holding gold, and holding gold only during the same stretches. The switching happens inside a stand-in return series for the seat, not as actual shares with fees and settlement, which is the usual simplification of these seat tests and is stated here. Inside the out periods: in the dot-com bear the energy sector lost -29% while gold made +30% and cash +7%. In 2008 the sector lost -29%; gold lost -2% and cash made +1%, so gold beat energy there but not cash. In 2022 the exit did not turn off until June, by which time the energy rally had already happened: from there to the re-entry in April 2023 the sector made +2%, the producers fund lost -14%, gold made +7% and cash +3%. The calendar year and the engine’s year are not the same thing, and the famous energy year falls in the part the engine was still in. Funds listed after a stretch began are measured from their listing inside it. No construction met the pass mark, which asks for a clear gain over cash, over gold, and over gold held the same way, at both sizes, with no deeper fall than gold. The midstream fund AMLP edged cash by a hair in several rows and never came near the gold seats; every other fund trailed all three. What this settles is narrow: these six energy constructions did not earn the seat. It does not settle energy’s long-term merit as an investment, and it does not close the watch on the young midstream income fund from the stone before this one. The book keeps what it has.

Most of these funds own an index, or a single asset like bitcoin or gold, and sell options on it every month; the option buyer pays cash now and takes most of the gains above a certain level. That cash is the distribution, so the fund hands you your own upside as income and keeps a price that barely moves. The long-short fund and the bond funds work differently and are read on their own terms. We wrote the rules and the pass mark down before the test ran, so nothing could be tuned afterwards. Every fund on the issuer’s page except SPYI and QQQI, which were tested last week, was measured two ways: against a reference for what it holds over the identical window (the plain index for the equity funds; SPY as a broad reference for the long-short fund and a pipeline index for MLPI, neither a replica of the fund), and as a 20% sleeve beside the engine on every dial, held two ways (always in, or in only while above its own 200-day average), at $25,000 and $250,000, judged against the same seat holding bills and holding gold. Against its reference, eleven of seventeen trail, most by two to eight percent a year with nearly the same worst fall; six are ahead (NEHI, XSPI, BNDI, CSHI, TLTI, NLSI), the bond ones by about one percent a year. Of the four older funds these designs copy, the Nasdaq and S&P covered-call funds since 2013 and the gold one since 2013 trail their index by three to eleven percent a year; the Treasury one, since 2022, beat a Treasury index that was itself falling. Three of the four had a shallower worst fall than their index, so the family read is: less rise, and only sometimes less fall. Beside the engine, nothing passed. Three met the numbers on some dials: SPYH, a hedged S&P fund, on every dial; NLSI, a long-short fund, on every dial; MLPI, an energy-pipeline fund, on SELECT and STEADY. All three are young, SPYH at seventeen months and the other two under a year, and have only ever traded in a rising market, where a fund that sells calls looks its best; the rule is that such a window can earn a watch, never a pass. SPYH’s daily moves also track the engine’s (correlation 0.83) and the Nasdaq fund’s (0.93), which fails the different-asset bar whatever its numbers say. NLSI trades with a spread of about a quarter of a percent, which no test here charged. The rest: two of the three boosted funds, launched this February, are behind their index already and the third is level with it; the bitcoin income funds have lost 26% and 35% of their price since listing (their worst falls 37% and 48%), the ether one 37% of its price with a worst fall of 50%; the gold income fund trails gold by about eight percent a year with the same fall; the bond ones sit near zero and add nothing a bills seat does not. QQQH’s price history reaches back to 2019 because the ticker belonged to a predecessor fund whose option strategy was later changed, so its record through 2022 is the predecessor’s, not the current fund’s. What stays open: NLSI and MLPI hold their watch until their first real fall, and neither is a candidate for the book before then. The book keeps what it has.

We took every fund from this weekend that had beaten a bills sleeve on at least one dial, twelve of them plus gold and bills, and let five different recipes build a blend from them: equal weight, lowest volatility, highest return per unit of risk, equal risk from each part, and one that directly minimises how much the blend moves with the engine. Each recipe could only look backwards, re-picking its weights every January from the years before, using only funds already listed by then, and only 2019 to 2026 was judged, which holds the 2020 crash and the 2022 bear. The mix was held as whole shares at real prices from the sleeve’s own dollars, cash in bills, monthly rebalances, next-day settlement and a dollar per fund traded, and then handed to the engine as a single price series standing in for the sleeve. This stone went through four passes. The Lab found five defects in the first run, the reviewer three in the second, and the third pass answered them; the Lab then tested the one thing the third pass seemed to show, that a nearly-all-bills sleeve beat STEADY’s gold, by putting bills in place of gold directly inside the engine: bills lose there, two points a year of return and about a tenth of Sharpe at both sizes, with a slightly shallower fall. So that reading was an artefact of the stand-in series, and it is withdrawn. Three accounting gaps in the method remain and are why it stops here rather than being patched again: the sleeve’s own cash and the book’s cash are kept separately; the mix sizes and fills its constituents at the same close, which is the timing flaw fixed elsewhere in the engine; and the least-correlated recipe treated a fund’s pre-listing days as zero returns when measuring correlation. Settling the blend question properly needs the constituents inside the engine as real lines, which is a feature the engine does not have and would have to be built and audited before any blend could pass. What can be said, and only this: on every dial and at both sizes, none of these five recipes improved the book over the gold sleeve, and the covariance estimates two of them rest on were unreliable (most yearly fits needed repair). It does not establish that no blend can, and it does not establish that only gold can. The book keeps what it has.

Mel sent the issuer’s whole shelf, five hundred and twenty-four products sorted by assets, and asked which could be sleeves. Most of it already had a stone: broad and factor equity (17, 19, 117, 149), sectors and themes (15, 16, 121, 150), countries (150), dividend funds (149), preferreds (7), option-income and buffer-style funds (146, 147), TIPS and the short Treasury funds (one of which is the book’s cash), gold and silver (the incumbent sleeve), bitcoin (its own gated seat), ether (worse than bitcoin, September 2nd), semiconductors (16, 121). What had never been run as a sleeve was the fixed-income and hybrid middle of the lineup: mortgages, investment-grade and short corporate credit, two high-yield funds, emerging-market bonds, municipals, convertibles, floating rate, the Treasury curve at three maturities, TIPS again, the flexible-income active fund, minimum volatility, global infrastructure, two REIT funds, and the systematic-alternatives fund launched last December. We wrote the rules and the pass mark down before the test ran, so nothing could be tuned afterwards on the same design as the country and theme stone: each a fifth of the book, held outright or on its own 200-day gate, on all three dials, against each dial’s bills and gold sleeves, corrected runner, frozen inputs. 114 constructions. None passes. The bond funds are the honest surprise for anyone who expects bonds to diversify a stock engine: the Treasury curve is the only thing on the shelf with a negative correlation to the engine (-0.13 for the long end), and it earned about nothing over the window (0.1% a year with a 48% fall) and lost 7% in the ten months of 2022 the engine was out, the one stretch it was held for. Credit and high yield carry equity risk (0.50, 0.46) and pay a bond’s return for it. Mortgages and municipals are near-zero correlation and near-zero return, which is what bills are, with a fall. Convertibles are equity in a costume (0.64). Minimum volatility beats bills gated on ULTRA by +0.015 and gold by +0.005, the best number on the shelf and a sixth of the gate. One line is not buried: the systematic-alternatives fund, a multi-strategy book of trend, carry and the like with no equity leg, is nine months old, has never seen a bear, and reads 23% a year with a 2% fall over that stretch; every construction of it clears the gate by a margin that means nothing on nine months. By the registered rule it was a watch, the only one on the shelf, until the Lab found the same strategy family with ten years of record in its mutual-fund siblings and we ran them through the same gate: correlation with the engine 0.17 and 0.12, 7.3% and 4.9% a year with falls under ten percent, and as a sleeve the better of them beats bills on SELECT by +0.056 and loses to the gold sleeve on every dial (-0.068 on SELECT, -0.088 on ULTRA). A low-correlation stream that earns less than gold is a slower bill with a fee. The watch is closed on its family’s record, with the Lab’s caveat kept that these are the siblings’ histories, not the new fund’s. The shelf is otherwise closed: on this book, a bond is a slower bill with a drawdown, and a hybrid is equity with a story.

Mel: emerging markets, robots, quantum, and so on. The record already said what a fixed slice of these would do (grave 17, higher-volatility assets generally; 15, sector rotation; 16 and 121, semiconductors; 19, spreading the rule across markets; the 44-ETF expansion screen of September 3rd, where every large edge turned out to be the same asset as the root), but one construction had not been run on them: the gated sleeve, held only while the fund is above its own 200-day average at the month’s first session and in bills otherwise, which is how bitcoin and gold are held. We wrote the rules and the pass mark down before the test ran, so nothing could be tuned afterwards: twenty-seven funds, ten geographic (emerging markets broad, India, China, China internet, Brazil, Japan, Korea, Taiwan, developed ex-US) and seventeen thematic (robotics twice, quantum, AI, innovation, space twice, clean energy twice, uranium, lithium, biotech, medical devices, defence, infrastructure, energy, semiconductors), each as a fifth of the book both fixed and gated, on ULTRA, SELECT and STEADY, against each dial’s bills and gold sleeves, corrected runner, frozen inputs. 162 constructions. None passes. The correlations say why before the simulations do: every one of these funds moves with the engine’s own returns, from 0.32 for energy to 0.81 for the AI fund, and the themes cluster around 0.72 to 0.76; they are the root with more volatility and a story. The gate lets a few of them beat bills, quantum and semiconductors and infrastructure on ULTRA, energy on SELECT (+0.048); none beats the gold sleeve, the best being energy gated on SELECT at +0.003 and semiconductors fixed on ULTRA at +0.001. The gate helps the worst of them (China, uranium, innovation) by taking them out of their falls and cannot make them a sleeve. The emerging-market funds are the weakest of all: 5.2% a year alone for the broad fund since 2013 with a 40% fall, and negative on every dial, gated or not. The lesson is the same one the income and dividend stones taught, now on the growth side: a sleeve earns its place by being uncorrelated with the engine, not by being exciting, and every theme is correlated. What robotics, quantum and AI do have is a claim on the root itself, the semiconductor root (grave 121) and this one; the engine already holds the Nasdaq-100, which is where those themes live when they work.

Mel had a feeling there was a dividend sleeve somewhere, and asked the Lab to scour on its own while the option-income shelf was being buried. The Lab took it in three passes on the corrected sleeve runner, each frozen before it ran, each reproduced independently. First the midstream names that pay the fattest distributions, an energy partnership and the pipeline fund: beside ULTRA over 2019 to 2026 they returned about what the gold sleeve returned and fell eight points deeper (-30% and -29% against -21%). Then the five dividend funds with the best-reasoned methods on the shelf: sustainable dividends with low price and earnings volatility (LVHD), global free-cash-flow screens (GCOW), high dividend with a low-volatility screen (SPHD), quality dividend growth (DGRW), global aristocrats (WDIV), each as a fifth of the book beside ULTRA over 2017 to 2026 against bills, gold, the S&P and the world index: every one had a lower Sharpe than both bills (1.158) and gold (1.194), the best of them (GCOW) at 1.145, and every one fell deeper than bills. Mel then asked the right question: was ULTRA the wrong host? So the same five ran on SELECT, with its lower exposure, and on STEADY, replacing its gold. On SELECT the best of them (LVHD) added +0.015 of Sharpe to the unchanged engine, below the declared benefit, with a lower return and a deeper fall, while the gold sleeve on the same book added +0.108. On STEADY every dividend replacement lowered Sharpe and deepened the worst fall, from 18.4% to between 22.6% and 25.2%. Ten more constructions, ten failures. The active dividend-value fund that had met the gate on a single-bear window (stone 147) was closed the same day: at ten and thirty percent it fails, and a block bootstrap of its twenty-percent result crosses zero against both the S&P and bills. None of this says dividend investing fails. It says a fixed slice of a dividend fund does not improve this book on any of its three dials, for the same reason the income funds did not: it carries the equity risk the engine already sizes, and the two sleeves that do not, bills and gold, are the comparison it has to beat. The feeling was worth testing. It is now tested.

Mel read the two income stones correctly: the only sleeve that has ever earned its place, gold, won because it is uncorrelated with the equity the engine already holds, so the riddle is to find the most uncorrelated thing with a return. Measured first, not guessed: every candidate’s correlation with the engine’s own daily returns on the full-history test, not with the index. The zeros with a positive record were bills and gold (both already in the book), pure trend futures with no equity leg, the dollar, a convex tail fund that has not bled, and the rising-rates fund already withdrawn; everything with a strongly negative correlation (tail funds, VIX products, anti-beta) pays for it in carry every year it is not needed, and Treasuries were a zero until 2022 and then were not. We wrote the rules and the pass mark down before the test ran, so nothing could be tuned afterwards: each candidate and two gold blends as a fixed 20% sleeve beside ULTRA on the corrected runner, over its own window, against the same 20% in bills and, this time, in gold, because a new uncorrelated sleeve has to beat the one the book has; plus the sleeve must be positive on the engine’s worst days and its own failure mode must be in its window. Every candidate lost to the gold sleeve: pure trend -0.058 of Sharpe, trend-plus-carry -0.080, the convex fund -0.059, the dollar -0.045, half gold and half trend -0.048. Trend beat bills and lost 10% in the ten months of 2022 the engine was out, the one stretch a diversifier is for, because trend reversed as the engine re-entered. The convex fund is the interesting one: positive on the engine’s worst days, a fall of a few percent, but three years old with no calm year in its record, the year that decays it; watched, not passed. The honest footnote is about the incumbent: over this same 2013-to-2026 window the gold sleeve beats the bills sleeve by only +0.011 of Sharpe on ULTRA. Gold earned its seat on the long record (1975 on, with its own twenty-year bear) at equal drawdown; on the last thirteen years it is a wash against cash beside this engine. The bar for the next sleeve is therefore not high, and nothing cleared it. The most uncorrelated thing with a return that we can buy is the thing we already hold.

Mel’s follow-up: never mind weekly, what about the JEPI tier, monthly funds, and the components of a model portfolio making the rounds (leveraged tech, efficient gold, gold miners, momentum, managed futures, commodities, small value, international value, a closed-end-fund fund, a rising-rates hedge, income funds)? We wrote the rules and the pass mark down before the test ran, so nothing could be tuned afterwards on a test whose rules were written down in advance’s design: thirty-two funds with a usable record, each against its own asset class over the identical window, then as a fixed 20% sleeve beside ULTRA on the full-history test against the same sleeve in the asset class and in bills, then the 2022 stretch the engine was out, then the tax arithmetic. This stone was regenerated after the Lab’s audit found four defects in the first runner (the sleeve book had run on SELECT’s settings; a sleeve whose underlying was the Nasdaq had pinned the core’s own signal; gates compared rounded figures; inputs were not frozen). The corrected runner fixes all four and the inputs are now archived; the numbers here are the corrected ones. The record is more mixed than the weekly shelf: the option-income funds still trail what they hold (JEPI -6.8 points a year, JEPQ -4.3, GPIX -2.6), while several of the portfolio’s components beat theirs (S&P momentum +4.2 over a decade, small value +4.9, international value +5.2, the gold funds +8.1 and +8.2 over gold and the miners). As a sleeve almost none of that survives, and the reason is the book it sits beside: ULTRA is already a volatility-sized equity position, so a fifth of anything with equity beta adds correlated risk the dial then has to cut, and a bills sleeve is hard to beat. Momentum beat its index by +0.031 and bills by +0.014, under the gate. Small value, international value, the gold funds, managed futures, the closed-end-fund fund: the same shape, ahead of their asset class, not ahead of bills by enough, or with a deeper fall. JEPI beats bills by +0.019; the rising-rates fund RISR beats them by +0.068 but with a deeper fall, on a window that is one regime. One fund met the gate: CGDV, an active dividend-value fund, +0.045 of Sharpe over the S&P sleeve and +0.034 over bills with an equal fall, on a window from February 2022 that contains one bear, the 2022 one, in which value beat growth by the widest margin in twenty years. That is the era a dividend-value fund is built for, and dividend-and-value ballast has a stone of its own (117). It went to the Lab with that caveat and came back closed the same day (stone 149): at ten and thirty percent it fails, and a block bootstrap of the twenty-percent result crosses zero against both comparators. While the engine was out in 2022, nothing on the list was a cash substitute; the managed-futures fund, sold as the crisis hedge, lost 12% in those ten months. A Roth removes tax drag; it changes no verdict.

Mel’s question, and a fair one: the weekly-pay option-income funds from Roundhill, YieldMax, REX and the rest are the fastest-growing shelf in retail; is there a sleeve in the best of them, and does a Roth change the answer? We wrote the rules and the pass mark down before the test ran, so nothing could be tuned afterwards with the whole shelf frozen: twenty-three funds, each against its own underlying over the identical window, then as a fixed 20% sleeve beside ULTRA on the full-history test against the same sleeve in the underlying and in bills, then the one episode the engine was actually out that any of them lived through (May 2022 to March 2023), then the tax arithmetic. This stone was regenerated after the Lab’s audit found four defects in the first runner (the sleeve book had run on SELECT’s settings; a sleeve whose underlying was the Nasdaq had pinned the core’s own signal; gates compared rounded figures; inputs were not frozen). The corrected runner fixes all four and the inputs are now archived; the numbers here are the corrected ones. The record: with every distribution reinvested, twenty-two of the twenty-three trail the thing they hold, by 11 points a year for QYLD over twelve years and 19 for ULTY, with the same worst falls; up-capture equals down-capture, which is what selling the upside for premium has to produce. The one exception, IDVO, holds a different portfolio from its benchmark, and the semiconductor fund’s 111% a year is the semiconductor rally minus fourteen points. As a sleeve, nothing passes: JEPI, the best of them, beats the S&P sleeve by +0.037 of Sharpe and the bills sleeve by only +0.019, under the gate; the four that beat both did so on windows with no bear in them and are watched, not passed. While the engine was out in 2022, none was a cash substitute: bills made +2.4% with no fall, the funds -2.2% to +3.9% with falls of ten to twenty percent. And the Roth: it removes the tax drag, which is enormous on the weekly payers (15 points a year on QDTE, 26 on ULTY at a worst-case ordinary rate), and it cannot remove the upside the fund has sold, which is the gap to the underlying that remains in a Roth (-1.4 QDTE, -10.8 YMAX, -10.6 QYLD). A Roth makes them less bad. It does not make them a sleeve.

Pre-registered with the gate fixed in advance: at least +0.03 Sharpe on the STEADY book, no worse drawdown, two of three sub-windows, and a passing mechanism check on both roots; every variant enters the full-history test through the signal input only, so sizing, band, exit confirmation and execution are the shipped ones. A 2025 note on gold since 1969 reports that gold’s momentum is reliable only when Treasuries are also rising. On STEADY’s gold sleeve, which already holds gold only above its own 200-day average with a two-month confirmation, the extra check runs through that same confirmation (it does not exit the instant Treasuries turn negative), and it is nearly empty: the gold months it drops did average less (0.38% against 1.24%), so the mechanism is there, but nearly all of those months were months the sleeve was already out. Sharpe 1.015 to 1.029, worst fall unchanged, the same number of trades. Below the quarter of the gate, on a window that begins with the Treasury fund’s listing in 2002. Not added.

Pre-registered with the gate fixed in advance: at least +0.03 Sharpe on both the Nasdaq and the S&P root, no worse drawdown, two of three sub-windows, and a passing mechanism check on both roots; every variant enters the full-history test through the signal input only, so sizing, band, exit confirmation and execution are the shipped ones. The trend literature says a vote across horizons of volatility-normalised returns beats a single moving average with less churn. The raw month-end vote does what it claims: the months it would drop averaged 0.44% on the Nasdaq against 1.33% for the months it kept, and the same on the S&P, so the mechanism check passes on both roots. That check counts the vote’s own months, not the smaller set the confirmed rule actually sat out, which the Lab’s audit noted. The book still loses, and badly: SELECT -0.136, ULTRA -0.119, with deeper falls, because a vote of three slow signals is out through long stretches of ordinary drawdown that the single rule with its two-month confirmation rides through, and the dial has already shrunk the position by then. Dropping bad months is not enough when the price of being out is paid in the good ones. The single 200-day rule with confirmation stays.

Pre-registered with the gate fixed in advance: at least +0.03 Sharpe on both the Nasdaq and the S&P root, no worse drawdown, two of three sub-windows, and a passing mechanism check on both roots; every variant enters the full-history test through the signal input only, so sizing, band, exit confirmation and execution are the shipped ones. The published claim (US factors over nine decades) is that the volatility of down days alone predicts next-month returns while total volatility does not, so scaling by it times the market as well as sizing it. On our roots the premise fails before the rule does: months whose downside volatility exceeded total volatility were followed by 1.28% on the Nasdaq and 1.27% on the S&P, higher, not lower, than the other months. The estimator as run is the root-mean-square of the negative days over the same twenty sessions, scaled by the square root of two so it equals total volatility on symmetric returns (the registration’s wording left that denominator ambiguous; the code is the rule). Built from half the days, it is noisier, and our full-history test paid for that in trades: 32 a year against 27. Sharpe fell on all three books. The twenty-day standard deviation stays.

Pre-registered with the gate fixed in advance: at least +0.03 Sharpe on both the Nasdaq and the S&P root, no worse drawdown, two of three sub-windows, and a passing mechanism check on both roots; every variant enters the full-history test through the signal input only, so sizing, band, exit confirmation and execution are the shipped ones. The published result (ten equity index futures over four decades) says continuous inverse-volatility scaling over-trades in ordinary months, and that only the extremes carry information. On our book it split the roots: the S&P version improved by +0.032 with a shallower worst fall, the Nasdaq version lost 0.046 and its worst fall deepened by 6.3 points, because the months the rule leaves unscaled are exactly where the Nasdaq’s fast vol changes live. The mechanism check agreed with the split: after top-fifth-volatility months the Nasdaq averaged 0.10% a month against 1.39% in the middle, but the S&P averaged 1.22% against 0.82%, the wrong way round. A rule that works on one root and not the other is a fit, as the wide-band lesson taught. The continuous dial stays.

The idea: instead of holding the two-times fund, hold the plain fund and borrow the extra from the broker. You would swap the fund’s fees for the broker’s interest, and the question is which costs less. We wrote the rules and the pass mark down before the test ran, so nothing could be tuned afterwards. The two-times fund charges a yearly fee (0.95%) plus the cost of the borrowing it does internally, which runs a little above the short-term interest rate. Borrowing from the broker costs the same short-term rate plus the broker’s markup, which shrinks as the account grows: 1.5% under $100,000, 1% up to a million, 0.75% up to three million, and 2.5% on the free-tier plan; plus you pay the plain fund’s small fee on twice the money. Everything else in the engine stayed the same, so only the cost of leverage changed. Result: borrowing loses under $100,000 (0.14% a year worse on ULTRA) and badly on the free tier (0.67% worse); it is a wash from $100,000 to a million; it edges ahead only between one and three million, by 0.07% a year, a third of what we would need to see to change anything. Two things the test does not include, stated plainly: margin interest can be tax-deductible in a taxable account and fund fees cannot, and the two ways of holding leverage drift slightly differently between rebalances. Neither is big enough to close a gap this size for a small account. The two-times fund stays.

The lot-selection rule paid over a point a year on SELECT, so the next structural idea was the classic one: when a lot is under water, sell it, keep the exposure through a substitute, and use the loss now. We wrote the rules and the pass mark down before the test ran, so nothing could be tuned afterwards on the engine’s own trades from the full-history test, 1999 to 2026, with first-in-first-out lots, the test’s tax rates and a pooled loss carry, and the strictest version of the answer: every deferred gain paid at the end of the window. Harvesting at five, ten or twenty percent under the basis improves the after-tax return by +0.09, +0.03 and +0.01 points a year on SELECT and +0.09, +0.04 and +0.01 on ULTRA, and the best of those went negative in the last decade. The reason is in the engine: a rule that cuts exposure when volatility rises sells its losers itself, so the losses a harvester would bank are mostly banked already. The gate asked for a quarter point on both dials; nothing came close. It stays a good idea for a buy-and-hold book and an empty one for this engine.

Every daily-reset leveraged and inverse fund on the Nasdaq-100 must rebalance at the close, in the same direction as the day, by an amount known by mid-afternoon. We wrote the rules and the pass mark down before the test ran, so nothing could be tuned afterwards on 3,411 sessions of one-minute QQQ bars from IBKR, 2012-12-28 to 2026-09-02, thresholds fixed in advance. The mechanism shows: the last hour continues the day’s direction and the effect rises with the size of the move, from about 3 basis points on half-percent days to about 12 on two-percent days, and a placebo of random sessions never matches it. Then the honest part: the version a cash account can hold, buying at three o’clock on up days and selling at the close, nets 1.3 basis points an event after a dollar and a basis point each way, 0.57% a year, and it lost money from 2013 to 2016. The both-direction version does better (1.97% a year) and needs a short. By the registered gate that is era-dependent, not a product. What the study did settle for us: after a session down one percent or more, the fund is on average 9 basis points above the prior close by 09:45, so the engine’s morning execution of a cut is not paying a late-day penalty.

The second-opinion review of the product asked the right question: is the engine’s edge over a plain diversified book real, and if a broad book beats it, which layer pays, the diversification, the sizing or the timing? Pre-registered before any number was run: a frozen universe of fourteen liquid ETFs from 2007, four arms measured separately. A static book with fixed block weights; the same book scaled to a 10% volatility target; that book with each ETF held only above its own 200-day average; and the engine. The first pass was audited the same day by the Lab, which found four execution defects (a warm-up read as an exit, a fixed cash yield for the engine only, a modelled levered leg, a risk match that did not preserve Sharpe); a second audit then found stale deferred orders, an extra session of delay for the engine and a fee denominator error; this stone is the corrected third pass, every arm on one executor and one timing convention (prior-close signal, fill at that session’s close): decision at the close, fill at the next close, whole shares, a dollar a leg, a cash account that cannot spend the day’s sale proceeds, bills bought and sold like any line, the engine run from its own signal and dial through that same executor with the real levered fund. The static book lost to a two-fund 60/40 (0.51 against 0.64): the real-asset block cost more than it diversified. Sizing to a volatility target made it worse (-0.07). The trend filter helped (+0.12), but by cutting volatility and drawdown, not by picking months: in stocks the months it dropped earned more than the months it kept. The engine at the same volatility beat the best of the four in the full window and in each of the three sub-windows. The precise conclusion, and only this: this frozen construction failed its declared gates under this execution model. It does not show that diversified trend is exhausted, nor that the engine has alpha; it shows that the registered falsifier, a static book matching the engine at the same volatility, did not fire. The window has no dot-com bear, which is stated so this is never read as a full-cycle claim. Status: independently reproduced by the Lab on 2026-09-12 after a five-round audit (stale deferrals, timing, fee accounting, calendar edges, all fixed and tested); the rejection stands under its documented model, with the window’s limits as stated. The one executor is proven by substitution in the archive, and the one deliberate difference between arms is decision frequency: monthly for the books as registered, every session for the engine as sold.

Hold the bitcoin seat in a 2× or 3× daily-reset fund instead of the spot fund, the gold seat in 2× or 3×, gate unchanged. The real 2× funds from their listings, modelled before that from spot with financing and fee; the 3× versions modelled throughout, since no US 3× bitcoin or gold fund exists. Against the shipped pair at equal drawdown, since 2018 and since 2022. Bitcoin levered is off the ladder in both directions: the gated 2× sleeve on its own earns 11% a year against 26% for the 1×, with a −92% drawdown and a −43% worst day, because a daily-reset fund on an asset with 60% volatility pays the volatility decay every day and the gate cannot save it from the days inside a month. The pair with 2× bitcoin has a −37% worst drawdown against −24%, and its extra return is the size dial paid for twice. Gold 2× is a wash since 2018 and half a point since 2022; the modelled 3× gold reads well, and it is a fund that does not exist, on a window that is gold’s bull market. The sleeve fractions are the leverage dial for these seats, and they are already where the frontier put them.

When a sleeve’s gate is off, hold the inverse fund instead of T-bills: −1× or −2× bitcoin, −1× or −2× gold, daily reset, real funds where they exist and modelled before. The NASDAQ version of this idea is grave 88; bitcoin’s bears are three times deeper, so it earned its own run. Against the shipped pair at equal drawdown, since 2018 and since 2022. Bitcoin short-while-out loses 1.2 points since 2018 and wins 1.6 since 2022, which is the 2022 bear and nothing else: the sleeve alone earns 11% a year with the inverse leg against 26% in cash, because the gate is off for months after every bottom while bitcoin doubles, and an inverse fund is on the wrong side of every one of those months. The −2× version loses six points since 2018. Gold short-while-out loses two to four points on both windows; gold’s time below its average is drift, not decline. The gate’s job is to not own the fall. Betting on the fall is a different job, and the same rule is not good at it.

The trade every crypto desk describes: buy the spot fund, sell the bitcoin future against it, and collect the premium as the future converges at expiry, with no view on direction. The version a margin account at a US broker can actually run, pre-registered: the front CME contract sold on the first day of each contract month from the exchange’s launch in December 2017, held to its last-Friday expiry, spot-fund fee, slippage, a tracking haircut, and $3 a round trip on the micro contract; capital is the spot plus the 40% margin the broker holds against the short. The entry basis averaged 0.22% a cycle, 2.7% a year, and 42 of 104 cycles were in backwardation, where the trade pays to be in it. Net of costs it earned 0.46% a year on capital against 2% for bills, with 2018 at −9% and 2022 at −7% of capital, the years bitcoin fell. It is uncorrelated with everything, which is true and does not help: beside the pair it loses at equal drawdown at every size and whether the capital comes from the engine or from the margin account. The famous double-digit basis lives in the quarterly contracts and in offshore perpetual funding, in the months bitcoin is running, and neither is on offer to this account. What is on offer is a small premium that disappears in the years you would want it.

The product’s one real relationship, realised volatility predicting next-month risk, applied to the gold seat: instead of a fixed fifth of the book in gold, run a volatility target on it with the 1× and 2× gold funds as legs, gated by gold’s own 200-day, sized like the core, weights decided a day ahead, turnover charged. Two targets, the core’s own 0.26 and gold’s expanding median so the average exposure sits near one. Pre-registered against the shipped pair at equal drawdown, since 2018 and since 2022. The median-target engine lands on the shipped slice to the hundredth since 2018 and loses a point since 2022; the hotter target loses a third of a point since 2018 and earns seven tenths since 2022, which is the dial. The worst day improves, −6% against −10%, and that is the whole effect: a vol target on an asset whose volatility does not predict its return is a smoother way to hold the same thing. The relationship is the NASDAQ’s, and it did not travel.

The same idea on the bitcoin seat: a volatility target on bitcoin with a spot leg and a 2× leg, the real 2× fund from 2023 and a modelled one before it with its financing and fee, gated by bitcoin’s own 200-day. At the core’s 0.26 target it holds a third of a bitcoin position on average, because bitcoin’s volatility is fifty to sixty, and gives up almost all of the sleeve’s return; at bitcoin’s own median it averages 1.2× and still loses three quarters of a point since 2018 and two since 2022, because bitcoin’s volatility rises into its best months as often as its worst, so the target trims the runs it is meant to ride, and the levered leg’s cost is paid on the way. The worst day improves from −21% to −16%, which is real and not worth the return. Bitcoin’s edge in this product is its drift with the 2022s removed by the gate; the gate does the work, and the sizing should stay fixed.

Instead of a fixed 20% and 20%, split the two seats by twelve-month-minus-one relative strength, read monthly and applied a day late: the winner gets thirty and the loser ten, or the winner takes all forty. The gates still apply. Pre-registered, expected to be a dial, and worse than that: it loses five points a year at equal drawdown since 2018 and two since 2022, and takes the worst drawdown from −24% to −30% and −36%, because the two assets alternate on a timescale shorter than the lookback. Bitcoin leads for a year, the rule concentrates in it, bitcoin’s 2022 arrives with the sleeve at thirty or forty; gold leads through the bear, the rule rotates in as gold’s run ends. Cross-sectional momentum between two assets is a coin that remembers the last flip. The fixed split is the answer precisely because it does not.

The engine sizes on twenty days of realised volatility. The option market publishes its own forecast every day, the VXN, forward-looking by construction. So size on that instead: the VXN as the estimate, the VXN scaled to the realised series’ average so the level is not the dial, a 50/50 blend, and the larger of the two. Pre-registered; 2001 to 2026, both dials, same target, cap, band and exit. Every arm loses at equal drawdown, by one to three points a year, in every era, on both dials. Implied volatility runs about four points above realised on average because it contains the insurance premium, and it moves later than the realised estimate on the way down, so the engine sized on it holds too little through recoveries and is not sized down any sooner into declines, since the two rise together. The scaled version, which strips the level, loses more, which says the shape is worse and not just the mean. VIX as a cap governor, VIX as a veto, VIX as a thermometer and a vol-of-vol dial were already on this wall; this is the last way to put the option market’s volatility number into the engine, and it is the worst.

The price of crash insurance relative to at-the-money insurance, the put skew, is the option market’s own statement about tail risk. From the real NASDAQ chains, daily since 2012: the implied volatility of the put 10% below the market, thirty to forty-five days out, minus the at-the-money put’s, each backed out of the quoted mid, ranked through its own expanding history and used a day late. When it is in its top decile, size down by a quarter or cap leverage at one. Pre-registered, both dials, placebo of the skew history shuffled by year. Every arm loses at equal drawdown, about a point a year, and the placebo is decisive: on ULTRA the shuffled skew beats the real one 98% of the time with the cap and 35% with the scaling, which means the real timing is worse than random timing of the same size. Skew steepens when the market has just fallen and dealers have repriced the tail, which is when the engine has already cut and the recovery is nearest. Nine ideas from dealer positioning, five from the VIX, and now the skew: the option surface describes the last move well and the next one not at all.

The textbook auxiliary layer: when current volatility is low but the volatility of volatility is spiking, something is cooking, so size down ahead of it. Pre-registered as written, with the condition defined before the run: twenty-day realised volatility below its own expanding median and vol-of-vol above its expanding 90th percentile, measured two ways — the CBOE’s VVIX from 2007 and a realised version, the twenty-day dispersion of the volatility estimate itself, on the full cycle. Action: target scaled by three quarters, or leverage capped at one. Placebo: the vol-of-vol history shuffled by year. The condition is true on 6% of days with VVIX and 0.4% with the realised measure. One arm of eight reads a point: ULTRA capped at one on VVIX days, +1.15 at equal drawdown, but it loses a fifth of a point on SELECT, the shuffled series beats it 12% of the time, and its three quarters live in 2017–2026. Every other arm is within a quarter point of zero either way. The realised version, the one with twenty-seven years, is inert. The volatility target already reads the volatility of volatility, because a rising estimate is a rising estimate; a dial on the second derivative fires too rarely to matter and, when it fires, does not know the direction.

The other textbook layer: markets fall faster than they rise, so when volatility spikes while price is still above its 200-day average, dampen the selling to avoid whipsaws. The engine already has the two ingredients — a band that adds at 25 to 50% and cuts at 5 to 7.5%, and the monthly 200-day exit — so the untested piece was widening the cut to 10% or 15% while above the line. Pre-registered, full cycle, both dials. The wider cut earns a tenth to a quarter of a point at equal drawdown, below the bar, with the eras disagreeing: it loses in 1999–2008 and 2009–2016 and wins in 2017–2026, where volatility spikes above the trend line were the ones that resolved upward. The cut fires when volatility has already risen, which is information; waiting because the price is still above a moving average trades that information for a bet that the spike is noise, and over twenty-seven years the spikes above the line were not noise often enough.

The engine sizes on twenty days of close-to-close volatility. Range-based estimators — Parkinson, Garman-Klass, and Yang-Zhang, which adds the overnight gap back — use each day’s high and low and are several times more efficient in the textbook sense. Pre-registered on the full cycle, both dials, same target, same cap, same band. Parkinson and Garman-Klass read the NASDAQ’s volatility at 19.4% against 22.8% close-to-close, because the range inside the session does not contain the gap and the gap is most of the NASDAQ’s variance; the engine then holds more and the extra return is the dial, minus 1.5 to 2.6 points at equal drawdown. Yang-Zhang matches the close-to-close level and still loses 2.4 points at equal drawdown, negative in the two later eras, because its weighting of the overnight and intraday pieces changes the estimate exactly on the days after a gap, which are the days the sizing has to be right. The noisy estimator is the one that measures the risk the engine actually carries.

With lots sold highest-cost-first, the setting the product ships, a cut sometimes sells shares bought within the year, and the gain is taxed at the higher rate. So when the shares a cut would sell are mostly young, wait: require the cut to be two and a half or five points deeper before acting. Pre-registered, full cycle, after tax with the 35/20 rates and loss carry. It loses on ULTRA at both settings and on SELECT at the shallower one; the deeper wait scrapes a fifth of a point on SELECT at a point and a half more drawdown, below the quarter-point bar. The cut exists to sell before a decline runs, and a cut that waits is a cut that is late; the tax saved on a few young lots is smaller than what the delayed sale gives back. The tax-lot setting itself is confirmed on this run at +0.6 and +1.3 points a year, which is the part of this idea that was ever real.

Gold’s textbook driver is the real interest rate, and its twenty-year bear ran alongside high ones. So gate the sleeve on the ten-year TIPS yield instead of gold’s own price: hold gold while the real yield is below its own 200-day average, read monthly, or require both signals, or accept either. The Treasury’s daily real-yield curve from 2003, the same one-day lag as every other signal, beside ULTRA at a fifth of the book. The real-yield gate earns a point and a half at equal drawdown against the price gate’s 2.2, requiring both is worse still, and accepting either merely reproduces the price gate. The two gates agree on 44% of days, and where they disagree the price knows first: in the 2005–2008 stretch gold rose while real yields were flat and the yield gate sat out most of it. Holding gold with no gate at all beats every gate on this window, which is the cost the price gate pays on purpose for the 1980–2001 bear that this window does not contain. Real yields are a good explanation and a slow signal.

Twenty-six years of the CFTC’s weekly Commitments of Traders report for NASDAQ 100 futures, the non-commercial net position as a share of open interest, ranked through its own expanding history so nothing is known in advance, and used from the Monday after the Friday release. Three pre-registered arms: cap leverage at one when speculators are in their top decile, scale the target down by a quarter there, and the mirror, cap when they are in the bottom decile. Every crowd-is-long arm loses at equal drawdown, by three quarters of a point to nearly two points a year on both dials, and the mirror does nothing. The placebo settles it: shuffle the positioning history by year and the real series does worse than 82% of the shuffles, so the rule is not merely useless, it is timed against the market. Speculators are long the NASDAQ when it is trending, and the engine already knows when it is trending. Crowd positioning is a description of the last quarter’s trend wearing the costume of a contrarian signal.

Pre-registered and run on the full 1999–2026 cycle with lots, holding periods, the 35/20 rates and loss carry: first the wash-sale rule itself, which the published after-tax numbers had never modelled — a loss is disallowed when the same fund is bought within thirty days either side and goes into the replacement lot’s basis — and then the fix, buying a twin fund with identical returns for thirty days after any loss sale so the loss stands. The wash rule costs the engine about a third of a point a year on both dials, which is how optimistic the after-tax figures on this site were, and it is now stated there. The twin recovers almost none of it: two hundredths of a point on ULTRA, eight on SELECT, never more than a fifth of a point in any era. The reason is that the engine’s loss sales are the exit and the cuts, and the buys that would wash them come weeks or months later, outside the window; the deferred losses are a rounding of the tax bill, not a lever on it. The tax-lot setting shipped in September is worth a point; this was worth checking and is worth nothing.

The idea: after a crash the twenty-day volatility estimate stays swollen for weeks, so the engine holds its smallest position exactly when recoveries are steepest, April 2020 being the poster child. So for one, two or three months after each re-entry, cap the estimate at its long-run median. Pre-registered, full cycle, both dials, with the placebo of the same cap in random months. The shipped exit re-enters three times in twenty-seven years — May 2003, June 2009, April 2023 — and each time it re-enters after two confirming monthly closes above the band, by which point realised volatility is already at or below its median. The cap binds on a handful of days and changes ULTRA by three hundredths of a point; on SELECT it earns a fifth of a point, all of it from the 2009 re-entry, which the placebo beats anyway. April 2020 is not on the list at all, because the confirm never let the engine leave that spring. The premise was true of a fast exit; the shipped exit is slow on purpose, and the slowness already does this job.

The same law, the same exit and the same dials on the semiconductor index instead of the NASDAQ 100: the sector fund as the base from 2001, the 3× fund as the levered leg from its 2010 listing and modelled before that from the index with financing and fee. Pre-registered against ULTRA on the NASDAQ at equal drawdown over the same window. It loses by eight points a year on the hot dial and near seven on SELECT at equal drawdown, its worst drawdown is −40% against −27%, and the first era is negative outright: the semiconductor index fell 80% from 2000 and did not see its high again until 2017, so an exit that sells after two bad monthly closes and re-enters after one good one spends that decade paying for whipsaws in a market that was going nowhere. It wins the last era, 2017 to 2026, by two and a half points at a deeper drawdown, which is the sector’s bull market and nothing the engine did. The futures work showed the engine’s edge is a high-volatility-root edge; this is the reminder that the root also has to trend for a quarter century.

Scale the target down by a quarter when the account is more than 10% below its peak and by half below 20%, restoring it at a new high. Pre-registered as the idea most likely to be a dial, and it is: at equal worst drawdown it loses 1.2 points a year on ULTRA and 0.7 on SELECT, and in the two later eras it loses seven to eleven points raw, because a drawdown from peak is where the engine’s recoveries begin and the throttle holds half a position through all of them. The engine already has a drawdown response: the volatility target, which cuts exposure when the market gets rough and restores it when it calms, on the market’s own information rather than the account’s. A rule keyed to the account’s own high-water mark knows nothing the market does not and forgets it later.

Sell the at-the-money NASDAQ straddle every month, thirty to forty-five days out, both legs at the bid, and hedge the direction away every day with shares using the delta from the implied volatility in the quotes, so the return comes only from implied volatility exceeding realised volatility, not from where the market goes. Fourteen years of real chains, daily marks, expiry settled at intrinsic, hedge commissions and half a cent a share. The premium is there — implied averaged 19.4 against 18.5 realised — and it is too thin to survive its own costs: run on a book of cash, the strategy earns 1.2 to 1.8 percent a year against 2.0 for the bills it sits on, at drawdowns of 3 to 15 percent depending on size, with two hundred and sixty hedge trades a year. The good years pay a point or two; 2018, 2020 and 2022 each take it back, because a short straddle is short gamma and the realised volatility of a crash month exceeds anything implied the month before. Beside the engine it is worse than useless: it lowers return, deepens the worst drawdown by one to three points, and makes the worst days worse, since the months it loses are the months the engine loses. As for the account: one NASDAQ straddle is $72,000 of notional and about $14,000 of margin, so a $25,000 account cannot hold one. A real premium, priced fairly, hedged honestly, and smaller than the cost of collecting it.

Pre-registered before the run, then run exactly as written: every month buy a NASDAQ put spread, long 5% below the market and short 15% below, thirty to forty-five days out, spending half a percent of the account in premium, paying the ask and receiving the bid, marked daily at the mid on fourteen years of real chains, on SELECT and on ULTRA with their shipped exits. The idea was not income; it was insurance that would let the engine run a hotter dial at the same worst drawdown, with the restored return paying for the premium. It does not. The spread costs two and a half to three points a year and shallows the worst drawdown by a third of a point, so raising the engine’s dial to match buys back almost nothing: minus 1.8 points a year on ULTRA at equal drawdown, and three points raw on SELECT. The mechanism is real and useless at once: shuffle the monthly payoffs across months and the real timing is shallower than 94% of the draws — the spread does pay in the crash months — but 2022 is the only year it paid at all, and it took every other year’s premium to pay for it. The engine’s monthly exit already leaves the market in those same months, for free. Insurance bought on top of an exit that already works is a second premium for the same event. Whole contracts change nothing at $25,000 or $100,000; the structure was simply too expensive at every size.

Seventeen dividend and value equity funds with at least eight years of history — dividend quality, high dividend, aristocrats, dividend growth, low-volatility dividend, international dividend, and the older covered-call index funds — each with distributions reinvested, put through the construction the bitcoin-and-gold pair came through: a fifth of the book in the fund, either held only while it is above its own 200-day average read monthly or held always as ballast, beside the engine dialled to the same drawdown, on each fund’s own window from its listing (2015 for most) and since 2022, then added at 5% and 10% beside the pair itself. Every one lands between minus two points and plus half a point a year at equal drawdown beside the engine. Added beside the pair, the best of them earns a third of a point on its own window and nothing since 2022, and every one deepens the worst days against the same money in bills, because a dividend fund falls with the market on the days that matter: held always, its worst-day correlation is three quarters, the same as the index; gated, it is a tenth, because it is in cash. The yield is paid out of the price, so a fund that yields 4% and returns 11% is an 11% fund. Beside an engine that already holds the NASDAQ at up to twice its size, a slower piece of the same market is dilution with a coupon. Not covered by this stone, because they are not the same mechanism and do not have the history: the premium-income funds JEPI (from 2020) and JEPQ (from 2022) were run on their own short windows only as a diagnostic — JEPI reads minus a point beside the engine, JEPQ plus half a point at a correlation of 0.96, which is the engine with its upside sold — and neither is validated either way. The single-stock option wrappers and the 2025 listings (the YieldMax funds, AMDW, CHPY, BLOX) are out of scope as ballast: leveraged or option-wrapped exposure to one stock, a semiconductor basket, or crypto, one to three years old, each of which trails the thing it wraps by ten to twenty points a year at a correlation above 0.95. They are a bet on the underlying with the upside sold, and there is no eight-year test they could have taken.

The engine sizes its core by trailing volatility, so give the bitcoin and gold sleeves the same treatment: each slot scaled down when its own twenty-day volatility runs above a target, capped at the full slot, four targets from 30% to 80%, on the lagged full-history test since 2018 and since 2022, against the engine dialled to the same drawdown. Every target lands on or just below the fixed pair’s own frontier: a tighter target holds less bitcoin and earns less at a smaller drawdown, a looser one converges on the fixed 20/20, and none of them beats the fixed sleeves at equal drawdown. Volatility sizing pays inside the core because the NASDAQ’s volatility forecasts its next month; a fifth of a book in bitcoin is already a small, blunt position, and shrinking it further in its noisy months buys nothing the trend rule has not bought. A mechanism that reproduces a dial is a dial.

A home-made managed-futures sleeve: one slice of the book spread across whichever non-equity assets sit above their own 200-day average read monthly — gold, bitcoin, broad commodities, 7-to-10-year and 20-year Treasuries, silver, copper, platinum — weighted equally, or equally among the ones that are on, or by inverse volatility, at a fifth and at two-fifths of the book, against the engine dialled to the same drawdown, since 2018 and since 2022. Every basket earns less than the plain bitcoin-and-gold pair at equal drawdown, the best by two points a year since 2018, because six of the eight assets go nowhere on their own and every dollar spread onto them is a dollar taken from the two that do. Run the basket without bitcoin and gold, to ask whether a third stream is hiding in the rest, and it reads between minus half a point and plus half a point. Carve a slice for it out of the engine beside the pair and it costs one and a half to three points. Diversification across trend streams is what a futures book does with leverage and forty markets; done with eight spot funds and no leverage it is dilution. Two streams was the answer; the weekly screen keeps checking whether a third one appears.

The bitcoin candidate’s construction with a currency in the seat: a fixed 20% of the book in a currency fund — the dollar long and short, the euro, yen, sterling, the Australian and Canadian dollars, the Swiss franc, a basket of emerging currencies — held only while it is above its 200-day average read monthly, cash otherwise, against the engine dialled to the same drawdown. Every one lands within a point of nothing, from −1.1 to +0.6 a year, since 2015 and since 2022 alike. Held outright the funds return between −3 and +2.5 percent a year with volatility a fifth of bitcoin’s, so a fifth of the book in one of them moves the book by a rounding error whichever way the trend rule leans, and the rule itself has nothing to catch in a series that mean-reverts around zero. Currencies earn their keep in a futures book with leverage and carry across a dozen pairs at once; that book was tested and buried as grave 102. As a spot sleeve they are neither a return stream nor a hedge. A seat needs a different asset that goes somewhere on its own.

The bitcoin candidate’s construction with oil in the seat: a fixed 20% of the book in a crude or energy fund, held only while it is above its 200-day average read monthly, cash otherwise, executed the day after the read, against the engine dialled to the same drawdown. The crude funds — WTI, Brent, the optimised-roll version — earn between minus one point and nothing over their own history since 2015; the trend rule cannot rescue an asset that loses 5 to 10 points a year to the futures roll and spends most of a decade below its average. The energy stocks do better, and one, the producers fund, reads +2.8 points a year; but almost all of it is 2022, the inflation year, and its correlation with the NASDAQ is twice bitcoin’s. Gold, the sleeve the product already has, earns more than every oil variant but one at a third of the correlation. A sleeve has to be a different asset that goes somewhere on its own. Oil is a different asset that goes in circles.

Use the Shiller cyclically adjusted P/E, through its own expanding percentile so nothing is known in advance, to govern the engine: step out when it is in its top decile, scale the volatility target down as it rises, refuse to add exposure while it is in the top decile, or cap leverage at one while it is in the top quintile. On the S&P fund pair from 1994, its own history, and on STEADY from 2001. Every arm earns less than the engine it governs, most of them by three to seven points a year, because the percentile has sat above 0.9 for most of the last thirty years and the rules simply hold less for decades. The one arm that improves Sharpe, scaling the target down with valuation, is exposed by the placebo: shuffle the P/E history by year and the shuffled version scores the same, so its whole effect is owning less, not knowing when. Valuation tells you the decade’s return. The engine trades the month, and the month does not care.

Replace the engine’s exit — two monthly closes 5% below the 200-day average to leave, one 5% above to return — with Wilder’s parabolic stop, the accelerating trailing stop that flips whenever price crosses it. Standard settings, slow settings, slower still, fast settings, and the same stop read only at month end, all on the full 1999–2026 cycle, on SELECT and ULTRA. Every version loses eleven to fourteen points a year against the engine dialled to the same drawdown, and the fast version loses so much the control cannot reach its drawdown. The reason is the exit count: the parabolic stop leaves the market five to sixteen times a year, the monthly rule about once every nine years, and each false exit is a round trip through the band paid in missed days. A trailing stop is a fine instrument for a single trade with a defined risk. A strategy that is supposed to be in the market for decades cannot afford an exit that is triggered by every ordinary pullback.

The engine already has an oracle for bad markets: the monthly exit that takes it to cash. So when it is out, hold an inverse fund instead of cash — minus one times the index, minus two, minus three — sized as a half or a third of the account, with the daily-reset decay and the fee of the real products modelled, on the full 1999–2026 cycle, on SELECT and on ULTRA. Shorting whenever the engine is out loses at equal drawdown on SELECT and on ULTRA and, above one times, blows the worst drawdown out from the low twenties to the high thirties, because the engine is out for a long time in years that turn out fine, and an inverse fund bleeds every one of those days. Add a second condition, short only when volatility is also high, and it scrapes a point past the control on SELECT and on ULTRA — entirely from 2001 and 2008, negative through 2009–2016, and gone if the volatility threshold moves from 30% to 45%. A gain that lives in two crashes and dies when its one dial is nudged is a fit, not a finding. Random out-days beat nothing here; the real timing is the exit’s, and the engine already owns it.

Give ULTRA, SELECT or STEADY a fifth of the account in a bitcoin fund, or a bitcoin-and-ether fund, run the way the gold sleeve is run: sized to volatility, never levered, stepped aside by the same monthly exit. Bitcoin priced as a spot fund less a 0.25% fee, because the real funds are two years old; ether joins when it lists in late 2017. On the whole window, a shorter window because bitcoin has no price before 2014, the bitcoin sleeve looks like a gift: every engine earns two to four points more a year at the same worst drawdown, and random-sign bitcoin beats it in only 2% of draws. Then split the years. 2015 to 2017, when bitcoin compounded at 85% a year, supplies all of it; 2018 to 2021 is flat to slightly negative; 2022 to 2026 is negative. Rerun from 2018 alone and the sleeve adds nothing at equal drawdown on ULTRA, costs a quarter point on SELECT and most of a point on STEADY. The ether version is worse in every window. A sleeve is only insurance if its good years are the engine’s bad years; bitcoin’s good year was everyone’s good year, and it has not had another. The halving cycle was checked too: the eighteen months after each halving do carry the sleeve’s return, and that phase ranks second or third among the twenty-four offsets the same rule could have used — but the gain in that window has shrunk from 22× to 7.5× to 1.7× across the three cycles the data holds, the current window closed in October 2025, and a rule that says ‘own bitcoin from April 2028’ is a forecast, not a strategy.

The futures engine’s one weak era is a calm, steadily rising market, where a daily-reset 2× fund compounds better than a constant-notional contract. So hold the fund when realised volatility is low and the contract when it is high, at five thresholds, on the full 1999–2026 cycle after tax, at three account sizes. Every threshold loses to simply holding the contract all the time, by half a point to nearly a point a year. Each switch sells fund lots, which is a taxable event, and buys or sells a contract, and the two costs together are larger than the calm-era gap they were meant to close. The lesson is the same one the rest of this wall keeps teaching: a real pattern is not a tradable one when the act of trading it costs more than it pays.

The day-trading canon on micro Nasdaq futures, tested on thirteen years of one-minute bars, 2013–2026, a shorter window because those are the only years the broker keeps minute bars — and again on two-, three- and five-minute bars: opening-range breakouts with 5-, 15- and 30-minute ranges, with and without a stop; the first half hour’s direction held to the close; the last hour in the direction of the day; fading two and three sigma from VWAP; a 9/21 moving-average crossover. Every position is flat by the close, every trade pays a micro contract’s commission and two ticks of slippage. The breakouts lose on every timeframe and every range. The VWAP fades lose. The crossover loses badly. Two things survive as positive numbers — the opening half hour’s direction held to the close, and the last hour with the day’s trend — and both earn less than half the cash rate at a Sharpe near half, with the first one owing its whole record to 2018–2021. Changing the bar size from one minute to five flips the sign on the breakouts, which is what noise looks like. The daytime session itself earned 3% a year over the period; the return of this index arrives overnight, and a rule that is always flat overnight is fishing in the empty half.

Sell the front VIX future and collect the roll-down, and use the shape of the curve to know when: only in contango, only when contango is steep, only when the future sits above spot, switch long in backwardation, stand aside when the VIX is high. Thirteen years of the exchange’s own settlement prices, 2013–2026, a shorter window because the exchange publishes nothing earlier, front and second month rolled at expiry, 15% of the account in contract notional, after costs. Every curve-based filter earns less than simply being short all the time, and short all the time earns a Sharpe of 0.47 while owning the S&P earned 0.86 over the same years. The size dial changes nothing but the drawdown: at 15% of the account the strategy lost 17% in a week in February 2018 and 23% in a month in March 2020. One contract is $17,000 of notional, so a $25,000 account cannot hold a fraction of one anyway.

The managed-futures classic: sixteen liquid futures — four stock indexes, two Treasury contracts, gold, silver, copper, three grains, the euro, coffee, sugar, cattle — each held long or short by the sign of its own past twelve months, each sized to the same risk, the whole book scaled to 15% volatility, rebalanced monthly, after commissions and slippage, 2001–2026 (a shorter window, because continuous futures histories begin in late 2000). Returns are the matching total-return funds minus bills where a clean one exists, so nobody is fooled by roll gaps. It is real: random signs on the same positions beat it in only 2% of draws. It is also worse than doing nothing clever: the same basket held long only at the same risk earns a higher Sharpe in every era, and plain 60/40 beats it too. Shorter lookbacks are worse still, so twelve months is a ridge, not a plateau. Breakouts and cross-sectional versions land in the same place. And a diversified futures book cannot be built small: bonds and grains have no micro contracts, so a $25,000 account can express under a third of the risk budget.

Most of the Nasdaq’s return arrives while the market is closed and the daytime session has lost money on net for a quarter century, so hold micro futures from the close to the next open and sit in cash all day; futures make it possible without day-trading rules and with the kinder tax treatment. Tested on the fund’s own session bars as the proxy, 2001–2026 (a shorter window, because the fund’s own bars begin in 2000), one round trip a night, commissions and two ticks of slippage each way. The overnight leg alone does beat buy-and-hold on a risk-adjusted basis, but by a hair: at the same volatility it adds under a point a year and 0.07 of Sharpe, which is inside the noise of one era. On the S&P it is negative at equal volatility, so the two roots disagree. Sizing the overnight leg to 20% volatility lifts the return but hands back a −16.6% night, because the one thing this strategy holds is the gap. Two hundred and fifty round trips a year to buy a sliver.

Own the S&P only for the last four and first three trading days of each month, flat the rest. The old effect is still there on 2001–2026: those seven days earn most of the month’s return at two thirds of the volatility, and the other fourteen days are worse in every era but one. But the whole strategy earns 4.6% a year at a Sharpe of 0.42, sitting in cash two thirds of the time, which is less than the engine earns while also being in the market. As a filter on when the engine adds exposure it would move a handful of trades a year by a few days. Not wrong. Not worth a rule.

Sell volatility and collect the roll: the inverse-VIX fund since 2011 as the benchmark for every ‘harvest the premium’ idea. It compounded at 11% a year with a 69% volatility and then lost five sixths of itself on one evening in February 2018, a −95% drawdown it has never recovered. Sharpe 0.15 over its life. It is here because three graves above it were milder versions of the same bet, and because anyone who wants a futures strategy that pays every month should look at this line first.

If micro Nasdaq futures beat the 2× fund for the engine, micro S&P futures should do the same for the S&P sleeve, and the contract is smaller, so the rung would open lower. Same full-history test as the Nasdaq study: the full 1999–2026 cycle after tax, the S&P fund pair with the 2× fund modelled before 2006, one micro contract held at today’s size plus the fund remainder, both dials, the contract both never outgrown and outgrown from a $25,000 start, at five account sizes. It loses at every size and on both dials, by a third to half a point a year, and fails the exposure-matched control at $25,000. The mechanism explains why: the Nasdaq edge is mostly the 2× fund’s daily-reset decay in violent markets plus the tax treatment on a rule that trades often. The S&P is calm enough that its 2× fund barely decays, the tax paid is the same either way, and constant-notional futures simply forgo the fund’s compounding in steady years. Only the dot-com era was positive. Futures earn their keep on the index that thrashes; on the one that does not, the fund is the cheaper vehicle.

If the put-selling arm loses in long bears, let the core’s trend exit switch it off, or cap it at single size, once a bear is confirmed. Both roots, SELECT and ULTRA alike, the same full-cycle ladder as the graves above: switching off removes the 2008 losses and also the recoveries the arm was built to sell into, so the Nasdaq ladder earns less than a constant of the same average size; capping at single size does the same with more drawdown. The reason is timing. The exit confirms a bear only after two month-end closes, and the arm’s worst trade of all, sold in February 2001, was entered four weeks before the exit fired; the 2008 losses were sold into a bear the exit had already called. Shuffling the arm’s results onto random dates still beats the switched-off version on return. A slow exit cannot protect a fast loss, and the fast loss is the one this arm sells.

The plan for larger accounts was to run the put-selling arm on top of SELECT or ULTRA, sized at a fifth of the account, so the premium adds to the core. On the full cycle, with the same modelled-then-real put prices as the grave above, it adds about half a point a year to either engine and costs SELECT four points of drawdown. Dialling SELECT itself up to that drawdown earns 1.3 points more than the overlay does. Shuffling the same trade results onto random dates beats the real timing two times in three on return and nine in ten on drawdown: the arm loses in exactly the months the core loses, 2001 and 2008 above all, so its timing is worse than chance. Selling one put a month at fixed size keeps the return and still loses the drawdown contest to chance. And at today’s prices one micro Nasdaq index put is 1.1 times a $25,000 account, so the size tested here cannot be built below roughly $140,000. The rung comes off the ladder.

The premium overlay scoped for larger accounts sells a monthly index put 5% below the market and sizes it up as realized volatility rises — sell more insurance after the panic, when it is richest. It passed its controls on 2012–2026 because those are the only years with full option quotes. That window holds no bear market longer than three months. To see a long one, the premium was modelled for 1993–2011 from the S&P and Nasdaq volatility indexes, fitted where real quotes exist (it reproduces real trade results to a correlation of 0.998), while every trade still settles on real prices. The rule sizes to its 2.0x maximum through 2001 and again through late 2008, and the market keeps falling: on the Nasdaq it loses 37% of its collateral in 2001 and 30% in 2008; on the S&P, 36% in 2008. Its worst drawdown is roughly double that of simply selling one put a month, at every premium assumption tried, from a quarter below the model to twice it. After the fire the premium is richest; in a long bear the fire is still burning. B because the early premiums are modelled — the losses are settlements, which are not.

A small account grows mostly by what its owner adds, so: should new money wait for a dip? Five thousand dollars plus $500 a month, run through ULTRA and SELECT from 1999 and from 2009, with four ways of holding the money back — until the index is 5% off its 20-day high, until the rule itself has de-risked, until the trend exit has fired, or for a random delay of up to four months as the control. Investing each contribution the day it arrives beat every rule in the modern window: the dip rule by 1.4 points of final wealth, the de-risked rule by 11, the exit rule by 58, and all twenty random delays. Over the full window the dip rule shows a 6% gain, but so does a random delay, which beats immediate investing in two draws of three — that is the 2000 crash sitting at the start of the sample, not a rule.

A cash account can do three things with options: buy them, sell calls against a hundred shares it already owns, or sell puts with the whole strike held in cash. Priced off the live order book on 2 September 2026, every one of those on the Nasdaq-100 or the S&P 500 is between 1.7 and 14 times a $5,000 account. One cash-secured QQQ put needs $67,200; the micro Nasdaq index put $27,600, and it traded one contract that day at a 12% spread; the 2x fund’s put $8,400 at a 30% spread. The two underlyings cheap enough to fit — a $30 large-cap fund and the inverse Nasdaq fund — had no bid at all. An option is a hundred shares or nothing, and the rule sizes in single shares. Arithmetic rather than a backtest, hence the B: no test can move a contract multiplier.

The monthly trend exit waits for two month-end closes below its band. Checking weekly with a matching number of weekly closes looked like the same rule with less lag. It loses at every setting: two weekly closes give up two points a year and five points of drawdown; four give up a point; eight match the return but add drawdown on each engine tried. A faster clock re-admits whipsaws faster than the confirmation can remove them. The month-end close is the right resolution for a trend exit, and the rule ships at that clock.

Before the exit existed, faster volatility estimators lost outright: quicker de-risking bought quicker re-risking and the whipsaws cost more than the lag saved. With the exit now handling the extended bears, the question was worth asking again, and the answer changed shape rather than sign. A ten-day estimator adds about half a point a year and a hair of Sharpe — and pays for it with a worse drawdown and a worst single day of −7.5% against −5.9%. It is not an edge; it is the tail budget being spent. The twenty-day window stays.

The exit sits in cash about a sixth of the time, mostly in bear markets. Long Treasuries and gold are the textbook things to hold then. Measured from 2005, where the funds exist, both lose to plain bills: bonds give up 0.7 points a year and fourteen points of drawdown, gold adds twenty points of drawdown for a rounding error of return, and even short bills trail. The exit’s cash is not a bet; it is the absence of one, and every asset with its own drawdown breaks that.

The premium overlay sizes up when volatility is high, and the exit steps the core into cash in exactly those markets — so the obvious refinement was to sell puts only while the engine is out and the collateral is free. There is nothing to test: over fourteen years of real option chains the core was out at the moment of entry for eight of a hundred and forty-five trades, and those eight earned a third of what the others did. The overlay itself passed its timing placebo on the exit core; the gating is a non-result and is filed as one.

The cleanest idea anyone brought this year: keep full exposure while the market is open and observable, hold less through the night, when nothing can be done about a gap. No forecasting, just a refusal to rent leverage during the blind window. The decomposition killed it before the contest did. Since 1999 all of the NASDAQ fund’s return arrived overnight — the open-to-close leg lost money over twenty-seven years — and three-quarters of the damage on its worst days happened while the market was open. The blind window is the paid one. Every version that lightened overnight exposure was beaten by simply running the standard dial, and the mirror — full leverage overnight, less by day — only reinvented that dial with three hundred extra trades a year.

Under the plain exit the ugliest episode was April 2020: out a month after the low, back in a month later, a 14% rebound missed. A veto that ignores the exit when volatility is already extreme seemed to target it exactly. But the two-close rule had already removed that exit — April 2020 was a single close below the band. Under the shipped rule the only high-volatility exit left is November 2000, the one worth +192%, and every veto threshold that fired vetoed that instead.

The conditional sleeve held a trend fund only while the rule’s target weight sat below a threshold — one magic number, most of its gain from 2022. A band around the threshold might have made it robust. Twelve threshold-and-band cells; all twelve lost to plain bills, and the best beat fewer than a third of random switch schedules. With the exit in the engine, the account is already in cash during the markets the sleeve was built to cushion. Measured on a managed-futures proxy from 2007.

The trend tier beat cash on the no-exit engine, and the site said so. Measured per engine for the first time, it lost on every engine that has the exit: one to two points a year and five to nine points of drawdown, on real fund history since 2019 and a proxy since 2007. The exit sits in cash in exactly the markets the sleeve was built for; a trend fund held through an exit adds its own drawdowns on top of a defence already in place. This one changed the product: the sleeve was withdrawn from sale.

The band and the two-close rule each passed alone. Stacking a faster weekly re-entry on top, or a different band per sleeve, looked like more of the same. The grid said otherwise: the stacked rule beat the simpler one in nine of eighteen cells, its best cell was the overfit, and its whipsaw count went back up — a faster way in re-admits the churn the patient way out had removed. The per-sleeve band moved the result by a tenth of a point, which is noise. Three parameters is where fitting starts; the rule ships with two.

The idea: cap the hottest engine lower whenever VIX is calm, since the damage in a levered book comes from the spike out of quiet. Ten thresholds, two cap levels, the mirror rule and a random-day placebo. Every VIX rule left the worst single day exactly where it was, and the risk-adjusted gain rose in lockstep with the fraction of days capped — capping on VIX just meant capping most of the time, because when VIX is low the engine is already at its cap. The mirror, capping when VIX is high, did nothing at all. The worst days themselves settled it: five of the eight came out of calm, three out of stress, and the worst of all was a band-lag day the cap could not reach. VIX and the engine’s own thermometer read the same fever at the same time.

Selling a spread instead of a naked put posts only the width, so an income strategy that normally needs a five-figure account should compress into a few hundred dollars — the one structure that could have filled the gap below the options tier. It returned more than the core and looked better on every risk-adjusted measure we had. Then we shuffled the same trades onto random expiry dates. The random version scored higher. The timing was worth nothing: the flattering number came from a 93%-win-rate stream whose losses all arrive at once, which is the oldest way there is to fool a Sharpe ratio. It also loses three times what the core loses on the core’s worst days, and the configurations that actually worked needed a bigger account than the strategy they were meant to undercut.

Widening the buy-back threshold beat the shipped rule on every axis at once — and survived the bootstrap and the exposure-matched control that killed a dozen other ideas. Then we asked why, twice. The re-entries it skips outperform the ones it keeps. The whipsaw rate it supposedly avoids is identical. And across eighteen assets the effect helps most exactly where this strategy works worst. An edge that passes every statistical test but cannot say what it does is a coincidence wearing a lab coat.

The market’s own forecast of volatility — implied vol off the full chain — is the best predictor this project ever measured, six times stronger than dealer gamma. Fed into the sizing rule it lost anyway. Its errors run in one direction for months at a time, fattening and thinning with the very regime it is supposed to call, wrong precisely at the turns where the money is decided. The humble trailing estimate’s errors are pure delay, which the rule forgives cheaply; the option market’s are structural, which it cannot.