September 6 update: All eleven live entry plans below have been superseded by Campaign 1 for the September 8 window. The new plan requires confirmation and company evidence; it removes scheduled accumulation. The original calls remain here as the historical record.

Summary

Screening the 1,163 names carrying a September 4 close for durable structure — trading above the 200-day average, beating the S&P 500 over both one and three years, not yet overbought — returns 291 candidates, and 85 of them sit 20% to 55% below their own highs. The strongest of them are the ugliest to buy. Applied Materials sits 38.5% below its 52-week high, Corning 43.2%, Western Digital 41.6%, and not one of the three has held a range tight enough for long enough to produce a defined breakout level. Structure and entry run on different clocks. A jagged chart is not a reason to pass, but the desk's own backtest says it is a reason to own the name as one small position in a basket, not at the size a clean entry earns.

The proven-but-jagged list

Excess return is stated in percentage points against the S&P 500 over the window named. Every row clears both windows and sits above its long-term average; every row is also well off its high with no base underneath.

The missing base is the actual information, not a defect in the reading. A defined breakout level requires roughly six weeks or more inside a narrow range with little net drift. These names have been travelling too far each month to build one. Waiting for that level means waiting an unknown number of weeks while the trend that qualified them keeps running.

Alphabet is the fragile row. At 0.7% above its 200-day it is the only name here whose structural qualification could be gone in a single session.

What a clean entry looks like, for contrast

TSMC clears the same durability test and has the entry the list above lacks: $428.91, 15.0% above its 200-day, 10.5% off its high, 281.9 points of three-year excess return. NVIDIA closed 2.6% off its high with a defined breakout level at $236.54 — 2.7% away. Both are single-tranche decisions. The eight names above are not.

What buying the jagged list has actually paid

The desk ran the setup back over 129 month-ends from 2015 to 2025 across 995 names with a full daily history: the same durability screen, split by how far each name sat below its 52-week high at entry. Twelve-month return in excess of the S&P 500:

The median jagged entry loses to the index by nine points over a year, and the hit rate falls the longer the position is held: 47.1% at one month, 43.0% at twelve. The mean looks generous and cannot be owned. Strip the top 5% of outcomes and it drops from +22.25 to −3.29; that slice carries the whole figure. 38.4% of jagged entries lost to the index by more than 25 points, the tenth-percentile outcome is −66.6 points against −33.3 for the clean cohort, and the setup fired most often in 2021, the year it was most dangerous (397 entries, 21.4% hit rate, median −39.8). This is a lottery distribution, not a compounding one. The edge is real and lives entirely in a thin right tail, which is only ownable across enough small draws to catch one.

Two caveats. The universe is the desk's current coverage, so names that fell 40% and kept falling until they were dropped are absent, and the true jagged distribution is worse than the table. And crowding is a caution flag, not a signal: 27.6% of today's durable names sit in the jagged bucket, the highest reading in the twelve years measured, but 2025 was crowded too and worked.

Seven of the eight rows in the first table sit in that bucket; Alphabet, 17.2% off its high, is in the middle band. They are basket candidates, not core positions — seven small positions, not three conviction-sized ones. TSMC and NVIDIA sit in the clean cohort, the one that behaves the way a multi-year hold expects: positive median, an even hit rate, half the downside tail.

The ETF version has one trap

SOXX and SMH cover nearly the same complex with nearly the same numbers. The difference sits inside the wrapper: NVIDIA is 9.4% of SOXX and 22.6% of SMH. Bought as semiconductor exposure, SMH is a concentrated NVIDIA position wearing a sector label — a distinction that only surfaces when NVIDIA and the rest of the complex stop moving together.

The same screen rejects more funds than it passes. IGV, the software fund, trails the index by 20.8 points over one year and has no base. Energy funds carry a +25.2 one-year excess return against −31.0 over three years: a cyclical bounce, not a compounder. Neither belongs in a position sized for years.

The edge case is IREN

IREN fails on one criterion only. It closed at $44.68, 2.1% below its 200-day average of $45.66. Everything else passes — 841.7 points of three-year excess return, RSI 57.2, $1.6B of daily turnover.

The filing is further along than the chart. IREN's FY2026 Form 10-K, filed August 27, reports revenue of $707.0M and a net loss of $702.6M for the year ended June 30 — against $2.10B of operating cash flow and $5.90B of cash. Read the cash-flow statement before treating that $2.1B as earned: $1.84B of it is an increase in contract liabilities — customers paying in advance for AI-cloud capacity not yet delivered, with deferred revenue going from nil to $1.80B on the balance sheet in one year. Strip the prepayments and operations generated roughly $0.26B of cash. The loss itself is mostly non-cash — a $638.8M impairment of long-lived assets, $417.7M of depreciation and $205.0M of stock compensation against a $1.05B operating loss, narrowed to $702.6M by $338.0M of non-operating income — so the business is not burning money, but it is not self-funding either. Company-reported AI Cloud Services revenue rose eightfold to $128.8M over the same year.

The honest risk is two lines further down the same statement — capex of $3.00B against about $0.26B of operating cash before customer prepayments. The buildout is running on prepaid contracts and the balance sheet, not on the cash the business generates, and both the $3.4B NVIDIA cloud agreement and the 800MW Australian data-centre commitment extend that spend.

The stock rose 26.1% in the week to September 4, and all ten names in the neocloud group rose with it. What IREN has not done is reclaim its 200-day. The gap between the story and the structure is 2.2%.

Nebius is the same lane with the inverted profile: 46.3% above its 200-day and 230.1 points of one-year excess return, priced accordingly at 24.5% off its own high. One has the better business; the other has the better chart.

Desk Call

Revised the day of publication, twice. The Applied Materials line was a plain scale-in buy before the desk measured how the jagged setup pays; the stance stands, the sizing does not. The backtest was then re-run the same day on daily, complete price histories after two input faults were found — every figure in the distribution section moved slightly, none changed the read, and Applied Materials' 200-day corrected from $406.28 to $407.06 — the first number had been averaged over a year of dividend-adjusted closes from one provider while the rest of the list ran on split-adjusted closes from another. The other six jagged names were added to the Desk Call as the rest of the basket, so the basket is tracked as one.

Sources

• IREN FY2026 Form 10-K, filed 2026-08-27 — revenue, net income, operating cash flow, the increase in contract liabilities, deferred revenue, impairment, depreciation, stock compensation, non-operating income, capex and cash from the SEC EDGAR XBRL company facts for CIK 0001878848 (accession 0001878848-26-000052); cached statement: key-financials, company-facts.
• IREN investor news for the AI Cloud Services figure and the contract announcements.
• Price, trend and index-relative return data: the September 4, 2026 close, *.
• The durability screen and its cohort split: the desk's durability screen on the September 4, 2026 close, the desk's own tooling (durability-screen over *).
• The jagged-entry distribution: the desk's September 5 backtest, 2026-09-05-markup-pullback-entry-distribution-study (method, full tables, by-year regime read, survivorship caveat, the same-day correction) and its sibling script 2026-09-05-markup-pullback-entry-distribution-study over max (daily histories only, repaired through the September 5 fetch); doctrine entry markup-pullback in concepts.