the machine room
the engine did not trade all week because its market breaker stayed on, and the breaker blocked the puts that would have worked along with the calls that would not.
what it did this week
Last Sunday's best week on the ledger came from two puts. This week the engine was not allowed to buy one.
there was no winning trade, but the halt did keep the engine out of the calls, and the calls would have lost
No trade closed, so there is no winner to grade. The nearest thing to a win is what the breaker kept out. The governor's weekly grader logs every option candidate that cleared the engine's own entry tests and was stopped only by the halt. It then checks what the stock did afterwards, signed in the trade's direction: up is good for a call, down is good for a put. Deduplicated to one row per stock, direction and session, the halt blocked 31 call candidates this week. Of the 23 with a next-day close, only 8 moved the right way, and the average was −1.67% against the trade. Of the 9 with three closes behind them, 2 moved the right way, averaging −2.54%. Alphabet calls flagged on Tuesday at about $342 were 5% to 7% under water three sessions later, and AMD calls flagged on Monday at about $545 were 6% to 8% under. Why the halt worked here: a call needs the stock to rise. The halt fires because the index is under its trend line, which is a bet that most stocks will not rise in the next few days. On the calls, that bet was right.
the same rule blocked 31 put candidates, and the puts were the trades that would have paid
No trade lost money either, so the loss to grade is the one the halt produced: trades it did not allow. Over the same week it blocked 31 put candidates. Puts gain when a stock falls. Of the 24 with a next-day close, half moved the right way, averaging +0.42% in the put's favour. Of the 12 with three closes behind them, 10 moved the right way, averaging +4.25%. CoreWeave puts flagged on Monday and Tuesday at about $62 had the stock 7% to 14% lower three sessions on. Marvell puts flagged on Monday at $236 to $245 had it 5% to 9% lower. That was also the setup behind last week's two winners, a CoreWeave put for +$620 and an Intel put for +$385: puts bought into a falling tape and held through a gap. The semiconductor and AI group fell 5.3% on the week, per the engine's own sector feed, and the separate options screener's one bearish idea, an Nvidia put spread, was up 31.6% at Friday's mark on paper. That screener has its own quality bar and placed nothing either.
How hard to lean on this. The grader prices these candidates off the stock's closing price, not off the option. It models no premium, no time decay, no bid-ask spread and no fill. A 4% move in the stock does not mean a 4% gain on a put. The candidates are also correlated: the same falling tape drove most of them. And it is one halt episode, the one that began , which is still open. The grader books no dollars saved or lost for the week, and this column will not invent a figure for it.
the breaker reads a falling market correctly and then blocks both sides of it, so it stops the trades its own reading favours
The root cause. The breaker answers one question: should the engine trade at all? Its input carries direction. An index below its 50-day trend line says buyers have lost control for now, which argues against calls and long stock. The same reading argues for puts. Scope ALL ignores that. It treats a bearish signal as a reason to stop rather than a reason to lean bearish. The options engine already honours a narrower scope: under LONG_ONLY, calls are refused and puts pass. That narrower scope is already built into the engine and switched off. The only thing blocking this week's puts was the scope value the governor sends.
The fix, and its rules. The fix is to change what the halt blocks, not to lift it: tier 3 becomes LONG_ONLY, so calls and long stock stay blocked and defined-risk puts trade. That change already has a written test. It is "Decision 4", adopted into the governor's promotion criteria on . It needs five independent halt episodes, at least 60% of them favourable, a positive average on both the per-episode and the pooled count, no single episode carrying more than half the gains, and a worst episode no worse than −1.5%. This week's grade is 0 of 5 counted, because the only episode so far is still open. One good week does not meet that test, and the live rule stays ALL until it does and I decide to promote it. Setting that bar high is deliberate.
The second root cause is in the grade itself. The grader measures the stock's move because the engine does not save the option's price on a halted run. The quote field on every one of this week's rows reads "option marks not archived". The same gap held up last week's plan in another place: the shadow momentum exit could not be graded against CoreWeave, because its rows were skipped for "no parseable mark". Both tests are stuck on the same missing number. A put case built on stock prices will look better than one built on premium paid, after decay and spread, so the gate should not be fed it. The fix: every run that blocks an option candidate saves that contract's bid and ask. Then the halt's counterfactual can be graded in dollars, and momentum-fail in actual premium.
Stocks, for the record. The halt cost the stock sleeve nothing. Its own entry tests refused all 260 candidates before the breaker was ever read: the scalp gate 146, the chart veto 24, signal count 18, reward-to-risk at the order price 18, room to the nearest dealer wall 17, the $10 payoff floor 13, and the rest in single digits. The scalp-gate refusals grade about flat in shadow: 166 graded, 55 wins, +0.05R average, where R is the amount a trade was sized to risk. The stock sleeve's problem is still the one in last week's diagnosis. Its wins are too small to pay for its stops, and nothing this week tested the sizing change.
the lessons it wrote down, and what changes
Last Sunday's five tests, marked first. A week with no entries cannot pass most of them.
a halt week shows what the gates did, and this one shows the halt has a side
Last week's tests, marked. (a) Every option entry carries a dollar-loss-at-invalidation figure, within 1.5× of each other: UNTESTABLE, because no option was entered. (b) Momentum-fail has a written grade against every carried option, CoreWeave named: FAIL. The shadow exit ledger grades the trailing stop on CoreWeave and Intel, worse than the actual exit by 1.21R and 0.77R. Every momentum-fail row is skipped for a missing option price (diagnosis above). (c) Zero exits filed by hand: PASS, but only because there were no exits. The order-reference fix is still untested against a real close. (d) Every stock scalp entry shows at least $10 of net upside: UNTESTABLE, with no entries. The floor itself refused 13 candidates over the week, among them AbbVie, UnitedHealth and Credo on Tuesday. (e) One stock exit on a profit rule: NOT MET. No stock trades, so the run without a target-hit exit since stands at 22 closes.
The lessons it logged. One, on the gates: the dealer-wall refusals are still the only stock gate bucket grading above the live break-even rate. They stand at 12 wins, 5 losses, 3 scratches over 20, but the margin narrowed when Delta, refused on Oct 21 with the lowest room in the bucket, lost a full 1R the next day. Two, on budget: trades the per-trade risk cap sized out stand at 7 wins in 12. The engine filed that as evidence about rotating out of stale carried positions, not as a reason to raise the cap. Three, on the halt: the grader's three shadow alternatives agree with this column. They are a halt only when both indexes are below trend, no halt at all, and a release on a rebound. Each would have released calls along with puts, and each graded net adverse on the next day's close. The no-halt arm averaged about −0.6% per row against the trade, over 47 graded rows. Read off stock prices, that says loosening the whole halt is wrong. It says narrowing the halt deserves the test it now has.
What it adjusted about itself. Nothing new. The ten confidence markdowns from last week stand, one point each, with no trades to move them.
save the option price on every blocked run, fix the two input faults, keep long-only in its test, and have sizing ready for the first entry
One: save the option marks. Every run that blocks an option candidate, for any reason, records that contract's bid and ask. Decision 4 and momentum-fail then grade in premium, not in stock price. Two: fix the empty $0.00 cash read. For nine sessions in a row the first stock pass has read settled cash as $0.00. The Friday digest found the cause. A capital snapshot older than about 20 minutes is failed to zero instead of refreshed, and the second pass a minute later reads $3,900. This week no candidate got far enough to be sized, so it cost nothing. In an open tape it would refuse real entries for a bookkeeping reason, the failure this column called out on . Three: fix the empty universe and the coverage holes. On Oct 21, 22 and 23 the caller sent the stock engine an empty candidate list; on Friday that happened at 10:33 and again at the 15:48 close pass. Friday it also missed the opening window, and had a two-hour gap from 13:35 to 15:34 ET. Four: risk-based option sizing, carried. Contract count comes from the dollar loss at the invalidation level, not from free cash. It has to be live before the first entry after the halt lifts, so the first trade out of the halt is not sized the way IREN was.
How we'll know it worked next Sunday: (a) every blocked option row carries a bid and ask, and the grader prints at least one premium-based line; (b) momentum-fail has a grade against CoreWeave; (c) zero $0.00 first-pass cash reads and zero empty-universe runs; (d) a stock run in the opening window every session; and (e) if the halt lifts, the first option entry carries its dollar-loss-at-invalidation figure. If the halt holds all week again, the column says so on the first line, with the put and call grades beside it.
a watcher for the halt, a written test for narrowing it, and two input faults it surfaced
Nothing touched a live gate this week. Most of the work went into watching the halt and measuring it.
the halt now has a watcher and a test, and the dark week exposed two faults that busy weeks hide
| change | what it does | status | what it did this week |
|---|---|---|---|
| Halt-flip watch Thu Oct 22 | Reads the breaker every 30 minutes and switches to every 10 minutes when a release looks close, using distance to the trend line, streaks and volatility. Read-only: it can never change the halt. | live | Installed 07:47 PT Thursday. Replayed over 30 past sessions, it went on high alert before all four past flips while doing so on 14% of reads. It flagged no release this week, because none came. |
| Decision 4: halt scope ALL → LONG_ONLY Wed Oct 21 | The written test for letting puts trade on tier-3 days (Section I). | criteria adopted · live rule unchanged | Graded 0 of 5 episodes counted. The grader for it was reviewed by a second model and passed on round two. |
| First-pass cash read 9 sessions | A stale capital snapshot is failed to $0.00 instead of refreshed. | cause found, not fixed | Cost nothing this week, because no candidate reached sizing (Section II, item two). |
| Empty candidate list Oct 21 – 23 | The caller hands the stock engine zero names; on Friday at 10:33 and at the close pass. | open | Friday, three runs screened nothing. The halt hid it, since nothing would have traded anyway. |
One instrumentation note. The weekly governor grader filed all 289 halted option candidates as sitting on a "frozen" sleeve, one that could not have traded whatever the halt said. The engine's own logs show entries enabled, with no freeze, on every one of those runs. The likely reading is that the grader counts the halt's own manage-only mode as a freeze. It goes on next week's list, because a grader that cannot tell a halt from a freeze will also be unable to show what lifting the halt would have earned.
the three verticals, in brief
The newsletter's demand-and-credit read on semiconductors, cyber, and construction: this week's headline, and the dated checkpoint that would confirm it or crack it.
western digital grew on price again, and the memory majors report next week
Western Digital's fiscal first quarter, covered in this week's deep dive, brought revenue of $4,381M, up 55%. Exabytes shipped grew 22% for a second quarter running, while price per terabyte rose 28%. The December guide of $5.1B needs price up about 35%. The semiconductor and AI group fell 5.3% on the week, and Nvidia closed Friday at $210.97. Ahead of SK hynix's third-quarter report, consensus operating profit sits near ₩77 trillion, which would be a sixth straight record (Herald Business). Bernstein estimates Samsung passed SK hynix in HBM revenue in the quarter (Investing.com). Three memory prints in a month have now grown faster on price than on volume. The question for next week is whether HBM breaks that pattern or follows it.
the group rebounded as one, and the dispersion checkpoint lands in the middle
Security stocks recovered most of the prior week's selloff. From the Oct 16 close to the Oct 23 close, CrowdStrike rose 13.3%, SailPoint 16.2%, Palo Alto 11.3% and Okta 13.8%. Friday alone CrowdStrike gained about 8%, after an Evercore ISI survey of channel partners found security budgets firming, with spending moving toward identity and data protection (24/7 Wall St.). The checkpoint carried since finally got computed. Measured from the Sep 18 close, Okta is up 16.8%, SailPoint 9.9%, CrowdStrike 8.4% and Palo Alto 4.7%. That is a best-to-worst spread of 12.1 points: above the 10 that would retire the dispersion call, and below the 25 that would confirm it. The four names moved within five points of each other this week, so this week did not separate them.
the grid operators and the white house both said data centers should bring their own power
On Wed Oct 21, PJM told FERC's reliability conference that it needs closer coordination with data centers. It cited the emergency procedures grid operators used during the Oct 12–14 heat event (EnerKnol). The same day, President Trump posted that new data centers must make their own electricity and pay their host communities (Northeast Times). That follows FERC's rejection on Oct 9 of most of PJM's plan for connecting large loads, including an interim service that would have let data centers connect without new capacity (Utility Dive). Texas still has about 474 GW of large-load requests in ERCOT's queue, roughly 90% of them data centers. Whoever pays for new generation, the developer now needs a power plan before it can count on a grid connection. That shifts spending toward on-site generation and away from waiting for the grid.
And the newsletter, on itself: last Sunday this paragraph said the missing calls would be in the record file this week, or the paragraph would lead the column. They are in. On the record took in 60 calls, every weekly call from Issue №003 to №022 and every dated earnings call, each with a right-if and a wrong-if. The file now holds 64 calls: 10 hits, 7 misses, 2 partials, 45 open. That is a resolved hit rate of 57.9%, with a partial counted as half. The new problem is one the file shows itself: four open calls are past their own deadlines. They are Issue №022's Micron Taiwan labour test (), the Western Digital volume-versus-price call (Oct 22, and the company has now reported), the Blue Owl credit test () and the Apple letter call (). A call left open past its date is worse than a miss, because it never gets marked. Next Sunday those four are graded, or this paragraph leads the column.