the machine room
the engine's only trade was a $40 scratch on its only open day, and the bigger change of the week was mine: new live puts are now blocked, the side that has paid for a month.
last sunday this column said the four overdue calls would be graded, or this paragraph would lead. they were not graded, so it leads
The record file has not changed since . It holds 65 calls: 10 hits, 7 misses, 2 partials, 46 open, a resolved hit rate of 57.9% with a partial counted as half. Of the 46 open calls, 12 are now past their own deadlines, up from four. The four named last week are still open: Micron's Taiwan labour test, the Western Digital volume-versus-price call, the Blue Owl credit test and the Apple letter call. The eight added since include Seagate (), Avnet (Oct 28), Arrow and Sandisk (Oct 29), and the phantom-HBM-orders call (Oct 30). The root cause: the record has a writer but no grader on the schedule. Calls go in automatically each Wednesday; they come out only when I sit down and grade them, and in a busy earnings fortnight I did not. The fix: a Sunday sweep that lists every open call past its deadline, pulls the source its own wrong-if names, and drafts a hit, miss or partial with the source attached for me to sign. Next Sunday the count of overdue calls is zero, or this block leads again.
what it did this week
Last Sunday's diagnosis said the halt should let puts through. This week the halt let one put through, and then puts were switched off.
no trade won, but the call candidates the halt blocked fell 5% over three days
The only closed trade lost money, so there is no winner to grade. The nearest thing is again 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. Counted once per stock, direction and session, the halt blocked 16 call candidates this week. The next-day read flattered them: 8 of 9 graded moved the right way, by +0.66% on average. Three closes on, only 3 of 7 had, and the average was −5.35% against the trade. Micron calls flagged on Monday near $910 were up the next day; Micron closed Friday at $884.50. Why the halt worked here: a call needs the stock to rise inside the option's life. The halt fires because the index is below trend, a reading that the next few days will not reward buyers. A one-day bounce inside that reading is the trap a call buys into.
two oracle puts, bought in the one open session, flat by the close after three and a half hours nobody was watching
On Thursday at 10:03 ET, the first run after the halt released, the options sleeve bought two Oracle $140 puts expiring Nov 6, eight days out, at $4.95 each: $990 of premium. A put gains when the stock falls; at a delta of about −0.51 this one moved roughly half a dollar per dollar of stock. The entry carried the engine's lowest conviction and no confirming signals on its row, only a flow read. Within the hour the puts were 15% under, and the shadow momentum-fail rule, which watches but does not act, would have sold at $4.20. The engine then went quiet: no options run between 11:32 and 15:02 ET, a scheduler stall (Section III). The position recovered on its own, to a peak mark of $5.60 at 15:31 ET, about +$130, and then faded into the 15:51 session-close flatten at $4.75. −$40.00, −0.04R, filed as a scratch. R is the dollar amount the trade was sized to risk; a scratch is a result inside ±0.1R.
How the guardrails behaved. The same-day flatten did its job: no eight-day option was carried overnight into Amazon's and Intel's prints. The trade itself was sized the old way. It had no dollar-loss-at-invalidation figure, and its $990 cost sits above the $750 ceiling that came back four hours later. The engine's own minimum-days-to-expiry shadow rule marked it as a trade it would have refused for being too close to expiry. That rule is still watch-only.
the options sleeve can now trade only calls, and only on days the index is above trend, so the side that has been working has no way back in
What changed on Thursday. At 15:12 ET, at my request, the deploy restored a code-enforced $750 ceiling per option entry and blocked new live put entries, in the options compute and again at the broker gate. Exits and open positions were untouched, and paper history was kept. On Friday that block refused 4 put candidates (Amazon, Tesla, Pinterest, Nvidia) before the halt had even re-engaged on them.
The root cause it creates. Put the two rules side by side. The halt blocks everything on tier-3 days. Decision 4, the written test from , would narrow that to long-only, so calls stay blocked and puts trade. The put block now refuses exactly the trades Decision 4 would release. With both in force, the options sleeve trades calls on clear days and nothing on bearish ones. The record says that is backwards. Over the last four weeks the sleeve is 6 trades, 2 wins, 1 loss, 3 scratches, +$609, and the puts alone are 2–1–1, +$703. The drawdown sits in the mid-delta bucket, contracts with a delta between 0.35 and 0.50, which are further from the money and more sensitive to time decay: 3 wins, 7 losses, 5 scratches over 15, −$1,444. The contracts nearer the money are 4–4–2 over 10, +$541.80. The losses come from how the sleeve buys options, not from which direction it buys them in.
The fix, and its rules. Puts come back by one defined route, not by lifting the block. A bearish entry goes in as a put debit spread: buy a put near the money and sell a cheaper one further down, which caps both the cost and the gain and offsets much of the time decay. It is sized from the dollar loss at the invalidation level, inside the $750 ceiling, with enough time to expiry to pass the engine's own minimum-days rule. That route runs in paper first, against the same blocked put candidates the halt grader already logs, and graded in premium. It gets its own written exit test before it touches money, in the same shape as Decision 4: five episodes, a positive average per episode, and a worst case written down in advance. The bar is set before the result, so the result cannot move it.
The halt test, for the record. The grader counted its first full episode this week: the halt that began Monday and released Thursday. All four policies, Decision 4 among them, read 1 of 5 episodes, insufficient, and the live scope stays ALL. Across the week the put candidates the halt blocked split: 4 of 12 right the next day, averaging −0.84%, and 6 of 8 right three closes on, averaging +2.51%. These grades are on the stock's close, not the option's price. The grader still prints "option marks not archived" on every row, so neither number is a premium figure.
the lessons it wrote down, and what changes
Last Sunday's five tests, marked first. They were mostly plumbing, and the plumbing mostly did not get fixed.
one open day was enough to show the sizing change never shipped, and the price archive still does not exist
Last week's tests, marked. (a) Every blocked option row carries a bid and ask: FAIL. All 97 rows in this week's halt grade read "option marks not archived". (b) Momentum-fail has a grade against CoreWeave: FAIL. The shadow exit ledger still skips CoreWeave, IREN and now Oracle for "no parseable mark". Oracle can be graded by hand from its own price journal: momentum-fail would have sold at $4.20, a −$150 loss against the −$40 the session-close exit took. That is one row, and it goes against the rule. (c) Zero $0.00 cash reads and zero empty-universe runs: FAIL. Monday was the tenth session running that the first stock pass read settled cash as $0.00. On Wednesday and Friday those $0.00 passes refused qualified names, MGM Resorts among them on Friday, before a re-run a minute later read the real balance and refused them for other reasons. Monday had four empty-universe runs and Friday one. (d) A stock run in the opening window every session: UNVERIFIED. Monday's and Tuesday's digests confirm one; the others don't state it either way. (e) The first option entry after the halt lifts carries its dollar-loss-at-invalidation figure: FAIL. Oracle carried none.
The lessons it logged. One, on sitting out: the tally of no-trade days now has 25 days where sitting out cost proxy profit against 24 where it saved. Cost days lead for the first time. The engine filed this as the caution costing money, not as a case for loosening any single gate. Two, on the stock gates: the dealer-wall refusals, which skip a stock when a large options position sits too close above the entry, are still the only stock bucket grading above live break-even: 14 wins, 6 losses over 20 decisive, 70% against a 58% break-even. Room to the wall still does not sort winners from losers, so the engine keeps its proposal to shadow-test the threshold rather than move it. Three, on the stock sleeve's open day: Thursday was clear and it still entered nothing. It refused 10 names at the scalp gate, which requires a volume surge with bullish flow behind it, 5 on stale strategy-map exits, and 1 on the $10 payoff floor.
What it adjusted about itself. Nothing new. The ten one-point confidence markdowns stand, and no new flag was raised, because one scratch moves no bucket.
stop the afternoon stalls, archive option prices, write the put-spread route down, and size by the stop
One: the scheduler stall comes first. Four afternoons lost and one live position left unwatched for three and a half hours is the most expensive fault in the stack, even in a $40 week. The fix for the file-lock failure is staged; it goes in, and every scheduled run reads its inputs from a local copy rather than a cloud-synced one. Two: archive the option price on every run. Any run that blocks, enters or shadows an option records the contract's bid and ask. Without it, the halt test, momentum-fail and the put-spread route all grade on stock prices. Three: write the put-spread route as a paper study (Section I), with its exit test in writing before the first paper trade. Four: risk-based option sizing, carried a second week. Contract count comes from the dollar loss at the invalidation level, under the $750 ceiling. Oracle showed what happens without it.
How we'll know it worked next Sunday: (a) every session has runs through the close on both sleeves, and a session-close row exists every day; (b) the halt grade prints at least one line in option premium; (c) the put-spread study has a written test and at least one paper entry; (d) zero $0.00 first-pass cash reads; and (e) any option entry carries its dollar-loss-at-invalidation figure and a cost under $750. If the halt holds all week, the column says so on the first line.
a put block and a cost ceiling went live, and the scheduler lost four afternoons
One deploy touched a live gate this week. The rest of the trouble came from files the scheduler could not open.
the only live change was a restriction, and the worst fault was the one that stopped anything from running at all
| change | what it does | status | what it did this week |
|---|---|---|---|
| Put block + $750 ceiling Thu Oct 29 | Blocks new live put entries in compute and at the broker gate; caps each option entry at $750. | live | Refused 4 put candidates on Friday. The weekly grader logs those rows as "unclassified", an accounting gap. |
| Afternoon stalls Mon, Tue, Wed, Thu, Fri | Scheduled runs refuse to start when a project file on the cloud drive is not downloaded to the Mac. | fix staged, not stopping it | No power hour or session close Monday, Tuesday, Wednesday or Friday; a 3.5-hour midday gap Thursday with Oracle open. The book was flat each time. |
| Halt release and re-arm Thu – Fri | The breaker cleared Thursday and re-engaged before Friday's open. | working | The first full halt episode, Monday to Thursday, closed and was counted by the grader. |
| Momentum-fail grading carried | Grades the shadow exit against each closed option. | blocked | Skipped CoreWeave, IREN and Oracle for a missing option price. |
Two notes from outside the engine. The separate options screener, which publishes paper ideas and never trades, sized one card this week: an Eli Lilly call spread that fired on Thursday with the stock already through the spread's upper strike, so most of its possible gain was gone before the paper fill. It is the third chased fill in three weeks, and the rule to void such cards was written twice and never wired in. Second, the weekly halt grader still marks every halted candidate as sitting on a "frozen" sleeve. That is the same fault flagged last Sunday, still open.
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.
sk hynix printed an 80% operating margin and said its memory is sold out to 2029
SK hynix reported on Wed Oct 28: revenue of ₩104.6 trillion and operating profit of ₩83.3 trillion, an 80% margin, up 38% on the quarter and ahead of the ₩76.7 trillion consensus (CNBC, LSEG). It said DRAM and HBM output is effectively sold out through 2029, with 14 customers on long-term agreements, and that HBM4 passes HBM3E in shipments and revenue this quarter (TrendForce). Server DDR5 sales rose about 40% on the quarter, and the board approved a ₩20 trillion buyback. The semiconductor and AI group then fell 6.98% on Friday. Last week's checkpoint asked whether HBM rose on volume or on price. The release gives no HBM volume split I could verify, and the analysis available credits price for most of the gain. The call is not confirmed.
fortinet cut its service forecast, and the group gave back the prior week's rebound
Fortinet beat third-quarter estimates but cut its service revenue forecast, and on Thu Oct 29 the stock fell about 20% (24/7 Wall St.). Palo Alto fell about 5% and CrowdStrike about 4% in sympathy, and both fell again on Friday, to about $368 and $250. Datadog lost about 12% the same day. On the engine's sector feed, cyber fell 5.50% for the week and software 5.88%, giving back the rebound from Oct 16 – 23. A firewall vendor guiding service revenue lower hit the platform names as hard as it hit Fortinet. That is the group moving together again, the opposite of the dispersion call.
pjm asked for a march auction for data-center power, and oracle flagged a project that may miss 2028
On Fri Oct 30, PJM filed for a special auction in March to buy new generating capacity for data-center demand, after FERC called the original plan deeply flawed and pushed it back five months (Bloomberg). The auction still needs FERC to approve new rules on who pays. Oracle sent a force-majeure notice on its Project Jupiter data center in New Mexico, citing a risk it will not be online by 2028 as planned (Insurance Journal). In Arizona, SRP now makes a data center pay 60% of the cost of its grid upgrades before it builds anything (KJZZ). Each of these moves the cost and the timing of power onto the developer, so the money is going to power that projects arrange themselves before the building goes up.
And the newsletter, on itself: the record's state is in the block at the top of this column, because last week's promise put it there. One thing to add. Three of the overdue calls are memory calls whose evidence arrived this week, in SK hynix's print and the other memory reports of the same fortnight. They can be graded now, not in a month. Most of the work is already sitting in the data.