the machine room
the first green week this column has graded, +$33.88, and the most expensive thing in it was an exit the engine decided on, logged, and never actually made.
what it did this week
After a week of zero fills, both sleeves traded. The profit is small and it is real, but the week is worth grading for a different reason: the stop-loss side of the engine behaved for the first time, and a new kind of failure showed up exactly where last week's diagnosis said to look, at the exit.
two lam research calls bought and sold inside one afternoon, and the reason it worked is that it never slept on them
At 11:55 ET the options sleeve bought two Lam Research $305 calls expiring that Friday at $6.05 each, with the stock at $302.99: $1,210 of premium, three days to expiry. The entry case was flow: two screeners ranked it bullish, the options tape agreed, and the semiconductor-equipment group was rotating in. The stock climbed about 2.3%, the calls marked as high as $9.225, and the take-profit rule (sell once the option is up 50% on what was paid) fired at 15:45 ET and filled at $9.10: +$610, within about $25 of the best price the position ever printed.
The logic of the exit is the whole trade. A three-day option loses a large share of its remaining value every night it is held, because the time left to be right is most of what the buyer is paying for. That decay, called theta, is the cost that has sunk this sleeve all summer. This position was bought and sold in the same session, so it never paid an overnight charge. The engine's own shadow rule for option expiries (a test arm that refuses anything under 30 days) would have blocked this trade, and it won. That is not a case against the rule. It is the case for the narrower version of it: a short-dated option is a same-day instrument, and the damage comes from holding it past the bell. The honest caveat: the name had been sized and ready since 10:07 ET, and a chain of repairs to the order path set the entry time rather than the signal. The full story is in Tuesday's build note.
And the stock-side win answers a question this column has asked for three weeks. Palo Alto, bought at $383.70 on Wednesday, exited at $392.59 for +$17.77, +1.82R, on its designed +1.5R profit target. Out of twenty-five stock trades it is the first one to leave on a profit rule rather than on the clock.
the engine decided to sell the nvidia calls at 3:49 on wednesday, the sell never filled, and it found out the next morning after the gap
At 10:08 ET Wednesday the options sleeve bought two Nvidia $230 calls expiring Oct 16 at $6.20: $1,240, 23 days to expiry, bought close to the day's high. The trade came with an invalidation line (the stock price that proves the idea wrong): $227.05 on Nvidia itself. Nvidia crossed it late in the day, and at 15:49 ET the rule fired correctly and sent a sell at $5.15. That order was cancelled unfilled. The engine's position record had already moved the trade to "closed, exit pending," and nothing reopened it, so the calls were carried overnight with no exit working. Thursday Nvidia opened near $222.10, well through the line. The rule re-fired at 10:04, a $3.80 sell was cancelled in turn, and a re-priced order filled at $3.65 at 10:31: −$510, −0.41R.
The split, as the trade review wrote it down: about $300 of the loss is the overnight strand, the gap between the $5.15 it meant to sell at and where it was able to sell Thursday. About $90 is exit timing: Nvidia held its opening low on the 20-day average, the whole market squeezed at 12:14, and the calls touched $4.10. The stock never got back above $227.05, so the exit rule was right on price. The rule was right; the order carrying it out was not.
The same night, on the stock side: Atlassian, bought at $194.56 Wednesday, had its end-of-day sell cancelled unfilled too. To place that sell, the engine had first cancelled the protective stop order at the broker, and it never re-armed it. So from Wednesday's close to Thursday's open Atlassian sat with no stop at all, and the engine's own record said it was protected. It was flattened at 09:38 Thursday for −$3.37. Two positions, one night, one defect: the engine treated "I sent the exit" as "I am out."
the trade passed its reward-to-risk check at the last printed price and failed it at the price actually paid
At 09:41 ET Friday the stock engine bought two shares of Eaton at $446.95, stop $441.08, target $452.30. The entry check asks that the distance to the target be at least as large as the distance to the stop. At the last trade price the ratio read 1.10, a pass. But buy orders go out a little above the ask, and at the actual fill the ratio was 0.91, a trade that should have been refused. Worse, the target sat just past $450, the strike where options dealers held the most call exposure, a level that tends to act as a ceiling. Eaton topped at $447.33 and hit the stop at 10:06; the resting stop at the broker filled at $440.87. That same morning the check was moved to measure at the price the engine can actually pay, and a second gate now refuses a quick trade whose target sits beyond that kind of ceiling. Eaton is the trade both were written for.
the loss side is finally being held; what fails now is the handoff between deciding to exit and being out
Start with what improved, because it is measurable. Last week this column said the engine's losing trades ran far past their stops: four of fourteen losses were worse than −1.5R. This week's four stock losses were −1.04R, −0.72R, −0.48R and −1.14R. The two that hit a stop, Lam Research on Monday and Eaton on Friday, were closed by a stop order resting at the broker rather than by a scheduled run noticing late, and both filled within a seventh of a unit of where the stop was drawn. That was fix two of last week's three. On the win side, Palo Alto's target exit moved the realised payoff (average win divided by average loss) from 0.56 to 0.76 through Thursday, which lowers the hit rate the engine needs to break even from 64% to about 57%. It is still right only 26% of the time, so the gap is narrower and still wide. The win side is six trades and one of them carries the whole move, so the grader treats 0.76 as provisional. That is correct.
Now the root cause of the week's biggest loss, which is not bad luck and not the entry. The engine models an exit as one event: the rule fires, an order goes out, the record says closed. At the broker it is two events, sending a sell and getting it filled, and anything can happen in between. Late in the session, a limit order priced at the middle of a wide option spread often does not fill before the close. On Wednesday that gap cost $300 on Nvidia. On Atlassian it was worse in kind, if not in dollars, because cancelling the stop to make room for the sell left the position with no protection and a record claiming protection. The same week, Palo Alto's record never showed its profitable exit, and a zero-price quote made a phantom stop-out appear in the log at −78R while the shares were still held. All four are one design fault: the engine's record reflects what it intended, not what the broker confirmed.
The fix, and it is specific. One: an exit is not complete until a fill is read back. Any sell cancelled unfilled has to put the position back to open with an alert, and any cancelled stop has to come back off the protected count. Both landed Thursday and Friday for the cases that bit this week (Section III). Two: late exits have to be priced to fill. An invalidation after 15:30 ET should go out at the bid, not the midpoint, with a re-price ladder that starts early enough to fill twice before 15:50. Paying one spread is cheap next to a gap. Three, on the options side, and the structural one. Tempus AI, bought Friday for −$50, is the clean example: the engine bought the $85 call with the stock at $84.72, and $85 was the strike carrying the most dealer call exposure, the level most likely to cap the move. A single bought call at a ceiling pays full price for upside the ceiling takes away. The right structure for that read is a debit spread: buy a lower strike and sell the $85, so the ceiling becomes the profit target instead of the obstacle, at a lower net cost. The new wall-room gate now refuses the naked version. The spread would let the engine take the trade. The order path still places one leg at a time, which is why the fix remains open after eight logs.
the lessons it wrote down, and what changes
Last Sunday this column set five tests. Here they are, marked first, because a plan nobody grades is a wish.
the tally that measures sitting out finally moved, and it moved against the gates
Last week's tests, marked. (a) No "no opportunity" row filed off an entirely stale input set: MISS. Seven of this week's accounting rows had zero fresh candidates and all fourteen input feeds stale or missing, and were still filed as "no opportunity": three on Wednesday morning, three Thursday afternoon, one Friday. Four other rows were correctly filed as input failures, so the category works when something chooses it. The precedence rule that should choose it automatically has not shipped. (b) The volume-surge reading as a field on the shadow row: MISS. It still lives only inside the text of the gate's refusal. (c) A halt released intraday: NOT VERIFIED. The intraday reading ran all week and was confirming the calmest regime tier by 07:57 PT Friday. This run did not reconstruct whether any morning this week started halted, so it claims nothing. (d) The counterfactual day tally moves past Aug 24: PASS. (e) A trade closes on a rule, not the clock: PASS, Palo Alto on its target. Atlassian and Fortinet still closed on the clock.
The lesson from (d) is the biggest one this week. The tally that asks whether sitting out was the right call had been stuck, because it depended on a flag a scheduled step stopped writing. On Sep 26 the grader started deriving that flag from the position files instead, checked the derivation against 48 days where the flag did exist (46 agreed), and backfilled. The result reversed the record: 39 no-trade days, 20 where sitting out saved money and 17 where it cost money, and a net of +10.59R in trades not taken. It used to read 17 saved, 9 cost, −6.29R. The engine wrote the right caveat next to it: that +10.59R is priced at the designed 3 : 1 payoff, and the engine realises 0.76 : 1, so much of it would not exist on live execution. That is a reason to re-price the scorecard at the real payoff, not to open a gate. This week's own shadows point the same way: of the 24 trades it declined Monday through Wednesday, graded on the following sessions, 9 would have won and 10 lost, +5.33R at the design payoff. Wednesday's six alone were +2.67R, including Palo Alto, which it also took for real.
What it adjusted about itself. Five confidence markdowns stay in place, one point each: the swing profile (14 trades, 23% hit rate), chips and AI (6), cybersecurity (8), the most common signal combination (7), and trading in a risk-off tape (24). No new flags. On the options side the grader logged the finding that matters most: of 12 losing or flat option trades, 11 died on the 50%-of-premium stop or on invalidation, not on slow decay. And options with under 29 days to expiry are 1 win, 8 losses, 1 scratch at −0.31R, against 2 wins, 3 losses at −0.09R for longer-dated ones. That is a structure problem, not a run of bad luck. The instrument turns ordinary moves in the stock into outsized moves in the premium, and the short-dated version does it fastest.
make the record agree with the broker, then let the stops move up
One: exits are read back, not assumed (live). An unfilled option exit now reopens the position and raises an alert that leads the next status card. A stop the engine or the broker cancels is now caught on the next run and replaced. Next: late invalidations priced at the bid with a re-price ladder, so no exit is waiting on a midpoint fill at 15:49. Two: the gates measure at the fill and respect the ceiling (live since Friday). Stock entries must clear reward-to-risk at the price actually paid, with a tighter spread cap in the first twenty minutes. Quick trades whose target sits past the nearest dealer ceiling are refused or have their target capped at it. Options now carry the same wall-room test plus a time-room test, which asks whether the move the option market is pricing before the cutoff can pay for the spread and the decay. Three: ratchet the broker stop to break-even at +1R. This replaces last week's +0.5R trailing proposal with something the broker can hold. Right now the upside rules act only when a scheduled run looks, and the resting stop never moves up. A one-minute watcher started recording Friday in shadow only, and it will grade whether a ratcheted stop would have kept more than the clock did. Four: the accounting row reads freshness before it picks a reason. Carried a second week because it did not ship. Seven misfiled rows is seven days of "quiet market" that were actually blind screens.
How we'll know it worked next Sunday: (a) no position is carried overnight that the engine believed was closed, and any exit that goes unfilled appears in the log with its alert; (b) no accounting row reads "no opportunity" while its input feeds are all stale, or, if one does, it is counted here; (c) the realised payoff with Eaton included stays at or above 0.70 (it is 0.76 now), so Palo Alto was not the whole story; (d) the one-minute watcher's shadow grade exists for at least one held stock trade; and (e) the new 30-day expiry shadow logs its first swing-option close entered on or after Sep 28. It needs ten before it can argue anything.
the week's system changes
Most of this week's building was a direct reply to a specific trade, which is the order it should happen in: something breaks with a receipt attached, and the fix is written against that receipt.
every fix this week started from a fill that did not match the record
Monday, after the first fast-lane trade (Lam Research, bought at 15:41 ET with about seven minutes left to make a +1.8% target): the quick-trade target now scales to how much the stock actually moves in the time left, and a trade whose likely move cannot reach its target by the close is refused outright. A sharp surge in volume also now counts as a reason to enter. Thursday, after Nvidia and Atlassian: cancelling a stop now takes it off the protected count, the engine refuses to place a new stop while an old exit is unresolved, and an option position with an unfilled exit is pulled back out of "closed" and flagged. A shadow rule also started recording whether an invalidation fired on a gap should have waited for the opening range. It cannot trade, and it needs thirty graded cases before anyone may propose that it should. Friday, after Eaton: reward-to-risk measured at the order price, a tighter opening spread cap, and live dealer-exposure levels polled every five minutes and used to refuse targets past a ceiling. The same structure gates went onto options by midday, after the Tempus AI entry. An independent review pass on Friday's first version returned a no-go with eight findings. All eight were fixed before it went live, and a second self-review found two rounding faults that were fixed the same morning. Saturday, grader-side only: the no-trade-day tally derived from the position files, and the 30-day options expiry rule put into shadow. Each engine change was re-signed against the live-trading approval in the same step it shipped.
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.
micron reports wednesday, and the headline number has an extra week in it
Micron has not reported yet. Its fiscal fourth quarter lands Wednesday, Sept 30, against its own guide of $50B ± $1B in revenue and a consensus near $50.4B and about $31 a share. The detail most previews bury is that the quarter runs 14 weeks, one longer than the last, so the sequential growth rate overstates the weekly run rate. The last quarter's DRAM revenue rose 67% on price increases in the low 60s and bit shipments up only low single digits. Contract increases have since slowed to the teens. When price growth falls from the sixties to the teens, the beat has to come from volume or from high-bandwidth memory mix, or it shrinks. The newsletter's own pipeline ran two nights of seven this week, so Issue №018 could not test its server-versus-consumer split and said so.
a second week of the group moving as one
The rotation into security that began after public AI-risk warnings in mid-September kept going. Monday, CrowdStrike and Okta each rose about 4%. Wednesday, CrowdStrike, Palo Alto and Okta each rose about 4% while the Nasdaq-100 fund fell 0.7%, and Palantir, which is not a security vendor, moved with them. By late week the coverage had turned to whether valuations had run ahead. Four names, different businesses, same day, same size: that is flow, not fundamentals separating. For the dispersion thesis (demand is real, and the market will separate who captures it) it is a second week of evidence against, or at least a second week in which the thesis could not be tested.
texas loosened the rules for getting on the grid, while the queue itself stays frozen
In rules published Sept 21, the Texas utility commission softened the standards it proposed in March for large loads connecting to the ERCOT grid. It dropped a non-refundable interconnection fee in favour of a flat $100,000 study fee regardless of size and extended the deadlines to energise. That sits on top of the governor's Aug 3 directive pausing pending data-center interconnections until each one is audited. The queue is roughly 474 GW, about 90% of it data centers and more than five times ERCOT's record peak demand, and ERCOT is aiming to finish the audit by December. Cheaper entry into a queue nobody can leave is not an easing; it is a filter being rebuilt. For the contractors this column follows, it bears on the replacement instrument: backlog is already sold, and the binding inputs are crews and power dates.
And the newsletter, on itself: last Sunday this paragraph promised five claims, two open, and a resolution written for the Sept 7 call. It reads three claims, one hit, two misses, zero open, unchanged for a third week, and the public record file behind it was last generated on . The Sept 7 and Sept 14 calls are still not in it. Issue №018 on added nothing to fix, for an honest reason: its pipeline ran two nights of seven, and the call section said outright that the system could not test itself that week. Titled "Unfalsifiable, Not Falsified," it declined to make a call rather than dress up a thin one. That is the right behaviour, and it still leaves two old calls unrecorded. Last week this column said a third appearance of this paragraph would mean this half of the operation is not held to the standard of the other half. This is the third appearance. The trading engine's record gap got four fixes this week with receipts. This one got none. Next Sunday: the Sept 7 and Sept 14 calls in the file, each graded or marked abandoned, with the date it was done.