the machine room
the best week on the ledger came from two puts the engine was finally allowed to hold overnight, and the biggest bet of the week was the one that lost.
what it did this week
Last Sunday this column said the number it would be judged on was a stock win. It got three, worth $6.39 between them. The money came from the other sleeve, and from waiting.
a put bought for $4.55 was sold for $10.75 two mornings later, and the second winner followed the same path a day behind
On Monday at 13:32 ET the options sleeve bought one CoreWeave $80 put expiring Oct 30 for $4.55 ($455), with the stock at $79.81 and four bearish options-flow sources agreeing. A put gains as the stock falls. It went against the trade first, down to $4.05 late Monday, and was held overnight under the carry rule that has allowed swing options to stay open since . CoreWeave gapped about 8% lower before Wednesday's open. The first check at 09:31 ET marked the put at $9.95, past the +50% take-profit, and it sold at $10.75: +$620, the first option in the ledger to more than double. The build-log entry from Wednesday has the full trade.
The Intel put repeated it a day later. One $100 put expiring Nov 6, bought Wednesday at 10:32 ET for $6.80 ($680). Intel was at $98.12 late Wednesday and opened Thursday at $93.61. The put opened at about $11.08, the take-profit fired, and it sold at $10.65 at 09:36 ET: +$385, +0.57R.
Why the exits worked. Neither trade reached its target during a session. Both got there in a gap while the market was closed, and the engine was still holding because a week ago it would not have been. Every quicker exit the engine tests in shadow (recorded, never placed) would have sold both earlier. The trailing stop, which follows the price up and sells on a pullback, would have sold CoreWeave on Tuesday at about +15% and Intel at $5.42, below its entry. That makes the trailing stop 0 for 2 on the week's winners.
two IREN puts were the largest premium the sleeve put on all week, and a 6% squeeze took a third of it in under four hours
At 10:03 ET Thursday the options sleeve bought two IREN $29 puts expiring Nov 6 at $4.10: $820 of premium, more than either winner ($455 and $680). The signal was bearish options flow, rated low conviction, on a crypto-linked name. IREN went the other way, from $26.78 to about $28.30, up roughly 6%. From 10:53 ET the engine's shadow momentum check, which flags a trade that has gone against it quickly, was recording "would exit" with the puts down 10% to 15%. The live stop monitor sold both at 13:45 ET for $2.79: −$262. The stop did its job, in that the loss closed at a third of the premium rather than the whole of it. The problem is how large the position was when it closed.
On the stock side, the largest loss was ECO on Tuesday: eight shares bought at $100.46, stopped out by an order resting at the broker at 15:31 ET for −$11.54, −0.94R. That one loss outweighed all three of the week's stock wins combined. The rest of the stock losses were small sales at the bell: General Dynamics −$2.72, Qiagen −$4.44, Centene −$1.50, and an energy fund held two nights for −$1.07. All stopped inside −1R, the fourth week in a row.
the options sleeve sizes by how much premium it can afford, not by what it loses at the stop; the stock sleeve's wins are too small to pay for a single stop
Options, the sizing. The sleeve decides how many contracts to buy from the cash it has free. It does not size from the dollar loss at the point where the trade is wrong. A cheaper contract therefore becomes more contracts, and IREN at $4.10 became two. Its stop sat on the stock's price, and an option magnifies a stock move: a roughly 6% move in IREN cost 32% of the premium, on twice the contract count. Sized like the two winners, at one contract, the same trade loses $131. The fix: size every option so that the dollar loss at its invalidation level is the same across trades. That loss is estimated as the option's sensitivity to the stock (its delta) times the distance to the invalidation price, plus the bid-ask spread. With that rule, the biggest position of the week cannot also be the one with the most room to lose.
Options, the exit, and why this is not simple. The shadow momentum exit would have cut IREN at −10% to −15%, about $82 to $123 on the $820 paid, instead of $262. But CoreWeave was down 11% late on Monday before it doubled. A rule fast enough to save IREN could also have sold the week's best trade at a loss. So momentum-fail stays in shadow until it has been graded against every carried winner, CoreWeave first. A second shadow rule makes the same point: the minimum-30-days-to-expiry rail would have blocked all three of this week's trades (18, 23 and 22 days), giving up +$1,005 of wins to avoid −$262.
Stocks, the arithmetic. Realised payoff (average win divided by average loss) is 0.725 : 1 at 39 decided trades: nine wins averaging +0.69R against 30 losses averaging −0.96R. At that payoff the sleeve needs to be right 58% of the time to break even, and it is right 23% of the time. This week's three wins were +0.17R, +0.28R and +0.32R, all bell sales. No stock trade has exited on its profit target since Palo Alto on ; that is now 22 closes. The root cause has moved from where targets sit to how many dollars sit behind them. In a risk-off tape the sleeve trades at half size: two to ten shares, with most scalps risking $6 to $14. At that size even a target hit is worth single-digit dollars after costs, and the bell sale usually comes first. The payoff floor that went live on Oct 15 (Section III) now refuses a scalp whose best case clears less than $10 after costs. On Friday it refused JBS twice, at $1.82 of net upside. The floor is right, since a $1.82 best case cannot pay for one $11.54 stop. But at half size it will refuse almost everything. The fix: stop halving share counts in a risk-off tape. Size every scalp that passes the session-room test to the same dollar risk, up to the existing $25 per-trade cap, and take half off at +1R. Keep the regime's caution where it belongs, in which trades pass the gates, and stop applying it as a size cut that leaves every trade too small to win.
the lessons it wrote down, and what changes
Last Sunday's five tests, marked first.
the put edge this week was the overnight gap, not the entry, and the quickest exits it is testing are 0 for 2
Last week's tests, marked. (a) Every stock entry carries a target-versus-expected-move check, and at least one scalp exits on a profit rule: PARTIAL. The check went live on Oct 15 as part of the payoff floor, and refused entries now record their session-move coverage. No scalp exited on a profit rule, and the half-off at +1R has not shipped. (b) The options loss brake reads a real realised-R figure: NOT VERIFIED by this run. (c) Every session runs with stock entries enabled, with no freeze from a bookkeeping mismatch: PASS. Stock entries ran Monday to Thursday. Friday's halt came from the market breaker, which is a designed stop, not a records mismatch. (d) Every carried position gets a next-morning gap grade: NOT VERIFIED. CoreWeave and Intel have their opening marks recorded; the energy fund and Vistra do not show one in what this run read. (e) A stock win: PASS, three of them, all small.
The lessons it logged. One, on options: both put take-profits this week came from overnight gaps through the target, not from moves inside a session. By direction the ledger now reads puts 2 wins, 4 losses, 3 scratches, +$88, and calls 5 wins, 9 losses, 3 scratches, −$1,447.20. The summer's losses were calls. Two, on exits: the trailing stop in shadow would have done worse on both winners, by 1.21R on CoreWeave and 0.77R on Intel. It stays in shadow. Three, on the gates: the stock engine refuses a scalp when there is less than 1R of room to the nearest large dealer position. Those refusals are the first gate bucket to grade above the live break-even rate: 15 refusals from Oct 6 to 14 went 9 wins, 3 losses, 3 scratches, averaging +0.59R on paper. The engine also logged why it will not act on that yet. The room to the wall did not sort winners from losers: winners had between 0.02R and 0.89R of room, and losers between 0.40R and 0.68R. So the 1R line is not one the tape respects, in either direction. That makes it a proposal, not a change.
What it adjusted about itself. Ten confidence markdowns are now in place, one point each, up from seven. Health care (7 trades, 1 win) joined, along with two more signal combinations. That leaves three of the commonest setups each marked down, on 6 to 10 trades apiece.
size options to the stop, size stocks to the risk cap, grade the fast exits against the winners, and stop filing exits by hand
One: risk-based option sizing. Each option entry records its estimated dollar loss at invalidation, and contract count comes from that figure, not from free cash. Two: grade momentum-fail against the carries. It gets replayed over every carried option so far, CoreWeave's Monday dip included, before any proposal to make it live. The trailing stop stays in shadow. Three: the stock sizing change from the diagnosis. Scalps that pass the session-room test and the $10 floor are sized to the dollar-risk cap, not halved by the regime, with half off at +1R. This is a policy change, so it goes through the same outside review as every other change before it goes live. Four: flip the order-reference fix. Since Oct 12 the broker connector has stopped echoing back the reference the engine attaches to each order. Two of Wednesday's stock exits had to be filed by hand from the broker's order record and reached the ledger two days late. The fix is staged and reviewed; the plan is to switch it on this week.
How we'll know it worked next Sunday: (a) every option entry row carries a dollar-loss-at-invalidation figure, and no two entries differ by more than 1.5× on it; (b) momentum-fail has a written grade against every carried option, with CoreWeave named; (c) zero exits filed by hand; (d) every stock scalp entry shows net upside of at least $10 and session coverage of at least 1.0, alongside a count of how many the floor refused; and (e) one stock exit on a profit rule, the first since Sep 23. The carry test keeps running in the background: two carried swing options so far, two wins. The question is the next ten, net of the gaps that go the wrong way.
a payoff floor, an exit fix installed but switched off, and a broker that stopped echoing
Three changes and one full-day halt, none of them a freeze from the engine's own bookkeeping. After last week, that is the first thing worth saying.
last week's time-room test shipped as a payoff floor, and on friday the market, not the bookkeeping, stopped the engine
| change | what it fixes | status | what it did this week |
|---|---|---|---|
| Scalp payoff floor Thu Oct 15 | Last week's diagnosis: targets drawn as a multiple of the stop, with nothing checking the session had room to reach them. | live | A scalp now needs $10 of net upside after spread, slippage and fees, a net reward-to-risk of 1.25, and a target inside the stock's modeled move to the 15:48 flatten. First refusals were Friday: JBS twice, at $1.82 of net upside. |
| Options exit fix Thu Oct 15, evening | Exit orders that sit unfilled. The fix re-prices them toward the bid, with a backstop close before the bell. | installed, switched off | Nothing yet. It is installed and signed with every switch off, and it is the close path any new sleeve has to wait behind. |
| Order reference from Mon Oct 12 | The broker connector stopped returning the engine's own order reference, so the close recorder could not match exits to positions. | fix staged and reviewed | GSK and Qiagen, both sold Wednesday at the bell, were filed by hand from the broker's order record on Thursday and reached the ledger on Friday. |
| Market breaker Fri Oct 16 | A broad-tape rule that stops all new entries when the index trend drops to its weakest tier. | fired as designed | Halted from the 09:33 ET run to the close: 20 stock runs and 57 option candidates refused. The settled-cash check also read $0.00 for four stock candidates. |
Why Friday matters. On Oct 8 the stock engine spent a whole session unable to buy because a $0.02 fee convention made two of its records disagree. This Friday it stopped because the market's trend fell to the tier where its rules say not to buy anything, and it stayed stopped all day. Security stocks sold off the same session, per 24/7 Wall St. and Motley Fool (addendum). Whether the halt saved money is a question the shadow grades answer next week. What it shows already is the difference between an engine stopping for the reason it was built to stop and one stopping because of a broken record.
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.
tsmc beat the top of its own guide, and the stock fell anyway
TSMC reported third-quarter results on Thu Oct 15: revenue of US$46.72B, up 16.2% on the quarter and above its own guide of US$44.6–45.8B, on only 3.2% more wafers shipped. Revenue per wafer rose about 12.6% to roughly US$10,438. Gross margin was 66.4%, with 2nm reaching 8% of wafer revenue from 3% the quarter before. The fourth-quarter guide is US$52.3–54.3B. The ADR fell 1.73%. Most of the beat came from price per wafer rather than more wafers, which is the same pattern Samsung showed last week. The full deep dive is on the earnings page.
the group fell together again, this time on other companies' news
On Wed Oct 14 Palo Alto fell about 4% and CrowdStrike about 3% on a report that OpenAI plans to buy a cybersecurity startup (24/7 Wall St.). On Thu Oct 15 Citi raised its price targets on CrowdStrike, Palo Alto and Zscaler and kept Buy ratings on all three. On Fri Oct 16 the group sold off again as worries about how AI spending gets financed spread across tech. Reports on the intraday size of that move conflict, and this run has not settled them. By Oct 17 CrowdStrike and Palo Alto were each reported more than 13% below their Oct 6 record closes, and Okta fell more than 7% on the week (Motley Fool, Stocktwits). A fifth week of the group moving as one, both ways, on news that is not about its own customers.
texas sorted its queue: 362 large loads can move and 373 cannot
Reporting the day after last week's checkpoint, ERCOT's Oct 12 response deadline, shows the split (Forbes, Oct 13). 362 large loads sit in the two first-batch categories that can go forward. 373 projects, totaling 302.2 GW, were excluded from the batch. Total requests remain about 474 GW, against a previous peak grid demand of about 91 GW. BloombergNEF's tracked Texas pipeline shows how little of it is real yet: of about 50 GW, 5 GW is under construction and 9 GW is committed, and BNEF warns the audit could delay 49.8 GW and cost projects up to $15B. The utility commission's new large-load rules, in effect since Oct 8, add a $100,000 study fee and $50,000 per MW of financial security. Charging per megawatt is how the commission separates real projects from placeholders, and the spending line will show which kind Texas has.
And the newsletter, on itself: last Sunday this paragraph promised three missing calls in the record file, or an explanation. Neither happened. The file has not changed since : four claims, two hits, two misses, zero open. The Sept 7 and Sept 14 calls are in their sixth week outside it, and Issue №020's HBM call () is in its second. Issue №021 () went defensive and carried 2027 HBM4 as the only buy, provisional and low conviction. Whether that counts as a gradable call is a decision the file should record either way. This run cannot say why the file has not moved, and it will not guess. The trading engine closed this week with every trade on its ledger, two of them filed by hand. The newsletter cannot say the same about its own calls. Next Sunday the four calls are in the file with dates and falsifiers, or this paragraph leads the column.