Three weeks ago Claude put a commitment on the record in this series. If CRWD reported earnings and the stock popped, it would trim twenty to thirty percent of the position. Not a vague intention. A specific plan, documented in public before the number arrived.
This week CRWD popped more than eleven percent after the close on Wednesday. And then Claude had to decide whether the reason it made that commitment still applied, or whether the earnings themselves had changed the picture enough to justify a different response.
That decision, and the reasoning behind it, is the most interesting moment this experiment has produced.
Let me give you the scoreboard first.
The scoreboard (August 27 close, $150 base)
One week ago the gap between the two portfolios was three cents. This week it is $3.55. Claude went from -0.37% to +4.06% in five trading days. The lead is back, rebuilt in a single earnings week by the two positions that Claude held through the summer’s doubt and doubled down on at the worst moment, August, when everything was falling.
Claude is beating VOO by 2.26 points, SPY by 2.17, and QQQ by 4.64. ChatGPT is positive at +1.69% but still slightly behind the S&P 500. QQQ, the Nasdaq 100, remains negative on the experiment.
NVDA: twelve weeks to vindication
In Week 3, when NVDA was Claude’s single biggest losing position, I wrote that Claude refused to sell and said its job was to follow the evidence rather than defend the pick. The evidence at the time was mixed. Chip rental prices were falling. China revenue was effectively zero. Insider selling was elevated. The thesis looked pressured.
Claude held. Then in Week 8, when NVDA had fallen further, it doubled the position at a lower price, upgrading conviction from 3 to 4 on the back of Microsoft’s AI revenue confirmation.
This week’s earnings report answered the questions that had hung over the position for three months. Revenue grew more than one hundred percent year over year and came in well above expectations. Guidance for the next quarter came in above analyst estimates. Margins held flat sequentially, eliminating the compression fear Claude had flagged since June. And the guidance assumed zero contribution from China, meaning the business printed those results with that entire market already written off.
Claude upgraded conviction from 4 to 5. It will not sell before the next capital injection, when it plans to add. The plan it documented in Week 3 turned out to be correct. That is worth saying plainly: twelve weeks of holding through doubt, a mid-experiment add at a lower price, and the earnings confirmed the thesis rather than challenged it.
CRWD: the test of the pre-commitment
This is the more complicated story, and I want to give it the space it deserves.
CRWD reported a genuine blowout. Revenue came in above estimates. Earnings per share beat. Guidance was raised. The metric that mattered most for the stock’s valuation, net new ARR, surged more than fifty percent and cleared the bar that bears had set as the minimum for the quarter to be considered satisfying, by a wide margin. The stock jumped more than eleven percent in after-hours trading.
For three weeks Claude had said: if the stock pops on earnings, I trim. The stock popped. A strict reading of that commitment says sell.
But Claude’s analysis surfaced a genuine tension that I think is worth naming directly because it gets at something real about how rules interact with new information.
The original reason for the trim commitment was specific: the position had outrun its fundamentals, the stock was trading above the average analyst price target, and a CTO departure had added uncertainty right before a binary event. The trim was pre-committed as insurance against a pop driven by sentiment rather than business strength.
What actually happened was a pop driven by business strength. Net new ARR that cleared the bear bar by a significant margin. A guidance raise that reflected genuine momentum. CEO language on the call that named the AI security opportunity as the largest in the company’s history. The specific worry Claude was hedging against, a valuation stretched beyond what the earnings could support, did not materialize. The earnings grew into the valuation.
Claude’s resolution was a partial trim. It sold roughly one-third of the CRWD position, locking in gains on the most expensively valued holding while keeping two-thirds to stay exposed to a thesis that just accelerated. The proceeds sit in cash, roughly $6.50, to be deployed into AVGO weakness after its September 3 earnings report or at the next capital injection.
The reasoning it gave is the part I want to put on the record. The pre-commitment exists to guard against rationalization. Every bull always thinks the thesis strengthened. The rule is a guardrail against that tendency. But the rule was designed for a specific condition that did not occur, and executing it mechanically when the underlying facts changed would be its own kind of failure. The honest response is a partial execution that honors the original intent while incorporating the new information.
I think that reasoning is correct. I also think it is the kind of reasoning that is easy to use as an excuse in the wrong direction, which is exactly why I am writing it down in full rather than just reporting the trade. The trim happened. The rationale is documented. The position is two-thirds intact at 4 out of 5 conviction. The experiment continues.
AVGO: next up
Claude’s largest position by dollar invested, $55, did not have an earnings event this week. It recovered modestly from last week’s selloff, closing at $371.54, still slightly below the blended entry price. September 3 is the next catalyst. Claude is holding at 4 out of 5 conviction with the cash from the CRWD trim available to deploy into weakness if the print disappoints or the stock falls before the report.
The thesis that drove the AVGO purchase has not changed: custom silicon and networking infrastructure for AI hyperscalers, with the Apple chip partnership running through the end of the decade. The competitive signal from a rival winning a Google chip deal is real and on the watch list. September 3 will tell us whether AVGO’s numbers reflect a thesis under pressure or a thesis that the selloff overstated.
ChatGPT’s quiet rebuild
ChatGPT made no trades this week, and I think that is the right call.
The portfolio it has built over the past month is genuinely more defensible than what it ran in June. VOO now represents 46.6% of the portfolio. PLTR, which has grown to 33%, is up more than thirty percent from the blended entry after its extraordinary earnings report. RKLB and ASTS together represent less than fourteen percent of the portfolio. Ten dollars of cash provides optionality.
RKLB is down 40.8% from entry. ASTS is down 40.5%. ChatGPT is holding both without adding, which is the right posture. The businesses are executing: Rocket Lab posted record quarterly revenue and a multi-billion dollar backlog. ASTS has launched thirteen satellites and is preparing beta service. Both companies are doing what they said they would do. The market is asking whether the capital requirements and execution timeline are compatible with the current valuation. That question will not be answered by the next quarterly report.
The gap between the two portfolios is $3.55. Fifteen weeks remain.
What the earnings week actually proved
I started this experiment because I wanted to know what AI systems actually do with real money under real conditions, documented publicly with no editing of history.
Week 12 gave the clearest answer yet. Claude held NVDA for twelve weeks through a period when every reasonable reading of the evidence suggested the thesis was under pressure. It documented its reasoning each week, downgraded conviction when the evidence warranted it, refused to sell on fear, and added at a lower price when a specific data point strengthened its view. The earnings confirmed the thesis. That is a twelve-week investment decision that played out exactly the way careful, documented investing is supposed to play out.
It also pre-committed to a trade it was not sure it wanted to make, then had to decide in real time whether the conditions that motivated the commitment still applied. It arrived at a partial execution with documented reasoning. That is not a perfect decision, because perfect decisions do not exist. It is an honest one.
The pre-commitment structure itself is worth keeping regardless of how this specific trade turns out. Writing down the conditions that would make you act, before you are emotional about the outcome, is the single most useful discipline either AI has demonstrated in twelve weeks. It does not guarantee the right decision. It does guarantee that you cannot pretend you planned whatever you did.
What I am watching
AVGO reports September 3. Claude has $6.50 in cash and a stated preference to deploy it into AVGO weakness or the next capital injection. That plan is on the record.
The next capital injection is the other event on the horizon. Both portfolios will receive another fifty dollars per side. For Claude, NVDA is the stated priority for adding. For ChatGPT, the question is whether any of the cash reserve or future injection goes toward the speculative positions or stays in the VOO anchor that has rebuilt the portfolio’s foundation.
The gap is $3.55. The experiment has run twelve weeks. It has fifteen more to go.
The most honest thing I can tell you about where this stands is that three weeks ago I thought Claude had probably won. This week I think the experiment is genuinely open. Both portfolios are positive. Both have theses intact. One has a more defensible construction. One has more upside embedded in positions that are still deeply underwater. December will tell us which one mattered more.
What would you have done with the CRWD trim? Full execution of the pre-commitment, or the partial execution Claude chose? Hit reply and tell me. I read every one.
Following the experiment? Subscribe to get the AVGO earnings reaction next week, the next capital injection, and every move both AIs make as the experiment heads into its final fifteen weeks.