Eight Days In, What Is Working for the AI Manager: Its Own Two Trades, the Semiconductor Spread, and the Day It Came Back

Wednesday was Day 8, and the AI manager was back at the desk. The portfolio rose 1.6 percent to $25,387, which leaves it down 2.3 percent since the handover, ahead of 4 of the 10 indexes. On Monday and Tuesday the decision model never ran: both runs failed on an API credit error, and the book held by default. Wednesday's run completed, checked all 17 exit conditions, found none triggered, and queued three adds. This is the "what is working" piece, so here is what is, with the losses kept in view.

## Working: the model's own trades are above cost

The AI has made two trades of its own, both on Friday October 16, both adds on a red day. ServiceNow is up 3.3 percent against cost and is 7.1 percent of value, the second-best position in the book by that measure. TSMC is up 2.0 percent against cost at 7.2 percent of value. Two trades is a sample, not a record, and both were bought into a selloff that has since reversed, so part of this is timing nobody predicted. But the two positions the model chose are above water while most of what it inherited is not.

## Working: the semiconductor spread

The index this portfolio was expected to fear most is SMH, the semiconductor index, and it is the one the portfolio leads by the most. Since the handover SMH is down 7.7 percent and the portfolio is 5.5 points ahead of it. The leads over the other AI-heavy benchmarks are thinner: 2.6 points over emerging markets, 1.6 over AIQ, 1.2 over the Nasdaq 100. The pattern after eight days is plain. The portfolio falls less than the AI indexes and more than the broad ones. The Dow is up 2.3 percent since the handover and the portfolio trails it by 4.5 points; Europe is up 1.8 and the gap is 4.1. Beating 4 of 10 means beating the four indexes most like itself and trailing every index that is not.

## Working: the rules ran when the model did not

The stop-loss, the cash floor and the position caps are code, and code does not need API credits. IonQ, the worst position at 19.1 percent below cost, is 5.9 points from the automatic 25 percent stop. Nothing was sold on the two silent days because nothing met its condition, and that was true whether or not the model was there to say so. On Wednesday, when it was, it checked all 17 exit conditions and found none triggered. The written exits did their job in both states: absent model, present model.

## What is not

The score is still a loss. Down 2.3 percent since the handover, trailing six indexes, and the week is flat: over the last five trading days the portfolio is up 0.1 percent. The best position against cost, Monday.com at 12.6 percent, is 0.1 percent of value, so the best pick barely moves the book. Snowflake is 12.5 percent below cost and AMD is 9.1 percent below, and AMD is one of Wednesday's three adds. The model is adding to a loser on new customer evidence, which is either conviction or the thing that goes wrong next.

## The three adds

Wednesday's update queues three orders for Thursday's open, about $2,405 in total against cash of 25.9 percent of value: TSMC from 7.2 to 11.2 percent on a Moody's upgrade that cited AI demand; AMD from 2.7 to 5.2 on Vultr's planned 20 MW deployment; Shopify from 0.7 to 3.7 on the Shop Pay integration into Meta's Muse agent. Each carries an exit condition and each cites one article. That is the method working as written: evidence, a size, and a way out.

## What to watch

First, Thursday's fills. The update will show whether all three orders filled and at what price, and cash should drop from 25.9 percent. If cash does not move, the fills failed, and the reason belongs on the page.

Second, the AMD add. It is the one position the model is increasing while it sits 9.1 percent below cost. If the Vultr plan converts into reported revenue, the thesis holds; if the next results cut the AI-chip outlook, the written exit condition should fire, and whether it does is the real test of the method.

Not investment advice.