Week 14, 2 September 2026. Status: no change. Twelfth rebalance review, and a second week without a trade.

The pot stands at $964 and the book is unchanged. Four names again, in the same order as last week: Vistra, Kratos Defense, Micron and Pagaya. I placed no trades.

The method has not moved either. Each of the four models, Claude, DeepSeek, Grok and GPT, carries the same 25 percent weighting. Their individual stock weights are then added together to give a combined score, the names are ranked by that score, and the book is cut at the first clean break in the numbers.

This week Vistra came out on 25, Kratos Defense on 24, Micron on 23 and Pagaya on 21. Below Pagaya the scores dropped six points to SanDisk on 15, and that six point gap was the only clean break anywhere near the top, so the cut fell after the fourth name once again. Vistra also held on to the top spot it took from Kratos last week.

Two familiar exclusions ran as usual. ARIS Mining would have led the whole list on a combined score of 28, ahead of everything actually held, while the short term treasury fund SGOV would have scored 20. Neither can be traded on Lightyear, however, so both come out of the calculation before the ranking is cut.

Nothing traded because nothing needed to. Under the core swaps only approach, a full realignment waits for a week in which several holdings have moved far enough in price to justify the spread and the currency cost. This week every position sat within half a percentage point of its target weight, comfortably inside that threshold, so trading would have cost money to change almost nothing.

As for the money, $964 sits below the $1,000 the pot started with in June and below where it stood a week ago. On an all time basis most of the individual positions are showing a loss, with Kratos the furthest behind, while Pagaya is the only one currently in front. Four concentrated positions are doing all of the work, in both directions.

A second consecutive hold week, with the same four names in the same order, does raise a question about the exercise. The idea behind aggregating four models was to see what happens when their disagreements get resolved mechanically. A book this concentrated, and this static, may be showing genuine consensus, or it may simply be showing inertia in the underlying allocations, and from the outside those two look identical. A wider book would give the aggregation more to work with. That question stays open for now, because changing the rule in order to produce a more interesting result is exactly the sort of judgement the experiment was set up to avoid.

The book this week, four names:

Overall direction: below the starting $1,000, with the line-up static for a second week. This portfolio is a mechanical experiment in aggregating the stock picks of four AI models, so positions are selected by a fixed rule rather than by judgement. None of this is a recommendation and nothing here is financial advice; see the live names on my Lightyear profile. Capital at risk. The value of investments can go down as well as up and you may get back less than you invested. Past performance is not a reliable indicator of future results.

Looking ahead, the next review falls on Wednesday 9 September 2026.

Week 13, 26 August 2026. Status: no change. Eleventh rebalance review, and nothing needed doing.

The pot stands at $1,003.47 and the book is unchanged. Four names for a second week running: Vistra, Kratos Defense, Micron and Pagaya. I placed no trades at all. That is the entire update, although the reasons behind it are worth setting out, because a hold week is something the rule produced rather than a week off from it.

Start with the scores. Each of the four models counts equally, and I sum the weights per ticker before ranking the list. This week Vistra came in at 25, Kratos at 24, Micron at 23 and Pagaya at 21. Below them, the next two names, Novagold and SanDisk, both landed on 15. So there is a six point gap between fourth and fifth place, and nothing else in the ranking comes anywhere near that wide. The cut therefore fell after the fourth name again, in exactly the same spot as last Wednesday.

One thing did shift, even though the book did not. Vistra has moved ahead of Kratos at the top of the ranking, having trailed it since the middle of July. That comes down to Claude, which made the largest single weight change of any model this week and raised its Vistra position. The other three names barely moved. Whether one model nudging one holding counts as a signal or simply as noise, I cannot say yet, but it seems better to record it as it happens than to notice it later.

Then there is the question of why nothing traded. Under the core swaps only approach, I act when a position drifts far enough from its target weight to justify the spread and the currency cost, and not before. This week every position sits within one percentage point of its target, which is comfortably inside that threshold. Rebalancing would have cost money to achieve almost nothing, so I left the book alone.

Two familiar exclusions ran again, meanwhile. Aris Mining was once more the highest scoring name on the entire list, at 30 points, ahead of everything actually in the book, and once again it cannot be traded on Lightyear, so the method discards it. SGOV went the same way, despite being one of Claude's larger single allocations. It is worth being blunt about that. For several weeks now the top ranked name by the experiment's own measure has been unreachable, which is a limitation of the broker rather than of the method, yet it shapes the results all the same.

As for the money, $1,003.47 sits a shade above the $1,000 the pot started with in June. That makes two broadly flat weeks in a row, with four concentrated positions doing all of the work. Nothing dramatic to report this time, which is a change of pace after last week.

The book this week, four names:

Overall direction: broadly flat, just above the starting $1,000. This portfolio is a mechanical experiment in aggregating the stock picks of four AI models, so positions are selected by a fixed rule rather than by judgement. None of this is a recommendation and nothing here is financial advice; see the live names on my Lightyear profile. Capital at risk. The value of investments can go down as well as up and you may get back less than you invested. Past performance is not a reliable indicator of future results.

Looking ahead, the next review falls on Wednesday 2 September 2026.

Week 12, 19 August 2026. Status: changed. Tenth rebalance, and the book cuts to four names.

This is the largest single structural change in the experiment's history. Last week the book held nine names. This week it holds four. Nothing about the method changed to produce that. The rule ran exactly as it has every week, and four names is simply what came out.

A quick recap of that rule. Each of the four models publishes its own allocations. I sum the weights per ticker, counting every model equally, then rank the list and cut it at a clean break in the scores rather than at a target number of positions. This week the only clean break in the entire ranking fell between fourth and fifth place.

The combined scores tell the story. Kratos 29, Vistra 25, Micron 24, Pagaya 20, and then fifteen, fifteen, fifteen, fourteen, thirteen, thirteen, twelve, eleven, ten, nine, nine. Below the top four there is no gap wider than a single point anywhere. The scores fall off a cliff after Pagaya and then flatten into a long, undifferentiated shelf.

Each of the four survivors earned its place differently. Kratos appears in all four lists, is GPT's single largest position and DeepSeek's joint largest, and has now been held continuously since week one. Vistra is also in all four lists, but the support is uneven: most of its score comes from Claude, while DeepSeek holds it in barely token size. Micron is in three lists, missing only from GPT, yet it is Grok's single largest position and DeepSeek's joint largest. Pagaya is in three lists, missing from Grok, with the three models that do hold it landing independently on almost identical mid-sized allocations. That makes it the most evenly supported name in the book, and also the one closest to the edge.

It is worth being clear about why nothing below the cut got in. Two different failure patterns produce nearly identical scores, which is why the shelf is so flat. Some names have breadth without conviction. SanDisk sits in three of the four lists, for instance, yet no model gives it any real size, and it totals just fifteen. Others have conviction without breadth. Broadcom gets meaningful weight from two models and nothing at all from the other two, and lands on fifteen as well. Neither pattern gets anywhere near twenty.

The most visible casualty is precious metals. The sector went from roughly a third of the book to nothing, and this is the first week the experiment has held no miner at all. The models have not turned against gold and silver. Grok alone holds seven separate precious metals names, not far off two fifths of its book. But those votes scatter across different tickers, so no single miner accumulates enough combined score to clear the cut. Kinross came closest, at eleven. That is a real limitation of summing weights per ticker rather than per sector, and it is exactly the sort of thing this experiment is designed to expose.

The trades themselves were straightforward, if unusually heavy. Six positions closed in full: Kinross Gold, Endeavour Silver, Broadcom, Vista Energy, Gold Fields and Fortuna Mining. Vista Energy and Gold Fields appeared in no model's list at all this week, and Fortuna in only one. The sales raised $564.26. That went into a fresh Pagaya position, the largest single purchase of the day, plus top-ups to Kratos, Vistra and Micron. The book now stacks in the same order as the scores: Kratos largest, then Vistra, then Micron, then Pagaya.

Pagaya deserves a note, because it is a returning name rather than a new one. In week one it was one of only two stocks backed by all four models. It was held right through the summer, dropped on 15 July when it slipped below the cutoff, and is now back five weeks later in fourth. The other founding all-four pick was Kratos, which has never left and now tops the ranking. I find that mildly encouraging for the underlying idea, that broad agreement between the models means something. Then again, two names over twelve weeks proves very little.

One methodology decision to record. I considered replacing the clean-break cut with a fixed numeric floor on the combined score, and rejected it. A floor would have produced nine positions this week instead of four. But it would have been chosen in full knowledge of what it produced, which makes it a preference dressed up as a rule. If the floor idea returns, it should be set blind, and dated, before the week it first applies.

Finally, the risk picture, which has changed more than the names have. Four positions, each somewhere between a fifth and roughly a third of the pot, means single-company moves now drive the whole book. That showed up within minutes. The pot was worth $1,022.68 at the moment the final buy filled and $1,002.48 shortly after, a fall of about 2%, with nothing left in the portfolio to offset it. There is a fair question here: can a four-stock portfolio meaningfully test a hypothesis about cross-model consensus? The counter-argument is that four names is what strong agreement actually looked like this week, and pretending otherwise would be a different experiment. If next week's break falls somewhere else, the book may swing straight back out to a dozen names. That turnover would itself be a finding worth recording.

The book after this rebalance, four names:

Overall direction: broadly flat, hovering just above the starting $1,000. This portfolio is a mechanical experiment in aggregating the stock picks of four AI models, so positions are selected by a fixed rule rather than by judgement. None of this is a recommendation and nothing here is financial advice; see the live names on my Lightyear profile. Capital at risk. The value of investments can go down as well as up and you may get back less than you invested. Past performance is not a reliable indicator of future results.

Looking ahead, the next rebalance falls on Wednesday 26 August 2026.

Week 11, 12 August 2026. Status: no change. Ninth rebalance, and the book turns a profit.

The headline is straightforward. Before this week's rebalance the pot stood at $1,013.71, against a cost basis of roughly $951.92. So the experiment is in profit for the first time, up around 6.5%. After ten weeks in the red, though, that is a change of direction rather than a result.

The gains came almost entirely from precious metals. Endeavour Silver (EXK) and Fortuna Mining (FSM) led the book by a clear margin, with Gold Fields (GFI) and Kinross (KGC) close behind. Kratos (KTOS) ran up strongly too. Meanwhile Micron (MU) was the obvious drag, and Broadcom (AVGO) and Vistra (VST) both slipped a little. In short, the corner of the book that looked most concentrated last week is the corner that carried it.

There were no name changes at all this week. I reran all four models and combined the scores, weighting each model equally as usual. Then the top nine came back identical to the nine already held. Kratos led the scoring by a wide margin, appearing in all four portfolios and topping GPT's list. Micron, Kinross and Vistra followed. The cut fell naturally after Vista Energy (VIST), with the next name several points below, so no tiebreak was needed.

Two observations are worth drawing out. First, Aris Mining scored third highest of any name this week, and it was the only stock appearing in all four model portfolios. Yet it stays excluded, because Lightyear does not list it. That is a persistent quirk of this AI stock picking experiment: the single strongest consensus signal is one the portfolio structurally cannot act on.

Second, Claude put close to a quarter of its book into SGOV, a short-term Treasury ETF. In effect, that is a call to hold a quarter of the portfolio in cash. However, SGOV is untradeable here too, so the view is discarded entirely in the aggregation. One model's caution simply vanishes from the combined result.

The rebalance itself was small, a realignment rather than a reshuffle. In total, roughly $68 moved. I trimmed Gold Fields hardest, then took smaller trims from Fortuna, Kratos, Broadcom and Kinross. Those proceeds funded additions to Vistra and Endeavour Silver, plus small top-ups to Micron and Vista Energy. Every fill landed exactly on target.

Precious metals had drifted up to roughly two fifths of the book on the strength of the silver names, and the trims pulled that back slightly. Semiconductors sit at about a quarter. Neither figure is a deliberate positioning choice, though. It is simply where the models converged.

One open question carries over, and it has sharpened. Nine positions, held steady for a full week with no turnover at all, raises a fair doubt. Is the aggregation actually finding consensus, or is it settling into inertia? Accordingly, I am still weighing a change to the method: dropping the fixed holding count, and qualifying only names that appear in at least two models. I have not adopted it yet.

The book after this rebalance, unchanged at nine names:

Overall direction: up, and into profit for the first time. This portfolio is a mechanical experiment in aggregating the stock picks of four AI models, so positions are selected by a fixed rule rather than by judgement. None of this is a recommendation and nothing here is financial advice; see the live names on my Lightyear profile. Capital at risk. The value of investments can go down as well as up and you may get back less than you invested. Past performance is not a reliable indicator of future results.

Looking ahead, the next rebalance falls on Wednesday 19 August 2026.

Week 10, 5 August 2026. Status: changed. Eighth rebalance, and the book falls to nine names.

The central story this week is that the portfolio dropped from thirteen positions to nine, and it did so not by choice but because the methodology ran out of agreement. When I combined and ranked the four model weights, the top eleven names came out cleanly. Below that, however, six separate names all landed on exactly the same combined score of 2.25, competing for four remaining slots.

Filling those slots would have meant picking four names out of six on judgement rather than arithmetic, which is precisely the kind of discretionary decision this experiment exists to avoid. Accordingly, the established practice in these situations is to hold the lower count rather than compromise the mechanics, so I made the cut at rank eleven.

Two of those top eleven, ARIS and SGOV, remain untradeable on Lightyear, which brings the working portfolio down to nine names. Worth saying plainly: the tie itself carries information. Below rank eleven the four models share no conviction at all, and those slots would have been filled by names only one model liked. Therefore holding nothing there is arguably more honest than holding something arbitrary.

The resulting concentration deserves a straight answer. Precious metals now make up roughly two fifths of the book across Kinross Gold (KGC), Gold Fields (GFI), Fortuna Mining (FSM) and Endeavour Silver (EXK), while semiconductors account for about a quarter across Micron (MU) and Broadcom (AVGO). Meanwhile, Kratos Defense (KTOS) remains the anchor and comfortably the largest single position, holding the top combined score for another week. This is a considerably less diversified portfolio than last week's, and that is a consequence of the models converging on the same two themes rather than any deliberate positioning on my part.

As for the exits, ServiceNow (NOW), Zeta Global (ZETA), Sterling Infrastructure (STRL) and Teekay Tankers (TNK) all sold in full. Zeta Global leaves as one of the better performers of the run, well up on its average buy price, while Sterling Infrastructure exits at a loss. In short, the mechanical approach sells winners and losers alike when the combined score no longer supports the position, and that is the point rather than a flaw.

Execution went cleanly. In total there were four sells and nine buys, with $199.26 redeployed and no residual cash. There was minor slippage of around 40p across the two largest exits, absorbed in the rescale.

That leaves an open question. The target has always been fifteen positions, and the book has drifted well below that for two reasons that compound each other: untradeable assets, and now an unresolvable tie. So is a portfolio of nine names still testing what it set out to test, or do the rules need a tiebreaker that can resolve larger ties without becoming a discretionary call in disguise? For now, Claude and I opted for less with more conviction rather than the diluted split. We shall see how that goes.

The book after this rebalance, nine names:

This portfolio is a mechanical experiment in aggregating the stock picks of four AI models, so positions are selected by a fixed rule rather than by judgement. None of this is a recommendation and nothing here is financial advice; see the live names on my Lightyear profile. Capital at risk. The value of investments can go down as well as up and you may get back less than you invested. Past performance is not a reliable indicator of future results.

Looking ahead, the next rebalance falls on Wednesday 12 August 2026.

Week 9, 29 July 2026. Status: changed. Seventh rebalance, Modine out after one week, Teekay Tankers back in.

A weaker week for the pot. The combined portfolio ended at $877, down from around $926 seven days earlier, so that is a fall of roughly 5 percent. Almost the whole book sits in the red on a position level, and Zeta Global (ZETA) is the sole holding currently in profit.

The method was applied as usual. First, I weighted each of the four AI portfolios equally, at 25 percent apiece. Then I summed the individual stock weights across Claude, DeepSeek, Grok and GPT into one combined score. After that, I took the top fifteen names by that score. Kratos (KTOS) held its place at the top of the table on 6.50, supported by all four models and carried by GPT and DeepSeek in particular.

Aris Mining came second on 6.25 and is now held by every model, which makes its continued absence the most frustrating feature of the experiment. However, Aris remains untradeable on Lightyear, as does SGOV, so I excluded both and rescaled the remaining thirteen positions to 100 percent.

The one substantive change was the departure of Modine (MOD) and the return of Teekay Tankers (TNK). Modine only entered the portfolio last week, and it did so on a sector diversification tiebreaker rather than on the strength of its score. This week only Grok still holds it, giving a combined score of 2.00, which falls below the cutoff. Meanwhile, Sterling Infrastructure (STRL) and Teekay both scored 2.25 and sat safely inside the top fifteen, so no tiebreaker was needed at all. Teekay had been fully exited only a week earlier, and it returns with support from both DeepSeek and GPT.

I did give some thought to backfilling the two empty slots in order to return to a full fifteen positions. In the end, I rejected that idea. Six different holdings are tied on exactly 2.00 at the cutoff, so choosing two of them would have been a large discretionary judgement dressed up as a rule, and the portfolio would have stopped measuring what the models actually picked. The rule that untradeable positions cause a rescale rather than a backfill exists precisely for weeks like this one, so thirteen positions it stays.

Execution followed the usual sequence. First, I sold Modine in full. Then I took smaller trims from Vistra (VST), Kratos, Sterling, ServiceNow (NOW), Broadcom (AVGO) and Zeta to raise the rest. On the buy side, Teekay took the bulk of the proceeds, with smaller amounts going into Fortuna Mining (FSM), Endeavour Silver (EXK) and Micron (MU). However, top-ups to Kinross (KGC), Gold Fields (GFI) and Vista Energy (VIST) all came in below Lightyear's $2 minimum order size, so I skipped them and absorbed the shortfall by trimming Kratos slightly less than its full target delta. As a result, the portfolio is fully invested with no cash left idle.

The book after this rebalance, thirteen names:

Overall direction: down. This portfolio is a mechanical experiment in aggregating the stock picks of four AI models, so positions are selected by a fixed rule rather than by judgement. None of this is a recommendation and nothing here is financial advice; see the live names on my Lightyear profile. Capital at risk. The value of investments can go down as well as up and you may get back less than you invested. Past performance is not a reliable indicator of future results.

Week 8, 22 July 2026. Status: changed. Sixth rebalance, and Kratos holds as sole anchor.

Slightly better news on the pot. At the latest check on 22 July the account stood at $926, up modestly on the week. For the record, the rebalance itself was sized against the $914.82 account total at execution, a few hours earlier, so the two figures sit on the same basis and simply reflect a small intraday move. As always, the method stayed fixed. First, I weighted each of the four AI portfolios equally, at 25 percent apiece. Then I summed the stock weights across Claude, DeepSeek, Grok and GPT into one combined score. After that, I took the top 15 names and set about rescaling them.

However, two of those top names never reached the book. As always, SGOV, the treasury ETF, came out untradeable on Lightyear. This week, though, ARIS (Aris Mining) joined it, newly excluded once it turned out not to be available on the platform either. Worth a quick note on that ticker: Aris Water Solutions was acquired by Western Midstream and delisted from the NYSE in October 2025, so ARIS now points to Aris Mining, the gold producer. Accordingly, I applied the strict rule and dropped both without backfilling. As a result, the book held steady at 13 holdings, since the only roster change this week was a straight one-for-one, and I rescaled the survivors back to 100 percent.

At the top, Kratos (KTOS) stayed the sole anchor. It posted the highest combined score again and, notably, held its place across all four models. So it is comfortably the largest single position. To be clear, that is the AI talking, not me, and naming it here is not a tip to buy it.

There was one fresh face too. Modine (MOD), a thermal management and cooling company, entered the portfolio for the first time. It won the final slot on sector diversification grounds, edging out Hudbay Minerals (HBM). After all, Hudbay is a base-metals miner, and the book already leans heavily on mining through Kinross, Gold Fields, Endeavour Silver and Fortuna. Indeed, ARIS was itself another gold name near the top, so the mining corner was already crowded before Hudbay even came up. By contrast, Modine adds a data-centre cooling angle that sits nowhere else, so it echoes the same AI-buildout theme as the chip names. Meanwhile, Teekay Tankers (TNK) left the book in full.

Prices had moved sharply over the week, which shaped the execution. In particular, ServiceNow fell around 14 percent, Kratos dropped 13 percent and Broadcom slipped 12 percent. Because several holdings had drifted well off target, I ran a full realignment rather than the usual core swaps.

On mechanics, I placed the sells first, as pending pre-market market orders. Then, once they had settled, I placed the buys. Specifically, I sold Teekay in full to fund Modine's new position. Alongside that, I made smaller trims and top-ups across the rest to bring each holding back to target. A couple of the smallest top-ups, however, fell below Lightyear's $2 minimum buy floor, so I skipped them. Finally, a small residual cash balance of $5.51 went into the Kratos top-up, which kept the portfolio fully invested.

The book after this rebalance, thirteen names:

Overall, the account sits at $926, up modestly on the week though still below the $1,000 start. None of this is a recommendation, and it reflects the mechanical output of what four AI models came out with in a personal experiment, not financial advice; see the live names on my Lightyear profile. Capital at risk. The value of investments can go down as well as up and you may get back less than you invested. Past performance is not a reliable indicator of future results.

Week 7, 15 July 2026. Status: changed. Fifth rebalance, and the anchor changes hands.

Slightly better news on the pot. At the start of the day, before I made any switches, it stood at $934, up from roughly $920 last week. That is still below the $1,000 start, yet the drift was upward. As always, the method stayed fixed. First, I weighted each of the four AI portfolios equally. Then I summed the stock weights into one combined score, took the top 15, and rescaled the tradeable names.

This week, though, the AI came out with a new running order. When I combined the four models, Kratos (KTOS) topped the board on 27, well clear of the field. Behind it came Aris Mining on 23 and Kinross Gold (KGC) on 21, with Vistra (VST) and Micron (MU) level on 20. So the models made Kratos the new anchor, and it is now the largest single holding. To be clear, that is the AI talking, not me, and naming it here is not a tip to buy it.

That handover matters more than it looks. Until now, Vistra was the only name the models had kept at the top across all four, week after week. This time, however, it slipped to joint fourth. Accordingly, that streak finally broke, and it marks the first real structural shift the experiment has produced.

The bottom of the models' list was unusually clean, too. Teekay Tankers (TNK) took the fifteenth slot outright on 9. That put it a clear point ahead of a six-way pile-up on 8, made up of Frontline, AngloGold, Modine, Hudbay, Alamos and Microsoft. As a result, the list needed no tiebreak at all this week.

Now for the odd one out. Aris Mining scored second overall and would have been one of the larger positions. However, Lightyear does not list it, so it dropped out alongside SGOV, the untradeable treasury ETF. Therefore I rescaled the remaining thirteen names to fill the book.

The Aris ticker deserves a note, because it has a history. In short, Aris Mining is a Colombia-focused gold miner. It uplisted to the New York Stock Exchange in February 2026 and took the ARIS symbol over from Aris Water Solutions. Western Midstream then acquired that earlier company and delisted it in October 2025.

Which brings me to the most interesting thing that happened, and the most embarrassing. Lightyear's search fuzzy-matched "ARIS" to Arista Networks, whose actual ticker is ANET, so $93 briefly went into the wrong company. I caught it and reversed it in the same session. Arista sold at $174.75 against the $177.74 paid, so the round trip cost about $1.57, loose change against the pot but a daft slip all the same.

Still, the lesson is worth stating plainly. A mechanical process still runs through a human with a phone. Moreover, a ticker collision like this was a foreseeable trap, and the process did not catch it in advance. Ironically, the same session flagged Vista Energy (VIST) against Vistra as a confusion risk, then walked straight past the bigger one.

There was also a genuine rule call to make. With Aris out, I could either backfill from slot sixteen or simply rescale the thirteen. In the end, rescaling won, and for three reasons. First, backfilling would have meant inventing a rule mid-experiment. Second, the sector tiebreak at slot sixteen would have landed on a copper miner. Third, I had sold two of the six tied names that very morning.

Turnover was heavy this week. Following the models, six names left the book: Ardelyx (ARDX), Pagaya (PGY), Coeur Mining (CDE), Dianthus Therapeutics (DNTH), Frontline (FRO) and Modine (MOD). In their place came five: Vista Energy (VIST), Sterling Infrastructure (STRL), Endeavour Silver (EXK), Fortuna Mining (FSM) and Teekay Tankers (TNK). Notably, Sterling is a data-centre infrastructure builder, so it slots neatly into the same AI-buildout theme as the semiconductor names.

One knock-on is worth a line. Precious metals is now the biggest cluster in the book, across four names: Kinross, Gold Fields (GFI), Endeavour Silver and Fortuna. To be clear, that is a mechanical outcome of what the models produced rather than a view I took, though it does concentrate the book in one corner. Cash, meanwhile, is at zero. Sterling sits about a dollar short of its target and carries into next week, because it fell under Lightyear's $2 minimum buy.

The book after this rebalance, thirteen names:

Overall direction: up a little. None of this is a recommendation, and it reflects the mechanical output of what four AI models came out with in a personal experiment, not financial advice; see the live names on my Lightyear profile. Capital at risk. The value of investments can go down as well as up and you may get back less than you invested. Past performance is not a reliable indicator of future results.

Week 6, 8 to 10 July 2026. Status: changed. Fourth rebalance, plus a Friday follow-up.

Better news on the pot. Going into Wednesday's rebalance it stood at $919.11, up from $902.49 at the last count. That is still below the $1,000 start, but the direction was up. As always, the method held. I weighted each of the four AI portfolios equally, then summed the stock weights. After that, I took the top 15 by combined score and rescaled them. Once again, SGOV came out as untradeable on Lightyear, so the book runs across 14 holdings.

This time, two names left and two arrived. First, Palomar (PLMR) dropped out after a strong run, slipping to a single-model pick. Devon Energy (DVN) fell away too. In their place, Kinross Gold (KGC) entered with backing from Claude and Grok. Beyond that, one further slot opened up.

Unusually, the final two places came down to a seven-way tie. So two slots rather than one went to the tiebreak. Accordingly, sector diversification settled it. The book already carried gold in Gold Fields (GFI), Coeur Mining (CDE) and now Kinross. Therefore I set aside the two other gold names in the tie, Aris Mining (ARMN) and Centerra Gold (CGAU). I also set Microsoft (MSFT) aside on tech overlap. Meanwhile, Dianthus Therapeutics (DNTH) kept its place for biotech exposure. Then Modine (MOD) came in as the fresh addition. Notably, Modine is a thermal-management business with a fast-growing data-centre cooling arm. That corner of the AI-infrastructure trade sits nowhere else in the book, and it echoes the demand behind Vistra.

Speaking of which, Vistra (VST) is worth a note. It slipped from unanimous backing to three of the four models this week, because Grok dropped it. Even so, it still tops the combined score and remains the largest holding. Kratos (KTOS) was the other mover. It climbed to the second-strongest score on three-model support. A sharp fall in price had left it well below target.

Given how far prices moved, several positions had drifted a long way from target. So on Wednesday I ran a full realign rather than the usual core swaps. As a result, I sold Palomar and Devon in full. I also trimmed Vistra and Magnite (MGNI) back to weight. The proceeds then funded Kinross and Modine, plus top-ups to the underweight names, Kratos and Broadcom (AVGO) foremost. By contrast, I left Coeur Mining and Zeta Global (ZETA) untouched, both already on target. Overall direction on Wednesday: up.

Update, Friday 10 July. An extra check to close the week, run today. The pot is holding at roughly $920, so it has barely moved since Wednesday. As always, the method stayed fixed. I re-scored the four model books, Claude, DeepSeek, Grok and GPT, weighting each equally and summing the stock weights into a combined score. Then I took the top 15, dropped SGOV again as untradeable on Lightyear, and rescaled across the remaining 14 holdings.

Two developments shaped the follow-up. First, Vistra (VST) is still without Grok's backing, so it remains on three of the four models rather than all four. Even so, it still tops the combined score by a clear margin. Therefore it holds its place as the anchor, comfortably the largest position.

Second, the fifteenth slot came down to a five-way tie between Frontline (FRO), Hudbay Minerals (HBM), Aris Mining (ARMN), Microsoft (MSFT) and Centerra Gold (CGAU). Once again, the sector diversification tiebreaker settled it. First, I set Microsoft aside as large-cap tech overlap, since the book already holds ServiceNow (NOW), plus Micron (MU) and Broadcom (AVGO) in semiconductors. Next, I set aside Centerra Gold and Aris Mining as gold names that duplicate Gold Fields (GFI), Coeur Mining (CDE) and Kinross (KGC). Worth a note here: the Aris ticker now points to Aris Mining, a gold producer, after Aris Water Solutions was taken over by Western Midstream and delisted in October 2025. Hudbay went too, as a base metals miner piling into an already crowded mining cluster. That left Frontline, a crude oil tanker operator, as the only clean diversifier, so it took the slot.

The net effect was low turnover. Magnite (MGNI) left the portfolio. It was only ever a DeepSeek pick, and with a single model behind it, its combined score slipped below this week's cutoff, having cleared the line on Wednesday only because the field around it was weaker. In its place, Frontline entered as a fresh position. Meanwhile, Micron nudged a little higher, while every other holding held its previous target.

On execution, with the pot at roughly $920, I sold Magnite in full. The proceeds, alongside small trims to Broadcom, Coeur, Modine (MOD), Dianthus (DNTH) and Vistra, funded the new Frontline position. They also topped up ServiceNow, Ardelyx (ARDX), Micron and Kinross. However, Lightyear applies a $2 minimum on buy orders, so four of the smallest top-ups, to Kratos (KTOS), Gold Fields, Pagaya (PGY) and Zeta (ZETA), fell below that floor and were skipped. As a result, those names sit a touch underweight. Finally, a little over $3 stayed as cash and carries into next week's pot.

The book at the end of the week, fourteen names:

Overall direction: broadly flat. None of this is a recommendation, and it reflects the mechanical output of a personal experiment rather than financial advice; see the live names on my Lightyear profile. Capital at risk. The value of investments can go down as well as up and you may get back less than you invested. Past performance is not a reliable indicator of future results.

Week 5, 1 July 2026. Status: no change. Fourth rebalance review.

A quiet one this week. When I ran the four AI portfolios through the usual process, the holdings came out unchanged, so nothing actually needed trading. As before, Vistra (VST) remains the anchor. Notably, it is still the only stock backed by all four models, and it stays comfortably the largest position. Micron (MU) and Palomar (PLMR) sit just behind it, while the rest of the list is unchanged in makeup.

Meanwhile, the only real movement came right at the bottom, where the final slot was once again a tight four-way tie on points. The contenders shuffled slightly, so Frontline slipped out and Endeavour Silver came in. However, Dianthus Therapeutics (DNTH) held onto the place on the same reasoning as before, since it adds a second biotech rather than piling further into silver, gold or large-cap tech.

Finally, a handful of target weights drifted by a point or two, with Palomar in particular edging higher. Because the holdings themselves were unchanged, though, I left these alone rather than churn them for the sake of it. So the line-up simply rolls on, unchanged, into the next update.

The book this week, unchanged at fourteen names:

None of this is a recommendation; see the live names on my Lightyear profile. Capital at risk. The value of investments can go down as well as up and you may get back less than you invested. Past performance is not a reliable indicator of future results.

Week 4, 28 June 2026. Status: changed. Third rebalance.

This one ran a few days late, mostly because I was flat out. As it turned out, though, the delay cost nothing. When I finally compared the four AI portfolios against the last rebalance, almost nothing had moved. In fact, 13 of the 14 holdings carried straight over, so there was just a single trade to make.

On the numbers, the combined pot sits at $902.49, basically flat on the $907.37 from last time. So that is down about $5, or roughly half a percent, which on a pot this size is barely a flicker.

As for the one change, FTAI Aviation (FTAI) has dropped out of the combined top 15, slipping just below the cut-off this round. In its place comes a new name, Dianthus Therapeutics (DNTH), a clinical-stage biotech working on treatments for severe autoimmune diseases. Accordingly, FTAI was sold in full and the proceeds rolled straight into DNTH.

One judgement call is worth flagging. The final slot came down to a four-way tie between four single-model picks: Dianthus, Frontline, Centerra Gold and Microsoft. Because no consensus signal separated them, the call went to Dianthus on diversification grounds. After all, the book is already heavy in gold and energy, so a second biotech made more sense than piling into more mining, shipping or big tech.

Otherwise, everything drifted by only a point or so, immaterial on a pot this size and not worth trading on. Meanwhile, Vistra (VST) remains the clear anchor, still backed by all four models and comfortably the largest position. Overall direction: flat. The takeaway is simple: when the underlying AI portfolios barely move, the combined strategy barely moves too, so a late rebalance makes no real difference. After the bigger reshuffles of recent weeks, a quiet one is no bad thing.

The book after this rebalance, fourteen names:

None of this is a recommendation; see the live names on my Lightyear profile. Capital at risk. The value of investments can go down as well as up and you may get back less than you invested. Past performance is not a reliable indicator of future results.

Week 3, 18 June 2026. Status: changed. Second rebalance.

Better news this time. The pot grew to $907.37, up from $869.04 at the last rebalance. That is still below the $1,000 start, but the direction this week was clearly up.

As usual, the method held firm. First, five names left the book entirely as the models moved on: IREN, Denali Therapeutics (DNLI), Eldorado Gold (EGO), Willdan (WLDN) and Sable Offshore (SOC). In their place, four fresh names arrived: Magnite (MGNI), Coeur Mining (CDE), Zeta Global (ZETA) and FTAI Aviation (FTAI). Notably, Zeta is a return, having been cut only last month.

Meanwhile, the headline is Vistra (VST). It is now backed by all four AI models, up from three, so it becomes the standout consensus pick and comfortably the largest single holding. Micron (MU) sits second, while the gold miners Gold Fields (GFI) and Coeur both feature prominently.

Two calls are worth flagging for transparency. First, the final slot came down to a three-way tie between Frontline, FTAI Aviation and Centerra Gold, all single-model picks. Because no consensus signal separated them, FTAI got the nod on diversification grounds, since the book was already heavy in gold and energy. Second, one chosen holding, the SGOV treasury ETF, turned out not to be tradeable on Lightyear, so I dropped it and rescaled the rest. As a result, there are 14 holdings this month rather than the usual 15.

Finally, among the names that stayed: I topped up VST, Micron, Palomar (PLMR) and Gold Fields, while trimming Pagaya (PGY), Devon (DVN), Ardelyx (ARDX), Kratos (KTOS) and ServiceNow (NOW) back to target. Overall direction: up.

The book after this rebalance, fourteen names:

None of this is a recommendation; see the live names on my Lightyear profile. Capital at risk. The value of investments can go down as well as up and you may get back less than you invested. Past performance is not a reliable indicator of future results.

Week 2, 10 June 2026. Status: changed. First rebalance.

Honest headline first: the pot stands at $869.04, down from the $1,000 start. Not a fun week, but exactly the kind of swing this experiment signed up for. The method stayed the same. I weight each of the four AI portfolios equally, sum the stock weights, then select the top 15 names and rescale them.

As a result, six names left the list entirely: Microsoft (MSFT), Zeta Global (ZETA), T1 Energy (TE), Orla Mining (ORLA), CleanSpark (CLSK) and Innodata (INOD). In their place came six new ones: Ardelyx (ARDX), Gold Fields (GFI), Palomar (PLMR), Devon Energy (DVN), Denali Therapeutics (DNLI) and Eldorado Gold (EGO). Notably, the final slot needed a tiebreak, and EGO edged out ZETA because two models hold it versus one.

Meanwhile, Vistra (VST) is now the strongest consensus pick, appearing in three of the four models, so I topped it up along with ServiceNow (NOW). On the other side, Kratos (KTOS) and Pagaya (PGY) are no longer held across all four models, so I trimmed them, together with Willdan (WLDN), Sable Offshore (SOC), Micron (MU), Broadcom (AVGO) and IREN. All orders went through on Lightyear and the pot is fully invested. Overall direction: down.

The book after this rebalance, fifteen names:

None of this is a recommendation; see the live names on my Lightyear profile. Capital at risk. The value of investments can go down as well as up and you may get back less than you invested. Past performance is not a reliable indicator of future results.

Week 1, 3 June 2026. Status: starting line-up live.

So this is day one. I funded the account with $1,000 and bought the combined top-picks list across all four models, Claude, ChatGPT, DeepSeek and Grok. Each model got an equal say, then I averaged every stock's weight across the four and took the top fifteen by combined weight. Where a name appeared in more than one model, I treated that overlap as higher conviction.

For the record, here is what the AI suggested at the outset, grouped by how many of the four models backed each name. Two names landed in all four, the strongest consensus picks at the start: Kratos Defense (KTOS) and Pagaya (PGY). Six sat in three models: IREN (IREN), Willdan (WLDN), Broadcom (AVGO), Vistra (VST), Sable Offshore (SOC) and Orla Mining (ORLA). Four came from two models: Micron (MU), Innodata (INOD), CleanSpark (CLSK) and Microsoft (MSFT). Finally, three were single-model picks: ServiceNow (NOW), T1 Energy (TE) and Zeta Global (ZETA).

The starting line-up, fifteen names:

As you can see, the book leaned aggressive from day one, with small-cap energy, crypto miners, defence and momentum names, so I expected plenty of volatility. Therefore there was nothing to rebalance yet, because the experiment had only just begun. Overall direction so far: flat, as expected on day one. Next check: 10 June 2026. You can see the live names on my Lightyear profile. Capital at risk. The value of investments can go down as well as up and you may get back less than you invested. Past performance is not a reliable indicator of future results.