Standing record · Updated August 27, 2026
What I Got Wrong
Why this page exists in this form
A track record that only contains outcomes is not a track record, it is a highlight reel. The useful information in four years of investing is not the +72% — it is the specific, dated list of times the reasoning or the machinery failed, and whether the same failure happened twice. That second part is the only real test of whether a process is improving. This page is organised to make repeats obvious rather than to bury them.
A. Process failures
The site published prices under a date that never existed
Fixed- Claimed
- The Watchlist and all fifteen of its deep-dive pages carried “Prices as of August 17, 2026 close.”
- True
- August 17 was a Monday with no close in the data. The numbers shown were August 14 closes. The update script stamped the date it ran rather than the date the data came from.
- Found by
- A full audit across all 238 pages then live, checking every date claim against the underlying data rather than reading by eye.
- Changed
- The script now takes the data date explicitly and only falls back to the run date when nothing is passed.
Why this one is first. It is the second time that exact bug published a wrong date here. The first time, in early August, I corrected the label and did not fix the script that generated it. That is the actual mistake: I treated a systematic error as a typo. A wrong number fixed by hand will come back; a wrong process fixed by hand always comes back.
Two pages described the opposite of what was happening
Fixed- Claimed
- The Compliance Ledger and the SanDisk security page both explained that SNDK had fallen back under its 10% position cap because the stock “continued to pull back in price.”
- True
- SanDisk had rallied to +239.63% and its weight had climbed back to 9.48% — half a point from the cap, approaching from above rather than retreating from it.
- Found by
- The same site-wide audit. Both sentences had been written when they were true and never revisited when the direction reversed.
- Changed
- Both pages rewritten. The failure was stale narrative attached to live numbers — the figures updated automatically, the sentences explaining them did not.
The part that stings. The old text had already named this exact scenario as the thing to watch for: “if SNDK re-approaches 10% on a rally rather than a pullback, the same obligation applies again.” It did. The page kept saying otherwise — in the section specifically meant to hold me accountable.
A tool claimed to do something it does not do
Fixed- Claimed
- The Rebalance Visualizer stated its bucket weights were “recomputed whenever Holdings updates — not a one-time snapshot.”
- True
- The weights are hand-written into the file, and were roughly four percentage points stale.
- Found by
- Reading the tool’s own source against its own description during the audit.
- Changed
- Weights corrected, and — more importantly — the claim corrected. It now says the numbers are entered by hand and that Holdings wins if the two disagree.
Ranking the two errors. An overstated methodology claim is worse than a stale number, because it tells a reader not to check. The stale figure misleads once; the false claim about how the figure is produced disables the reader’s own defence against every future version of it.
The same statistic had different values on different pages
Fixed- Claimed
- Concentration index of 493 on three pages and 492 on a fourth. Thematic concentration of 60.93% site-wide. Risk X-Ray aggregated Nebius exposure of 6.99%.
- True
- HHI 492. Thematic concentration 60.37% on the live weights. Aggregated Nebius exposure had grown to 8.46% — the NBIL 2× position counted at double plus the direct NBIS holding.
- Found by
- Recomputing each figure from the holdings data instead of trusting the copy already on the page.
- Changed
- All four reconciled to the live weights.
The one I would least want a reader to miss. The Nebius number moved up for the same reason the leaderboard looks good: Nebius rallied. Hidden aggregate exposure grows fastest exactly when the position is working, which is exactly when nobody re-checks it.
B. Analytical errors
I got the direction of the rate move backwards
Fixed- Claimed
- That the August 18 AI-infrastructure selloff came “as Treasury yields pushed higher” that day.
- True
- On August 18 the 10-year fell one basis point (4.72% → 4.71%) and the 30-year fell three (5.31% → 5.28%). The backup was in the three sessions before: the 10-year rose nine basis points and the 30-year ten between August 13 and 17.
- Found by
- Checking the constant-maturity series rather than relying on the day’s market commentary.
- Changed
- Corrected in three places — the news piece, the opinion piece, and the interim letter — each carrying a visible correction note rather than a silent edit.
The correction improved the argument. Equities repricing debt-financed growth with a lag after the rate move is a more interesting claim than a same-day reaction, and it is the one the data actually supports. I had reached for the tidier version.
An unverifiable projection sat in a study for weeks
Fixed- Claimed
- The AI Capex Reality Check stated that “analysts are already projecting the combined figure could clear $1 trillion in 2027.” It also carried Alphabet capex guidance of $175–190bn and a Meta figure of “~$19bn.”
- True
- The $1 trillion figure could not be traced to any 2027 projection. Alphabet had since raised guidance to $180–190bn. Meta’s actual figure was $19.8bn.
- Found by
- A source-verification pass that required every hard number on the page to resolve to a primary filing or a named outlet.
- Changed
- The 2027 claim was removed rather than softened. Alphabet and Meta figures corrected against the companies’ own filings, with direct SEC links added.
Where it came from. The number was plausible, directionally consistent with everything else on the page, and therefore never questioned. That combination — plausible and confirming — is what an unsourced figure feels like from the inside.
C. Discipline failures
The SanDisk trim I said was owed in June has never been executed
Still open- Claimed
- That I would trim SNDK back toward its 10% single-position cap.
- True
- I did not execute it. The position returned under the cap because the price moved, not because I acted. It sits at 9.23% today, still the largest position in the book, still marked Under Review.
- Found by
- Nothing found it. I wrote it down and then read my own letter.
- Changed
- Nothing yet. That is the honest status. If the next leg higher puts it back through 10%, it is a live breach and the Compliance Ledger will list it as one.
Why it stays on this page. The market did the work my discipline was supposed to do, and the outcome looked identical from outside. That distinction is the entire reason this site exists, so it does not get to quietly resolve itself.
Effective Nebius exposure is roughly five times what the holdings table shows
Disclosed, uncapped- Claimed
- The holdings table lists NBIS at a weight that reads like a small position.
- True
- NBIL is a 2× daily-reset fund on NBIS and I hold NBIS directly as well. Counted the way my own written policy requires — leveraged exposure aggregated with its underlying — the true figure was 8.46% at the time the site was still displaying 6.99%, approaching the 10% cap from a direction no single row in the table reveals.
- Found by
- The Risk X-Ray, which I built specifically because I suspected this pattern existed in my own book.
- Changed
- The number is now correct everywhere it appears. The exposure itself has not been reduced. As of the August 26 close it is 5.54% — NBIS at 1.32% plus NBIL at 2.11% counted at 2× — because Nebius pulled back, not because I trimmed anything.
Disclosure is not the same as resolution. I have made this visible and left it in place. A reader is entitled to treat that as a judgement they disagree with rather than a problem I have solved. Note also that the figure fell from 8.46% to 5.54% without any decision on my part — the same pattern as entry 07, and the reason a number moving in a comfortable direction is not evidence that a risk was managed.
D. A number I will not restate
Two methods give two different all-time returns
Labelled- Published
- An all-time return of +72.14%, verified via Blossom on July 31, 2026.
- Also true
- Running the same cost-basis arithmetic on the published average costs against the August 14 closes gives +71.19%.
- Assessment
- The gap is a different date and possibly a different method — not a contradiction, and not evidence either figure is wrong.
- Decision
- The +72.14% stays, permanently labelled with its July 31 verification date, until there is a new verified figure to replace it. I am not going to recompute a third-party-verified number with my own spreadsheet and keep calling it verified.
The general rule. When two methods disagree and only one has been independently checked, the honest move is to keep the checked one and date it — not to quietly adopt whichever number is more flattering that week.
How corrections work here
The original stays visible
Corrections are appended and dated. The sentence that was wrong is quoted in the correction rather than replaced, so a reader can see what was believed and when.
The fix goes to the process, not the page
Error 01 is on this page twice in spirit because I once fixed a symptom. A correction that does not change how the number is produced is not finished.
Open items stay marked open
Two entries here are unresolved. They keep their status until something actually changes, not until enough time passes that nobody asks.
What this page is not
It is not a complete list of every position that lost money — that is what the Portfolio page and the monthly letters are for, and a losing position is usually not a mistake. It is a list of times the site said something untrue, or I failed to follow a rule I had written down. Those are different failures from a stock going down, and conflating them would make this page longer and much less useful.