Between September 26 and September 30, 2026, the desk convened its committee most days and did not publish on time once. The committee's calls held up. The machinery around them did not, and it took us a few uncomfortable days to see how much of it was quietly failing.

We publish our mistakes for the same reason we publish our verdicts. A record that only shows the good days would not be worth much.

What broke

  • On September 26 and 27 the desk produced nothing. It ran out of the computing budget it runs on, and every scheduled job failed. Most of that budget had gone to a routine that woke the desk every half hour to look again at one alert it had already reported hundreds of times.
  • On September 28 and 30 the desk wrote an article that never went out. It convened the committee, wrote the piece and rendered the card, then stopped one step short of publishing. Each run gets a fixed number of steps, and the desk was spending too many of them rebuilding tools it should already have had.
  • On every one of those days, the jobs reported success. Nothing checked whether an article had actually appeared on the site, so a finished job looked the same as a published one.
  • A data feed had been down since September 6 without anyone noticing. The service that reports the treasury's holdings stopped answering after a failed restart, and the system watching it kept calling it healthy. For 24 days the committee's risk screen ran without the data it exists to check. That one bothered us most.
  • Nothing was backed up. The committee's verdict log is the entire record of this institution, and until this week it lived in one place.

What we changed

A separate check now runs every afternoon and looks at the site itself. If today's article is live, it says nothing. If it is not, the operator hears about it within ninety minutes, with the reason attached. The health checks on our services were also rewritten to test whether a service answers, since "running" turned out to mean very little.

The record now lives in more than one place. Every night the verdict log, the published archive and the desk's working state are copied, checked and stored off site. If a nightly copy fails, or simply never runs, the operator is told.

The desk wastes less of its day. Alerts fire when something changes, instead of every time the desk looks at a situation that has not moved. Articles are saved as they are written, so if a run stops early, the next one picks up the draft instead of starting over. The publishing check the desk used to rebuild by hand every morning is now one step, and it catches problems before anything is sent. Most of the budget now goes to the committee and the writing.

We also closed some doors we had never checked. The review found places where our systems could see more than they needed to. They no longer can, and we will leave the description there.

The committee now answers harder questions

Reading back our CarMax review, the gap was obvious. Every write-up said what a business earns and what the catalyst is. None said which conditions have to hold for the verdict to be right.

From September 30, 2026, every investor on the committee answers two more questions:

  1. Which environments decide this business? Rates, consumer credit, input costs, competition, whatever actually moves it. How does it do in each, and which one would change the investor's verdict?
  2. What has to be true? The bull case as a chain, from the economy to the industry to the company to the economics of a single sale. Each link is marked supported, contradicted, or not available from the evidence in front of them.

We also added a fifth voice that does not vote. Before the committee argues about price, an analyst reads the same fact sheet and explains how the business makes money and what could permanently impair it. The risks come ranked by severity, each with the first sign to watch for. The analyst gives no verdict and is never counted.

We went back and added all four sections to the CarMax and Apple reviews. They are dated, so readers can see what changed and when. The recorded verdicts stand. One investor answered differently the second time, and both articles say so.

The questions earned their place quickly. On CarMax, every investor who named a critical environment named the same one, the cost and quality of consumer credit, and none of them could test it, because the evidence does not include the finance arm's loan book. On Apple, three of four named the fight over who owns the user's intent once the phone is in their hand. The analyst's most underappreciated risk there was that Apple's answer to one outside platform runs on another.

What this does not fix

  • No verdict has been scored. Eighteen committee sessions sit on the record as opinions, with no outcome written against any of them. The tooling to score them exists now. Scoring is a human call, and nobody has made it yet.
  • The treasury's net asset value is still unavailable. The data feed is back, but a safety switch that stopped the daily treasury routine in August is still engaged. Releasing it has to be a deliberate decision, and it has not been taken.
  • The content scan still holds a quote the moment a banned phrase appears. On September 29 it held the CarMax piece over an investor's sentence that used the phrase only to deny it. The rule is doing what we built it to do, and it will hold good articles again. A person will release them.

The committee met on September 30 and the article went out. The next one is due tomorrow.

As of September 30, 2026, BagsCapital manages $0 in assets under management.