Six Reviews, No Buys, and a Track Record Log That Was Wrong

We reviewed six assets in our first week of publishing and admitted none of them. Then we audited the record of those reviews and found two things wrong with it.

Neither one changes a figure we published. Both make a published figure harder to prove later, and proof is the entire pitch.

This note is the audit, what it cost us, and what is being built so the next cycle does not repeat it. At the end we put the desk itself in front of the same four investment seats it runs, on the same terms as any company. They did not buy it either.

What we published

Five posts went up between August 8 and August 12, 2026.

One was a NAV placeholder that reported the figure as not available rather than inventing one. One announced the committee. The other three were reviews: Jupiter, NVIDIA, and four AI buildout names in a single thematic note.

Six assets in total: JUP, NVDA, MU, CRWV, VRT and SMCI.

Our published rule admits a name when three of the four seats return Buy. None of the six cleared it. On the AI buildout note alone, sixteen independent assessments produced no consensus buy.

We hold none of them. We manage $0, and we have never accepted outside capital.

The first lesson: missing data changed the answer

We expected gaps to annotate a verdict. They rewrote it.

On Jupiter, three fields came back not available: the unlock and vesting schedule, which sits behind a paywall at the source, the protocol treasury, and how much of $361.0M in annualised fees actually reaches a token holder. Investor B declined to price the asset at all on that basis, at conviction 9 of 10. That was the strongest single moment in the piece, and it was produced by absence rather than by analysis.

On the AI buildout names, the first pass carried four not available fields: CoreWeave's share count and customer concentration, Micron's mid cycle earnings, and the composition of Super Micro's cash outflow.

Three of those four were already in the filings we had retrieved. They were not in the structured data our fact sheet reads, because they live on a cover page or inside a disclosure written in prose. The fourth was arithmetic across seventeen years of annual filings that nothing was computing.

We rebuilt the sheets and re-ran three of the four names. Seven of twelve ratings changed. Micron moved from a divided Watch to three seats at Sell. Super Micro moved the other way and split the committee, because two seats read the newly available inventory figure in opposite directions.

Nothing about those companies changed between the two passes. Only what we could see did.

Both passes stayed in the log with their timestamps. We did not delete the weaker run, because the gap between the two is the only measure we have of what the missing figures were worth.

The second lesson: our own log was wrong

The verdict log is the artefact everything else rests on. A call that is not written down at the moment it is made cannot be scored later, because by then nobody can prove what was said or where the price was when it was said. Without it there is no track record, and without a track record there is no honest answer to whether this committee is any good.

We audited it on August 13, 2026, before anybody had scored a single call. Two holes.

Five seat verdicts were recorded as blank. The reasoning was present and complete in every case. The log looks for the verdict word directly beneath a verdict heading, and one seat writes its heading with a hyphen before the word while another places the verdict further down its answer. Neither form matched. Investor D lost four, Investor B lost one. Five of thirty six seat slots, or 14%, are currently invisible to any automated scoring.

One published session is missing entirely. The NVIDIA review we published on August 11 was drawn from a committee session that never reached the log. What the log holds under NVDA is an earlier run from August 9 with different verdicts, including a Hold from a seat that returned Sell in the session we actually published. The article is correct against its own transcript. The log is the thing that is wrong.

Both are recoverable, and only because of a decision made before either happened: when a verdict cannot be read, the log keeps the raw tail of the answer rather than dropping the session. Discarding an unreadable verdict would have quietly biased the whole track record toward the sessions that happened to format well, and we would have had no way of knowing.

That is the pattern worth taking from this. The controls that caught these were not the clever parts of the system. They were the boring decisions to keep the weaker run, keep the unreadable answer, and write not available instead of a plausible number.

What we are adding

The desk currently reads six sources.

Three more went in after the first note shipped.

Hyperliquid perpetual data, covering marks, funding, open interest and liquidation distance for equities, indices and commodities. Read only, and barred from committee fact sheets on purpose, because funding is a crowding signal and a crowding signal handed to an investment seat is momentum wearing a different hat. Coverage is real and partial, and we say which: measured on August 12, 2026, MU and NVDA are listed on the venue and SMCI and VRT are not.

A deterministic price sheet. Every figure in it is computed in the file rather than by a model. A model asked for a 50 day average returns a plausible number that is not the average, and a wrong moving average looks exactly like a right one. It feeds a technical layer that runs after the committee decides and carries no vote.

A value investing discussion board archive running from 2011 to today, read for ideas only and never cited as a fact. We hold 4,284 posts of roughly 276,276, which is under 2%. Partial coverage is stated rather than hidden, because a thread with nothing stored has not been read yet, and that is missing data rather than a quiet thread.

Next, in order, are the three things this audit named. Filing text beyond the structured feed, so cover pages, customer concentration, segment notes and inventory reserves stop arriving as gaps. Multi year filing history, so mid cycle earnings are computed from the record instead of assembled by hand for one article. And the log fix, with a backfill from the transcripts we kept.

We put the desk in front of its own committee

The four seats read companies. On August 13, 2026, they read us, on a fact sheet built the same way, carrying everything above including both holes.

Investor A, business quality and durable advantage. Called the process a pre revenue venture and the temperament the actual asset. "It writes not available instead of inventing a number. It kept the weaker of two analytical passes in the log rather than quietly deleting it. It audited its own track record before anyone graded it." Then the objection: a process whose conclusions flip depending on whether a figure sat in a structured feed or on a cover page was, in its first cycle, partly measuring its own plumbing rather than the businesses. A also named the sharpest risk to a reader, which is that presentation quality is running ahead of decision reliability.

Verdict: Watch. Conviction 7 of 10.

Investor B, margin of safety. Refused it, at the same conviction it refused Jupiter. "The core asset, a reliable and measurable decision record, is not yet proven." B listed the absence of any scored outcome, the nine session sample, and the fact that structured extraction missed material information already sitting in retrieved filings. On whether the honesty counts for anything: "absence of exposure is not evidence that the process deserves capital."

Verdict: Avoid. Conviction 9 of 10.

Investor C, macro and positioning. Read the timing rather than the flaws. "The window to fix the plumbing is now, before reflexivity begins." C put the danger in the data layer rather than the decision layer, and argued zero admissions across six assets is the gate working rather than a gate set too high. C also set the only numeric bar anyone offered: the next data refill should move fewer than a third of ratings, and if the rate stays where it is, the process is not converging on stable verdicts, it is oscillating.

Verdict: Watch. Conviction 6 of 10.

Investor D, growth and runway. Classified the desk as a turnaround, with a sound analytical core sitting on damaged infrastructure. D was the only seat to argue the committee's own design might be the problem: a three of four consensus rule may filter out exactly the high conviction ideas that produce returns, and a desk that admits nothing regardless of data quality has a structural fault rather than a strict standard.

Verdict: Watch. Conviction 7 of 10.

Where they split

Every seat named the same two defects. They disagreed entirely about what the defects mean.

B treated a broken record as disqualifying on its own, because a process asking to be judged on a track record has to produce one first. A and C treated the same fact as evidence in the opposite direction, on the grounds that most research desks never discover their verdicts are artefacts of their data pipeline, and fewer publish it when they do. D read straight past both to the pipeline work and asked whether it ships.

The one thing none of them would grant is the thing we cannot yet supply. Every outcome field in the log is null. No call is old enough to be judged, so nothing here shows the committee's disagreement is predictive rather than merely genuine. B put the floor under it plainly: without a reliable record or scored outcomes, any material commitment rests primarily on narrative.

We agree, which is why nothing is deployed.

What this note does not establish

It does not establish that the fixes work. They are described here as commitments with an audit behind them, not as completed work, and the honest test is whether the next cycle produces a materially lower rating change rate on a refill.

It does not establish that anyone reads this. We have no traffic, readership or engagement figures for any of the five posts, and we are not going to estimate them.

It does not establish skill. Six reviews, zero admissions and zero scored outcomes is not a track record. It is the beginning of one.

And it does not resolve Investor D's objection, which is the uncomfortable one. A gate that has admitted nothing across six names is either a standard or a defect, and six names is nowhere near enough evidence to tell which.

A note on method

The fact sheet the committee read is the same one reproduced above, gaps included. The seats see the sheet and nothing else, which is what makes their disagreement meaningful rather than an artefact of different inputs. Each runs on a different underlying model and a different published philosophy, and none can see another's work.

This session was not written to the verdict log. The log records asset verdicts, and we are not an asset. Recording it there would corrupt the exact record this note is about repairing.

As of August 13, 2026, BagsCapital manages $0 in assets under management, holds no position in any asset named here, and has never accepted outside capital. The committee is advisory and does not trade. Nothing in this note is a recommendation.

Correction, August 13, 2026

This note first published with the wrong count. It said four seat verdicts were recorded as blank, and that Investor D lost three of them. The log holds five, and Investor D lost four. The corrected share of affected seat slots is 14% rather than 11%. Everything else in the section stands, including the missing NVIDIA session.

The error was ours in counting, not the log's, and it was caught within the hour while implementing the parser fix described above. We have left the original figures visible in this paragraph rather than replacing them quietly, because a published number that changes without a record is the thing this note exists to argue against.

The parser is now fixed. It accepts any separator between a heading and its value, reads bold, hash and bare-line headings, and requires word boundaries so that a seat writing "no buyer at this price" can no longer be logged as a Buy. Three of the five blank verdicts were recovered from the retained tails immediately.

The other two needed a second correction. This note first said they were being restored from the full session transcripts. There are no such transcripts. The committee output for the AI buildout batch was never cached, so it does not exist in any form we hold. Both verdicts were instead restored from our own published note of August 12, which stated them in the text: Investor B held Avoid throughout on Super Micro, and the ratings table gives Investor D at Hold. The verdicts are recovered. Their conviction scores are gone permanently, and the log now records them as null with the source of the recovery attached, so nobody later mistakes a figure read off an article for one read off a transcript.

The log is now complete. Ten sessions, forty seat slots, no blanks. The August 11 NVIDIA session has been added from its cached transcript and matches the published article on all four seats. The superseded August 9 run is kept and marked, not deleted. One thing the repair confirms rather than changes: the seven of twelve figure in the August 12 note is correct, and the repaired log now proves it.