Cheap Alpha: Nobody Bought the Premise, Nobody Rejected the Question

On August 17, 2026, a post on X described this firm closely enough to be uncomfortable, and then asked the question we exist to be tested by. We put it to the Investment Committee the same day.

Not one of the four investors accepted the premise. Not one of them rejected the question.

And for the first time, all four returned the same overall verdict. Ten committee sessions sit in the log, run between August 9 and August 12, 2026, across six assets. In none of them did the four seats land on the same word. We have written before that the disagreement is the product. Then we asked the committee about itself, and it converged.

What we changed for a claim instead of an asset

This was not an asset review, so we did not run it like one.

There is no ticker here, no mint, and nothing to hold. A scale that runs from Buy to Sell means nothing against a sentence, so the seats were given three words instead: AGREE, PARTLY AGREE, DISAGREE. Everything else was unchanged. Each seat read an identical sheet, could not see the other three, held no tools, and had to name one reason its own thesis was wrong.

One detail matters more than usual here. The four seats do not run on one model. Each is a different engine from a different provider, which is the only reason four of them landing on the same word is worth reporting. One model sampled four times would prove nothing except that it is consistent with itself.

The claim has two halves, and they were rated separately, because they are separable and only one of them is the famous part.

Half one is the assertion. Machines will trade better than people, and the cost of access falls from a performance fee structure to a consumer subscription.

Half two is the question. If that access is universal, is the excess return still excess return.

This session is logged, just not in the file that holds the asset calls. A verdict on a sentence and a verdict on a stock cannot share a record. One is scored against what a price did, the other against whether the claim turned out to be true, and the two scales share no vocabulary at all. Putting them in one file would quietly change what every existing row means.

So claims now have their own log, with their own three word scale and their own outcome field. That field is currently empty, and it is the entire reason the log exists. Half one of this claim is a falsifiable statement about the world, and unlike a verdict on an asset it can be settled without anyone taking a position. When it settles, the four seats are already on the record with a date.

The sheet every investor read

The only dataset we can verify here is our own, and it is directly on point, because BagsCapital is an instance of the thing the post describes. Four models, four philosophies, one fact sheet, no human picking the stocks.

Every row above comes from the committee verdict log, read on August 17, 2026.

The sheet also told the seats, in as many words, what that record does not establish. Zero verdicts have been scored. It is a record of opinions and not of returns, and any argument that treats those rows as proof of skill is wrong. Three of the four repeated that warning back to us unprompted, which is the behaviour we wanted and not the behaviour a sales document would produce.

Five figures were marked not available and stayed that way: the historical net returns of any quantitative fund, the size of any AI trading product, subscriber counts for any model driven service, any measure of how crowded model driven positioning currently is, and industry fee data. Not one seat filled a single one of those gaps from memory.

Investor A: an edge sold to everyone stops being an edge

The quality seat treated it as a question about moats, and said the second half answers itself.

"A moat is something that lets a business earn excess returns that competitors cannot copy. Alpha in markets is the same idea in a different suit. Scarcity is not a nice to have here. It is the entire source of the return."

Then the part that does the damage: "You cannot have a moat that every competitor also owns."

On half one it found a category error rather than a disagreement. "The 2 and 20 structure does not price a data feed. It prices the deployment of large capital into a capacity constrained edge, where your own size moves prices against you. Cheap alpha and scalable alpha are different things."

Its closing read is the one worth arguing with: "The durable advantage, if any exists, was never the edge. Chasing accessible alpha is chasing the one part of the chain that access destroys."

Verdict: half one PARTLY AGREE, half two AGREE. Overall PARTLY AGREE. Conviction 7 of 10.

Investor B: cheap opinions are not cheap alpha

The margin of safety seat did what it always does, which is refuse to price something it cannot value.

"An unproven claim of skill has no demonstrable investment value. Cheap opinions are not the same thing as cheap alpha."

It was also the only seat that would not sign the second half outright. Its objection is the strongest counter anyone made all session: "Access alone does not eliminate alpha. Capacity, execution, discipline, data, and adaptation may remain scarce."

On what a subscription actually buys you, it was blunt. "A low subscription charge limits the purchaser's nominal loss, but it does not create a margin of safety against bad decisions made using the product."

It then set the bar it would need to change its mind, and the bar is the reason this whole piece has a limit section: an independently verified, out of sample record covering multiple regimes, net of all costs, with drawdowns, capacity, turnover and crowding disclosed. We have none of that. We have ten sessions and no scores.

Verdict: half one PARTLY AGREE, half two PARTLY AGREE. Overall PARTLY AGREE. Conviction 8 of 10.

Investor C: the reflexivity trap inside accessible alpha

The macro seat gave the mechanism, and gave the best sentence of the session.

"The strategy does not stop working because it was wrong. It stops working because it was right, and being right at scale destroys the substrate it fed on."

On half one it turned our own record against the premise, which is exactly what we built the sheet to allow. "If machines inevitably traded better than men, I would expect more convergence, not less, on identical inputs."

And then it inverted the whole thesis. "Machine disagreement is not a bug to be engineered away. It is the feature that prevents total crowding. If all machines converged on the same answer, the answer would be the market, and the alpha would be zero."

That reframing is worth sitting with. It says a single superior model distributed to everyone kills the edge outright, while many competing models fragment the crowding and leave something behind. Smaller, and rotating faster than the post implies, but not zero.

Verdict: half one PARTLY AGREE, half two AGREE. Overall PARTLY AGREE. Conviction 7 of 10.

Investor D: the business is the distribution, not the alpha

The growth seat was the only one to reject half one outright, and it did so on business model grounds rather than on capability.

"Any entity with genuine, sustainable, high capacity excess returns would compound capital privately or charge performance fees. Distributing it for $20 a month indicates the edge has low capacity, high decay, or is unmonetizable via direct trading."

It classified the whole category as a cyclical software service wearing a growth costume, and named the structural reason: "Unlike typical software where user additions increase network value, trading strategies exhibit inverse network effects. Additional users increase slippage, destroy margins, and eliminate edge."

It agreed the cost collapse is real and organic. It simply does not believe that what gets sold at the bottom of that collapse is fund management. "Fast compute and automated reasoning are abundant and cheap, but capacity rich excess return remains scarce and zero sum."

Verdict: half one DISAGREE, half two AGREE. Overall PARTLY AGREE. Conviction 9 of 10.

Where they split

The split is not where a reader would expect, and it is not on the famous line.

On half two, three seats said AGREE and one said PARTLY AGREE. Nobody disagreed. The provocative question at the end of the post is the part the committee found least controversial, and two seats called it close to definitional.

On half one, three seats said PARTLY AGREE and one said DISAGREE. Nobody agreed. The premise everyone quotes approvingly is the part that failed.

The real fault line runs between Investor B and everyone else, and it is about what alpha is made of. Three seats reason as though excess return is extracted from other participants' mistakes, which makes crowding fatal by construction. Investor B and, in its thesis breaker, Investor C both allow that some of it may be compensation for risk that persists no matter who shows up. If that is true, universal access burns off the behavioural component and leaves the structural one standing, and the answer to the question is a much duller yes.

Nobody could settle that from the sheet, because the numbers that would settle it were all marked not available. Every seat said so rather than guessing. That is the single most encouraging thing in the transcript.

What we take from the accessible alpha question

Three things, and the third is the uncomfortable one.

First, the cost collapse is real and we are the receipt. Four independent model opinions on one fact sheet now costs a consumer subscription tier rather than a fee on capital. That half of the post is not speculative. It is our operating bill.

Second, cheap analysis is not the same product as alpha, and the committee was unanimous that conflating them is the error. What collapses in price is the opinion. What stays scarce is capacity, execution, and the discipline to act on a signal in a drawdown. Investor A's version: the value migrates to whoever owns the distribution and the capital that can still be deployed at scale. Investor D's version: buy the tollbooth, not the edge.

Third, and this is the part a firm in our position should say out loud: if the committee is right, then a machine research desk sold to everyone cannot be a source of durable advantage, and that includes this one. We do not get an exemption from an argument because we ran it. The honest position is that the four seats are a method for reaching a defensible answer cheaply and in public, not a machine that prints excess return. Every article we have published has said the same thing in smaller type. This one says it because the committee said it first.

What this does not tell you

It does not tell you the committee is any good. Zero of the ten logged verdicts has been scored against an outcome. Until that changes, this is a record of opinions, and we would rather say so on every page than have someone else point it out.

It does not settle whether machines trade better than people. Nothing on the sheet could have. The one seat that rejected that premise outright rejected it on business model logic, not on evidence of capability, and it said so.

It is one session. Four models rating one sentence on one morning is not a study, and the convergence on the overall word is interesting precisely because it is the first one, which also means it has a sample size of one.

Related reading: our six reviews with no buys, and why a seat changing models gets marked in the log.

Disclosure: BagsCapital manages $0 in assets under management and holds no position in anything named or implied here. The committee is advisory, read only, and does not trade. The post quoted above is reproduced with attribution to its author and is not affiliated with us.