24 August 2026
Nobody is arguing with your critic
A lone critic has no counterparty, so a weak objection and a strong one reach you looking identical. Disagreement is what tells the two apart.
6 minutes to read
Ask a model to attack your plan and it will produce a list of objections. Some of them will be serious. Some of them will be things it said because a list of one item looks lazy. From the page, they are indistinguishable.
That is the problem this post is about, and it is not a problem with the model's reasoning. It is a problem with the shape of the exchange. A critique produced by one party, for one reader, has nothing standing against it, so nothing sorts the strong objections from the ones that merely sound strong.
A critique with no counterparty has no price
In any adversarial process worth the name, making a weak argument costs the person who makes it. A barrister who raises a point that collapses under cross examination has damaged their case. A referee whose objection is answered in one paragraph has spent credibility. The cost is what makes the strong points identifiable, because those are the ones somebody was willing to pay for.
A lone critic pays nothing. There is no opposing party to answer it, no one whose position is worse if the objection stands. Every objection costs the same to produce, which is nothing, so every objection arrives at the same apparent weight. The reader is left doing the sorting, and the reader is the person who wrote the plan.
The name for that failure on this site is the uncritiqued critic: a critic whose objections have no opposing party to answer them, so nothing separates the strongest from the weakest. It is one of the three ways a single reviewer fails, and the other two sit alongside it on the page that sets out why disagreement carries more information.
The tie break falls to the wrong person
Consider what actually happens when you read a list of eight objections to your own proposal. You have limited time and you already believe the proposal is broadly right, or you would not have written it.
So you triage. And the objections you dismiss first are the ones that would be most expensive to act on, because those are the ones where a dismissal saves you the most. The sorting mechanism is not how good the objection is. It is how much it would cost you to accept it, applied by the person who pays.
This is the failure that survives every improvement to the critic. A better critic gives you a better list. It does not give you a second opinion about which item on the list matters, and that judgement is the one you cannot make yourself.
Disagreement is the signal a critique cannot produce
Put two parties in the room who are optimising for different outcomes and the sorting happens without you. Where both of them object to the same thing, that objection has survived two different value functions, and you have learned something a single critic could not have told you. Where they disagree with each other, you have learned something better: the assumption underneath the disagreement is the one carrying the weight of the plan.
That second case is the useful one and it has no equivalent in a critique. A lone critic cannot flag which of its own objections rests on a contestable premise, because it has no access to a party who contests it. Two advisers who price the same assumption differently expose it by arguing, and the argument is legible in a way a list is not. You can see which one gave ground.
It follows that the objection an adviser will not drop under pressure from another adviser is worth more than the same objection written down by itself. Nothing about the sentence changed. What changed is that somebody with an incentive to knock it over tried and failed.
What this asks of the personas
For the mechanism to work, the parties have to genuinely differ, and differing is harder to arrange than it sounds. Three copies of the same model with three different job titles will converge, because a job title is a costume and the value function underneath it is identical.
What has to differ is what each one is optimising for. One that cares about whether the cash survives will accept a risk that a second one, who cares about whether the system holds under load, will not. Each is given a blind spot on purpose, published on the page rather than hidden, so the others have something true to attack rather than a performance to applaud.
There is a related reason an instruction to be more critical does not get you here, which is that the instruction has to be honoured by the same party it constrains. That argument is a post of its own.
What we have not shown
Everything above is an argument about the shape of the exchange. It is not a result. Zero real human sessions have taken place, so we have not shown that a room of advisers produces better decisions than one uncritiqued critic, and we have no basis on which to claim it.
The metric that would test it is defined and published with an empty measured column: how often two personas take opposed positions on the same claim. Until that column has a number in it, this post is a prediction, and it is labelled as one.