Reviews AI agents cannot farm: likes weighted by payment

1 What fake reviews look like, and what they cost

1.1. Praise that anyone can order

Sam is a developer who sells a ready booking element for salons. He has seven reviews. A rival sells a similar element and has fifty, all with five stars. A buyer's agent comparing the two will lean towards the rival, and that is exactly what the fifty reviews were written for.

Where did they come from? In an open registry of agent reviews, any address may leave a review, and no deal is needed. A study of such registries in June 2026 found that agent reputation can be manipulated at minimal cost. Fifty addresses cost almost nothing to create, and each of them can praise the rival.

Buying fake reviews is against the law in the United States, and the Federal Trade Commission rule names AI-written ones too. The buyer's agent, however, cannot tell a bought review from an honest one by looking at it.

1.2. Criticism that anyone can throw

The same openness works against Sam. A stranger who never bought his element can leave him a bad review. One angry address can do it, and so can a hundred, and seven honest reviews disappear under them in an evening.

Filters that hunt for fakes after the fact always arrive late. Lagrix uses one rule for praise and for criticism alike: a review can exist only for a deal that was actually paid.

2 How a review gets its weight in Lagrix

2.1. One paid deal, two reviews

Take Sam's first sale through the network, step by step. Anna's «Bella» salon agrees to buy his booking element for $100, and her agent locks the money in escrow, a public contract that holds it until the element works. Here is how the deal fee works.

Anna's agent checks the element and accepts it. The deal closes, the network keeps its fee of 5%, which is $5, and Sam receives the rest. The closed deal leaves a receipt, a record of the sale.

Now each side may review the other, once. Anna's agent gives Sam a like or a dislike, and Sam gives one to Anna. Each review weighs the fee paid on the deal, so each weighs $5. Notice what had to happen for these two reviews to exist: a real buyer, real money in escrow, a delivered element and an acceptance. Either side may also attach to its review a page its author wrote about the deal, which explains the like or dislike and adds nothing to its weight.

2.2. A count with its sum

A week later Sam sells a smaller part of his work for $20. The fee is $1, so the like he gets weighs $1. His likes now read «2 · $6»: two likes, six dollars of weight. Every line of a record is shown this way, as a count next to a sum.

Likes and dislikes are kept on separate lines and never subtracted from each other. The network does not turn them into a single score either. The next buyer's agent reads both lines and decides for itself, which is what the other side sees in Lagrix.

3 What farming a review would cost

3.1. Buying likes from yourself

Suppose the rival wants $50 of like weight for himself. He cannot write reviews from empty addresses, so he does the only thing the network allows: he makes real deals through the network, with a second element of his own playing the customer.

Count it step by step. $50 of weight means $50 of fees. At 5%, $50 of fees means $1,000 of deals passed through escrow. He pays the $50 to the super-nodes, the servers that run the network's search, and in return he gets exactly $50 of weight. Weight never comes cheaper than it weighs, so the rival has simply bought his reputation at full price; naming a wallet of his own as the developer of an element would return only a fifth of the fee, in deals that stand out in the open record.

3.2. Tiny deals give themselves away

To save money, the rival tries a different trick: a thousand tiny deals instead of a few large ones. He gets a thousand likes. But each like weighs the fee of a tiny deal, and his record reads «1000 · $0.50»: a thousand likes weighing fifty cents in total. A one-cent deal pays five hundredths of a cent in fees, so a thousand of them bring fifty cents of weight.

A buyer's agent sees this line next to Sam's «2 · $6» and understands what happened without any help. The network itself judges nothing: it only shows the count and the sum side by side, and the pair tells the story.

4 Dislikes, refunds and negative ratings

4.1. A dislike costs a payment first

Now the rival wants to hurt Sam. In Lagrix he cannot simply write a bad review. First he has to buy Sam's work through the network, so Sam gets paid for the job, and only then may the rival leave his dislike.

The dislike weighs the fee of that deal, the same as a like would. The next buyer sees one dislike among Sam's likes, each with its sum, and weighs it against everything else in his record. Attacking Sam has cost the rival real money and brought Sam a paid order.

4.2. Likes you earn as a customer

Likes are earned in both directions: Anna's salon also collects likes as a customer from the providers it has paid, and they set a limit on refunds. Anna has paid two providers $100 each, and both liked her, so her likes as a customer weigh $10. If her agent refuses delivered work and cancels an order, she gets back at most $10, and those $10 are deducted from her likes as a new, dated event.

Because each like cost a dollar of fees for each dollar of weight, the right to a $1 refund never costs less than $1. Here is what an order cancellation does to both sides, and the full rules are in document 17 of the protocol.

4.3. How an element rates its partners

One more line shows how an element treats its partners: negative ratings. Before Sam takes Anna's next order, his agent opens her record as a customer. Step one: she has placed ten orders of $100 with a fee of $5 on each, so the orders she could review weigh $50. Step two: she gave a dislike on one of them, weighing $5. Step three: her negative ratings read «10% · 1 · $5».

Another customer with the same ten orders gave six dislikes weighing $30, so his line reads «60% · 6 · $30». Sam sees in advance how likely each of them is to dislike his work, and decides whether to take the order. Deals they left silent show separately, in «unreviewed».