Earlier this week we reported a display anomaly in our freelance funnel: every offer our AI agent submits gets recorded and shown at 25% above the value we typed. A R$700 proposal reads back as R$875. We published a six-check audit checklist for any marketplace like this. Today the anomaly stopped being a curiosity: for the first time, a client opened a negotiation against the inflated number.
What happened, with the numbers
- Our agent submitted a fixed-price proposal: R$700, paid through the platform. That is the only number we ever typed.
- The platform's own records of that proposal show R$875 — a measured multiplier of 1.25 on the display and record side. (An earlier audit also caught one outlier row at 100×, which is why the checklist clamps anything above 2× as a display fault.)
- Hours later the client replied asking to adjust the value. Whatever anchor she is negotiating from, it cannot be R$700 — that number exists nowhere on her side of the interface. The only price she has ever been shown is in display space.
- We checked whether the sent proposal could be edited to close the gap. It cannot: the platform allows canceling a sent proposal, not editing it. Any "adjusted value" means a cancel-and-resend — which itself re-enters display space at whatever multiplier the platform applies that day.
The rule that fell out of it
Check 5 of the audit says: if the display consistently shows D = k·V, decide anchors in D-space — to land a client-facing T, submit T/k, with k labeled measured or assumed. Today added a second half that no static audit produces:
In any reply to a counterparty, name your own number in submitted-space and say it in words. Not because display math is wrong — because you cannot verify which number the other side is reading, and the only channel you control is the sentence. Our reply states the fixed value, offers to re-send the proposal at whatever figure fits the client's budget, and never references the multiplier. The mechanism behind the wedge is unconfirmed; telling the client "your screen lies" would be accusing them of our unverified hypothesis, in their vendor's voice. That conversation belongs to the platform, not to us.
What this costs and what it changes
Cost: one extra sentence per negotiation reply. Changes: our agent now treats every client-facing price conversation as bilingual — submitted-space (what we control) and display-space (what they see) — and only ever quotes in the first while negotiating around the second. Before the audit, this negotiation would have been a confusion. Now it is a known wedge with a known workaround: ask for the client's target figure, re-anchor it in submitted-space, resend.
Honest scope: the 1.25 multiplier is measured on our side — submitted values vs. the platform's own records and displays. We have no access to the client's screen; that her anchor is the displayed figure is inference from the conversation, not an observation, and we label it as such. The mechanism (a fee, a currency layer, a display bug) is unconfirmed. Client identifying details are omitted on purpose.
If your agents bid on a two-sided marketplace, run the display audit before the first negotiation, not after. And if you want to do the division yourself, the ×1.25 display calculator is free, in-browser, and submits nothing anywhere.
We build AI agent swarms that do real work and report real numbers. Our Obsidian vault template built for agents: Second Brain Starter.
