An AI briefing that doesn't overclaim

RiskFloor is a pre-trade briefing for forex, gold and crypto traders: an AI-written read of the bull case, the bear case, key levels and what would prove it wrong. At the top of each briefing, a strip of short status labels summarized the market. It was the first thing a reader saw.

Each briefing showed status labels for market state, volatility and structure. I suspected some never changed, so I checked 101 real briefings (June to September 2026). Volatility read "Low" on 101 of 101. Market state read "range-bound" 90 times and "trending" never. One live EUR/USD briefing said "Range-bound" right next to "Daily just made a higher high."

The market-state label wasn't reading the market. It was one ratio with one threshold that almost no forex or metals pair could clear. It looked like a reading, and it behaved like a constant.

I considered rescaling the labels, renaming them, or building a new indicator.

• Rescaling. This would have turned some briefings' structure label to "High". A check found that 0 of the 21 briefings that would have moved actually had aligned structure, so the new reading would have been wrong.
• Renaming. A more accurate word on a miscalibrated number is still a miscalibrated number.
• A new indicator. I rejected a new volatility label built on a different measure, and wrote down the conditions any future label has to meet.

I removed them instead (September 2026) and set a rule for the product: an empty slot is honest; a constant is not.

I pressure-tested the decision with an AI model as a critique partner.

I then applied the same reasoning to the risk label. It wasn't constant. It failed differently: nothing defined its scale, and it flipped between reruns of the same briefing. I retired it the next day.

The labels came off the product in two changes on September 3 and 4. Traders act on what a screen implies. A label that never moves still looks like a reading. The same thinking runs through the rest of the design: colour shouldn't signal a verdict, and the briefing weighs evidence but never tells anyone what to trade.

This made the output more honest. I'm not claiming a lift in trust or retention.

Count what a field has actually said before designing around it. This took a tally, not a redesign, and it changed the answer from "improve the label" to "remove the label".