Claims are cheap. Recomputation is the check. We took five claims circulating this week and ran each against the same three checks: recompute it from the raw numbers, compare it against a dumb baseline, and walk it forward. Four fail. One holds.
A grid bot made 300% APR. The bot made about 3% in two weeks. The dashboard annualised a short lucky window into triple-digit APR and omitted the conditions, the costs and the benchmark. The 3% is real. The APR is arithmetic, not a yield. Verdict: fail.
A 70% win rate. Over 20 trades, a 70% win rate has a Wilson 95% interval of 48-85%. Over 100 trades it is 60-78%. Same win rate, four different truths. A win rate without a trade count is not a claim, it is a costume. Verdict: fail without N.
32 of 56 strategies were profitable. Before costs. After a realistic 0.8-pip spread, 0 of 56 made money. The average winning strategy made about $2.48 per trade and paid about $7.13 in spread per trade. The backtest was not wrong about direction. It was wrong about what a trade costs. Verdict: fail.
AI trading bots with 10-year backtests. We checked 18 advertised AI trading claims. 17 were unverifiable — no data, no method, no record. One had numbers. A claim you cannot check is not a claim, it is an ad. Verdict: fail.
92% fail FTMO. This one holds. The maths is the daily drawdown rule: FTMO measures the cap from the day's highest equity point, not the close, so one bad morning ends a challenge even on a net-profitable fortnight. Risk 2% and three losers in a row ends it. The lesson is the rule, not the trader. Verdict: holds.
The 10-second check, for any claim: ask for the prediction file, not the screenshot. Recompute the accuracy from the file. Compare it against always-bullish and no-change baselines. Split it into chronological folds and watch the last fold. If it fails any of the three, the equity curve was the product.
You can run that check free, in your browser, nothing uploaded: https://startbuildit.online/lie-detector/
A claim you can't check isn't a claim. It's an ad.
