How Truth Radar works

Truth Radar polls the public timelines of the accounts it monitors. The interval is adaptive rather than fixed: while an account is actively posting it checks every two seconds, and it stretches to twelve seconds, a minute, and eventually two minutes as an account goes quiet. The schedule is driven by the account's own historical posting pattern by hour of day.

The trade-off is cost. A permanent two-second poll would be marginally faster and roughly thirty times more expensive, and the extra speed would be spent almost entirely on hours when the account reliably does not post. Backing off is what makes a seconds-latency product affordable at this price.

There is also an optional institutional push feed. Where polling measures its detection in seconds, a push feed delivers in roughly 150–250 milliseconds because the platform notifies us rather than us asking. It is not enabled by default.

Every new post is classified by Claude into four things: whether it is market-relevant at all, which tickers or assets it implicates, the direction a desk would expect, and a confidence score.

Images are read, not skipped. A large share of posts on the monitored accounts carry no prose — they are a screenshot of a headline, an article, or a chart. Attached images are downloaded and passed to the classifier, so a screenshotted headline is classified on what it says. A monitor that treats those as empty posts misses a substantial part of the signal.

Machine classification is fallible, and the confidence score exists so you can weight it. A low-confidence bearish tag on an ambiguous post is a prompt to look, not a conclusion to trade.

Your watchlists filter what actually reaches you. A watchlist is a set of tickers, keywords or phrases, optionally restricted to market-relevant posts only. Tickers match exactly; keywords match on whole words and phrases, so “chip” does not fire on “chipping.”

The same matcher runs for live alerts, for the historical search in the dashboard, and for the CSV export. That is deliberate: if back-testing a watchlist used different logic from the live feed, what you tested would not be what you receive.

Matches go out over four channels, in any combination: Telegram, an outbound webhook into your own systems, email, and the in-app notification feed. The webhook payload carries the matched terms, the full classification and the measured detection latency, so a downstream system can act on it without re-parsing the post.

Delivery is recorded per channel, including failures. If a channel fails, the alert is still in your dashboard and you are told the delivery failed — the in-app feed is the one channel that cannot be taken down by the outage it is reporting.

It reads only public posts. It does not access private accounts, direct messages, or anything behind a login. It is not affiliated with Truth Social or Trump Media & Technology Group.

It is not investment advice, and it does not place trades. It tells you what was said, what it appears to touch, and how fast we saw it. Every decision after that is yours. See the latency methodology for exactly what the speed figure does and does not claim.