How CrowdTemp works

Every topic gets a score from 32° (total hostility) to 212° (total devotion). The Fahrenheit display follows a straight line through three anchors: 32° at total hostility, 80° at neutral, and 212° at total devotion. Neutral is room-temperature water: no opinion. The warm half of the scale is longer, as on a thermometer.

The levels are frozen from 32° to 41°, cold from 42° to 70°, lukewarm from 71° to 105°, warm from 106° to 132°, hot from 133° to 187°, and boiling from 188° to 212°. Frozen means about 80% negative. Cold leans negative. Lukewarm means the crowd is mixed or indifferent. Warm leans positive. Hot means about 60% positive. Boiling means about 80% positive. Internally, scores live on a 0 to 100 scale. The score maps the crowd's mean opinion through a compressive curve so the middle of the scale has room and unanimity reaches the ends. The Fahrenheit scale is the display layer.

A temperature measures expressed opinion in live public conversation. It does not poll anyone and it does not predict outcomes.

CrowdTemp currently reads five public sources, grouped into three classes. More sources coming soon.

Each topic is fetched on a cadence tied to how actively it is tracked. Only social posts vote. News and markets are read, shown on the topic page and counted as mentions, but they do not move the temperature: a headline is coverage, not an opinion. The source mix varies by topic and over time, and every topic page shows its own live source breakdown.

1. Ingestion. Posts matching a topic's search terms are fetched from each source. Before anything is stored, posts pass through deterministic filters: NSFW content labels, a promotional spam filter (launch announcements, referral links, hashtag walls, ticker stuffing), and word boundary matching against the topic's terms.
2. Opinion scoring. An LLM reads each post and scores the direction and strength of its opinion, plus a confidence and a relevance flag. Posts judged off topic or promotional are set aside, and so are posts that only pass information along: a headline, an announcement, a result, a stat, a quote with no comment. Good news is not an opinion; someone has to react to it. Each scoring run samples recent posts, so the temperature follows the conversation as it moves rather than only its most shared moments.
3. Aggregation. Every post that expresses a view counts, including mixed and undecided ones; each counts in proportion to how sure the reading is. Stronger opinions move the temperature more than mild ones. Widely shared posts carry more weight than quiet ones without being able to dominate, and no single account can either: each account gets a small fixed number of posts in the pool, and the pool follows where the conversation is. A reading built on only a few posts is held closer to the middle and marked as still settling until more come in. The temperature settles over time rather than jumping.

Four layers, in order:

1. Deterministic filter at ingestion. Known promotional patterns are rejected before storage. These are plain string checks. Posts and accounts that Bluesky itself has labeled as bots, spam or inauthentic are dropped at the same point.
2. Account filter from our own verdicts. An account whose posts have been marked off topic almost every time over the past week (at least ten scored, nine in ten set aside) is treated as automated and its posts are no longer stored. The list is rebuilt nightly from scoring results, not from handles or follower counts, and an account comes off it on its own once it starts posting like a person. Price tickers, deal feeds and headline aggregators are what this catches.
3. LLM relevance gate at scoring. Promotional posts and posts with no opinion that get past the pattern filter are marked irrelevant by the scoring model and excluded.
4. Aggregation caps. A per-account limit on the pool bounds the influence of anything that gets through both, and duplicate posts count once.

None of this is airtight. A coordinated campaign using many accounts and natural language would register here the same way it registers on the platforms themselves. We tighten these layers as we find new patterns.

Topics are scored live as new posts arrive. Every topic page shows when it was last scored.

• Source mix. Every source has its own demographics and mood. Today most social posts come from Bluesky, because Bluesky publishes an open feed and the other sources are rate limited. Bluesky's crowd is not the whole internet, and its lean shows most on political topics. Each topic page shows its exact mix so you can judge the blend yourself.
• Smoothing lag. Scores trail breaking events by hours. This is a design choice: the temperature tracks sustained opinion, which makes it resistant to single viral moments and manufactured spikes. Where it leads is the slower kind of move, a crowd warming or cooling on something over days before coverage catches up. Where it lags is the sudden kind.
• LLM judgment. Opinion scoring is done by a language model, and reading tone at scale is imperfect for machines and for people.
• Language coverage. Posts are scored in any language, though the current source set leans toward English language conversation. Coverage broadens as sources are added.

If a topic is scoring strangely, the scored posts on its page show what the number was built from.