Methodology and error rates

Every rule in this system has a design document with the measurement behind it. This page lists each measured quantity with its number and date, and names what is not measured yet. A number without a date on this site is a live count.

Who is in the system

Accounts with at least 5,000 followers at the time a post is captured; nobody else is recorded (ADR 0007, second amendment, 2026-10-17). An account that crosses the threshold later starts its record then.

Catching posts

Three paths: a search for every newly discovered token's address, the watchlist polled every five minutes, and sweeps of launchpad links (designs 0001, 0003, 0005). Recall, measured 2026-10-17 against X itself over the 50 most-followed accounts with calls and their posts of the previous 24 hours: 811 of 853 original address posts recorded (95.1%), 46 of 87 address replies (52.9%; replies are not polled by design). Five accounts hit the measurement's 200-post cap, so their counts are lower bounds. The 42 misses were traced on 2026-10-17: 30 were posted during the 16 October outage and fell outside the one-hour look-back after the restart, 2 were cut by the poll page cap, 5 did not match the wide tier's address filter, 4 were not yet due for their daily poll, 1 arrived four minutes late (design 0005, section 3.7). Polls cut at the page cap in the last seven days: 0; the posts in those holes were not fetched. Known gaps: addresses only in images, ticker-only posts, chains other than Solana, Ethereum, BNB Chain, Base, and Robinhood Chain (design 0008), posts after a token's last scheduled search by accounts not yet watched, bare addresses from accounts polled once a day (13% of such posts on 2026-10-16).

Labeling posts

A post is scored only when its author takes a side on the token at the time of the post. Every post gets one of five labels: call (the author pushes the token now: says to buy or hold it, gives a target, says they bought, or presents news as good for it), warning (to avoid or sell it, a short, or that it is a scam), mention (the post brings the token up without taking a side: an informational list, a poll, a genuine question, news relayed plainly, a digest of updates, a short price note, or a brag about gains already made, also when it posts the address), ad (content written by software, spam, giveaways, or the token's own account), or unclear (the token is only the chain, venue, or yardstick of another trade). A person promoting a coin they launched makes a call like anyone else. Two models label every post independently (ADR 0019); when they agree, that is the label, and when they disagree about a call or a warning a third model decides. A label a person decides after reading the post replaces the models' label. These rules apply from 2026-10-19; measured that day on 200 live posts labeled by hand before any verdict was seen, 87.5% were labeled as by hand (the rules before them, 59.5% of the same posts). Labels given earlier stay as they were until they are labeled again. Only calls and warnings enter a score.

Accounts that are not influencers

Reviewed daily over the last 30 days of each account's posts (design 0006). Currently 3,131 accounts reviewed, 52 marked automated feeds, 111 marked advertising accounts. Since 2026-10-18 a mark takes effect as soon as the review gives it, except on the owner's own list, where it waits for a person; a person can reject any mark, and can mark an account the review missed. Hand check of 2026-10-17 over 110 reviewed accounts: every one of the 45 marked accounts was a feed or an advertiser (precision 1.00); 14 feeds or shills were not marked (recall about 0.72). A repeat-poster rule was tried the same day and withdrawn within the hour after marking 11 people among 21; it now runs in shadow, applied nowhere, until paired with engagement and measured again. Engagement is being collected (design 0007); its rule has no threshold yet. A marked account is no longer read: its watch row is closed, its new posts are not recorded, and its calls held from before stay unprocessed; a rejection releases them. Of the last 24 hours' captured posts, confirmed feeds and advertisers account for 0%, marked accounts nobody has read yet 0%, reviewed and unflagged accounts 95%, accounts not yet reviewed 5%.

Prices and outcomes

One price source (ADR 0011): the close of the minute the post was made, and the price at 1 hour, 24 hours, 7 days, and 30 days after the call of record (6 hours and 3 days until 2026-10-17). Price-source deviation is not measured since the second source was removed on 2026-10-16.

Scores

One formula, v3 since 2026-10-18 (design 0004, docs/scoring-formula.md), built on one question: what did a follower earn. Every published call and warning counts, one per person and token, whatever confidence the labeling models reported, measured at four moments after posting: 1 hour, 24 hours, 7 days, and 30 days. At each moment a call is worth what a follower made on it, any gain above +900% counted as +900%, so a big win can pay for several losses and several losses outweigh a small win. Hits are counted as in an exam where wrong answers cost: five losing calls cancel one winning call. A warning whose token rose loses the rise in full; one whose token fell further than the day's typical call earns half what a right call earns, since most tokens fall anyway. An account's record at a moment is half its gain, three tenths its net hits, and one fifth the calls that did not collapse 90%, blended with the average caller's record by its measured calls against ten, so one lucky call cannot carry anyone. The leaderboard ranks one score that weighs the moments: 1 hour 5%, 24 hours 10%, 7 days 65%, 30 days 20%. Chosen on backtests (2026-10-17): following every call for a week lost a follower 38.0% per call, and 12% of calls were in profit after 7 days; the twenty leaders of each day's ranking made calls the next week that lost 3.0% per call, 29% of them in profit, better than the field on 14 of 15 days. With more weight on the first hours, accounts whose calls run up and collapse within the week rose, and the fifty leaders did worse than the field. The formula's own error is not estimated beyond this; every input is shown beside the score.

Logos

The 200 largest coins by market cap show logos this site serves itself, taken once on 2026-10-19 with the list of the largest coins (ADR 0023). A token shows such a logo only when its contract and its symbol are the coin's; a coin kept on a wrapper, such as Bitcoin on WBTC, shows the coin's own. Every other token shows the logo its creator published. Data provided by CoinGecko.