How matching works

Example: Sam and three new coins

Of Sam's 100 posts, 30% are jokes; of his 40 coin posts, 35% are about Animal mascot coins, 25% give “Lore” as the reason, and only 3% bring up “Tech depth”. Each new coin is compared with these numbers question by question; the four cards are weighted into one overall with the default weights (5/15/30/50).

B is unlike on two cards, so its overall only reaches 60%: Standard, not Strict. C has no launcher account, so the personality and posting cards can't be told and their weight is re-spread: only the base and coin-picking cards count (coverage 50%), just enough. Any less coverage and the overall reads can't tell, passing nothing.

The rules: the same questions, compared one by one

The same questions: on the creator's side each of his posts is asked once and counted into the four cards' shares, distributions and score bands; on the coin's side the coin's name, ticker, pitch and optional fields, plus the launcher's X bio and posts, go to Jev as one package — same questions, same options, word for word, with the subject changed to “this launch”.

No post is cut short: only when the package is too long are posts dropped, fewest likes + reposts + replies + quotes first, and the coin page says how many. The coin's own fields are never dropped. An answer Jev is unsure of (no option clearly ahead, or a score pulled both ways) counts as can't tell.

A yes/no question looks at his share: he often has it (≥ 10%) and so does the coin → alike; he often has it and the coin doesn't → unlike (lacks what he usually wants); he almost never has it (< 5%) and the coin does → unlike (he stays away from it); neither has it → listed, not scored; in between → can't tell.

A one-choice question looks at how much of his distribution the coin's answer takes: ≥ 20% alike, < 5% unlike, in between can't tell; his “none” is left out of the count, and a coin answering “none” is can't tell. A 0–3 score question is cut into low / mid / high and compared as a three-way choice. The account question (bot-like) uses 50% as its yes line.

Each card: similarity = alike ÷ (alike + unlike). Can't tell and neither-has-it stay out of the count but are always listed; a card with nothing comparable can't be told.

The overall = the four cards weighted (default 5/15/30/50; each subscriber group sets its own with sliders, always adding up to 100). Cards that can't be told, or weigh 0, drop out and the rest are re-spread by their weights. The comparable cards' weights add up to the coverage; under 50% the overall reads can't tell, so one lightly weighted card never stands for the whole.

Not compared: subject over time, following up on calls, posting rhythm, titles, token-talk share, coins called, repeat calls, chain, launchpad (they need a long record, or are the same for every coin); the five personality factors are sums of the markers and are shown, not counted twice. Market-cap range, followers, likes / reposts / replies / quotes are shown only. Five questions — main field, topic field, sentiment, conviction, falsifiable — are asked and listed with both sides and the verdict, but not scored: creators in the same field answer them almost alike, and the self-identification backtest showed they only raise everyone's similarity together.

⚠️ Every threshold is provisional, waiting for calibration (yes/no 10% / 5%, one-choice 20% / 5%, coverage 50%, tiers 75% / 60% / 50%): the numbers will change, and this page follows the same spec file.

Three thresholds

When you set up a group you pick one, which decides how high a creator's overall similarity must be to pass; then how many of the group's creators must pass before you are notified.

Strict

Overall similarity ≥ 75%
E.g. Sam at 80% overall → passes; 70% → does not

Standard

Overall similarity ≥ 60%
E.g. Sam at 65% overall → passes; 55% → does not

Loose

Overall similarity ≥ 50%
E.g. Sam at 55% overall → passes; 45% → does not

Whichever tier: a creator whose comparable cards weigh less than 50% in all (coverage too short) does not pass.

Notify only when at least this many of the group's creators each pass the threshold above. A creator with no comparison profile yet (module waiting to be recomputed) does not pass.

Coins that miss the threshold are still charged (the judging was done), but they are not pushed and no webhook is sent; they sit under “Below threshold” on the notifications page, with what fell short.

Every question links back to the source

Every question carries evidence from both sides: his posts read as “yes” (each opens the post on x.com), and which parts of this coin Jev read (the coin, the launcher's bio, the launcher's posts). Only what was there at launch is read — never later prices or chatter — so the outcome cannot leak into the reading.

What the four cards measure

The comparison profile comes from the analysis cards — counted from the same readings, with no new Jev call; a trait has the same name on the card, the creator page and the coin page.

• Base profile: what he talks about, his sentiment and depth, and how human the account reads; the card's title comes from here.
• Personality: behaviours visible in single posts, grouped into Honesty, Emotionality, Social, Agreeable, Rigor, plus Risk style (risk appetite, holding period).
• Posting: Tone (Humor, Emotional intensity, Irony, Checkable) and Rhythm (posts a day, hours, threads, emoji — computed, not asked of Jev).
• Coin picking: coin posts only: tiers, sectors, pick reasons, call types, the 11 themes, the 10 reasons and rejecting tokens, plus Habits (chains, launchpads, market caps, following up — shown only).

The four cards always run together and are not sold separately. The three ways to run them differ only in how many posts are read and the price:

• Free run (once per person): last 10 days, up to 10 posts, free; cannot be listed as a creator module — 10 posts hold too few coin posts, so coin picking would almost always be can't tell.
• Free seat (verified X accounts): last 365 days, up to 1000 posts, free; can be listed as a creator module.
• Paid: last 365 days, up to 1000 posts, 1 USDG; can be listed as a creator module.

For programs

The API and webhooks carry the same numbers:

• GET /api/v1/creators → compare — whether each creator can be compared, and per card how many questions he has a number for (null = too few posts for that card)
• GET /api/v1/launches/{id} → result.match — the group's threshold result (with the group's weights), and each creator's four card counts (scored questions only; each question carries scored, false for the five shown-only ones), overall similarity and coverage, and whether they pass each tier
• webhook → data.match — the group's threshold result when pushed, with each member's four card counts, overall (group weights) and pass; coins below the threshold are not sent

Full reference: API docs →

Limits

• This compares style. It does not predict price, or whether he will buy: in our research a profile built from posts could not tell what a person actually buys — creators bought only about 14% of the coins they called. Alike means close to how he usually is, nothing more.
• Every question was checked against blind human labels: Jev agrees with the labellers at κ ≈ 0.55–0.85; theme classification is weaker (≈ 0.45). A single item can be read wrong — the evidence is there so you can check it.
• With too few posts a whole card can't be told (under 10 posts, every card; under 5 coin posts, coin picking); with little launch material, many questions are can't tell.
• Until an older card or module is recomputed into a comparison profile, that creator passes no threshold, and no old number is used in its place.