How scoring works
CheckDragonfly reads what an X account shows in public, weighs it, and gives a score from 0 to 100 with an evidence-based conclusion and a confidence level. This page explains, in plain language, what we look at, how much each part counts, and what a result can and cannot tell you.
In short. We answer two questions separately: is this account what it presents itself as? (the score) and does it show known risk patterns? (the risk check). A serious concern decides the conclusion on its own. When there is too little history to judge, we say so instead of guessing.
What we look at
Every check looks at the same four areas, for every account, in the same order. We only use what X returns for that account at the time of the check. Anything X does not return is treated as unknown, never as a failure: it lowers our confidence, not the score.
Identity
- Does the account exist, and is it the handle you asked for?
- How long has it been on X?
- Does X show it as verified?
- Does it copy the handle, name or profile picture of a better-known account or brand?
Completeness
- Display name, bio, profile picture and location.
- A pinned post.
- Follower, following and post counts, read together rather than as one number alone.
Activity
- Its public posts from the last 30 days: how often it posts, its mix of original posts, replies and reposts.
- Whether posts are original or repeated and templated.
- Whether what it posts fits what its bio and pinned post say it is.
- Whether its activity is consistent with its history.
Risk
- Asking for money or crypto while steering people to private chats.
- Asking for a seed phrase, login or “verification fee”.
- Links to known phishing or malware sites, or to very new websites.
- Pictures that match known risk patterns, copy-paste posting, and accounts that were quiet for years and suddenly active.
How much each part counts
The score combines four independent readings of the account. Each produces its own 0–100 score.
| Reading | What it covers |
|---|---|
| Account profile check | Identity and completeness: whether the account exists and matches, its age, verification, profile fields, public counts and pinned post. Fixed rules, no AI. |
| 30-day evidence review | Activity: posting rhythm, originality, interaction, engagement, how the bio fits the posts, and history. Fixed rules, with AI-made text comparisons for the bio-versus-posts fit. Pictures read as AI-made or AI-altered, and a profile picture read as AI-made, count against the score; AI pictures used to present a persona that isn’t real cost identity and content points, while disclosed AI art, memes and parody do not. |
| AI reading of the posts | Language and vision models read the same 30 days of posts and pictures and score consistency of voice and profile, claims versus posted content, behavior, alignment and conversation. Every point must cite a real post. |
| Grok review | An overall assessment of the account by xAI’s Grok model, also required to cite posts. |
How much each reading counts is part of our proprietary method. In general, a reading that saw little counts for less, and the others make up the difference. A reading that could not run is left out; it is never counted as zero. If the AI reading and the Grok review run on the same model, the AI reading counts for less, so one model’s opinion is not counted twice.
Inside the account profile check, identity (the account exists and matches, its age, verification) counts the most, followed by profile completeness and public counts, then lifetime post count. Risk observations such as a protected account are recorded but earn no points.
The 30-day evidence window
We look at the account’s public posts from the 30 days before the check, up to a budgeted number of posts. Older history is used only in summary form (account age and lifetime counts from X). This keeps the check about how the account behaves now, and it means a result can change: re-checking later is a new observation of new evidence, not a correction of the old one.
Scores are calculated by a fixed method. The AI readings are set to be repeatable (temperature 0, fixed seeds), but language models do not promise identical output, so a reading can vary slightly between runs. Larger differences between checks come from the account’s posts, counts and age changing.
The six conclusions
Higher scores mean fewer concerning patterns were found in the public evidence. The score is one input; the conclusion also weighs confidence, risk signals and how much history there is.
The conclusion is decided in this order. The first rule that applies wins, and later rules can only make it more cautious.
| Conclusion | What it means for you | When it is given |
|---|---|---|
| A pattern worth a closer look | A specific measured pattern is listed with its evidence and ordinary explanations. It is an observation, not a conclusion about the person. | Any serious concern: resembling another account or brand, an active known risk pattern, money requests with off-platform contact, payment details in pictures, phishing links, or requests for credentials. |
| Many measured patterns observed | Many of the patterns this report measures were observed. Each is listed with its evidence and ordinary explanations; none is a conclusion about the person. | X could not find the account or returned a different one; or the combined risk is high; or the score is in the lowest band and a reading cites a specific concern, a risk signal was found, or a fact-based reading is weak. |
| Limited public history | There is not enough public history in this window to describe the measured patterns. This is not a negative finding. | The account is under 30 days old, wrote fewer than 5 posts of its own in the last 30 days, or its posts were not available. If risk is also elevated, the conclusion is Many measured patterns observed instead. |
| Several measured patterns observed | Several of the patterns this report measures were observed. Read them together and in context; many have ordinary explanations. | A score in the middle range; or a higher score where some risk is present, or where Grok’s review scored the account low (unless the verifiable facts are strong, Grok cites no concrete problem and names no farm, bait or spam pattern, and nothing else points to a risk). |
| Some measured patterns observed | Some of the patterns this report measures were observed. Each one is listed below with its evidence and ordinary explanations. | A fairly high score; or a high score where confidence is not high. |
| Few measured patterns observed | Few of the patterns this report measures were observed in the public activity that could be read. This describes the measurements, not the person. | A high score, with high confidence and low risk. |
Verifiable facts count for more
Some evidence is a matter of record: how long the account has existed, the profile and verification X reports, the measured posting rhythm, the follower sample and the account’s history. When these facts are strong and nothing undercuts them (no serious concern, no elevated risk, a substantial history, and no machine-like posting schedule), the AI reading of the posts counts for less of the score. An AI opinion that scores the account well below what the facts show, without citing a concrete problem in the posts, cannot pull the score far below what the measured facts show. A cited concern always counts in full. A schedule that is too regular to be a person (most posts on the same minute, or at one fixed interval) lowers the posting-rhythm reading.
The account profile check also acts as a ceiling: if it alone would be more cautious, its conclusion is used. Each conclusion comes with the main reasons behind it.
A harsher conclusion needs a reason you can see. A score in the lowest band is called Many measured patterns observed only when a reading cites a specific concern, a risk signal was found, or a fact-based reading is itself weak; otherwise the conclusion is Several measured patterns observed. A pattern worth a closer look is given only for a specific, cited warning sign.
Confidence levels
Confidence tells you how much the result had to go on. It is separate from the score: an account can score well with low confidence.
- High: nearly all four readings ran, they were confident in what they saw, the Grok review was available, and the account has substantial history (at least 180 days old and at least 15 posts of its own in the last 30 days).
- Medium: most readings ran, or the account has a moderate history. A result without the Grok review is never above medium.
- Low: much of the evidence was missing, or the account’s history is thin.
Only a high-confidence result can be Few measured patterns observed.
The role of AI
AI is part of how the score is made, and we want to be clear about where. The AI reading of the posts (which counts for less when the verifiable facts are strong) and the Grok review are automated AI opinions, and an AI comparison also feeds part of the 30-day evidence review. Some serious concerns, such as a resemblance to a brand or a known risk pattern, can come from Grok’s cited cross-check. AI findings are accepted only when they cite posts or profile text that actually exist; anything else is discarded.
The AI Analysis write-up shown after a check is different: it is written from the finished result and never changes the score, conclusion or confidence.
What an evaluation is not
- Not identity verification. We do not confirm who runs an account, and we never use facial recognition or identify anyone from their face. Impersonation is checked by comparing the account’s claimed name and handle with verified accounts, and whether its picture is an exact or near-exact copy of the verified account’s image. The banner and bio are compared the same way (a copied banner image or copied bio wording). These are whole-image and text comparisons, not face comparisons: a different photo of the same person is not matched.
- Not a background check, a credit or KYC check, or a statement about any person’s character or intent.
- Not a guarantee. The pattern index is not a probability. A high index does not mean an account is safe, and a low index does not mean anyone did anything wrong; it only summarizes how many of the measured patterns were observed.
- A point in time. It reflects what was public on X during the check. Accounts change hands, get hacked, or change behavior.
- Limited to public data. Protected posts, deleted posts and anything outside the 30-day window are not seen.
- Not yet calibrated. The method and the thresholds on this page are set by hand. They have not yet been validated against a large set of accounts whose real status is independently known, and every result says so. When that validation is done, the method may change and this page will be updated.
- AI can be wrong. The AI readings are automated opinions about public content and can misread sarcasm, other languages, or unusual but genuine accounts.
Think a result is wrong?
If you run an account that was evaluated, or you ran a check you believe is mistaken, email kellyhorak@authentix-account.com with the account and what you think is wrong, and it will be reviewed. You can also ask for an evaluation of your account to be removed. See also the Privacy Policy and Disclaimer.
Last updated: October 15, 2026 · scoring version AAE-FINAL-5.1 (scores are calculated by a fixed, published method; the AI readings are set to be repeatable but can vary slightly between runs)