A string does not reveal its selection process
Passphrase Strength: Random Selection and Checker Estimates
A word-based candidate can produce dictionary findings even when its words were chosen by a sound random process. That is not a contradiction. The process and the checker answer different questions: one describes how choices were made, while the other estimates guessability from the text it receives. Reading them separately avoids both dismissing random words because they are readable and trusting a familiar sentence because it is long.
The Password Strength Checker evaluates one single-line candidate using an offline common-and-English zxcvbn-ts model. Begin with its public example; anything printed in a guide or demonstration is not for real use. You do not need to provide an actual account password to understand the result, and you should never ask another person to share a secret for this exercise.
Ask what chose the words before interpreting their appearance
The word passphrase describes a format, not a selection method. A line from a popular story, an encouraging sentence, a private joke, and independently sampled words can all look like several ordinary words separated by spaces. Their appearance does not make their selection processes equivalent. A person composing a sentence usually follows meaning, grammar, memory, or preference. A uniform random procedure follows specified choices and probabilities instead. The checker receives the resulting characters, not a recording of those decisions. It cannot determine which procedure occurred simply by recognizing words or counting the spaces between them. Describing words as unrelated is also a judgment, not a probability rule. Two people can disagree about whether words connect while looking at exactly the same text; a documented selection procedure avoids depending on that impression.
Separate knowledge about origin from assumptions about appearance. If you watched an appropriate random generator select the words, that is information about the process outside this checker. If someone merely says the phrase looks random, the claim has not established independent uniform selection. Equally, a phrase that happens to suggest a small story is not proof that a person composed it. Chance can produce memorable combinations. The relevant evidence is what performed the selection and whether the output was kept intact, not how amusing, elegant, unusual, or easy to visualize the final sequence seems to a reader.
This guide is about reading the result, not teaching a memory routine or walking through generator settings. Those topics belong in the separate Password Generator hub. Keep a simple distinction here: known generation conditions can support a theoretical calculation, while unknown conditions remain unknown after checking. There is no origin questionnaire, personal context field, or account connection that fills that gap. Do not put a note such as randomly generated into the password field; it would become part of a different candidate rather than evidence about the one you meant to evaluate.
Understand the conditional six-word EFF calculation
Suppose six words are selected independently and uniformly from the 7,776 entries in the EFF long wordlist, with replacement. The number of possible ordered sequences is 7776 raised to the sixth power. Taking its base-two logarithm gives roughly 77.5 theoretical bits for that distribution. The conditions are essential: each entry must have the stated chance at each draw, and earlier selections remain eligible. This is arithmetic about the selection process, not a true-entropy reading taken from a particular string. Merely counting six words in a sentence does not establish any of those conditions.
A guessing model works differently. It can recognize individual words, rank them by its dictionaries, consider common arrangements, and combine its estimates. Different familiar words can receive different ranks even though a uniform wordlist procedure gave every entry the same selection probability. The library does not know the generation process and cannot recover it from the spelling. A checker score is therefore not proof of the source process, and its estimated guesses are not a substitute for the conditional calculation. A result above or below your expectation does not change what procedure actually produced the words. Do not work backward from the score to invent a likely wordlist size or random word count. Different constructions can produce the same estimate, so that reverse calculation would not identify the method used.
Repeated words illustrate the distinction especially clearly. Independent sampling with replacement permits a repeat; the next draw still has the full list available. A model can also recognize repetition in the resulting text and account for it as a guessing pattern. Those observations refer to different models and can coexist. Do not edit an honestly drawn sequence solely to force the checker toward a preferred result, then continue attaching the original distribution claim to the edited phrase. Choosing only outputs that meet an extra score condition changes the selection procedure being discussed and needs its own analysis.
Read score, estimated guesses, and type counts separately
The displayed score runs from 0 to 4 and is explicitly an estimate. It is a small set of bands, not five levels of verified account protection. Model-estimated guesses appear as 10^estimated guess log10 when a finite exponent is available. An exponent of 9, for example, expresses about one billion model-estimated guesses, not nine bits or nine attempted logins. Do not compare the exponent directly with 77.5: one is a base-ten model output and the other is a base-two calculation under specified generation conditions. Converting notation alone would not make the underlying models equivalent. The top band also has no extra room to show every modeled difference between long candidates. Two phrases can both score 4 without having the same guess estimate, and neither receives a certificate that it exceeds every real guessing strategy.
Pattern findings are aggregated by type. A dictionary count does not expose which words matched, and it need not equal the number of words you see. The output contains no actual secret segments, start or end positions, or sequence trace. A repeat finding is not a report of which word repeated. These limits keep the explanation about categories instead of reproducing pieces of the candidate. If the display uses an unfamiliar type, read it as part of the engine’s selected explanation rather than a request to reveal your phrase so another person can decode it.
Feedback and suggestions are fixed English messages supplied locally, not AI advice tailored to a personal history. Common and English dictionary coverage can miss phrases from other languages, specialist vocabularies, and local culture. A long quotation missing from those data can look less guessable to the model than it deserves. Conversely, recognizable words need not make a genuinely uniform sequence human-chosen. An unavailable finite guess exponent is simply unavailable, not a claim of infinite strength. There is no numerical crack-time display, breach search, account check, API call, or actual attack behind these fields.
Read the model on its own terms
Inspect a public word-based example
Use Check password to see an estimate without treating dictionary findings as evidence about how the words were selected.
Evaluate the exact single-line phrase you intend
Spaces are characters in the candidate, not decoration around the input. Accepted leading and trailing spaces remain part of the exact string, as do repeated internal spaces and the chosen punctuation. Nothing is trimmed, normalized, or truncated to simplify evaluation. A phrase with one space between words differs from a phrase with two, even if both look similar at a glance. Keep autocorrection, inserted punctuation, and accidental spacing in mind when privately reviewing what you entered. This checker does not decide how a separate destination form will handle those characters or whether it will accept the phrase.
The field accepts single-line candidates only. Carriage returns and line feeds are rejected, including attempts to paste or drop line breaks. A paragraph, a wrapped quotation containing actual newline characters, or a multiline note is not supported input. This is distinct from text wrapping visually because the screen is narrow: a visual wrap does not itself add a newline to a string. Do not remove line breaks from an intended password and call the resulting estimate a check of the original. Joining the lines creates a different candidate, even when the words remain in their former order.
All accepted candidates must be nonempty and within 128 code points, 256 UTF-16 code units, and 2 KiB of UTF-8 data. Unpaired surrogates are rejected as malformed text. The result reports code-point and UTF-16 length because these counts can differ. One displayed symbol may comprise several code points, while a code point outside the basic multilingual plane occupies two UTF-16 units. Accented text can also have different underlying representations despite a similar appearance; the checker does not normalize them into one. These are exact-text and size issues, not evidence that a longer encoding contributes more random choices.
Use one short session to learn what the result says
Start with Load example, which fills the field without checking it. The candidate stays masked unless you select Reveal; use Hide to mask it again. Check password is the explicit action that starts evaluation. A public example is useful for understanding the scale and pattern summary without putting a private phrase on screen. If you change that example, request a new check for the changed string. There is no side-by-side comparison board, saved history, or collection of previous candidates to return to. Treat each result as an observation about the one submitted value.
The engine is bundled for offline checking in a terminable worker, with a six-second limit. A job that cannot finish within that bound does not provide a successful estimate. Waiting out the limit is not a strength test and does not demonstrate that guessing would fail. The checker runs inside an opaque-origin isolated document with no WordPress bootstrap, no-store responses, and a connect-none restriction. Candidate evaluation does not require a remote query. That description concerns the local checking document; the surrounding parent page and initial asset loading can still produce network traffic.
Use Clear when you finish. Hidden visibility, pagehide, pageshow, and freeze also remove the input and derived DOM values and terminate active work, so switching away is not a way to keep a private result waiting. There are no Copy buttons, exports, PDFs, or stored reports. A manual browser copy, an earlier clipboard value, or clipboard history remains outside that clearing behavior. A browser extension, compromised device, or hostile parent replacing the interface is outside the isolation protection too. Removing working interface values does not promise erasure of every memory copy or of material captured elsewhere.
Make a decision without treating either number as a certificate
For a phrase you wrote yourself, a low estimate can draw attention to familiar structure, but a high estimate cannot supply the missing random selection process. Do not repair a quotation by adding a few symbols and then describe it as six uniformly drawn words. The checker cannot validate that story. If the concern is obtaining a new independently selected candidate, use the generator separately and judge its documented process. This checker only evaluates supplied text. It does not create a replacement, register it with a service, confirm account ownership, or tell you whether the destination accepted a password change. Form acceptance would answer another separate question anyway: whether that service permits the submitted text. Acceptance is not evidence about its origin, and rejection is not a mathematical assessment of the random procedure.
For a phrase known to come from independent uniform draws, read dictionary and repeat findings in that context. They describe how the configured estimator interprets the text, not a discovery that the generator secretly used human preferences. The conditional process calculation can remain valid while the checker ranks individual outcomes differently. At the same time, knowing the process does not prove the phrase stayed private. An example published in a guide is public regardless of its original random origin. Reusing it, displaying it in a meeting, or sharing its words changes the practical situation without changing the arithmetic of the original draw.
End with a conclusion you can defend without quoting the candidate: the model recognized dictionary structure, the process is known, or the source of the words remains unverified. Avoid recording the complete phrase just to preserve that explanation. For common human recipes, continue with common password patterns. The checker hub covers the tool as a whole, and all free tools provides the wider directory. None of those next steps requires retaining a candidate in this page. Clear it and keep any real credential handling separate from an educational comparison.
Questions about Passphrase Strength
Do six words always mean roughly 77.5 bits?
No. That figure assumes six independent uniform draws from the 7,776-entry EFF long wordlist, with repeats permitted. It describes the specified distribution, not any six-word sentence. The checker cannot verify those conditions or measure the true entropy of the supplied text.
Can a random passphrase have dictionary or repeat findings?
Yes. Randomly selected words are still words, and independent draws can repeat. The engine ranks recognized structures under its own model without knowing their origin. Its score is not proof that a phrase was human-composed or that a claimed random process was actually followed.
Should I keep generating until the checker gives my preferred score?
Filtering outputs by their score adds a selection condition, so it is not the original unrestricted draw process. Understand the documented generation method and the checker model separately. Neither a top score nor a theoretical process figure proves that a particular account or device is protected.
Continue with another password-checking guide
Check a proposed password in the Password Strength Checker or choose a related task below. Explore more practical utilities in all free tools.
Separate process from observation
Review a candidate with the right question
Read the score, model-estimated guesses, and pattern types as local feedback, then clear the temporary session.