# Classifier rules & scores

> **Sentinel v2 (beta):** This page documents Sentinel v2, a beta rewrite that may change before the stable release. It does not apply to Sentinel v1, the current stable version: see the [Sentinel v1 documentation](/docs/sentinel.md).

This page lists all 53 classifier rules and 22 email (EML) rules, with their scores and triggers, and the formulas that turn rule scores into a score and verdict. For the concepts behind them, see [Scoring](/docs/sentinel/v2/concepts/scoring/) and [Classifier](/docs/sentinel/v2/features/classifier/).

## Score computation

| Step | Formula |
| --- | --- |
| Rule score | Each rule reports `score`: `0` when it didn't trigger, `null` when it is disabled or doesn't apply (for example, no input for it). |
| Category score | Sum of the category's non-null rule scores. Text and HIS sums are rounded to 2 decimals. Similarity uses `max(SIMILARITY, NEAR_DUPLICATE) + CAMPAIGN` instead. |
| Counted categories | A category counts if it ran and has at least one enabled rule. **Injection**, **Similarity**, **Rate limit**, and **AI** count only when their score is `> 0`, so a clean check or a limit that wasn't exceeded doesn't lower the mean. |
| Overall score | Mean of the counted category scores, rounded down to 2 decimals (`0` when no category counts). Then raised by the [floors](#floors). |
| Verdict | `GOOD` for a score `< 1`, `NEUTRAL` for a score `< 2`, `BAD` for anything higher. |
| `triggeredRules` | Every rule that scored `> 0`, highest score first. Ties keep the order shown on this page. Each rule appears once. |

Categories and when they run:

| Category | Response field | Runs when |
| --- | --- | --- |
| Location | `location` | A country is known (IP geolocation, otherwise the request's `timeZone`) |
| Device | `device` | Device input (request headers) is present |
| Email | `email` | An email address is present |
| IP | `ip` | An IP address is present |
| Text | `text` | Non-blank text is present (the body's `text`, otherwise text from the form fields) |
| Injection | `injection` | Form fields, text, or an email address are present to scan |
| Similarity | `similarity` | Text is present and at least one similarity rule is enabled |
| Rate limit | `rateLimit` | The body has `rateLimit` and `RATE_LIMIT` is enabled |
| Human Interaction Signature (HIS) | `his` | `POST /v2/verify` only, with an HIS score in the challenge, outside HIS `monitor` mode, and at least one HIS rule enabled |
| AI | `ai` | The security group enables `classifier.ai`, `[classifier.ai]` has a provider, text is not empty, and at least one AI rule is requested |

## Floors

After the mean is computed, these conditions raise the overall score to at least a minimum value. Floors only raise the score; they never lower it.

| Condition | Minimum score | Resulting verdict |
| --- | --- | --- |
| Any injection rule triggered (injection score `> 0`) | `2` | `BAD` |
| A similarity match (`SIMILARITY` or `NEAR_DUPLICATE`) against a sample with weight ≥ `1` | `2` | `BAD` |
| `URL_PHISHING` triggered | `2` | `BAD` |
| `HIS_INTEGRITY` triggered | `2` | `BAD` |
| `HIS_REPLAY` triggered, or `HIS_AUTOMATION` ≥ `1.8` | `1` | at least `NEUTRAL` |
| Highest AI answer probability ≥ `0.9` | `2` | `BAD` |
| Highest AI answer probability ≥ `0.7` (and below `0.9`) | `1` | at least `NEUTRAL` |

## Enabling and disabling rules

Every rule in the classifier tables below can be disabled. To do so, use the security-group settings `classifier.disableRules` / `classifier.enableRules`, or the request body's `disableRules` / `enableRules` on `POST /v2/classifier` (each body list replaces the group's list of the same kind). If a disable list is set, every rule except those runs and the enable list is ignored. Otherwise, if an enable list is set, only those rules run. Otherwise all rules run. Because the lists are replaced separately, a body `enableRules` has no effect while the group sets `classifier.disableRules`; send `"disableRules": []` with it to clear the group's disable list. Rule names are case-insensitive and can be separated by `,`, `;`, or whitespace. See [Rules](/docs/sentinel/v2/reference/rules/).

## Email address

| Rule | Score | Trigger | Can be disabled |
| --- | --- | --- | --- |
| `DISPOSABLE` | `1` | The domain is on the disposable-email list and not on the trusted-provider list | Yes |
| `DMARC` | `0.5` | The domain has no DMARC record | Yes |
| `FREE_PROVIDER` | `1` / `0.5` | The domain is on the free-email list: `1`, or `0.5` if it is also on the trusted-provider list | Yes |
| `MX` | `5` | The domain has no MX record | Yes |
| `PRIVACY_RELAY` | `1` | The domain is on the email privacy-relay list | Yes |

## Device

| Rule | Score | Trigger | Can be disabled |
| --- | --- | --- | --- |
| `BOT` | `2` | The request is signed by a verified Web Bot Auth agent. Never triggers on `POST /v2/classifier`, which has no signed request to check. `BOT` is also an IP rule (see [IP](#ip)). | Yes |
| `ACCEPT_HEADER_MISSING` | `1` | No `Accept` header. `null` when the header is present but there's no `User-Agent`. | Yes |
| `ACCEPT_LANGUAGE_HEADER_MISSING` | `1` | No `Accept-Language` header. `null` when the header is present but there's no `User-Agent`. | Yes |
| `USER_AGENT_HEADER_MISSING` | `1` | No `User-Agent` header | Yes |
| `USER_AGENT_HEADER_UNCOMMON` | `1` | The User-Agent isn't on the common User-Agent list. Never triggers while the list is empty, or for a trusted app (`classifier.trustedApps`). `null` without a User-Agent. | Yes |
| `CLI` | `1` | The User-Agent is a command-line HTTP client, unless it's a trusted app. `null` without a User-Agent. | Yes |
| `LIBRARY` | `1` | The User-Agent is an HTTP client library, unless it's a trusted app. `null` without a User-Agent. | Yes |

## IP

| Rule | Score | Trigger | Can be disabled |
| --- | --- | --- | --- |
| `BOT` | `2` | A threat list classifies the IP as a bot, crawler, AI training/search agent, or user-fetch agent | Yes |
| `HOSTING` | `2` | The IP is a hosting/data-center IP. On `POST /v2/eml`, it's ignored when the sending server is authorized by the sender's domain. | Yes |
| `MALICIOUS` | `5` | A threat list classifies the IP as malicious | Yes |
| `PROXY` | `0.5` | The IP is a proxy | Yes |
| `TOR` | `1` | The IP is a Tor exit node | Yes |
| `THREAT_SCORE` | `0`–`10` | The IP's threat score (the highest of the matching threat-list entries); `0` without one | Yes |

## Location

| Rule | Score | Trigger | Can be disabled |
| --- | --- | --- | --- |
| `HIGH_RISK_COUNTRY` | `2` | The country is in the security group's `classifier.highRiskCountries` (default list `high_risk`) | Yes |
| `LOCATION_DISTANCE` | `0`–`5` | Distance between the IP's coordinates and `lastKnownLocation`: `0` up to 100 km, otherwise `log2(km / 100 + 1)`, capped at `5` and rounded to 1 decimal. `null` without `lastKnownLocation`. | Yes |
| `TIMEZONE_MISMATCH` | `1` | The request's `timeZone` belongs to a different country than the IP. `null` unless both are known. | Yes |

## Rate limit

The rate-limit category counts toward the mean only when the limit is exceeded (score `> 0`).

| Rule | Score | Trigger | Can be disabled |
| --- | --- | --- | --- |
| `RATE_LIMIT` | `10` | The request exceeds a `rateLimit.limit` from the request body | Yes |

## Text

The text engine scores each rule from `0` to `1`. Sentinel multiplies that by `5` and rounds to 2 decimals, so these rules range from `0` to `5`. `URL_PHISHING` is the exception. "Relative to text length" means the same number of hits scores lower in a longer text.

| Rule | Score | Trigger | Can be disabled |
| --- | --- | --- | --- |
| `SPAM_WORDS` | `0`–`5` | Spam dictionary terms and phrases, weighted, relative to text length | Yes |
| `PROFANITY` | `0`–`5` | Profanity dictionary terms, each distinct term counted once | Yes |
| `URL` | `0`–`5` | URLs (up to 5 counted), relative to text length | Yes |
| `URL_PHISHING` | `10` per URL | Each distinct URL on the phishing list. `0` while the list is disabled or not loaded. Triggers the `BAD` [floor](#floors). | Yes |
| `LINK_MARKUP` | `0`–`5` | Link markup (`href=`, `[url`, `[link`), relative to text length | Yes |
| `CURRENCY` | `0`–`5` | Currency symbols next to a digit (`$5`, `5 €`), relative to text length | Yes |
| `EXCLAMATION` | `0`–`5` | Runs of two or more `!`, relative to text length | Yes |
| `REPEATED_CHARS` | `0`–`5` | The same letter four or more times in a row, relative to text length | Yes |
| `INVISIBLE_CHARS` | `0`–`5` | Zero-width or other invisible characters between letters or digits | Yes |
| `MIXED_SCRIPT` | `0`–`5` | Words that mix Latin with Cyrillic or Greek letters | Yes |
| `RANDOM_CHARS` | `0`–`5` | Words with erratic letter case, or random letter-digit codes, relative to text length | Yes |
| `EMOJI` | `0`–`5` | Emojis, relative to text length | Yes |
| `OBFUSCATION` | `0`–`5` | Dictionary terms found only through obfuscation (leet, separators, repeated letters, `*` masking, look-alike characters) | Yes |
| `CAPITALIZATION` | `0`–`5` | Mostly capital letters: at least 12 cased letters, 60% of them uppercase | Yes |

## Injection

The injection category counts toward the mean only when it scores `> 0`, and any hit forces `BAD` (see [Floors](#floors)). `POST /v2/eml` doesn't run injection checks.

| Rule | Score | Trigger | Can be disabled |
| --- | --- | --- | --- |
| `SQL_INJECTION` | `5` | An SQL lexer reads a value as SQL, either as is or after a quote the value breaks out of: comment truncation, tautology, `UNION SELECT`, stacked query, or SQL function | Yes |
| `HTML_INJECTION` | `5` | A dangerous tag, an `on…` event-handler attribute, or a `javascript:`, `vbscript:`, or `data:text/html` URL | Yes |
| `EMAIL_HEADER_INJECTION` | `5` | A line break in an email value, or a line break followed by a mail header in a single-line field | Yes |
| `TEMPLATE_INJECTION` | `5` | A JNDI lookup, an arithmetic probe, or code access inside template delimiters (`{{ }}`, `${ }`, …) | Yes |
| `NOSQL_INJECTION` | `5` | A MongoDB operator key or object, or a JavaScript injection | Yes |

## Similarity

Matches the text against the training groups selected by `classifier.similarityGroups` or the body's `similarityGroups`. The category score is `max(SIMILARITY, NEAR_DUPLICATE) + CAMPAIGN`, rounded to 2 decimals, and counts only when it is `> 0`. See [Similarity](/docs/sentinel/v2/features/similarity/).

| Rule | Score | Trigger | Can be disabled |
| --- | --- | --- | --- |
| `SIMILARITY` | `0`–`10` | Semantic match: the highest `5 × weight × cosine similarity` among the matched groups, capped at `10`. `null` when no selected group has embedded samples or the text couldn't be embedded. A match against a sample with weight ≥ `1` forces `BAD`. | Yes |
| `NEAR_DUPLICATE` | `0`–`10` | Near-duplicate (SimHash) match: the highest `5 × weight` among the matched groups, capped at `10`. `null` when the selected groups have no samples. A match against a sample with weight ≥ `1` forces `BAD`. | Yes |
| `CAMPAIGN` | `0`–`10` | At least `min_sources` distinct sources sent this text (or a near-duplicate of it) within the campaign window: `5 × sources / min_sources`, capped at `10`. Otherwise `0`. `null` for short texts or when campaign detection is off. | Yes |

## Human Interaction Signature

Runs only on `POST /v2/verify`. See [Human Interaction Signature](/docs/sentinel/v2/features/human-interaction-signature/).

| Rule | Score | Trigger | Can be disabled |
| --- | --- | --- | --- |
| `HIS_AUTOMATION` | `0`–`3` | Bot probability of `0.5` or higher: `(probability − 0.5) × 6`, rounded to 2 decimals. Otherwise `0`. A score ≥ `1.8` forces at least `NEUTRAL`. | Yes |
| `HIS_INTEGRITY` | `3` | Inconsistent signature (fabricated or tampered samples, or a webdriver browser). Forces `BAD`. | Yes |
| `HIS_NO_INTERACTION` | `1` | The visitor didn't move, touch, scroll, click, or type | Yes |
| `HIS_REPLAY` | `2` | The same signature was already submitted. Forces at least `NEUTRAL`. | Yes |

## AI

Each AI rule asks the model a yes/no question. The rule scores `5 × p`, rounded to 2 decimals, where `p` is the model's probability of "yes", if `p` ≥ `0.7`. Otherwise it scores `0`. Rules that weren't asked are `null`. The highest probability also drives the AI [floors](#floors). If the provider fails, `ai.error` is `AI_UNAVAILABLE` and no AI rule scores. See [AI decider](/docs/sentinel/v2/features/ai-decider/).

| Rule | Score | Trigger | Can be disabled |
| --- | --- | --- | --- |
| `AI_SPAM` | `0` or `3.5`–`5` | Is the text spam? | Yes |
| `AI_SCAM` | `0` or `3.5`–`5` | Is it a scam or fraud attempt? | Yes |
| `AI_ABUSE` | `0` or `3.5`–`5` | Is it abusive, harassing, or hateful? | Yes |
| `AI_GIBBERISH` | `0` or `3.5`–`5` | Is it gibberish? | Yes |
| `AI_OFF_TOPIC` | `0` or `3.5`–`5` | Is it off-topic for the form? Asked only when `classifier.aiPurpose` is set. | Yes |
| `AI_CUSTOM` | `0` or `3.5`–`5` | The group's own question. Asked only when `classifier.aiInstructions` is set. | Yes |

## Email (EML) rules

`POST /v2/eml` scores a raw message with the EML rules below. It also runs the classifier on the sender address, subject and text, and originating IP. Email analysis is aggregated differently from the classifier:

| Aspect | EML behavior |
| --- | --- |
| Score | **Sum** of the non-null EML rule scores (not a mean), rounded down to 2 decimals, with a minimum of `0` |
| Classifier result | Its overall score becomes the `CLASSIFICATION` rule. The classifier's own rules appear in `classification.triggeredRules`. |
| Verdict | `spam` is `true` when `score` is `2` or higher. There are no floors. |
| Classifier differences | No device, rate-limit, HIS, or injection checks. `HOSTING` is skipped when the sending server is authorized by the sender's domain. |
| `triggeredRules` | EML rules that scored `> 0`, highest first. `LIST_UNSUBSCRIBE` never appears, because it only lowers the score. |
| Disabling | `X-Disable-Rules` request header: up to 200 EML and classifier rule names, separated by `,`, `;`, or whitespace. Classifier names in it replace the security group's `classifier.disableRules`. |

| Rule | Score | Trigger | Can be disabled |
| --- | --- | --- | --- |
| `ARC` | `4` / `0.5` | ARC result `fail`: `4`; `neutral`: `0.5`; any other result: `0`. `null` when not checked. | Yes |
| `ATTACHMENT_RISK` | `0`–`4` | The highest risk among all attachments: deceptive name or executable content `4`; archived executable, disk image, or executable `3`; macros `2.5`; encrypted archive `2`; HTML/SVG or type mismatch `1.5` | Yes |
| `BULK_MAILER` | `1.5` | `X-Mailer` or `User-Agent` names a bulk-mailing tool | Yes |
| `CLASSIFICATION` | classifier score | The classifier's overall score for the sender, subject, text, and originating IP | Yes |
| `DATE_ANOMALY` | `1` / `0.5` | `1` when `Date` is missing, invalid, or more than a day after the topmost `Received` time (or now, without one); `0.5` when it's more than 7 days before that time | Yes |
| `DELIVERED_TO_MISMATCH` | `1` | The first `Delivered-To` differs from the first `To` address | Yes |
| `DKIM` | `4` / `0.5` | `4` when any DKIM signature fails; otherwise `0.5` when any is `neutral`; otherwise `0`. `null` when not checked. | Yes |
| `DMARC_ALIGNMENT` | `4` / `1` | DMARC `fail`: `4` under a `quarantine` or `reject` policy, `1` under `none` or an unknown policy. `null` when not checked. | Yes |
| `FROM_SPOOFING` | `2` | A `From` display name contains a different address than the sender's (`"admin@paypal.com" <evil@attacker.example>`) | Yes |
| `HIDDEN_TEXT` | `1.5` | At least 200 characters of hidden HTML text, making up more than half of the HTML's text | Yes |
| `HTML_ONLY` | `0.5` | An HTML body without a `text/plain` alternative | Yes |
| `LINK_PHISHING` | `10` per link | Each distinct link target on the phishing list that the text's `URL_PHISHING` didn't already score | Yes |
| `LINK_TEXT_MISMATCH` | `1` | An anchor's text shows a URL or domain on a different site than its target (subdomains count as the same site) | Yes |
| `LIST_UNSUBSCRIBE` | `-1` | A `List-Unsubscribe` header on mail that passes DMARC. Lowers the score. | Yes |
| `MALICIOUS_ATTACHMENT` | `10` per attachment | Each distinct attachment (by checksum) on the malicious file hash list | Yes |
| `MESSAGE_ID_ANOMALY` | `1` / `0.5` | `1` when `Message-ID` is missing; `0.5` when it's malformed or its domain is unrelated to the sender and relays | Yes |
| `NO_SUBJECT` | `0.5` | No subject, or the subject `(No Subject)` | Yes |
| `NO_TEXT` | `2` | No text body | Yes |
| `REPLY_TO_SPOOFING` | `2` | The first `Reply-To` differs from the first `From`: by address for a free-provider sender, otherwise by domain | Yes |
| `SPF` | `4` / `0.5` | SPF result `fail`: `4`; `neutral`: `0.5`; any other result: `0`. `null` when not checked. | Yes |
| `SUSPICIOUS_LINK` | `2` / `0.5` | The most suspicious link: `2` for a `javascript:` URL, a `data:` URL outside `img`, or an IP-address host; `0.5` for a link shortener | Yes |
| `UNDISCLOSED_RECIPIENTS` | `1` | No `To` recipients | Yes |

See [Email analysis](/docs/sentinel/v2/features/email-analysis/) and the [API reference](/sentinel-api-docs-v2.html).
