Scores

Tag scores are where an administrator tunes the spam filter most directly. Every finding of the spam filter is recorded as a tag, and the score of a message is the sum of the scores of its tags. Changing how much a tag counts therefore changes how strongly the filter reacts to that finding, without touching the check that produces it.

The score of each tag, or the action it triggers, is defined by a SpamTag object, found in the WebUI under Settings › Spam Filter › Scores. A tag without a SpamTag object contributes nothing to the score.

Score, reject and discard

Most tags contribute to the score, but some findings are decisive on their own. SpamTag therefore has three variants:

Variant	Fields	Effect when the tag is added
Score	tag, score	Adds score to the message score. Positive values count towards spam, negative values towards ham.
Reject	tag	Refuses the message at the end of DATA with 550 5.7.1 Message rejected due to excessive spam score., whatever the score.
Discard	tag	Accepts the message with 250 2.0.0 Message queued for delivery. and drops it, whatever the score.

A Reject or Discard tag cuts the analysis short. Once such a tag has been added, the AI classifier is not consulted, and the rules that depend on it do not run. The decision is applied when the scores are added up. When a message receives both a Reject and a Discard tag, which of the two applies is not defined.

The tag field accepts a glob pattern, so one object can set the score of a whole group of tags.

The following object lowers the score of the classifier’s medium-confidence ham verdict:

{
  "@type": "Score",
  "tag": "PROB_HAM_MEDIUM",
  "score": -3.0
}

The following object discards every message sent to a spam trap:

{
  "@type": "Discard",
  "tag": "SPAM_TRAP"
}

Default scores

The default scores are part of the rule set that Stalwart downloads on first start, together with the rules that produce the tags (Rule set updates). The rule set defines about 400 scores.

Changing a score in the WebUI, or with x:SpamTag/set, takes effect for the next message. Local changes are also safe from updates: a later rule set update adds scores for new tags, but never replaces a score that already exists.

The current score of a tag is shown in the WebUI under the spam filter scores. For a particular message, the ClassifySpam action reports the score and the disposition of every tag the message receives.

The rule set uses the decisive variants sparingly. The only tag that it sets to reject is BLOCKED_DOMAIN, added when the sender’s domain is in the blocked-domains list.

Tag families

Tag names start with a prefix that identifies the check that produced them, so the prefix of an unfamiliar tag shows where to look for its cause. The main families are:

Prefix	Produced by	Examples
RBL_, RWL_, DBL_, DWL_, SURBL_, URIBL_, RECEIVED_	DNS blocklists and allowlists, for the client IP, the IPs of the Received chain, domains and URLs	RBL_SPAMHAUS_SBL, DBL_PHISH (7.5), RWL_MAILSPIKE_GOOD
DMARC_, SPF_, DKIM_, DKIM2_, ARC_	Sender authentication; see below	DMARC_POLICY_REJECT (4.0), SPF_ALLOW (-0.2)
PROB_	The statistical classifier	PROB_SPAM_HIGH (8.0), PROB_HAM_HIGH (-8.0)
AI_	The AI classifier	AI_PHISHING_HIGH (7.0), AI_LEGITIMATE_HIGH (-3.0)
TRUSTED_	Trusted domains	TRUSTED_DOMAIN (-7.0)
FROM_, REPLYTO_, TO_, RCPT_, SUBJECT_, MID_, HELO_, RCVD_, FORGED_	Header analyses	FORGED_SENDER, RCVD_NO_TLS_LAST
MIME_, HTML_, HAS_	MIME structure, HTML and attachment analyses (MIME types)	MIME_BAD_EXTENSION (2.0), MIME_LIMIT_EXCEEDED (8.0)
FREEMAIL_, DISPOSABLE_	Addresses at free and disposable mail providers	FREEMAIL_FROM
PHISHED_, HOMOGRAPH_URL, REDIRECTOR_URL, URL_	Phishing protection and URL analyses	PHISHED_OPENPHISH (7.0)
SPAM_TRAP, PYZOR	Spam traps (15.0) and Pyzor (3.5)

Tags ending in _BLOCKED or _BLOCKED_OPENRESOLVER score 0. They mean that a DNS blocklist refused to answer the server’s resolver, not that the sender is listed (DNS blocklists).

Authentication tags

The MTA checks the authenticity of every message it receives, and the authentication tags carry the results of those checks into the score. Four mechanisms are involved:

• SPF verifies that the sending IP address is authorised by the domain of the envelope sender.
• DKIM verifies the cryptographic signatures in the message.
• DMARC requires one of the two to pass for a domain aligned with the From header, and lets the domain owner publish a policy for messages that fail.
• ARC preserves authentication results across forwarders, which would otherwise break SPF and DKIM.

Sender authentication describes the checks, and the settings that make them reject messages at SMTP time. The tags are added whatever those settings are, so a failed check that does not reject a message still counts against it.

Tags	Result	Default scores
DMARC_POLICY_ALLOW	DMARC passes.	-0.5
DMARC_POLICY_REJECT, DMARC_POLICY_QUARANTINE, DMARC_POLICY_SOFTFAIL	DMARC fails, and the domain’s policy is reject, quarantine or none.	4.0, 1.5, 0.1
DMARC_NA, DMARC_BAD_POLICY, DMARC_DNSFAIL	No DMARC record, an invalid record, or a DNS error.	1.0, 0.5, 0.0
SPF_ALLOW, SPF_FAIL, SPF_SOFTFAIL, SPF_NEUTRAL, SPF_NA	SPF passes, fails, soft-fails, is neutral, or the domain has no SPF record.	-0.2, 1.0, 0.0, 0.0, 0.0
DKIM_ALLOW, DKIM_REJECT, DKIM_NA	A DKIM signature verifies; no signature verifies and the first one fails; or the message has no signature that could be evaluated.	-0.2, 1.0, 0.0
DKIM2_ALLOW, DKIM2_REJECT, DKIM2_PERMFAIL, DKIM2_TEMPFAIL, DKIM2_NA	The DKIM2 chain verifies, fails, is invalid, could not be verified because of a temporary DNS error, or the message has no DKIM2 signature.	-0.2, 1.0, 0.0, 0.0, 0.0
ARC_ALLOW, ARC_REJECT, ARC_INVALID, ARC_NA	The ARC chain verifies, fails, is malformed, or is absent.	0.0, 1.0, 0.5, 0.0
TRUSTED_DOMAIN	DMARC passes and the From domain is in the trusted-domains list.	-7.0

One case never reaches these tags. A message whose From header holds addresses in several domains cannot be evaluated by DMARC. When DMARC verification is enabled, the MTA rejects it before the spam filter runs (DMARC).