Skip to main content

​
Spam-Keyword Deliverability Classifier

Carrier filters and email inbox providers score every message against a vocabulary of classic spam tells — “FREE!”, “click here”, “act now”, ”$$$”, prize claims, and all-caps subject lines. A message stacked with these triggers gets blackholed, junked, or rate-limited before it reaches your recipient. Orbit runs a SpamAssassin-style classifier over your content to catch those triggers before the carrier does. Every outbound message on SMS, WhatsApp, email, and RCS is scanned in line with the send, and the same scan runs live as you type in the composer. This page explains what the classifier checks, how the score is computed, and what to do when a verdict comes back as warn or block. All endpoints below are rooted at https://api.orbit.devotel.io.
The classifier is a tenant-owned hint, not a platform mandate. A warn or block verdict tells you carrier filters are likely to hurt deliverability on this content; your organization’s policy scan mode decides whether a block verdict stops the send or is only recorded. Orbit never uses the classifier to gate your traffic globally. This page is not legal advice — for the legal spam statutes in your send regions, see AU Spam Act, CAN-SPAM, and CASL.
Two different pre-send scans run on your content, and they answer different questions:
  • The spam-keyword classifier (this page) scores carrier-deliverability and inbox-reputation vocabulary — the words and patterns that make carrier filters and mail providers treat a message as junk.
  • The policy scanner (Pre-Send Policy Scanner) covers the legal and regulatory lanes: TCPA quiet hours, SHAFT content, missing opt-out phrases, GDPR sender identity, country sender registration, and DLP (payment card numbers, SSNs, IBANs in the body).
The spam-keyword classifier is the scoring engine behind one row of the policy scanner’s rule table; both run together on every send, and a high spam score appears alongside the legal findings in the same verdict payload.

​
Channels covered

The classifier runs on SMS, WhatsApp, email, and RCS. Push notifications are excluded — that content is delivered through the OS provider, whose own filtering is upstream of Orbit. For email, the subject line is scanned in addition to the body, and subject hits carry extra weight (see How the score is computed) because the subject is the gatekeeper for inbox placement.

​
How the score is computed

The classifier evaluates your content against a table of keyword and pattern rules derived from SpamAssassin’s public rule set and mainstream mailbox deliverability guidance. Scoring works like this:
  1. Each rule carries a weight from 1 to 25. Higher-signal patterns (prize/winner language, pharmacy spam, account-verification phishing) weigh more than lower-signal ones (aggressive percent-off claims).
  2. The weights of every matched rule are summed. Extra matches inside the same rule add a smaller bonus each, so repeating “click here” five times does not stack five full weights.
  3. Subject hits count 1.5× (email only) — the subject drives inbox placement more than any body line.
  4. SMS and RCS hits count 1.25× — US carrier 10DLC filters are stricter about urgency, “FREE”, and call-to-action language than email providers, so the same vocabulary is penalised harder on those channels. WhatsApp is kept at 1× because Meta’s own filtering is upstream.
  5. The total is clamped to 0–100 and rounded to a whole number.
The score maps to a verdict: A single flashy phrase (“you’ve won”) never blocks by itself. A stack of mid-signal phrases (“FREE!” + “click here” + “act now” + ”$$”) is what crosses the warn and block thresholds — the same shape carrier filters penalise.

​
What a hit looks like

Every matched rule is returned as a structured hit, so you can pinpoint the exact words to rephrase: Rule identifiers are kept stable across classifier revisions, so a trend line on urgency_act_now remains comparable over time.

​
Rule taxonomy

The rules fall into broad families:
  • Urgency / pressure — “act now”, “limited time”, “expires today”, “final notice”.
  • Free-offer language — “100% free”, “no obligation”, “free gift”, plus a dedicated rule for a shouted FREE!.
  • Money claims — multiple dollar signs ($$$), “earn money”, “work from home”, “passive income”.
  • Prize / winner language — “you’ve won”, “claim your prize”, “congratulations”, lottery and sweepstakes terms.
  • Discount pressure — “save big”, “lowest price”, and a % off rule that only fires at 50% and above (ordinary retail discounts like “20% off” are ignored).
  • Generic click-here calls to action — “click here”, “tap to open”.
  • Spam denial — “this is not spam”, “we honor all removal”. Denials are one of the strongest spam tells there is.
  • Health / pharmacy — weight-loss claims and drug-name vocabulary.
  • Adult / dating vocabulary.
  • Investment / crypto pump — “guaranteed returns”, “next bitcoin”.
  • Credit / loans / debt — “bad credit”, “pre-approved”, “payday loan”.
  • Account-verification phishing — “verify your account”, “unusual activity”. Legitimate transactional alerts from your own service are rarely phrased this way; if your traffic needs these words, use your own verified sender identity and avoid urgency framing.
For email, the subject line additionally runs four format rules:

​
Compose-time live linter

You don’t have to send to see the score. The composer calls POST /messages/lint — the same endpoint that runs the full policy scanner — as you type (debounced), and it renders the classifier’s output inline:
  • A score gauge shows the 0–100 score with the pass/warn/block bucket.
  • Each matched substring in spam_hits is underlined in your draft body or subject, labelled with its rule name and category.
  • Rewriting the flagged phrase re-scores the draft on the next debounce, so you can watch the score drop below 60 before you send.
The lint endpoint accepts channel, body, and an optional subject (for email), so the live scan sees the same inputs the send-time scan will.

​
Send-time verdict

At send time, every outbound message on the covered channels is scanned by the same engine as part of the pre-send policy scan. The result lands in the send response and on the message metadata:
  • spam_score — the 0–100 score.
  • spam_hits — the structured list of matched rules, in the shape above.
  • A violation entry (spam_score_elevated for 60–79, spam_score_high for ≥80) inside the violations array and the X-Policy-Violations response header.
Whether a block verdict actually stops the send is governed by your organization’s policy scan mode (warn, strict, or off — default warn, which records but never stops). See Pre-Send Policy Scanner for the mode table and the POLICY_VIOLATION error shape.

​
Worked example

An email marketing draft:
200
Rephrasing to “Start your trial today — save 20% on annual plans” with a calm subject line removes the prize, click-here, free, aggressive-discount, and punctuation hits and brings the draft back under the pass threshold. The classifier is deliberately a deliverability hint, not a regulation. Nothing here decides whether you are legally allowed to send — it predicts how carrier and inbox filters will treat the content you are sending. The enforcement decision (record, block, or skip) lives in your own policy scan mode setting. Contrast this with the per-country spam statute pages — AU Spam Act, CAN-SPAM, CASL — which document the legal obligations (consent, identification, unsubscribe handling) that apply independently of any scoring. A message can pass the classifier with a score of 0 and still violate CASL if consent is missing, and a legally compliant message can still get junked if it reads like spam. Check both.