pdmsc.com, pdmsc.net and pdmsc.org all reach this same site.
PDMSC · staged mail classifier

Mail classification that never leaves your network.

PDMSC is a cascade of small, specialised models. Cheap, precise layers settle most mail early; the expensive layers only see what survives. It runs beside your existing rspamd on hardware you already own, with no per-seat metering and no message content sent to anyone.

In production on live mail · BSD, Linux and Solaris · x86 to SPARC
98.88%
of spam caught, over a 5,030-message holdout set that is never trained on
0
legitimate messages lost, across 4,953 known-good messages
58 ms
median scan time; 133 ms worst case observed, on a production host
1.37 GB
resident memory for the English set, with 557 MB on disk

01What PDMSC does

Cascade

Decide early, defer when unsure

The current English set is nine models arranged as a cascade, each able to decide or defer. Peer groups require agreement before a single specialist can act alone.

Honest

Undecided is an answer

Mail the cascade cannot score, through no detectable language or too little text, is returned as undecided rather than guessed at. The surrounding rspamd signals decide it.

Integration

Drops in beside rspamd

PDMSC is an external_services backend over HTTP. It can run in observe-only mode first, scoring live mail at zero weight so you see what it would have done before it decides anything.

Footprint

Runs on what you have

FreeBSD, OpenBSD, BSDSUniX, Linux and Solaris 10, on AMD64, Intel, ARM64 and SPARC64, on bare metal, virtual machines, or logical domains.

Languages

Language sets load independently

The on-disk format is structural rather than fixed to one shape. Language sets sit alongside one another, and models tuned to a particular site load the same way.

Safe failure

Fails open, never blocks mail

If the classifier stops for any reason, the verdict is neutral and mail flows on unfiltered. A licence problem can stop the classifier; it can never stop your mail.

02How we measure

Every figure on this page comes from the real daemon on production hardware, not from an offline scoring harness, because offline replay reads about 4% optimistic here. Rates are over messages the cascade can score, and that exclusion is stated rather than folded into the percentage. Corpus labels are audited before a result is believed; one message counted against the engine in an earlier run turned out to be spam that had been mislabelled as legitimate. Figures are dated August 2026 and are reproducible on request, including on your own corpus.

03Part of a larger stack

Proxy

Scorpion

A protocol-aware mail proxy with click-time URL protection. See scorprox.com.

URLs

MISUIC

URL intent classification that judges what a link is for. See misuic.com.

Reputation

SURIBS

Hostname and domain reputation, in beta. See suribs.com.

Get started

See what it would have done with your mail.

Tell us about your rspamd deployment and we will scope an observe-only pilot on your existing infrastructure.

enquiries@misullivan.com