lukasandClaude Opus 5
95ca1b49f6
Second-pass spam filtering per SPEC Appendix C. The engine is Rspamd in its own container; this plugin is glue and policy, never a spam engine of its own. Classification (§C.4) hooks message.new, acts only on inbox arrivals of enabled accounts, and runs deny list -> allow list -> engine -> thresholds. Every failure path ends in "deliver normally, unclassified": an unreachable engine, a missing blob, an account with no Junk folder, a failed verdict write. Nothing here may delay or lose mail, so no service declares depends_on: rspamd. Two containers, started with everything else rather than behind a profile: rspamd, and a redis of its own. The separate redis is not incidental — the app's redis holds sync leases and arq queues under noeviction, and an ever-growing Bayes corpus must never be able to fail a job enqueue. This one takes RDB snapshots and evicts, because its data is derived and retrainable. Network checks (RBL/SURBL/fuzzy) are off: they leak message-derived hashes and URLs to third parties, which a self-hosted server's users did not ask for. The engine makes no outbound connections in the default configuration. The controller password is generated into .env like MASTER_KEY. It is internal; humans will reach Rspamd through the admin panel, authenticated as the ordinary administrator, rather than typing it. local.d/groups.conf zeroes six SMTP-time symbols — DATE_IN_PAST, PREVIOUSLY_DELIVERED, HFILTER_HOSTNAME_UNKNOWN, RCVD_NO_TLS_LAST, ARC_NA, DMARC_NA — and greylisting is disabled. Measured against 398 real delivered messages, those checks were firing on 44-71% of ordinary mail and pushing 11.3% of it over the banner threshold; scoring mail at rest for facts true of every delivered message is simply the wrong question. After: median 3.09 -> 0.70, p95 7.00 -> 3.40, banner rate 11.3% -> 1.0%. The spec's 12.0/6.0 thresholds were never the problem and are unchanged. Symbols are zeroed rather than removed so verdicts stay explainable, and they matter again at the M6 edge (§C.7). Still to come in C1: the §C.6 router and UI slots. C2 adds the learning loop — until Bayes is trained, genuine scams and legitimate bulk mail overlap in the 5-8 band and no threshold separates them. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_018SqJYt7jD7oUEkoY68sBZo
35 lines
1.8 KiB
Plaintext
35 lines
1.8 KiB
Plaintext
# Second-pass reality check (SPEC §C.1: this filter runs on mail the provider has
|
|
# already accepted and delivered).
|
|
#
|
|
# Every symbol below is a legitimate SMTP-time signal that is meaningless — or
|
|
# actively wrong — when scoring a message at rest, days after delivery, fetched
|
|
# over IMAP. Measured on a 398-message sample of real delivered mail, they fired
|
|
# on 44-71% of it and pushed 11.3% of ordinary mail over the banner threshold.
|
|
#
|
|
# They are zeroed, not deleted: they still appear in the symbol list, so a verdict
|
|
# stays explainable, they just no longer contribute score. When M6 lands, the edge
|
|
# scanner (§C.7) sees real SMTP state and these matter again — there, not here.
|
|
#
|
|
# Measured on 398 real delivered messages, before -> after this file:
|
|
# median 3.09 -> 0.70
|
|
# p95 7.00 -> 3.40
|
|
# >= 6.0 11.3% -> 1.0% (the banner threshold)
|
|
# max 11.66 -> 9.15 (move threshold is 12.0, so: headroom)
|
|
# The thresholds themselves were left at the spec's 12.0/6.0 — they were never
|
|
# the problem. Re-measure after the C2 learning loop has a trained Bayes: until
|
|
# then genuine scams and legitimate bulk mail overlap in the 5-8 band, and no
|
|
# threshold can separate them.
|
|
symbols {
|
|
# "This message is older than the scan" — always true for us.
|
|
"DATE_IN_PAST" { score = 0.0; }
|
|
# "Already delivered once" — by definition, for a second-pass filter.
|
|
"PREVIOUSLY_DELIVERED" { score = 0.0; }
|
|
# Properties of the connecting host, which we never saw.
|
|
"HFILTER_HOSTNAME_UNKNOWN" { score = 0.0; }
|
|
"RCVD_NO_TLS_LAST" { score = 0.0; }
|
|
# Absence-of-authentication penalties. We cannot verify SPF/DKIM/DMARC/ARC at
|
|
# rest anyway, so "not available" says nothing about this message.
|
|
"ARC_NA" { score = 0.0; }
|
|
"DMARC_NA" { score = 0.0; }
|
|
}
|