Word lists

DaisyFilter ships verified rule packs and copies reference files into plugins/DaisyFilter/lists/ on first run. The pinned CSV inside the plugin JAR is the runtime baseline; deleting or damaging a reference copy cannot disable it.

Rule sources

Rules are kept separate by source and never merged ambiguously:

• GamerSafer — an imported, attributed third-party list (see below).
• Master list — a merged, multilingual word list (English plus Spanish, Portuguese, French, German, Italian, and Russian) built from openly-licensed sources and deduplicated against GamerSafer at build time. See Master list below.
• Daisy defaults — DaisyFilter’s reviewed core-safety pack, active before asynchronous third-party loading starts.
• Runtime overrides — disabled rules and exceptions written by the GUI.

Each rule has a stable ID, one or more categories, a severity (1–5), a match mode, the normalized forms it applies to, the surfaces it applies to, and optional action overrides.

GamerSafer attribution

DaisyFilter bundles a pinned snapshot of the MIT-licensed GamerSafer word list at commit d426da8da414892e0efee0381f0480220182df09, including its LICENSE, README, and source manifest with file checksums. GamerSafer remains attributed under its own licence; DaisyFilter itself is proprietary.

Rule IDs are salted hashes of the normalized entries. No source term is ever printed in status output, logs, masked previews, or rule IDs. The versioned default-unclassified overlay gives every upstream-uncategorized row an explicit strict censor policy; reviewed severe and ambiguous exceptions override that fallback.

Master list

The master list broadens coverage beyond the English-centric GamerSafer set with multilingual profanity (Spanish, Portuguese, French, German, Italian, Russian). It is generated at build time by the list curator, which merges openly-licensed sources, normalizes and deduplicates them (including against GamerSafer so nothing is redundant), and writes a pinned, checksum-verified CSV bundled in the JAR. Flat sources carry no per-word category, so each entry is tagged swearing (censor) — a safe default; the curated core and structural rules still block the severe slurs.

Sources and their licences are recorded in the bundled NOTICE file (LDNOOBW under CC BY 4.0, GamerSafer and bad-words under MIT). To regenerate after updating the gitignored tools/list-curator/sources/:

list-curator masterlist           # rebuild filter-lists/master/master-list.csv (+ SHA-256)
list-curator confusables          # rebuild the Unicode confusables map

After regenerating, update the pinned MASTER_LIST_SHA256 in DefaultRules to the printed checksum.

Reviewing GamerSafer into the overlay

Most GamerSafer rows arrive uncategorized. The strict fallback prevents those rows from silently broadcasting while the development-only list curator records individual exceptions through the daisy-category-overlay.yml, which is keyed by hashed ids and therefore safe to commit (it never contains source terms):

# coverage report (no terms shown)
list-curator stats

# write the rows still needing a decision (ids only; add --reveal to include terms
# into a worksheet under the gitignored exports/)
list-curator worksheet exports/worksheet.csv --reveal

# after filling in category/severity/disposition, merge decisions into the overlay
list-curator apply exports/worksheet.csv

The plugin loads the overlay on startup and reload and applies each decision to the matching rule. The coverage command treats the explicit fallback as a decision and must report 100%.

Customising

Do not edit the copied GamerSafer CSV to change runtime policy; it is an attributed reference copy. Use the hashed daisy-category-overlay.yml to review imported rows, and allowlist.yml for benign exceptions. Invalid decisions are reported using only their opaque rule ID.

Server owners can add RE2/J-safe regular expressions and literals in custom-rules.yml. Invalid entries are rejected during reload without replacing the previous valid rule snapshot.

Built-in safety patterns

Beyond word lists, DaisyFilter ships structural detection rules that run against the raw message (so dots, slashes, and @ are preserved):

• URLs and www. links, and bare domains on common TLDs — blocked by default.
• IPv4 addresses — blocked. Any dotted-quad literal is treated as an address (the filter cannot tell a server IP from a four-part version number), so allowlist your own server/LAN addresses under allowlist.yml → domains if you reference them in chat.
• Discord invites (discord.gg/..., discord.com/invite/...) — blocked.
• Email-like and international phone-like patterns — monitored and alerted only (advisory “potential personal information,” never auto-punished).

Add permitted hosts to allowlist.yml → domains. These patterns are precise by construction and treated as high-confidence; tune category dispositions in categories.yml.

Match modes and forms

Short entries use exact token matching to avoid accidental substring hits. Strict matching (leetspeak, confusables, invisible-character removal) is enabled only for curated rules. See categories.md for severities and dispositions.