How it works

Four steps between a message and a decision.

Nothing here is a black box to you. Every step is inspectable in your dashboard, every number is yours to change, and every decision arrives with the reasoning attached.

Step one

Undo the disguise before reading the message.

A filter that only reads what was typed loses to anyone willing to type it differently. Wardnox resolves each message past the families of disguise below before a single rule looks at it, and keeps the original untouched, because the original is the evidence.

Letters that are not letters

Characters from other alphabets, and marks stacked on top of ordinary ones, chosen because they look identical to you and read as different to a filter.

Characters you cannot see

Invisible and direction-controlling characters dropped inside a word so that nothing matches it, while it still reads normally on screen.

Text pulled apart

Spacing, punctuation and formatting used to break a word into pieces that a reader reassembles without noticing and a filter does not.

Things wrapped in other things

Layers of encoding, and links wearing a label that says somewhere else entirely.

Resolving a message aggressively is how you catch a disguise, and also how you catch an innocent person by accident. So the more aggressive readings count for less than the plain one, and your moderators are always told which form of the message a rule actually matched. A catch is never just an assertion.

SEEING THROUGH ITLIVE

A message, as a member sees it

What Wardnox found

  • A word broken up by a character with no width
  • A letter borrowed from another alphabet
  • Support phrasing pushing people into direct messages
  • A link arriving alongside both
Risk
High

Removed before anyone read it, and the sender timed out.

Step two

Collect evidence. Do not jump to a verdict.

Twenty detectors run across four stages, and each returns weighted evidence rather than a decision. What any one of them can contribute is capped, so the loudest signal cannot simply decide the outcome by itself.

Why the ceilings matter

A link can look odd in several unremarkable ways at once: long, oddly structured, slightly unusual. Added up naively that arithmetic reaches a removal. Held under a ceiling it reaches a warning, which is the honest answer for a link that is strange rather than dangerous.

Coincidence is not a pattern

Some pairs of signals mean much more together than apart, and Wardnox recognises those specific combinations rather than simply adding everything up. Weak signals that merely happen to co-occur stay weak.

Step three

The first thing that decides, wins.

Questions are asked in a fixed order, and the first one that settles the matter ends it. That is what makes “why did this happen” answerable rather than speculative.

1Is this in scope at allServer level switches, and any channel or category you have told it to leave alone. A channel you exempt is exempt completely.
2Does someone here outrank the rulesYour staff, recognised by the permissions they already hold in Discord rather than by a list you have to maintain, plus anyone you have chosen to trust.
3What do your own rules sayAnything you wrote, scoped to where you wanted it, including rules you marked important enough to apply even to trusted accounts.
4What does the weight of evidence sayEverything the detectors found, combined, and turned into the outcome you configured for that level of confidence.

Staff are exempt by permission, not by a list

Owners, administrators and moderators are recognised from the permissions they already hold in Discord, resolved for the specific channel they are speaking in. There are no IDs to paste and none to go stale when somebody changes role.

Trust without losing the evidence

Fully trusting a role skips detection, and with it the record: nothing is examined, so your logs can never tell you whether that role is being abused. You can instead mark a role as watched but never acted against, which keeps the evidence while leaving the people alone.

Step four

Do only what was configured, then show the working.

What happens at each level of confidence is yours to set. Guard rails then remove anything that should never have been possible, and record what they removed, so an administrator who switched something on and saw nothing happen is told why rather than left guessing.

1

Undo the disguise first

Attackers beat naive filters by writing the same thing a different way. Wardnox resolves a message past its disguises before any rule looks at it, and keeps the original untouched, because the original is the evidence.

2

Combine signals

No single weak signal, not a link, not a new account, not one alarming word, causes enforcement on its own. Detectors return weighted evidence, and the engine only escalates when a genuine pattern appears rather than a coincidence.

3

Explain every decision

Every action carries a reason your moderators can read: what fired, what it matched, and how the result was reached. No black boxes, and nothing you cannot argue with.

4

Administrators keep control

The profile, the sensitivity, what each outcome does, and every rule are yours to change. The example rules ship clearly labelled as examples, so nobody mistakes them for a curated ruleset doing the real work.

Read a day of real decisions first.

Every server starts in monitor mode. Nothing is enforced until you have seen what would have been.

No spam, no newsletter. One email when your server can be added.

Go deeper into the engine