Image scam OCR / QR checks, Dangerous-file layer decisions, trusted-link mode, and Anti-Spam policy stay deterministic first. Independent user, role, and mass-mention rules use explicit per-message and rolling limits under /shield-rules mentions with bb!shield-rules mentions prefix parity. OCR scam scoring covers fake Nitro, Robux, Steam gift, wallet, QR-login, and MrBeast-style crypto lures; moderator review can use the configured Shield log channel while Babblebox stores sanitized OCR text only.
Babblebox is a private bot available only in its support server.
Shield AI is a second-pass, live-message-only assist after local rules flag content. Every routing tier uses GPT-6 Luna (gpt-6-luna), ordinary guild access can be gated to Guild Pro plus provider/runtime readiness, and the support server is enabled by default through a separate support-AI switch that either of the two system owners can disable or restore.
Manage Messages is the moderator baseline for review surfaces, moderators are exempt from Anti-Spam by default, the mandatory punishment-DM policy is on for new guilds, temp-ban expiries are restart-safe, and public /info server stays separate from private /modinfo user.
Pure join volume can alert without punishment. Containment waits for stronger raid evidence, trusted-role suppressors, event mode, and recovery paths.
Standard is the default public Confessions mode and blocks vulgar language and explicit adult sexual text. Mature is available only while an administrator-selected public destination is currently marked NSFW by Discord. A language-only block does not add a Confessions strike or apply an immediate restriction. Both modes keep exploitation, coercion, credible threats, self-harm encouragement, severe hate, doxxing, malicious links, and other serious-harm protections active.
Recoverable Confessions denials can offer Copy Draft, Edit, Show Draft, Discard, and a preview-only safe fix. A language-only Standard block can offer a currently eligible Mature destination to open a private editor. Nothing moves or posts automatically. Encrypted recovery lasts up to 15 minutes, uses a domain-separated author binding, and never retains attachment bytes or raw CDN URLs; uncertain delivery waits for explicit reconciliation. Regular members, server administrators, and moderators cannot identify the author through normal Babblebox Confessions surfaces. The service operator remains a trust boundary.
Setup Hub, Role Tools, Role Rules, Setup Sync, logs, appeals, permissions diagnostics, and owner control are clear, narrow surfaces instead of a giant settings dump. Open /setup panel -> Logs for route setup. Use Set Security Alerts for the recommended private security bundle covering high-risk security, anti_nuke, and raidguard events. Audit Trail is separate, broader, and noisier for Discord audit-log observations or admin evidence.
Admin lane stays small with reversible locks, permission truth, who can use the commands, what /giverole may add, what /removerole may remove, Role Rules, /admin role-rules, /admin role-history, Role Tools, /admin role-tools, /giverole, /removerole, exclusions, new-member auto-role, and seven days of visible revoke history.
The panel-first rules flow keeps Actions, Options, and Exemptions together with a dedicated timeout profile, focused Mentions controls, separate GIF Flood / Media Pressure controls, and smart or never-ping alerts for compact moderation logs.
Dangerous-file layer checks use optional reputation without uploading files; Babblebox never executes, decompresses, inspects macros, or uploads files to VirusTotal or MalwareBazaar. The bounded private immunity lane and private feature-surface checks keep AFK reasons, reminders, Watch keywords, and configured local Shield checks separate from live moderation alerts.
Admin proof
Powerful flows, separated by risk.
Safety, moderation, and setup each have their own lane. That keeps routine operators away from risky actions while still giving admins fast recovery tools when the server needs them.
