[Rivmedia] Content Guard 1.0.3

Inherited a forum you are not sure is legal or questionable content? Reopening after a takedown? Tightening up before an Online Safety Act review?

Content Guard reads every thread on your forum against a policy you write in plain English, scores it 0 to 3, records the evidence, quarantines the worst automatically, and tells you forum by forum whether to keep, pre-moderate or hide. It turns "we should probably go through all of it" into a job that runs overnight and a dashboard you can act on in the morning.

It is triage with a human in the loop, not legal advice. Nothing is deleted automatically, every decision is logged with who made it, and the policy text is yours to set with your own advisers.

What it does

• []Bulk scan. Choose forums (or everything), optionally only threads active in the last N days. Runs in the job queue, resumable, cancellable. Each thread is judged on its opening post plus the most recent replies. []A written policy, not a black box. You describe what is legitimate, what is restricted and what is not permitted; the model scores against that. Ships with a UK-focused example policy for a fitness community that you can replace entirely. []Scores with evidence. Severity 0 to 3, categories, up to five short verbatim quotes that justify the score, and a two-sentence rationale a forum owner can act on. []Automatic quarantine at a threshold you set. Default: severity 3 (sourcing, supply, vendor contacts, evasion, minors, poisons) goes straight to the moderation queue, hidden from the public and reversible. Lower it to 2 to catch protocols and advertising too, or set 0 to keep it manual. []Forum verdicts. Per forum: scanned, flagged, severe, open, and a verdict of Keep, Review, Pre-moderate or Hide forum, with a one-click Hide that revokes the view permission for guests and registered members while keeping the content. []Flagged threads dashboard. Filter by forum, severity, status or category. Quarantine, soft delete, restore, dismiss as a false positive, whitelist, or rescan, singly or in bulk. Permanent delete exists, asks twice, and is never automatic. []Live screening. Once you reopen, each new post queues a scan of its thread a minute later and quarantines at your threshold. Staff posts are skipped. []Audit trail and cost. Every review stores the model, tokens, timestamp, action taken and the admin who took it. The dashboard shows totals so you can see spend as it goes.

Models and cost

Any model in the AI Framework list. Claude Sonnet 5 is accurate enough for a full pass at roughly a penny per thread; Claude Haiku 4.5 at about a third of that for a first sweep, with a Sonnet rescan of whatever it flags. Forty thousand threads is a few hundred pounds on Sonnet, around a hundred on Haiku.

Suggested workflow

1. []Keep the board closed. Install the AI Framework and add your key, then install Content Guard. []Edit the policy in Options to match your legal advice. Leave automatic quarantine at 3 for the first run. []Hide by permissions any forum whose purpose is itself the problem (a sourcing forum, a classifieds board for restricted goods). No scanning needed. []Bulk scan the rest. Read the forum verdicts, action the flagged list from severity 3 down.
2. Reopen with live screening on.

Requirements

XenForo 2.2 or later (tested on 2.3.6), PHP 8.0 or later, Rivmedia AI Framework 1.1.2+ with at least one API key.

Changelog 1.0.3

Admin navigation and route moved to the add-on's data files so they survive data syncs. Cancel scan now confirms before stopping. Forum list indentation fixed.

Changelog 1.0.0

First release.