Roadmap & Changelog

Roadmap

Dashboard control for which text engine runs on a policy. llm_config and ai_text_config already exist together on the API
JSON editor to set up the LLM by pasting a config
Policy versioning: revert to an older config from the policy or evaluation screen
Stronger scam and phishing detection (AI on top of URL allowlists, with user scoring)
LLM judge for sensitive labels
Vertical policy packs

Several countries per check

Check and Analyze accept country_codes, a list of up to 300 ISO 3166-1 alpha-2 codes. The first is the primary country
country_code is deprecated. It still works as a one-entry list and is ignored when country_codes is sent
Every country is sent to the AI text and image harm detection engines as country context
content_country_code rules match when any of the content's countries matches (eq, in) or none does (ne, not_in)
Review queue items return country_codes, and the country_code filter matches any of an item's countries. $exists: true now means the item has a country, and $ne / $nin also return items without one. This applies to saved queues and bulk actions too
Analytics count an item with several countries under each of them
Details: Check and Analyze and Rules

LLM context from custom data

AI Text (LLM) policies can read selected moderation_payload.custom keys as background context, e.g. the article or video a comment is posted under and the earlier replies in its thread
Set context_custom_keys (up to 10 keys) and an optional custom_keys_context_description on llm_config
Only texts are moderated; the context is never scored, labelled or flagged, and instructions inside it are ignored
Details: Context from custom data

Text normalization before moderation

Chat, Feeds, and Check now decode HTML entities and strip zero-width / invisible characters before filters, AI, and the review queue run
Encoded URLs and words (https://…, missing semicolons, &#104; chains, http) match the same plain rules as the rendered text
Stored message bodies are unchanged; the review queue shows the decoded copy
Details: Text Normalization

Analyze API

New analyze method on the server-side SDKs (POST /api/v2/moderation/analyze) for per-field text and image screening
Documented on the Content page next to Check, with a comparison of when to use each
Named texts via the server-side SDKs. Raw image:<label> bytes are HTTP multipart only (not in the SDK)
Stateful (review queue + rules) or stateless (inline verdicts only). Sync HTTP body or async completion webhooks
New webhook events: moderation.text_analysis.complete, moderation.image_analysis.complete, moderation.analysis.failed
Details: Analyze

Keyframe video moderation

Uploaded videos can be sampled into frames, run through image AI, and scored worst-frame-wins
Results present as ai_video, so existing video rules, billing, and queue filters keep working
Configure sampling in the dashboard; flagged videos play in the review queue
Call review: the timeline shows which rule fired and highlights the keyframe or caption that caused it
Details: AI Video File Moderation

Rules and flood detection

user_reaction_count: act on reaction velocity, or on the number of distinct messages a user reacted to
user_identical_image_count: the same image matches by bytes, not by CDN URL
Identical flood can count the same text across users (track_across_users)
Flood windows accept 30s and 15s. Optional min_text_length skips short copy
Optional country_code on Check and Analyze (ISO 3166-1 alpha-2)
Feeds activity and comment custom now reaches content custom-property rules
Domain allow lists win over AI Link / URL detection for those hosts
Details: Rules and Flooding

Review queue and dashboard

Dedicated Moderation API and webhook logs, no longer mixed into Chat → Logs
Usage → Moderation: org-wide AI text, image, and video charts plus a per-app table
Content queue: min-confidence filter, keyword search, and filter by rule name
Labels explorer: keyword search on stored content
Custom queue actions for standalone (non-Chat / Feeds) entities
Restore is the single reversal action (replaces a separate Unblock)
Context modal shows the action taken, and nests image L2s under their L1
Appeals: translate the reason and linked content; appeal one stream:user ban at a time
Server-side shadow enforcement: shadowed Chat messages and Feeds activities stay visible only to the author
Details: Review Queue, Appeals, and Logs & Analytics

Policy evaluations

New Evaluations page under Configuration in the Moderation dashboard
Create a fixed batch of production or pasted content, run it through a policy, and compare later runs to a baseline
Server-side REST API under /api/v2/moderation/policy_tests/

New Moderation Dashboard + Analytics

Advanced Call Moderation Controls

Expanded API & SDK Moderation Support

User Profile Moderation
Added support for delete actions within the queue with a corresponding webhook event