Grox: the LLM that reads every post

"Banger" quality screening

CODE grox/classifiers/content/banger_initial_screen.py + grox/tasks/task_banger_screen.py + grox/plans/plan_initial_banger.py

An LLM pass that screens posts for quality — X literally runs a 'is this a banger' classifier over content. Output feeds quality/slop scoring used in content understanding.

Implication: Content quality is machine-read by a language model, not inferred from engagement counts alone. Slop is detectable and detected.

Spam detection

CODE grox/classifiers/content/spam.py + grox/plans/plan_spam_comment.py

LLM spam classification including reply/comment spam — the reply-guy economy is machine-screened.

Implication: Template replies, copypasta, and AI-generated reply-farming are exactly what an LLM classifier catches that keyword rules never could.

Safety screens

CODE grox/classifiers/content/post_safety_screen_deluxe.py + safety_ptos.py + grox/plans/plan_post_safety.py

Multi-tier safety screening (including a 'deluxe' pass and PTOS policy screening) whose verdicts feed the withheld visibility-filtering layer.

Implication: The enforcement thresholds are not public; no external tool can tell you your safety verdict.

Reply ranking

CODE grox/classifiers/content/reply_ranking.py + grox/plans/plan_reply_ranking.py

Replies under a post are themselves ranked by an LLM pass — reply visibility is a judged competition, not chronological.

Implication: A thoughtful reply can outrank earlier/louder ones; low-effort first-replies are demotable on content alone.

Multimodal post embeddings

CODE grox/embedder/multimodal_post_embedder_v2.py / _v5.py + summarizer/

Posts (text + media, with ASR for audio) are embedded and summarized by the VLM — this is the ONLY path by which your content's meaning reaches the recommendation system.

Implication: The system reads what your post is about, including what's inside images and video audio. Alt-text games and keyword stuffing don't reach the ranker; meaning does.

Why this matters more than any weight

The 2023 algorithm counted engagements. The 2026 stack reads content — a Grok VLM embeds and summarizes every post (including what’s inside images and video audio) and screens for spam, safety and quality before ranking ever happens. Optimization theater aimed at counters (pods, bait, tag-stuffing) now runs head-first into a language model. The checker’s slop and bait flags exist because of this layer.