Public MIT Grok-only

X Posts auto-publish × SlimeTree-RLM μ-prefilter

▸ Stop posts that could blow up before they go live

The draft is pre-judged in-browser by the RLM ― if μ, warn / suppress before posting; if D/R, publish via X API v2 (POST /2/tweets). R verdicts are delegated to Grok (xAI) only (X API access policy).

1. X API v2 credentials

Create an App at developer.x.com → obtain a User Access Token (OAuth 2.0).
Required scopes: tweet.read + tweet.write + users.read. Posting requires X Basic ($200/mo) or higher API tier.

Not authenticated (paste your token, save, then "Check user")

2. Grok (xAI) API key ― R verdict only

R-verdict responses use Grok (xAI) exclusively. Issue an API key (Bearer) at console.x.ai. The key stays in localStorage and never transits the javatel server.

Grok API key

Why Grok-only? X (xAI-owned) is tightening API governance against X-data processing via competitor LLMs (Gemini/Claude/OpenAI). Keeping X-native = Grok preserves access integrity ― this is the "Platform-native LLM" principle of this library (details: X Integration hub).

2.5 LLM polish + flame-risk check ★ NEW: Pattern B (within-Grok tier)

Have Grok "judge flame risk + improve tone" for the draft. Because X-native = Grok-only, cross-vendor is impossible; the same Pattern B works within a single vendor by stepping grok-3-mini / grok-4-fast → grok-4.

Single LLM ― only the single Grok model chosen above Cost-tier (B) ― cheap polish → judge → escalate only the insufficient ones to premium for detail

cheap (B): 0 escalated (B): 0 esc rate (B): ―

3. Post draft

Up to 280 characters (X Premium allows 25K; this demo caps at 280). Before sending, the RLM classifies as D / μ / R.

0 / 280 chars

RLM state: not judged

Final verdict: ―

4. Post / API call log

(nothing has been run yet)

5. Why RLM works for X Posts

μ verdict stops flames before they start: PII / speculative predictions / sensitive terms are detected locally before posting → prevents account suspensions from X-side moderation violations
D verdict deterministically clears routine announcements: fact-based product announcements pass deterministically → raises the safety rate of auto-publishing
R verdict is Grok-only: borderline drafts go to Grok ― the X-native LLM keeps access policy aligned
WAL audit chain: every post candidate is logged in a SHA-256 chain → "who approved which draft into that post?" can be reproduced air-gapped after the fact

6. Source / related

All code on this page = MIT
RLM mock: slimetree-rlm-mock.js (shared with Meta hub)
X API v2 docs: developer.x.com/en/docs/x-api
xAI Grok API docs: docs.x.ai
Sister hub: Meta integration (multi-provider edition)