Experimental X-inspired audience and distribution simulator. Comparative, not predictive. Not X production. No live X feed.
A creator submits a draft (caption, up to 5 images, or a short video). Fifteen curated behavior profiles emit Phoenix-style affinities, which Python maps onto assumed per-impression priors. Public RankingScorer defaults and synthetic candidate-slate competition then drive a Monte Carlo spread. The headline blends 35% persona prior with 65% median sampled population/ranking outcome; Groq never emits it.
What you can do
- Score a hook against tech, fitness, finance, or comedy packs, sampled into 40 / 100 / 320 / 500 coherent audience members.
- Optional Hook B compare on the same media and seed.
- Verdict with full-cascade p10–p90, Niche Index, audience fit, negative risk, stability, and rewrite suggestions.
- Save owner-isolated runs with verified provenance/config hashes and replay stored probabilities by id.
- Optional cluster-aware BluePrint heads apply persona-preserving favorite (40%) and retweet (25%) log-odds lifts.
- Compatible actions are sampled independently; the shown graph is the run nearest median exposure and score.
Local run
Python 3.11+, Node 20+, Groq key in server .env. Visitors do not enter a key. Backend on :8000, frontend on :3000. Full commands live in the GitHub README.
Disclaimer
Prior-mapped research prototype, not empirically calibrated. Ranking weights follow public X defaults, not runtime experiments or the full production stack. p10–p90 measure variation across full-cascade simulator runs, not model confidence. Treat ranges as comparative scenarios, not forecasts.