Blossom Bifurcation Threat Model

An open, free alternative to gacha mechanics. Players pay for regret-free convenience, not for luck. This live demo scores how much a player looks like a "meta-gamer" probing the choice-permutation, so the game can adapt randomization per player.

MIT licenceMLP 7-16-8-4-1trained on simulated playersWritistic Studios

1. Describe a player

2. Threat score (live model)

...

Randomization applied

Smoothed score (EMA 0.7 old + 0.3 new)

Sample stage permutation from /permutation for this player. Same click, different outcomes when the threat is high.

3. Bands from the research paper

How it works

Player choices→7 behaviour features→MLP (5-model ensemble)→threat 0-1→EMA smoothing→randomization band→stage permutation

Honest caveat: no real player logs exist yet, so the model is trained on 30,000 simulated players (7 archetypes). Accuracy numbers in the repo are measured on simulations, not real players. Retrain on real beta data before relying on it.
Inference: runs entirely in your browser (the 5-model ensemble is loaded as JSON). Hosted API: bbm-threat-api.vercel.app. Or self-host service/ from the GitHub repo.