Sahek/velocity-gated-ranking-relief

OON Relief Simulation

A reference simulation of a velocity-gated out-of-network (OON) ranking relief mechanism, built as a companion artifact to an RFC proposal inspired by the publicly visible structure of xai-org/x-algorithm (home-mixer/scorers/ranking_scorer.rs, 2026-05 snapshot).

What this is

This is an independent design simulation, not extracted production code. It does not contain real internal parameter values, the actual Phoenix ranking model, or any non-public source. It exists to let the mechanism be reasoned about and stress-tested in isolation:

Should candidates that are accumulating organic engagement unusually fast relative to post age and author baseline, and have already passed safety/policy classification, receive a small, bounded reduction in their out-of-network ranking penalty — without opening a gameable path for coordinated inauthentic engagement?

What it demonstrates

Three synthetic candidate cohorts are scored:

• —breaking_event — organic, accelerating engagement, low prior exposure
• —normal_oon — steady-state out-of-network baseline content
• —coordinated_burst — engineered to have breaking-event-level raw velocity, but with network impressions already far past a macro-exposure threshold (simulating a click-farm/botnet amplification pattern rather than organic discovery)

The core question this prototype tests: does the impression-ceiling (CIB) guard actually suppress relief for the coordinated_burst cohort, or does it just get waved through because velocity alone looks high?

In local testing, the burst cohort saturates the velocity score (same as genuine breaking events) but the CIB guard's exponential decay brings its average score uplift back down to roughly normal-OON levels — i.e., the guard does meaningful work rather than being a stub. This is implemented logic, not a TODO, unlike the impression-ceiling tracker in the original RFC draft, which left it as a follow-up.

Structure

• —oon_relief_sim.py — core scorer: velocity hydration, normalization against author-tier baselines, the impression-ceiling CIB guard, and the bounded relief calculation
• —synthetic_data.py — generates the three test cohorts
• —app.py — Gradio demo with sliders for max_relief, base OON weight, and the CIB guard's threshold/decay parameters

Dataset

The synthetic cohort data used in this simulation is published separately as a HuggingFace dataset: Sahek/oon-relief-synthetic-cohorts.

Open questions this doesn't resolve

• —Real-world baseline velocity distributions (this uses illustrative constants, not measured data)
• —Whether normalization should be per-author-tier (as here) or per-account historical baseline
• —Whether content classifiers passing (is_policy_clean) is sufficient gating, or whether engagement-pattern anomalies (e.g. low reply-to-repost ratio, as modeled in the burst cohort) should themselves feed back into the policy-clean determination rather than only the impression guard

Author

Alice (Kay) — HuggingFace · X