PanoptOS — public research artifact

PanoptOS investigates early detection of automated Old School RuneScape accounts using only public Hiscores time series and delayed labels obtained from account disappearance after Jagex enforcement. This repository is the public companion to the technical report; a less formal adaptation is also included.

This is a research artifact, not a running system. PanoptOS does run as a live continuous-crawl pipeline with model inference deployed on a Raspberry Pi 5, but that production system is not what this repository contains. What is published here is the research subset: the methodology, model architecture, feature construction, training code, and a reference implementation of the active-discovery policy against a synthetic environment.

What is deliberately not published, even though it exists and works:

• The production orchestration: live HTTP client, async worker loops, rate-limit tuning, proxy configuration, database schema and writes, Docker deployment scaffolding, health/restart logic.
• Trained model weights and ONNX artifacts.
• Raw data, account names, and per-account scores.
• Account inference tooling.

Nothing in this repository will, by itself, score named players or crawl the live Hiscores.

Two framing notes for readers and downstream users:

• The model produces a statistical risk score, not an accusation about any individual account. The numbers reported in the technical report are aggregated.
• This work is complementary to detection systems Jagex operates internally — Jagex has access to client telemetry, payment metadata, IP history, and behavioral signals at a scale that no external observer can approach. PanoptOS demonstrates only that some signal is recoverable from public Hiscores alone.

What's here

panoptos/
  discovery/        # Active-discovery policy + offline simulation
    policy.py       #   Probability matching over Bayesian utility estimates
    simulation.py   #   Synthetic environment for offline evaluation
  training/         # Model, features, and training pipeline
    constants.py    #   Skill/activity inventory + character encoding
    features.py     #   Per-snapshot feature construction
    model.py        #   Dual-branch transformer + name encoder + Platt head
    dataset.py      #   Streaming Parquet IterableDataset
    train.py        #   Training loop with online Welford scaler + Platt fit
docs/
  report.pdf        # Technical report
  informal.md       # Informal adaptation

Quick start — discovery simulation

The active-discovery policy can be exercised against a synthetic environment with no external dependencies and no network access:

python -m panoptos.discovery.simulation

This runs the policy against a small toy population of tables with varying bot rates and saturation, and prints the per-table pull allocation that the bandit converges to.

Citation

If you build on this work, please cite the technical report (see docs/).