Sentinel-VC: Capability-Aware Event-Driven Moderation for Telegram Communities and Voice-Call State

Telegram community moderators face an observation gap between application events, voice-call state, and media transport. We present Sentinel-VC, an event-driven Node.js reference framework combining tenant-scoped token buckets, decaying exceedance evidence, collective call-join monitoring, and permission-aware mitigation. A bot-token deployment observes community actions and applies temporary chat restrictions with user-bound verification. An optional consenting user-admin MTProto adapter observes delivered user-identity call transitions in supergroups and broadcast channels and requests participant mute or guarded speaking admission. Neither deployment measures raw UDP generation or trustworthy account creation dates. The retained individual-detector evaluation comprises 24,000 synthetic episodes per policy across eight workloads and 30 seeds. Rapid-message and rapid-call-churn episodes are flagged, with virtual median onset latencies of 0.993 and 1.208 seconds. Slow churn and invisible media abuse are missed; an overlapping legitimate workload is fully flagged. Across the balanced mixture, precision is 66.7%, recall 50.0%, and false-positive rate 25.0%. Six separate deterministic follow-up fixtures show that the collective extension flags distinct-identity join bursts missed by the individual detector, but flags an identical legitimate audience surge as well. Preemptive join-muted admission and optional call-invitation-hash rotation provide an explicitly authorized response path. A fluid-queue illustration distinguishes cooperative admission effects from unresponsive traffic. These results establish reproducible policy behavior and limitations, not live audio recovery, crash prevention, UDP filtering, or production availability. Publication status: This is a preliminary research preprint that has not undergone peer review. Operational Telegram evaluation remains pending. Code and reproducibility artifacts: https://github.com/thecmjhb/Sentinel-VC Preparation disclosure: The implementation and manuscript were developed with substantial assistance from OpenAI Codex, including code generation and revision, test development, research-draft composition, and execution of the local evaluation scripts. Numerical artifacts were produced by the disclosed replay/model code rather than live Telegram measurements. The named human author is responsible for verifying references, methods, claims and results before submission; AI tools are not authors.