Protection Protocols

Three deterministic protection protocols with no AI judgement in the loop — pile-on detection, absent-person protection, and pattern-break detection.

What is it?

Protection Protocols (COS-003) are three deterministic safety systems that protect community members from specific patterns of harm. Each protocol operates as a state machine with defined rules — when these fire, it's because the rules said so, not because a language model had a feeling about it.

The three protocols cover pile-on behaviour, absent-person discussion, and behavioural pattern breaks that precede coordinated attacks or community disruption.

Why does it exist?

Some of the most damaging community dynamics are predictable in their structure, even when they're unpredictable in their specifics. A pile-on follows a recognisable pattern. Talking about someone who isn't present follows a recognisable pattern. The behavioural shift that precedes a coordinated attack follows a recognisable pattern.

These patterns are too important to leave to AI judgement. A language model might get it right most of the time, but "most of the time" is not acceptable when the consequence of getting it wrong is a real person being harmed. Protection Protocols use deterministic rules precisely because the stakes require certainty.

Why is it different?

Most moderation systems rely on AI classification to detect harmful behaviour. This means they're probabilistic — they might catch a pile-on, or they might not, depending on how the language model interprets the conversation.

Protection Protocols are deterministic. PP1's 10-stage escalation ladder fires based on defined thresholds, not AI interpretation. PP2's absent-person detection uses structural analysis, not sentiment guessing. PP3's pattern-break detection runs as a state machine with explicit transition rules. The decision to intervene is never left to a model's judgement.

How does it work?

PP1 — Pile-On Detection: A 10-stage escalation ladder that detects when multiple users converge on a single target. The system tracks message volume, participant count, sentiment direction, and temporal clustering. Each escalation stage has defined thresholds and defined responses, from quiet monitoring through to active intervention. The ladder is a deterministic state machine — the same inputs always produce the same escalation level.
PP2 — Absent-Person Protection: Detects when a community discusses someone who is not present and cannot defend themselves. The system identifies references to specific individuals, checks whether those individuals are active participants in the conversation, and triggers protective intervention when someone is being discussed without the ability to respond.
PP3 — Pattern-Break Detection: A deterministic state machine that identifies behavioural shifts preceding coordinated attacks or community disruption. It tracks baseline behavioural patterns for the community and individual members, then flags statistically significant deviations that match known pre-attack signatures — sudden topic shifts, unusual posting patterns, coordinated messaging, or structural changes in conversation flow.

Cross-Platform Enforcement

Protection Protocols are enforced identically across Telegram, Discord, and WhatsApp. All three platform adapters wire through the same fixed-order safety pipeline: scam detection, safeguarding, community standards, job-posting checks, NSFW/advances, community health, then deterministic safety. Edit handlers on all platforms run the full pipeline — closing the edit-evasion attack vector where a user posts a benign message then edits it to contain harmful content after initial checks have passed.

Every detection is metered to Prometheus via recordSafetyTrigger and recordModeration calls, ensuring no platform and no detection type is metrically invisible. Operators can query safety detection rates per-module and per-platform from Operational Assurance dashboards.

Example

Scenario: A discussion about a proposed rule change becomes increasingly hostile. Seven members begin directing criticism at the member who proposed the change. PP1 pile-on detection tracks the escalating message volume, participant convergence on a single target, and declining sentiment. At stage 5 of the 10-stage escalation ladder, Warburton intervenes — redirecting the conversation toward the substance of the proposal. The intervention, escalation level, and full decision chain are recorded in the decision provenance trail.

Why should anyone care?

Pile-ons can destroy community members. Absent-person discussions can damage reputations without recourse. Coordinated attacks can shatter communities that took years to build. These aren't edge cases — they're common failure modes in every online community of meaningful size.

Protection Protocols provide deterministic defence against these specific patterns. No AI deciding whether to intervene. No probabilistic guesswork about whether a pile-on is "really" a pile-on. The rules fire, every time, without exception.