Thelastpoet/Sentinel

Election-safety moderation API for detecting incitement, hate speech & election harms in Kenya’s multilingual, code-switched contexts. # Sentinel Open-source election-safety moderation API built to help protect Kenya's 2027 general election from ethnic incitement and election disinformation. Sentinel currently supports code-switched moderation across English (`en`), Swahili (`sw`), and Sheng (`sh`) in the live moderation hot path. Language-pack artifacts for additional languages (including Luo and Kalenjin) are included for operator evaluation and staged rollout, but are not active in default hot-path enforcement. Sentinel returns deterministic moderation decisions (`ALLOW`, `REVIEW`, or `BLOCK`) with full audit evidence, so every action can be explained and appealed. ## Who is this for? **Platform integrators** — You run a forum, news platform, or civic tech tool and need a moderation API. You send text, Sentinel returns a decision. Start with the Integration Guide. **Self-host operators** — You want to deploy and manage a Sentinel instance for your organization. Start with the Deployment Guide. Both audiences should begin with the Quickstart. ## Capability Snapshot - **Available now (default runtime)**: moderation routing and enforcement for `en`, `sw`, `sh` - **Included but not default-active**: additional language-pack artifacts (e.g., Luo, Kalenjin) - **Contract stability**: `POST /v1/moderate` is the stable integration surface; operator endpoints (`/admin/*`, `/internal/*`) can evolve ## What Sentinel returns ```jsonc { "toxicity": 0.92, "labels": ["INCITEMENT_VIOLENCE"], "action": "BLOCK", "reason_codes": ["R_INCITE_CALL_TO_HARM"], "evidence": [ { "type": "lexicon", "match": "kill", "severity": 3, "lang": "en" } ], "language_spans": [ {"start": 0, "end": 26, "lang": "en"} ], "model_version": "sentinel-multi-v2", "lexicon_version": "hatelex-v2.1", "pack_versions": {"en": "pack-en-0.1", "sw": "pack-sw-0.1", "sh": "pack-sh-0.1"}, "policy_version": "policy-2026.11", "latency_ms": 12 } ``` Every response includes the evidence that drove the decision, the versions of all artifacts invo …