Add raw repeat spam check with threshold 6

Batch EN detox inference and safer gibberish early-exit

Refine blacklist matching and strengthen spam heuristics

Improve spam heuristics and reduce short-token false positives

Use /analyze-batch for bulk UI requests

Add early-exit and Detoxify call counters to moderation logs

Handle quoted spaced-letter obfuscation in normalization

Remove regex spam patterns from spam detection

fix: update API_URL to use environment variable for flexibility

Normalize spaced-letter obfuscation in text cleaning

fix(ui): map INSULT/HIGH verdicts correctly

fix: preserve natural double letters in normalization

Use remote Detoxify loading in sentinel-api

Delete models

Create models

Prefer local Detoxify checkpoints with download fallback

Add startup diagnostics for model load and int8

Collect garbage after quantization to reduce peak RAM

Add int8 quantization and batch moderation endpoint

Revert float16 casts for faster CPU inference

Use float16 model loading with safe fp32 fallback

Add system-status endpoint for RAM/CPU monitoring

Optimize moderation fast paths for blacklist and Detoxify

Fix verdict logic to include insult hits

Add offensive model files for HF runtime

Fix HF metadata emoji

Add HF Space metadata

Deploy Sentinel API to HF Space

initial commit