Modern AI deployments need guardrails that are both fast and deep. A single-layer classifier is too brittle; a full LLM judge on every request is too slow.
TG-Sentinel splits evaluation into three tiers:
- L1 Precedent Decision Cache — deterministic HMAC hash and high-confidence vector matching for ultra-fast fast-path denial of known violations without upstream LLM calls.
- TG-SENTINEL:FAST — lightweight classifier (4B) for high-volume, low-latency decisions — benchmark probes measured sub-25ms time-to-first-token (p50 21ms, p99 23ms) on our deployment.
- TG-SENTINEL:MAX OR ULTRA — deep LLM reasoning (9B/27B) and TG-Sentinel:G1 (46.24B MoE) for high-stakes adjudication, edge cases, and asynchronous precedent verification.
This layered design lets operators tune the latency-vs-certainty trade-off for each workload while keeping a single audit trail across every decision.