AI-Sentinel

Layered Defense Pipeline · Live Fire
0
Intercepted
0
Sanitized
0
Delivered
L8 · Token Optimizer · non-security · runs on ingress

The layer that
pays for itself.

semantic cache — repeat prompts answered without touching the LLM
complexity router — easy prompts sent to cheaper models
prompt compression — every request trimmed before it leaves
before
$100K/mo
after
$62K/mo
−38%30–50% typical
same traffic, same answers — cache hits and routing do the saving. Never blocks: any optimizer error degrades to a plain pass.
AI Oversight · above the pipeline, never inside it

Four tiers
of judgment.

T0
The deterministic pipeline — the spine. The only tier that ever blocks. Byte-identical with every judge switched off.
T1
Inline classifier — wakes only on a gray-zone verdict. Advisory, never blocks, fails safe to benign.
T2
Async judge — samples redacted traffic off the hot path and turns suspicions into rule evidence.
T3
Frontier overseer — rare, budgeted, writes human-readable reports. Advisory by constitution.
The thesis: AI is an untrusted evidence generator. A deterministic validator and a shadow canary decide what survives — and turning a rule on still takes a human signature. The judges make the dome smarter; only the dome decides.
Full throughput · the dome at line rate
0 requests / second
P99 added latency · 20 ms · every one of them runs every layer
Audit Chain · L7Attested
    Patrol

    The dome is up.

    drag take control · click a layer to inspect
    1–8 run a scenario · esc resume patrol
    Presenter Control · esc resumes patrol

    AI-Sentinel · Inspection

    Real-world precedent · what this costs when it lands
    AI-Sentinel
    Arming the gauntlet