Tiers

Every detector sits in a tier. A tier decides when a detector runs and whether a policy may switch it off.

A tier is not a cost

This is the part that is easy to get wrong. Every model-backed detector in the set is the same XLM-RoBERTa base, so at the same input length they all cost the same. They share one budget rather than splitting into a fast tier and a slow one.

gibberish is T1 because unreadable input should short circuit the tiers above it rather than be scored by them, not because it is cheaper. It is not.

What a full scan costs

Cost is per token and close to linear within a window, about 1.66 ms per token on one thread, and it steps at each window boundary. A window holds 94 content tokens, so 94 tokens is one forward pass and 95 is two: the step is about 33 ms for one more token. A document of n tokens costs roughly ceil(n / 94) passes, which is proportional rather than catastrophic, and that is the property worth relying on.

A full output-side scan with every detector wired would be one rule check plus seven encoder passes, roughly 360 ms at the reference input.

Keeping that off the common path is the entire job of the tier system, and it is a scheduling property rather than a cost one.

This said "about 0.6 ms per token" until 2026-08-14. That rate belonged to piiguard's published 266 MB INT8 export, withdrawn on 2026-08-12 for losing an entity entirely on 13 of 120 texts, and it also averaged a slope across the window boundary, so it was wrong twice in the same direction. See benchmarks/latency_sweep.py, which records the sweep and refuses to run on a busy machine.

Threads

Measured at the 87-token reference input: 157.32 ms on one thread, 79.50 on two, 42.99 on four, 25.63 on eight. The default stays at one. A library that quietly commandeers the machine it is embedded in is worse than one that is honestly slower, and a policy can raise it deliberately.

These read 54.7, 29.8, 17.8 and 12.4 at 96 tokens until 2026-08-14, on the same withdrawn export. 96 tokens is also two windows, so that table measured parallelism across two passes, one of them 19 tokens long, which is not the shape anyone wanted to know about.