The Filter AI / Labs
Labs.
Two tools that make an abstract argument concrete. Both run entirely in your browser — no request is made, nothing is stored, and there is no key to enter. View the source if you would rather check than trust.
Payload inspector.
Paste any text and it reports what a human reader cannot see: invisible and tag characters, mixed-script words, encoded runs, forged role delimiters, instruction shapes, and outbound link targets. Every finding names the legitimate traffic that produces the same signal, because that is the part a real guardrail has to live with.
Read the findings the right way
A finding is a reason to look, not a verdict. Almost every signal here has an innocent source: developers paste encoded blobs, multilingual text mixes scripts, documentation quotes injection examples on purpose. Treating any single one of these as a block rule is how you generate the false alarm bill in section 06. Use the second tool to see what that bill actually looks like.
Base rate calculator.
The arithmetic that decides whether a guardrail is worth deploying, and the one almost nobody runs before tuning a threshold. Put in your traffic and your classifier's two numbers, and it tells you how many real people you block for every attack you catch.
Try the last preset. A classifier with a 99.9% detection rate and a 0.1% false positive rate would be state of the art, and at a one-in-a-thousand attack rate it still blocks roughly one real person for every attack it catches. That is not a bad model. That is what rarity does to any detector, and it is why the pattern library puts classification fifth.
What these are not.
- Not a guardrail. The inspector is a teaching aid built from readable rules. It has no model behind it, it is trivially evaded by anyone who reads its source, and it should never sit in a production path.
- Not a scoring service. Nothing you paste is transmitted, logged or retained. The page makes no network requests after it loads.
- Not a benchmark. The calculator does arithmetic on numbers you supply. If your classifier's real-world rates differ from its datasheet — and they will — the output differs too.
- Not a substitute for the practice. Tools show you the signal. Catching it under time pressure, mixed into ordinary traffic, is a separate skill.
The tools show the signal. The exercises build the reflex.
35 exercises across seven units, every one scored on both numbers — the attacks you missed and the real people you blocked. Unit one is free and runs in your browser.