Nyaya · a scam filter you can argue with · project site
This filter learned 47 rules from 5,574 real text messages in 2.13 seconds on a CPU — no GPU, no network, no weights. Below is all of it. Read a belief, disagree with it, switch it off, and watch the verdict change. That is the whole product claim, and it takes about ten seconds to check.
Editing this text rescores it immediately. The verdict and the list of rules that fired are announced as they change.
Forty-seven learned beliefs. Each is a toggle button: activate it to switch that belief off, and the verdict is recalculated and announced.
Scroll the ledger and three things are visible that no accuracy score would ever have told you. None of them is a design choice. They are what the learner actually found, and they are only findable because every belief is a sentence.
Every one of those is a limitation, and every one is legible in about a minute because the beliefs are written in words instead of stored as weights. That is the argument: a filter you can audit is one you can correct. The thing protecting a person should be a thing that person is allowed to read.
On the identical split, a textbook algorithm from the 1990s scores higher. I am publishing that here rather than leaving it for you to find in five minutes, because it is the actual result and it is more interesting than a win would be.
| method | precision | recall | F1 | fit | what you can read |
|---|---|---|---|---|---|
| nyaya readable rules | 100.0% | 77.8% | 0.875 | 1.85 s | 47 sentences |
| naive Bayes | 98.7% | 94.4% | 0.965 | 0.06 s | 7,930 weights |
| hand-written keyword list | 63.4% | 71.6% | 0.672 | 0.00 s | 14 words |
| majority class | 0.0% | 0.0% | 0.000 | 0.00 s | 1 constant |
Readability costs about nine points of F1 and sixteen points of recall. That price had never been published before. It is a real cost, paid for a real thing: 7,930 naive-Bayes weights cannot be opened, understood, argued with or edited by the person they are protecting. Forty-seven sentences can — you just did it. Closing that recall gap while keeping the rules readable is an open research question here, not a claim already won.
No API key, no GPU, no dependencies beyond the Python standard library. The page you are reading uses the exact rules and scoring the repository produces.