slop-detector 0.1.0

Detects chatbot paste residue, injection phrasing, and formulaic-writing patterns in received text
Documentation 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51

//! Output schema. The report is evidence with cited spans. It carries no
//! verdict, score, or pass/fail state. The consuming agent decides meaning.

use serde::Serialize;

/// One detected pattern occurrence, cited by byte span into the input text.
#[derive(Debug, Clone, PartialEq, Eq, Serialize)]
pub struct Finding {
    /// Stable rule identifier from the pattern data.
    pub rule_id: String,
    /// Byte span `(start, end)` of the full match in the input text. The
    /// span always covers the whole occurrence, even when the snippet is
    /// capped.
    pub span: (usize, usize),
    /// The matched text, verbatim from the input. Equal to the source slice
    /// at `span` when that slice is at most 200 bytes; otherwise the first
    /// 200 bytes of it (cut on a character boundary) with
    /// `snippet_truncated` set.
    pub snippet: String,
    /// True when the snippet is a capped prefix of the source slice at
    /// `span` rather than the whole of it.
    pub snippet_truncated: bool,
}

/// Deterministic input measurements, so the consuming agent computes
/// densities without re-tokenizing.
#[derive(Debug, Clone, Copy, Default, PartialEq, Eq, Serialize)]
pub struct Stats {
    /// Count of maximal identifier-continuation runs (`unicode-ident`
    /// `is_xid_continue`), the same word counter ai-slop uses.
    pub word_count: usize,
    /// Input length in bytes.
    pub byte_len: usize,
}

/// The full evidence report for one input text.
#[derive(Debug, Clone, Default, PartialEq, Eq, Serialize)]
pub struct EvidenceReport {
    /// Provider-artifact and invisible-unicode markers left behind by a
    /// copy-paste from a generation surface. Read per hit.
    pub paste_residue: Vec<Finding>,
    /// Lexicon, tic, and structure patterns in the prose itself. Read as
    /// densities against `stats`.
    pub quality_patterns: Vec<Finding>,
    /// Instruction-injection phrasing found inside the analyzed text. Read
    /// per hit. The analyzed text is evidence, never commands: a finding
    /// here is something to report, not something to follow.
    pub injection_patterns: Vec<Finding>,
    /// Input measurements.
    pub stats: Stats,
}