From source bytes to a gate decision.

Extraction picks the prose

Each format has a grammar, and the grammar says which regions hold English. Markdown contributes headings, paragraphs, list items and GFM table cells. A source file contributes its comments and its string literals. Plain text contributes paragraphs.

Code is never checked. Comments and strings are.

Inside Markdown, inline code, fenced and indented code, raw HTML, image syntax, link destinations and front matter are protected. Link labels stay prose. Quoted blocks are left out unless analysis.include_quotes turns them on.

Inside source files, several regions are recorded as deliberate exclusions rather than checked. A shebang or a tool directive comment such as nolint or eslint-disable is a directive. An indented example inside a comment is comment-code. The tag and example sections of a JSDoc or Javadoc comment are doc-tag and doc-example. A shell literal that holds a program for another interpreter is an embedded-program. A Python byte literal is a byte-literal. An expression inside a string is an interpolation, and it inserts a protected boundary so the words on either side cannot form a false phrase.

Every excluded region keeps its byte span and its reason, and a saved report carries that list. Import paths, re-export module paths and Go struct tags are excluded as technical metadata; strings in exported declarations stay checked.

One language check runs per block. A block with at least 20 letters, more than half of them outside the Latin script, is recorded as a non-latin-prose exclusion and the rest of the document is still checked. The block keeps its position, so block numbering and finding fingerprints do not move.

Units

Extraction yields blocks. The NLP backend splits blocks into sentences and tokens, tags parts of speech and derives chunks. The default backend is pure Go, English only, and embedded in the binary; no model is downloaded at run time.

Two unit kinds carry a score: the sentence and the paragraph. Rules declare which one they measure through scope, and a document-scope rule such as repetition.near-sentence compares candidates across the whole file.

Sentence

Phrase matches and single-sentence structure: announced importance, chat preambles, long sentences, noun stacks.

Paragraph

Density and length: connective overuse, hedges, paragraph length, reading grade, em dash density.

Document

Repetition across blocks: exact and near-duplicate sentences, repeated openers, heading and summary echo, POS templates.

Rules and activation

A rule is a versioned descriptor plus an implementation. The descriptor names the contexts it applies to, the NLP capabilities it needs, its correlation group, its default severity and gate, its weight and cap, and the limits its author recorded. The rules catalog lists all 40 as the engine reports them.

Evaluation produces evidence with original source locations and an activation in integer thousandths. An exact phrase match activates fully. A density or length rule ramps linearly from its onset to its saturation, so a text just past the threshold does not score like a text far past it.

A rule that exhausts its own analysis.max_candidates budget on a document abstains with the reason budget_exhausted. Its partial findings are dropped, the other rules keep theirs, and the report names the abstention. The run stays complete.

Scores

For each sentence and paragraph the engine multiplies activation by weight. It sorts the contributions by descending raw value, and stable identifiers break ties. Within one correlation group, identical or nested evidence contributes only its strongest signal. Rule caps and group caps then limit the rest, and the total stops at 100.

The trace records the raw and effective points of every contribution, and the reason for each reduction. Sentence and paragraph assessments stay independent: a sentence total never joins its paragraph again. A clean paragraph appended to a document cannot lower the index of an existing one.

Document statistics report the raw paragraph maximum, the nearest-rank median and P90, and the eligible flagged fraction.

The index is an editorial-policy index from 0 to 100, not a percentage and not a probability. Without a calibration pack, slop_probability is null with the status calibration_unavailable.

The gate

Three things can fail a run. A rule set to gate: forbid fails whenever it matches. A local score gate fails a unit whose effective score reaches the configured threshold and which has at least min_words words. An operational error or an incomplete analysis fails regardless of policy, and --no-gate cannot pass either.

Gate decisions use effective values, which means the values that remain once source permissions apply. Document summary statistics stay raw. Derived gate diagnostics arrive after suppressions, no permission covers them, and they never feed back into scoring.

What the alpha does not do

Thresholds are initial editorial policy, not measured precision. Every builtin rule is marked experimental, and no editorial model has met the criteria for use in the product. POS-based chunks are surface candidates, not grammatical dependencies. Lexical overlap is not a claim of semantic equivalence. An exact phrase match establishes a configured condition; it does not prove that an editor would reject the sentence.