How a file is judged

Split. adhere splits the file into sections of whole statements, in runs of at most 30 lines.
Judge. One request per file asks Jev, for each rule, how likely the file is to break it.
Locate. A second request, only for rules above their threshold, asks which section and lines break the rule, and whether Jev sees enough to decide.

A file too long for Jev’s context is skipped and counted in the summary.

The steps in detail

Split. adhere splits each file into sections of whole statements, such as its imports, constants, and functions, packed in order into runs of at most 30 lines. A class or function longer than that splits into its methods or the statements of its body. Jev reads the file as these sections.

Judge: one request per file.

The state is code, the file’s sections under their numbers from 1, as { code: { "1": "…", "2": "…" } }, without their comments (see Comments), beside what the config reads.
Each rule is one noul (yes/no probability) question that carries the rule’s description and details, and its code under its words, must and never, or a guideline’s should and should_not. It asks whether the file diverges from the pattern must shows, with never as an example of diverging. For a rule with only code that must never be written, it asks whether the file contains that code.
Criteria for each answer draw the line at the rule’s scope, so a file with no code the rule is about is a no.
No rule sits in the state, so a rule’s probability depends only on the state and that rule, not on which other rules share the request. Every question shares the state cost of the request.
On up to 10 files per project rule, the rule’s question has a second one beside it, the linter check (see Validating rules), with the same fields.
A rule with appendState has requests of its own, here and below, whose state adds to this one.

Locate: a second request, only when at least one rule’s probability is above its threshold. Per flagged rule:

When the file has more than one section, adhere asks one choice question among the sections’ numbers, which yields the section to report. The options carry no text of their own, since each is a key of the state. On the eval, options that described each section, by its first line or by what it declares, located violations no better than options without text.
For each section, adhere asks two choice questions among its lines, each option the line’s text: the line the violation starts on and the one it ends on. The report underlines the lines chosen in the reported section. On the eval, lines asked for in this request were named as accurately as lines asked for in a request of their own, and they held a violation for 94% of real findings, most often as one line (study).
adhere asks one noul carrying the rule: whether you can tell if code breaks it from the state alone, without knowing what other files, libraries, services, or configuration do beyond what it shows. It yields the context score and its warning.

Limits. Jev reads at most 32k tokens of state and one question together, and 64k tokens in a request. adhere estimates tokens from the JSON it sends, at about three bytes a token, and leaves 1k of room for what Jev wraps around it.

When a file’s questions would not fit in one request, they are split across several.
If a file’s state and longest question would not fit together, or a provider reports a context overflow, adhere retries once without references when that map is present. Target code, symbols, and other context remain intact; no declaration excerpt is shortened to fit. Removing references may still leave the request too large, in which case the file is skipped and counted in the summary. A file with more than 255 code sections is also skipped, and the rest of the run goes on.
A file is also skipped when Jev itself counts it as over its context. Text denser than the estimate, such as CJK, can cause that.