Imperative detection — the hand-classified oracle

Classified 2026-09-06 for WI-061, before any matcher existed, from the candidate lines imperative-corpus-2026-08-17.md collected. The grep is the corpus document’s own — \b(must|never|always|shall|required|mandatory|do not|don't)\b, case insensitive, over each repo’s AGENTS.md and CLAUDE.md — and the counts it produced today match the corpus document’s exactly, at the commits below. Every verdict here was written by reading the line, not by running anything. The detector is measured against this document, never the reverse.

Verdicts

Two verdicts and nothing in between. rule: the line instructs the agent (or constrains an artefact the agent produces) — imperative mood, or a modal stating what must or must never be done. not-a-rule: everything else — a heading whose qualifier says MANDATORY, a statement of fact about the project or its history, a description of a tool or a document, a quoted or explanatory sentence. A line is classified as a whole; a line that carries both a rule and narrative is a rule.

Repository	Commit	Candidates	rule	not-a-rule	not-a-rule share
`hexguard`	`51b25da`	10	9	1	10%
`hexguard-templates`	`d24cf0a`	10	10	0	0%
`ai-cli` (this repo)	`6acfc39`	24	19	5	21%
`rewind`	`71b0ef9`	31	25	6	19%
`rift-forge-candidate`	`b1f4dd4b7`	134	40	94	70%
`axiom-mesh`	`3e1508a`	0	0	0	—
`gridforge`	`2645fbe`	0	0	0	—

The corpus document’s impression was wrong for the one repository that dominates the count. It said “almost all of them are genuine imperatives”; that is true of the four small files and false of rift-forge-candidate, whose CLAUDE.md runs to 1,368 lines and whose second half is project history in which never and always are narrative (“Meraki has never served them”, “a seed was never a corpus reading”). A matcher tuned on the small files would have shipped at a 70% false positive rate on the largest one — the WI-042 shape again, seen before the engine exists this time.

Not-a-rule lines, by repository

Every candidate not listed here is a rule.

`hexguard`

Line	Why not a rule
`AGENTS.md:101`	design statement — “recording one must cost almost nothing” describes the register’s purpose

`ai-cli`

Line	Why not a rule
`AGENTS.md:56`	design statement (same sentence as hexguard’s)
`AGENTS.md:75`	fact — “they do not prove actor identity”
`CLAUDE.md:16`	heading qualifier — `## The evidence rule (MANDATORY)`
`CLAUDE.md:65`	heading qualifier — `## Editing files from the shell (MANDATORY)`
`CLAUDE.md:83`	explanatory prose — “the earlier steps never ran”

`rewind`

Line	Why not a rule
`AGENTS.md:42`	fact about setup — “Two directories are required to work here”
`AGENTS.md:52`	heading — `### 2.3 Apply the principles; do not re-ask what they answer`
`AGENTS.md:62`	fact inside a quotation — “they almost never do”
`AGENTS.md:199`	fact — “Codex and Claude do not share memory”
`CLAUDE.md:4`	description — “specifics that do not belong there”
`CLAUDE.md:12`	fact about setup — “are gitignored but required”

Kept as rules, and worth stating because they are the arguable ones: AGENTS.md:46 (a table cell that instructs — “do not fabricate one”), AGENTS.md:76 (a table cell — “Answer it. Never surface it”), CLAUDE.md:38 and CLAUDE.md:44 (table rows constraining what a command and a subagent may report — “structure only, never values”).

`rift-forge-candidate`

The rule lines, since they are the minority (40):

AGENTS.md:11, 13, 15, 18 · CLAUDE.md:22, 35, 38, 40, 42, 44, 45, 52, 53, 64, 71, 86, 100, 109, 112, 115, 127, 130, 132, 136, 145, 146, 153, 173, 194, 223, 248, 257, 286, 287, 288, 292, 322, 869, 880, 1162.

The 94 not-a-rule lines, by shape:

Shape	Lines	Example
Heading qualifier	`CLAUDE.md:20, 94, 171, 229`	`## Engineering principles (MANDATORY for all development)`
Description of a document or process	`AGENTS.md:8` · `CLAUDE.md:84, 92, 218`	“Build-internals are never mirrored”
Project history and design narrative	`CLAUDE.md:325` through `CLAUDE.md:1365`, every candidate not listed as a rule (85 lines)	“Meraki has never served them”; “a value that was always wrong”; “never a second table” (as history, `837`; the same words at `869` are a rule)
Dev command line	`CLAUDE.md:1368`	`pnpm build:web · pnpm build:server …`

What the classification says about the matcher

Read before the matcher, these are the shapes it has to separate, in order of weight:

Past-tense and narrative modals. “was never”, “had never”, “never served”, “never a X” (noun phrase), “is always reversible”. Every one in this corpus is history. Rules use never and always at the head of a clause or after must.
Headings. Any # line is out.
do not with a subject. “they do not”, “Codex and Claude do not”, “the tests do not” are facts; “Do not restate” and ”; do not lay out options” are rules. The subject pronoun or noun before do not is the separator.
required and mandatory as adjectives are almost always headings or setup facts here (“are required to work here”, “gitignored but required”). The rules that use them also carry a stronger modal on the same line, so dropping the two words loses nothing this oracle contains.
Design statements about the system (“summing the parts must reproduce the previously published figure”) are classified not-a-rule when they describe an invariant of the product rather than instruct the agent. A matcher cannot separate these from rules by form; they are the residual error to measure, not to engineer away.

What this cannot tell you

One operator’s repositories, one classifier’s reading. The four small files were read twice; the 134 lines of rift-forge-candidate once, with the narrative region judged by shape.
The verdicts are about lines, because the corpus was collected per line. A rule spanning two lines counts twice; a heading and the rule under it count separately.
Zero candidates in axiom-mesh and gridforge remain the two repositories any “instruction files always carry unenforced rules” claim is falsified by.