# Critique log - course alignment (2026-08-04)
Two instruments, sets taken **independently and in full before any merge**: Claude with a browser,
and Codex read-only through the plugin (`write: false`, verified in the job record). The point of the
second instrument is not more eyes - it is that a reviewer who did not produce the material does not
reproduce its blind spots as agreement.
**Radius.** Codex owns what is falsifiable in the source: contradiction between files, orphan without
a parent, promise with no executor, rule violation against `CLAUDE.md`, number without a source.
Claude owns what needs a browser: breaks at 360, overflow, whether a rendered page still says what
its md says. Asking Codex about layout would return a confident invention.
**Verification before every fix.** A finding is re-read in the file before anything is changed; one
that does not hold up stays in this log marked **withdrawn**, with the reason. Deleting it silently
is how the same wrong finding comes back next time in the same words.
---
## Mechanical checks (the instrument, not an opinion)
Run first, because a class that a script catches should never be walked by hand.
| Check | What it asks the OUTPUT | Result |
|---|---|---|
| em dash in output files | the **sign** `—`, not "is the dash correct" | **3 681 found** -> 3 621 replaced, 60 kept |
| markdown surviving in a render | the sign `](` in rendered html | 0 |
| broken relative links | resolve every `href`/`src` from its own directory | 3 345 checked, 0 caused by this work |
| registry idle control | a `page:` row whose file does not exist | 2 -> both **withdrawn**, see D6 |
| section idle control | a declared `NAV_SECTIONS` id with no anchor on the page | 218 declared, 0 orphaned |
| documented token drift | every value in `DESIGN.md` against `kit.css` | 20 checked, 0 drift |
The em dash check earned its keep twice over: it asked for the **sign**, and the sign was in 67 files
nobody would have opened looking for it. The zone breakdown is the finding, not the total - **the
product surfaces were already clean** (8 in `wireframes/*.html`, 9 in `design/*.html`) and the rule
had simply never reached the documentation (1 261 in `ia/*.html`, 961 in `voice/docs`, 603 in
`ia/docs`, 406 in `wireframes/docs`).
**Named exception, with idle control.** A lone em dash inside a table cell (`| — | `, `| — |`) is
the "no value" mark, not prose - a bare `-` there reads as a hyphen. 52 spans matched the exception,
so the exception covers real cases and is not a decorative line in a rules file.
> **SUPERSEDED at step 8.2, and the reasoning above is not withdrawn - it is answered.** The
> exemption weighed two characters, `—` and `-`, and picked the one that does not read as a minus in
> a column of numbers. That is correct and it is exactly why a third character ends the exemption
> instead of documenting it: `–` U+2013 is not the hyphen, is not banned by `CLAUDE.md`, and already
> lives in the product as the range mark (`А–Я`). All **60** exempted spans became `–` - 34 in
> ``, 26 in `| … |` - which is the same 60 this log counted, arrived at from the other side.
> The five em dashes left in the repository are inside backticks in this file and in
> `design/kit/docs/consolidation.md`, where the character is the subject rather than the text.
>
> **And this exemption was nearly overwritten without being read.** The owner was asked which sign
> an empty cell should carry, answered «я не знаю, дай решение», and the answer was decided from
> first principles - the same two candidates, the same reasoning, the same conclusion about the
> minus. The exemption was found only because the sweep reached the md files and the search results
> printed this paragraph. A rule that lives in a critique log and not in `CLAUDE.md` is a rule the
> next pass re-derives; the dash rule is in `CLAUDE.md` now, in full, and this entry is what it
> replaced.
---
## Findings
Full form: class -> found -> **who found** -> fixed -> withdrawn -> deferred.
| # | Class | Finding | Who found | Status |
|---|---|---|---|---|
| **D1** | rule violation | The "no em dash in output files" rule had never been applied to documentation: 3 681 occurrences in 71 files, while the product surfaces were already clean. | Claude (script) | **fixed** - 3 621 replaced, 60 exempted by a named rule |
| **D2** | contradiction | The inventory published **50 components / 14 organisms** on three surfaces; counting the rows gives **51 / 15**. | Claude | **fixed** in `inventory.md`, `kit.html`, `overview.html` |
| **D3** | number without a source | "**12** components exist only as render functions" - counting gives **9** only-JS, plus 2 that are also in markup, so 11 come out of that file. The number had been estimated, not counted. | Claude | **fixed** on all three surfaces |
| **D4** | contradiction, inflated | "the footer stands on **all 142 screens**" - `wfFooter` is called on **129**; dialogs, system pages and the hub carry no page chrome. The claim was the right shape and the wrong number, which is the hardest kind to notice. | Claude | **fixed**; also `142 screens` -> `142 files (141 screens + the hub)`, because one of the files **is** the hub |
| **D5** | orphan / promise | 18 IA node pages carry 3-14 `` each and declared **no** `NAV_SECTIONS`, so the panel could not show them. The rule "under the active page - its sections" was unenforced on the largest group of pages in the project. | Claude | **fixed** - sections wired from each page's own ` |