The proof

The pixel comparison, and why it has no photographs

The obvious form of this page is a wall of before-and-after screenshots, and this project decided against exactly that on 2026-08-11. 120 state snapshots came out of this kit that day, framed as measurements and honest about it, on one argument: a value moves and every picture of it is quietly stale until somebody re-shoots it, and a stale picture is worse than no picture, because it looks checked. A page of proof built from the one artefact the project has already rejected would be the least trustworthy page in the folder. So the comparison here is the thing a screenshot never carried: the element count, the shape of every difference, and the named row that explains it.

The raw material is on disk and re-runnable: screens/pixel-before.json, screens/pixel-after.json, screens/pixel-brand.json, screens/ledger.txt and screens/brand-ledger.txt, each beside the script that produced it.

Round one, step 8: the sample moves onto the system

The proportions are the story. The stage that was not allowed to change the look moved 4074 values, and every one of them is two scales the founder adopted by one rule each: nearest step, ties to the larger, applied mechanically. Nobody decided value by value, which is what makes the result auditable rather than arguable.

Round two, 2026-08-12: the brand

The chips moved because margin-left: auto in chip.css absorbs the slot's 25.235px, and for no other reason. The rows were re-aligned by an LCS on (tag, text) rather than by document order, because the lockup inserts three elements per screen and a fingerprint keyed by position cannot survive an insertion.

Round three, 2026-08-12: the rail from 240 to 220

The two height changes are on subscription-detail-unrecognized, where the line "All we have is how it appears on your statement" stopped wrapping. A block 21px shorter and the statement descriptor 21px higher. That is what the width was for.

Nothing on this page is unpaired. All 28 coloured screens were compared in all three rounds. There is no screen without a comparison and therefore no row here reading "unverified".

The critique, by class, and who found what

Two instruments, taken independently and merged afterwards, which is this project's rule rather than a flourish. Codex read the source in read-only and owned everything falsifiable in a file; this pass owned what only a browser can answer. The column that matters is the last one.

The overlap between the two instruments was ZERO, and that is the finding this table exists to carry. Codex returned 24 rows and this pass returned 4, and not one of them was the same row. A browser pass cannot see a contradiction between two files; a source pass cannot see a computed value. The rule that says "two instruments" is not redundancy, and the proof is that the single most serious finding of the step, a violated freeze, was invisible to the instrument that runs every session: the standing check looks for UNCOMMITTED changes in wireframes/, and this change had been committed three steps earlier.

What was dropped, and why that number is healthy. Of 45 computed-style disagreements this pass raised, 44 did not survive re-reading: 25 were min-height: auto against 0px, which is whether an element is a flex item and not a paint; two caught the skeleton mid-pulse; the rest compared a host-scoped class against a different host. Seven page-format warnings matched the word :hover INSIDE a comment that says no hover is declared. An instrument that returns the same answer at every input is not measuring the input.

The token audit, was and became

--color-focus, the token an entire keyboard depends on, measured in both themes rather than assumed: 6.23:1 on surface and 5.56:1 on canvas in the light theme, 6.82:1 and 7.51:1 in the dark. The non-text threshold is 3:1. Had the pair been left for later, the dark ring would have been the light petrol at 1.9:1: present in the file, invisible on the screen.

Consolidation, in numbers

The counting rule has a blind spot and it is named rather than hidden. This inventory reads what the product already draws, so it can only ever find what has been drawn. A brand, a splash screen, a first-run tour and an error page that nobody wireframed are invisible to it by construction. The brand is the proof: it occurs zero times in the grey corpus and no amount of reading it would ever have produced the mark.

The control census, both measurements

This reads as the opposite of the pixel comparison, and the pair is the point. The comparison says nothing changed by accident. The census says the set of forms was closed on purpose. One is a promise about the past and the other is a promise about the future, and neither is worth much without the other.

Six forms were tracked across three corpora until they came clean. Re-counted at step 9 and the answer changed: two rows still credited the frozen stand with 2 inverse and 1 .is-disabled that had been deleted underneath them. A census is a measurement and a measurement has a date; those two carried the number and not the date.

The dark theme, and the three instruments that caught it

The theme is not one check, and the three numbers below were produced by three different instruments looking at three different things. That is why they are reported separately rather than summed.

The two collisions are open and they are here rather than in a backlog file, because this is the page that claims the theme works. Three roles that a person can tell apart in daylight and cannot tell apart at night are a defect the light theme can never show you.

The dry run: the rollout that has not happened yet

The rollout moved to stage 12, and the main test of the system's completeness moved with it. This is the cheap replacement: take the densest screens the system has never been shown, and check on paper whether it could build them. Nothing was built.

Four of the five need nothing. The system covers the densest screens it has never seen, which is the answer this test was built to give and the good one. All 27 still-grey pages were ranked by distinct class tokens and these are the top five.

The whole gap is one page, and the re-run corrected two things this page claimed. It is 36 uncovered, not 37: .stage-flush is a modifier of the reviewer chrome class .stage and counts as chrome. And not all of them are the lp- namespace: 24 of the 36 carry the prefix and 12 do not, so retiring the lp- twins closes two thirds of the gap rather than all of it. All 36 are landing-only, so "the gap is one page" survives. It is not undeclared: the inventory names it twice and ends with the sentence "the kit takes the two-host pattern as the rule and retires the lp- twins". That retirement has not happened, and the number on the other side of it is 48 of 57 component files carrying a .landing selector that the COLOURED corpus never wears (0 of 29). The grey landing does wear it, at wireframes/index.html:14, and that page loads _wf.css rather than the system. Half of forty-eight files is written for a page that does not exist.

The accessibility half, measured on the screens

28 coloured screens, both accepted viewports, both themes. The eleven ARIA findings raised at the molecule round are not counted here: they were moved to stage 12 by decision, and they are in docs/backlog.md.

Stage 09, the whole of it: three patterns, one screen the system had never seen, two instruments

This section was added on 2026-08-13 and it is the second round of proof on this page, not a second edition of the first. Stage 08 proved that a sample of screens could be moved onto the system without moving a pixel. Stage 09 asked two harder questions and both have a number: can a composition be lifted out of the components without the screens noticing, and can the system build a screen nobody built for it.

Was and became, step 1: eight coloured pages move onto three patterns

Re-runnable, beside the script that made it: screens/ledger09.cjs, screens/ledger09.txt, screens/pixel09-before.json, screens/pixel09-after.json.

Step 5: the screen the system had never seen

The self-sufficiency test has no before, so it cannot have a pair, and the number that replaces the pair is what was NOT added. Alerts and its three states were built on 2026-08-13 out of design/system/ alone: 44 distinct system classes across four pages, 0 new CSS files, 0 new tokens, 0 new component variants, 0 new lines of microcopy, 0 style attributes and 0 page-level style blocks. Structure and copy are character-identical to the grey originals, which Voice owns and the wireframes freeze.

Step 6: the critique, by class, and who found what

Two instruments again, taken independently and merged afterwards. Codex read the source read-only and owned everything falsifiable in a file; a browser and scripted pass owned what only a running page can answer. Twelve findings held, two were dropped at verification, and the eleven that look like eleven defects are one defect with eleven addresses.

Two findings were dropped at verification, and the reason is on the record so neither returns. range-picker.css:114 writes font-weight: 600 inside a disabled-and-pressed rule, which reads as a state written with a value: it is not, because this system has no weight scale at all. Weight is a literal in every one of the 57 files, by the rule that colour goes through a role and geometry through a primitive and weight is neither, so tokenising one occurrence would have opened a third level to serve one line. The detector's em-dash-overuse, raised on 34 files, counts -- in running text, and the running text of this project is made of custom property names: the colour page holds 378 double hyphens and zero em dashes.

The overlap between the instruments was four rows out of twelve this time, against zero at stage 08, and the change is not noise. At stage 08 Codex read files and the browser read pixels, so they could not meet. At stage 09 both were reading the same thing, a set of claims in prose, and prose is falsifiable from either side. Where they met they agreed; the eight rows they did not share are still split the old way, with the source findings on one side and the running page on the other.

The numbers this stage leaves behind

Stage 10, Responsive: the promise is asymmetric, and both halves are measured

Added 2026-08-13, and it is the first round of proof on this page whose promise is not "nothing moved". Stages 08 and 09 owed zero difference at both viewports. A stage about width cannot owe that: at 360 the mobile layout is the base of a mobile-first product and it must not move at all, and on a wide screen the whole point is that things move. So the promise splits, and the second half is enforced differently: every desktop difference must be explained by a named row, and an unexplained one is exactly as much a defect as any difference at 360.

The narrow end: 32 pages, zero difference, box and ink

It failed once, and that is why it is worth running. The first pass lit up on six pages. Two of the four reading-measure fixes had been written as a comment appended after an already-closed comment block, which left prose in the stylesheet and a stray */ after it; the browser dropped what followed. The 360 check caught a broken stylesheet that the eye read as a normal page. A comment-balance check now runs over all 65 stylesheets and is clean.

The wide end: 32 pages changed, and every change has a row

The instrument: a sweep, because a defect lives between the points

Two instruments, and what each one was blind to