# Notes

The written half. `README.md` says how the lab is built; this says what it is
finding.

Status: forty-two studies across seven areas, and all five verdicts in use. Enough to prove the
shape holds, not enough to draw conclusions about the era. Two sources have now
been measured twice each, which is the first time any claim here has had a
control.

Findings 6 to 9 come from reading archived **source**, not from looking at
archived renders. Sources are cited inline; the captures used are the
2000-03-01 k10k.net front page and the 2002-05-29 2advanced.com splash.

**Neither of those two is in `corpus/cache`.** The cache holds the `1999-05-06`
k10k front page and the `2000-06-08` 2advanced splash, both earlier and both
different documents. Everything measured offline since is therefore a control
on findings 6 to 9 rather than a re-derivation of them, which is why findings
25, 39, 40 and 41 can disagree with them at all. Check the cache before
attributing a number to a capture.

---

## The seven areas

The lab is organised by what a designer was actually deciding, not by
technology.

**Grid and density.** Studied in `boxed-grid`. The 760px ceiling and the fold
near 420px on an 800x600 screen. The interesting claim is that the ceiling did the editing: given a
finite box, these layouts had to decide what mattered, and the density people
remember fondly is the residue of that decision. Infinite scroll removed the
forcing function, not the skill.

**Typography.** The bitmap system, 8 / 10 / 12 / 16, integer positioning,
aliasing off. Hierarchy with one weight, no italic and four usable sizes, which
pushes the work into spacing, case, letterspacing and color. Read `text-as-gif`
first: findings 7 and 39 say most of what this area studies as type was not
type at the time, it was exported pictures of type, and the other three studies
here are about the look that practice produced rather than the practice.

**Color.** The web-safe 216, dithering, and the palettes that came out of it.
Studied in `two-value-dark` and `palette-split`, from declared `bgcolor` values
rather than from screenshots. See finding 8. The description above said "the
web-safe 216 as a real constraint" for most of this lab's life, and finding 41
withdrew the phrase: `websafe-216` shows the dark wing had already left the
palette by 1999, because the palette could not express a ground and a panel.

**Navigation.** Rollover states as the primary feedback channel, the persistent
left rail, sub-nav that expands in place, and the intro-then-menu pattern.
Navigation carried a site's personality then in a way it rarely does now. Now
the largest area after imagery, at seven studies, and the only one holding
three different verdicts at once: `rollover` ABSORBED, `image-map` and both
gates ARTIFACT, `popup-window` and `status-caption` REVOKED. Finding 40 is why
that spread is not untidiness. A single handler ran three feedback channels and
they did not share a fate.

**Construction.** How pages were physically made: the Photoshop comp sliced
into table cells, `cellpadding=0 cellspacing=0 border=0`, spacer GIFs,
image-swap rollovers, no sprites, no alpha. Mostly ARTIFACT, and worth building
once to feel why.

**Composition.** Asymmetry, the offset panel, layered transparency, the
diagonal. The part that was never constrained by technology and is just design.

---

## Findings so far

**1. The palette split is structural, not cosmetic.**
The era is not one look. The studio wing is dark; the chrome and vernacular
wing is dark-on-light. Nearly every technique behaves differently across that
split, which is why `palette` is a declared catalog field rather than an
assumption. Anything that gets built on top of the studies later (a grain
overlay, a contrast switch, an audit) has to handle both or it silently only
works on half the lab.

**2. In `light-chrome`, two techniques from the same page have opposite fates.**
The bevel and the sliced table produce nearly the same picture. The bevel is
four border colors on one element and zero requests. The sliced rebuild is a
3x3 table shimmed with a real 42-byte `spacer.gif`. The page counts both live;
the ratio is the finding, and it is why the study sits at CONTESTED rather than
taking a single verdict for the whole page. A study is allowed to contain both
a survivor and an artifact, and pretending otherwise would flatten the era.

**3. Crisp sizes should be measured, not remembered, and the first two
attempts to measure them were both wrong.**

The result, at `devicePixelRatio` 1:

| face | on grid at | measured |
| --- | --- | --- |
| Silkscreen | multiples of **8** | 8, 16, 24, 32 pass; 22 and 33 fail |
| Departure Mono | multiples of **11** | 11, 22, 33 pass; everything else fails |

Neither number was asserted from memory. 22 and 33 were added to the ladder
specifically to try to kill the multiples hypothesis after a first pass
suggested it, and they confirmed it instead.

Two failures got there, and both are worth keeping:

- **Measuring the element instead of the text.** `.sample` is a `1fr` grid
  item, so its bounding box is the column width, identical at every size. The
  ladder reported the same advance twelve times and looked like a finished
  table. A `Range` over the contents measures what was actually laid out.
- **Averaging a proportional face.** Dividing a mixed string by its character
  count gives an average advance, which is not any glyph's width and cannot be
  checked against a pixel grid. It produced plausible numbers for Departure
  Mono, which is monospace, and silently meaningless ones for Silkscreen, which
  is proportional. The ladder now probes one repeated glyph at a time across a
  six-glyph set and takes the worst offender, and it detects and prints
  proportional versus monospace rather than assuming.

The general lesson is the one this lab keeps meeting from a different angle: a
measurement that returns a well-formed number is not the same as a measurement
that is measuring the thing.

**3b. The 2x display changes the answer, which is the point.**
Every number above is for `devicePixelRatio` 1. At 2, the grid halves and
Silkscreen comes on grid at multiples of 4. Two people reading this lab on
different monitors are looking at different results, so the ratio is printed at
the top of the instrument rather than assumed.

**4. The 2x display is the strongest argument against the whole aesthetic.**
Bitmap type existed because a normal face at 10px was mud. That is no longer
true. `type-ladder`'s antialiasing switch is the honest version of the
argument: turn it on and legibility improves while the character disappears.
Both are real, which is what CONTESTED is for.

**5. The weight budget inverted.**
Measured against a 60 KB / 56k budget: dark-studio 38.7 KB, light-chrome
17.6 KB, type-ladder 41.4 KB. All three fit, but almost none of that is the
page. Departure Mono alone is 22 KB and the two Silkscreen faces add ~7 KB, so
the fonts are the overwhelming majority of every study that uses them.

The era spent its budget on sliced images and had type for free, because type
meant whatever the machine already had. A period-weight page built today
inverts that exactly: the images are gone and the budget goes to webfonts, the
one thing the era could not buy at any price. `light-chrome` is the lightest
study here purely because it sets most of its text in Tahoma, which is to say
because it makes the period choice.

That also means the 60 KB ceiling is not yet doing any work. Nothing here
carries images. The budget only starts to bite once an image-built study
lands, and until then the readout is a fact about fonts.

**6. The markup is the primary source, and it is the part that survived.**
The captures are mostly rot. Images 404, Flash is gone, and a screenshot of an
archived page is a picture of the rot rather than the design. But `bgcolor` is
text, so it came through intact: `bgcolor="#253048"` is still sitting in the
2Advanced splash exactly as somebody typed it in 2002. So is the font stack,
the window.open geometry, the rollover preload list and the keywords meta.

Reading the source turned out to be far more productive than looking at the
renders, and every value in `two-value-dark` and `boxed-grid` is a citation
rather than an eyedropper sample.

**7. Forty images, two font tags.**
Census of the 2000-03-01 k10k front page: 18 tables, 37 rows, 62 cells, 40
images, 22 links, 45 rollover images preloaded as `_off`/`_on` pairs, 8 spacer
GIFs, and exactly **two** `<font>` tags. Their own keyword for the layout,
sitting in the meta tag the whole time, was `lotsofpeopleinboxes`.

Almost every piece of text on that page was a GIF. This reframes the whole
typography area: the tiny type the era is remembered for was largely not type
at all, it was pictures of type, set in Photoshop and exported. What little
live text existed was `size="1"` because that was the floor. The bitmap face
was a Photoshop asset, not a webfont, and `type-ladder` is therefore studying
something closer to what the era wished it had than what it actually shipped.

**8. Dark did not mean black, and this lab had it wrong.**
`dark-studio` was originally built on #05070a and #0a0f16, on the assumption
that "dark studio site" meant near-black. The source says otherwise.

| site | ground | panel | separation | ground luminance |
| --- | --- | --- | --- | --- |
| k10k, 2000 | `#404642` | `#2f3431` | 1.31:1 | 5.9% |
| 2Advanced, 2002 | `#253048` | `#2d3851` | 1.13:1 | 3.0% |
| Praystation, 2002 | `#333333` | (not declared) | – | 3.3% |
| modern default | `#0a0a0a` | `#111111` | 1.05:1 | 0.3% |

The step size is not shared: 2Advanced at 1.13 is nearer the modern default
than it is to k10k at 1.31. What is shared is a floor. Every measured ground
sits between 3% and 6% luminance, and the modern default is ten times darker
than the darkest of them. `dark-studio` now uses the measured 2Advanced pair
unmodified, and says so in its own source.

**9. Every prose number in this lab has been wrong at least once.**
`boxed-grid` shipped with a hand-typed "1.19:1" separation that was actually
1.31:1. `two-value-dark` shipped with an argument claiming the era pairs shared
a structure and that near-black collapses, and its own instrument only half
agreed. Both times the fix was the same and it should now be the default:
**delete the number from the prose and have the page compute it.** A claim a
study can check is a claim that cannot rot.

**10. The corpus is blocked on a rate limit, not on code.**
`scripts/corpus.py` is written, correct and unrun. The plan was to census 45
domains across four constituencies so findings 6 to 9 could become
distributions instead of anecdotes. What actually happened, in order:

1. Intermittent **503** "Temporarily Offline". Real, and transient.
2. Sustained **429**, caused by my own request rate rather than by the archive.
3. The **CDX index timing out entirely** while ordinary snapshot fetches still
   returned 200. Since every domain needed a CDX query before any of its
   pages, one endpoint was gating all 45. Fixed by dropping CDX completely:
   `/web/<timestamp>/<url>` redirects to the nearest capture and serves it in
   one request, and the real timestamp comes back in the final URL.
4. By then the address was **hard limited**: every request, snapshot endpoint
   included, returned 429 in 1.2 seconds flat.

The lesson is not about retries. Backoff cannot recover a budget that is
already spent, and retrying extends the block rather than clearing it. The
throttle is the only number that matters, and it is now 10s with the whole
incident written into the source above it.

Two consequences for the lab:

- **`portal-density` carries unverified numbers and says so**, in its source
  and in its catalog entry. Its 91 links and 74-above-the-fold are designed,
  not measured, and stay flagged until the corpus confirms or corrects them.
  That is exactly how `dark-studio` got its palette wrong and then right.
- **The corpus stays the highest-value next step.** Nothing about the approach
  is disproven; it needs a cooling-off period of some hours and one patient
  run. Everything caches, so it is built once.

**11. The dark studio look is an import, and the lab was drawing it second hand.**
`tdr-numeric` argues, and I think correctly, that the HUD vocabulary is not a
native web aesthetic. Product codes, circled indices, hazard rules,
registration marks and text set as specification came out of British graphic
design of the late eighties and nineties, The Designers Republic above all,
with Attik alongside. It reached the web fully formed from record sleeves and
game packaging.

That explains something the lab could not: **why this strand aged so much
better than everything else here.** It never depended on a browser limitation,
so nothing broke when the limitations lifted. It is the one SURVIVES in the lab
that was never a workaround in the first place.

Two corrections it forces:

- **The source is bright, not dark.** `dark-studio` and `hud-console` are both
  dark because 2Advanced was, and 2Advanced was a web-era narrowing of a much
  louder language. The lab had been treating a dialect as the language.
- **In the source, the numbers count something.** A catalogue number on a
  sleeve is real. In most pages that borrowed the look, including this lab's
  own `hud-console`, the numbers are decoration. That is the difference between
  the vocabulary and a costume of it, and it is a fair charge against our own
  work.

**12. The era was not uniformly dense and dark; the lab's sample was.**
Eleven studies in, everything was dense and most was dark, which implied a
claim about the period that the period does not support. `jp-minimal` is the
counterexample: the Tokyo wing around 1999 to 2001 was white, nearly empty, and
built around one small thing that responded to you. Measured on the page: 4.6%
of it occupied, 95.4% empty.

It is also the only work here where **motion is the content rather than the
chrome**. Take the interaction away and no page remains. That is why the
archive serves it so badly, and why the source census that recovers k10k
recovers nothing at all here: no bgcolor to read, no font tag to count, only a
lost `.swf`. A lab built on reading markup has a blind spot exactly the shape
of this work, and should say so.

**13. The aesthetic control inverts the lab's own question.**
`net-art-broken` is to aesthetics what useit.com is to usability. It is also
the only strand where **the constraint was the material rather than the
obstacle**. Everywhere else somebody is beating a limit; here the limits are
the paint, and a page that renders differently in two browsers is a bug to
every other study and a feature to this one.

So "does the technique survive the removal of its constraint" quietly assumes
the constraint was unwanted. This is the case where it was chosen. Worth
holding onto the next time a verdict looks obvious.

One deliberate infidelity, recorded rather than hidden: the originals used
seizure-rate flashing. That is a genuine harm and not a stylistic choice to
reproduce for accuracy, so the flicker is slow, low contrast, stoppable, and
honours `prefers-reduced-motion`. **Faithfulness stops where injury starts.**

**14. Twice now, a failure has been invisible because only the success path
had a counter.**
The corpus run reported "domains done: 0" for twelve minutes and I read it as
slow. It was crashing: expanding the domain list had sliced out the `BACKOFF`
constants that lived between `DOMAINS` and `def get(`, so every fetch raised
`NameError` before making a single request. The progress line counted successes
and never once mentioned that the error log was filling up.

Same shape as finding 3, where the type ladder printed one number twelve times
and looked finished. **A measurement that only counts successes cannot tell
"nothing is happening yet" apart from "everything is failing."** Every progress
readout in this repo should report failures beside successes, and the monitor
now does.

**15. The corpus is blocked again, and this time it was avoidable.**
The rate limit cleared overnight. I then spent it: repeated probes, a crashed
run, more probes to diagnose the crash. By the time the code was correct the
address was 429ing in 1.2 seconds flat again. Finding 10 already said retrying
extends the block and I did it anyway while debugging.

The rule that follows, and it is a discipline rather than a code change:
**diagnose against the cache, never against the network.** The 78-domain list
and the CDX-free fetcher are both correct and untested. `scripts/shoot.py` is
written and waiting on the same budget. Next attempt gets one patient run,
after a real cooling-off, with no probing.

**16. The foundry page is the only layout here that needed no rehabilitation.**
`foundry-store` studies the problem YouWorkForThem posed when Cina and Young
founded it in 2001: sell an aesthetic object whose presentation IS the product.

Its rule is still the right rule and it inverts everything else in this lab:
**when every cell in the grid is shouting in a different voice, the grid has to
shut up.** A strict neutral frame, one restrained accent, all the character
coming from the goods. That is the exact opposite of a studio front, where the
chassis IS the pitch, and it explains something the lab kept noticing without
being able to name: a foundry page from 2002 reads as more modern than a studio
page from 2002, because it was never expressing anything, so nothing in it
could date.

It also poses a requirement no other study has. Everywhere else the page is
ABOUT something. Here the page is MADE OF the thing, so the specimen has to be
honest: the two vendored faces are shown only at the sizes finding 3 measured
them crisp at. A foundry showing a bitmap face at a size it breaks at would be
lying about the product.

**17. Do not automate access to archives built to resist automation.**
Two services throttled in one day. archive.org 429ed the address after I spent
its budget probing (findings 10 and 15). Then a single `navigate_page` to
webdesignmuseum.org tripped Cloudflare's bot challenge and, because it is the
same browser and IP, blocked Erin's own tabs mid-browse.

The pattern is the problem, not either incident. These sites are curated by
people, served at their expense, and defended accordingly. The working
arrangement that actually holds:

- **Erin browses them by hand.** That is the supported path and it is not slow.
- **I read what is already loaded** in an open tab with a read-only
  `evaluate_script` returning small JSON. That cost nothing and worked
  perfectly, which is how the Golden Age list came back.
- **I never navigate them myself**, and I do not fetch them in a loop.

The corpus fetcher stays for archive.org, once, patiently, after a real
cooling-off. Everything else is manual. Getting a collaborator blocked out of
their own research is a worse outcome than any dataset is worth.

**18. The closest thing this lab has to a law, and it arrived twice by accident.**
`foundry-store` concluded that when every cell in a grid shouts in a different
voice, the grid has to shut up. `community-board` was built afterwards, from a
completely different problem (hosting other people's submissions rather than
selling type), and arrived at the identical rule: white boxes, a plain grey
frame, no character anywhere in the chassis.

Two unrelated commercial problems, no connection between them, one answer:

> **A frame's job is inversely proportional to how much its contents have to say.**

That is worth more than either page alone, because neither was built to prove
it. And it retroactively explains the thing this lab kept noticing without
being able to name: **the studio fronts here date hardest.** `dark-studio`,
`hud-console` and `splash-gate` are the pages where the frame had the most to
say, and every one of them is legible as period work at a glance. The foundry
and the board are not, because they were never expressing anything, so nothing
in them could go out of date.

It also quietly reframes the lab's verdicts. SURVIVES has been reading as
"this was good", but there are two very different routes to it: a technique can
survive because it was excellent, or because it was self-effacing enough to
have no surface to age. Those deserve to be told apart, and currently are not.

**19. Two captures of the same site, ten months apart, and the palette is
identical.**
Everything measured so far came from a single capture per site, which meant
every value in this lab was technically a sample of one. `corpus/cache` also
holds a k10k front page from **1999-05-06**, ten months before the 2000-03-01
one the rest of the lab is built on. Censusing it was free, offline, and cost
the archive nothing.

| | 1999-05-06 | 2000-03-01 |
| --- | --- | --- |
| tables | 17 | 18 |
| rows | 34 | 37 |
| cells | 57 | 62 |
| `<img>` tags | 39 | 40 |
| links | 19 | 22 |
| **`<font>` tags** | **2** | **2** |
| **spacer GIFs** | **8** | **8** |
| ground / panel | **#404642 / #2f3431** | **#404642 / #2f3431** |

The content churned and the structure did not. The two values in finding 8 are
not an artifact of which Tuesday the crawler visited: they are the site's
declared identity, held across ten months and a content cycle. The same goes
for "exactly two font tags", which was the most surprising number in finding 7
and is now the same number twice.

It also recovers something finding 8's table did not have, because nobody
thought to look for it: the text and link colours, `text="#FFFFFF"` and
`link="#cacaca"`. The type on that ground is not white. It is one step down,
and the white is reserved.

Method note worth keeping: the census had to be run twice. The first pass
counted 39 `<img>` tags but only 3 with a `src`, and reported that **zero**
preloaded images were also displayed, which is impossible when the resting
states are the preloaded `_off` files. The cause was a filter dropping absolute
URLs while the preload list used relative ones, so the two sets could never
intersect. Finding 3 and finding 14 again, from a third angle: the number was
well formed, internally consistent, and measuring nothing. The tell was that it
was impossible, not that it looked wrong.

**20. The rollover was three channels wearing one handler, and they did not
share a fate.**
`rollover` is the first study to take ABSORBED, and building it showed why the
verdict had to exist. The sixteen `onMouseOver` handlers in the 1999 source are
not doing one thing:

| | channel | handlers | fate |
| --- | --- | --- | --- |
| A | the hovered element changes | 14/16 | absorbed into `:hover`, early and completely |
| B | a *different* element changes | 9/16 | not expressible in CSS at all for twenty years |
| C | a sentence in the status bar | 16/16 | disconnected |

Filing that under SURVIVES would have hidden the interesting half. **Channel B
is the one that breaks the tidy story.** Nine of the fourteen image handlers
change something the pointer is not on, and k10k put the caption panel *above*
the icon strip, where no combinator CSS had could reach it. `:has()` finally
made it reachable, and the study demonstrates that rather than claiming it:
switch the rebuild to CSS and the caption works with no script at all. It costs
six stylesheet rules where the script needed one function, because a selector
cannot take an argument, and the payload has to be `content`, which means the
caption is no longer text. The page proves that last part on itself, by trying
to select it and getting an empty string.

The cost of the original is measured, and it is the number worth keeping: the
page pulls **52 distinct image files and 22 of them are states nobody sees
unless they point at the right rectangle**. Forty-two percent of the image
payload, over a line doing about 5 KB a second, for feedback most visitors
never triggered.

**21. A channel can die without its API dying, and that is worse.**
This study's first draft asserted that `self.status` "reads back empty, channel
removed", and put a probe on the page to show it. The probe said the opposite.
`self.status` accepts your string and hands it back unchanged, exactly as it
did in 1999. What was removed is the surface it wrote to, not the interface.

So the 1999 line still runs, still returns, still throws nothing, and does
nothing. All sixteen handlers write to it and two of them write to *nothing
else*, which means for those two the entire feedback was living in browser
chrome that has not existed for twenty years, in code that never once
complained. **"It still runs" is not evidence that it still works**, and a
silent no-op is a worse failure mode than an error, which is the same lesson
findings 3 and 14 keep arriving at from the measurement side.

The correction matters more than the fact. The claim was wrong, the page was
built to check its own claim, and the check caught it. That is the only reason
this lab gets to make assertions at all.

**22. The dark studio was not always dark, and the lab was citing a name when
it meant a date.**
Finding 8 corrected `dark-studio` from invented near-black to 2Advanced's
measured `#253048`, 3.0% luminance, and that correction still stands. What it
quietly implied is that the pair was a fact about the studio. It is not. The
same studio's splash, captured **2000-06-08**, two years earlier, declares:

| | ground | rule | text / link | ground luminance |
| --- | --- | --- | --- | --- |
| 2Advanced, 2000 | `#999999` | `#003366` | `#FFFFFF` (all four) | **31.9%** |
| 2Advanced, 2002 | `#253048` | | | 3.0% |

`#999999` is mid grey. It is more than five times brighter than the brightest
ground in finding 8's table, which was k10k at 5.9%, and more than ten times
brighter than the same studio two years later. The dark look was **arrived at**,
somewhere between these two captures, and finding 8's floor is a claim about
what dark pages did in 2002, not about who these designers were. The lab had
been treating a measurement as an identity.

That capture had been sitting in `corpus/cache` unopened the whole time.

**23. The three-channel rollover is a k10k fact. The industry tool shipped one
channel.**
Finding 20 counted sixteen handlers running three feedback channels and
concluded things about "the rollover". The 2000 2Advanced page is the control
that finding needed, and it disagrees:

| | k10k, 1999 | 2Advanced, 2000 |
| --- | --- | --- |
| authored in | Adobe GoLive | Macromedia Dreamweaver |
| swap function | hand-written `hiLite()` | stock `MM_swapImage()` |
| rollovers | 16 | 1 |
| preloaded images | 37 | 1 |
| remote swaps (channel B) | 9 | 0 |
| writes to `self.status` (channel C) | 16 | 0 |

Dreamweaver's generated rollover does channel A and nothing else. It has no
idiom for changing a second element and never touches the status bar. So
channels B and C were not conventions of the era, they were **craft**, written
by hand at one site, and the thing that actually became universal is precisely
the thing the standard tool emitted.

This is the sharpest argument yet for the corpus, and against the lab's habit
of generalising from k10k: k10k is the most-cited front page of the period
because it was exceptional, which makes it the worst available proxy for
typical. Two sites is not a distribution either. It is just enough to know the
first one was not one.

**24. Some techniques were not outgrown. The permission was withdrawn, and the
lab's question has no slot for that.**
The whole lab asks: does the technique survive the removal of its constraint?
Two studies now answer with something the question did not anticipate.

`popup-window` measures six `window.open` calls. On all five k10k ones, every
one of the seven chrome flags is `false`: not resizable, no scrollbars, no
menubar, no toolbar, no directories, **no location bar, no status bar**. The
designers were not opening a browser window, they were removing the browser
from it. `rollover` measures sixteen handlers writing sentences into
`self.status`.

Neither of those died of progress. **Nothing about either constraint lifted.**
If browsers still honoured the flags, the chromeless popup would still work
today, exactly as written. What happened is that both capabilities let a page
lie about where you were, so both were taken away on purpose, as anti-phishing
measures, by the people who make browsers.

The live probe in `popup-window` shows precisely where the line was drawn.
Asked for 236x134 with every flag off, a browser today returns a drawable area
of 236x133, within a pixel of the request, wrapped in 67px of chrome the page
asked to have none of and cannot remove. **The canvas was granted and the frame
was refused.** That gap is the entire modern position on this technique, in two
numbers.

So there is a category here the four verdicts do not name: not SURVIVES, not
ARTIFACT in the ordinary sense of being outgrown, not CONTESTED, not ABSORBED.
Call it revoked, provisionally, and do not add it to `VERDICTS` yet. ABSORBED
earned its slot by turning up repeatedly before it was named, and two cases is
not that. Both are filed ARTIFACT with the reason written out, and if a third
turns up the vocabulary changes.

**Superseded by finding 29.** The third case turned up: `ambient-audio`. REVOKED
is now the fifth verdict and `popup-window` is refiled into it. The rule this
finding set worked exactly as written, which is the main reason to trust it the
next time the vocabulary is under pressure.

**25. k10k had no ground colour. It had a random background image, and the
lab has been calling two cell fills a ground for six studies.**
Findings 8 and 19 both present k10k as declaring a ground of `#404642` and a
panel of `#2f3431`. Reading the rest of the 1999 source kills the first half of
that. Every one of the nine `bgcolor` attributes on that page sits on a
`<td>`:

| colour | cells | declared sizes |
| --- | --- | --- |
| `#404642` | 5 | 203x32, 203x164, 7x164, 7x164, 203x7 |
| `#2f3431` | 4 | 144x28, 144x173, 81, 63x4 |

There is exactly one `<body>` tag on the page and it is not in the markup. It
is written at runtime:

```js
var inxback = 0 + randNum(7);
document.write("<BODY BACKGROUND='images/dynamicfront/baggrunde/k10kbackground"
  + inxback + ".gif' text='#FFFFFF' link='#cacaca' ...>");
```

So the actual backdrop of the most-cited dark page of the era is **a tiled GIF
chosen at random on every load**, from a set of seven. `baggrunde` is Danish
for backgrounds. The two colours the lab has been quoting are the fills of the
boxes sitting on top of it.

What survives the correction: the two values are real, both measured, and the
separation and luminance numbers in finding 8 are unaffected, because they were
always a comparison between two declared colours. What does not survive is the
word **ground**, and with it the tidy picture of an era that chose a flat dark
backdrop on purpose. At least one of these sites did not choose a backdrop at
all. It rolled one.

Two caveats, because the temptation is to over-correct:

- This is verified for the **1999-05-06** capture. The 2000-03-01 capture,
  which is where finding 8's row actually comes from, is not in the cache and
  cannot be checked offline. It may well declare a body colour. Until it is
  re-read, finding 8's k10k row should be read as "two declared values", not
  as "ground and panel".
- `two-value-dark` was already more careful than the prose around it: its own
  note said "five cells declared at #404642". The study had the right fact and
  the surrounding argument had the wrong word, which is its own small lesson
  about where errors hide.

Two more details from the same block, both worth keeping:

- **`randNum` is not a random number generator.** It is
  `Math.round(num * Math.cos(now.getTime()))`, the cosine of a millisecond
  epoch. Because cosine is flat near its extremes, the results are arcsine
  distributed rather than uniform: the first and last backgrounds came up far
  more often than the middle ones. The page had seven backgrounds and did not
  show them equally.
- **Two thirds of the randomisation is dead code.** `inx = randNum(7)` and
  `inx2 = randNum(26)` are both computed at load and never interpolated
  anywhere. Only `inxback` is used. Whatever the other two were for was
  removed, and the calls were left behind.

**26. The most honest artwork is the least like the era.**
Every study with imagery had been shipping grey rectangles labelled with their
own dimensions. Honest, and it makes a period page read as a wireframe. Four
imagery vocabularies now have studies of their own, and putting them side by
side settles something the lab could avoid while its pages carried no art:

| study | vocabulary | measured marks | reads as 2000 |
| --- | --- | --- | --- |
| `art-diagram` | rule stacks, outlines to scale, registration marks | all | no |
| `art-texture` | halftone, scanlines, gradient wash, arc | some | yes |
| `art-studio` | exploded panels, tech-organic curve, reticle | **none** | most |
| `art-pixel` | ordered dither, isometric cubes, fixed grid | one | yes |

**The plate with no evidence in it is the one that reads as evidence of the
era.** `art-diagram` is built entirely from real measurements of a real page
from 2000 and could have been drawn in any decade. `art-studio` cannot be
checked against anything and places instantly.

That is not a flaw in the drawing, it is the shape of the problem. The era's
imagery was not encoding data, it was expressing a studio, and a lab that only
permits itself measured marks cannot reproduce work whose entire purpose was
expression. Finding 7 pointed here without following it: the k10k front page
ran 29 distinct images against 2 font tags, so those pages were overwhelmingly
pictures, and every study in this lab is text. The grey rectangles were hiding
how far from the era that puts us.

Three consequences worth keeping:

- **A seventh area.** The six areas were "what a designer was actually
  deciding", and choosing the artwork was one of those decisions, arguably the
  dominant one. `imagery` now exists. Its absence had been letting the lab
  describe an era of pictures entirely in text.
- **The verdicts strain here.** SURVIVES, ARTIFACT, CONTESTED and ABSORBED
  were built for techniques answering constraints. Three of these four are
  CONTESTED, not because the vocabularies are alike but because "defensible
  now only as deliberate style" is the only slot decoration fits in. Related
  to finding 24: the vocabulary is built for one kind of question and keeps
  meeting others.
- **`art-texture` found something the lab has not seen elsewhere.** Halftone
  exists because ink cannot print continuous tone. Scanlines exist because a
  CRT draws in lines. Neither is a property of a web page, so that plate is
  a page **simulating the limitations of media it is not**. Everywhere else in
  this lab somebody is beating a constraint. There, somebody is importing one
  for the look of it.

Two rules held throughout, and they are why this did not become a shared
graphics layer: each study owns its drawing code outright, with the four
imagery studies containing near-identical helpers on purpose; and every mark
declares whether it is measured or designed, per study, in a table the page
builds itself.

**27. The imagery this lab was missing was not a style, it was the picture.**
Finding 26 added four imagery studies and they were all still *marks on a
surface*: rules, dots, panels, pixels. Put beside actual captures of the
period, that set turned out to be missing the thing that covered the most
pixels on nearly every page, which is a **rendered or composed image filling
the frame as the page itself**.

Four more studies close that gap, and each one turned out to carry a finding
rather than just a look:

- **`art-render`.** A pre-rendered 3D still, made in Lightwave or 3DS Max over
  hours and shipped as one flat JPEG. The verdict is ARTIFACT and it takes
  care to say why, because the look is obviously alive: the technique is not
  the look. The technique is "the machine cannot draw this, so bake it
  elsewhere", and that constraint is gone completely. The specimen projects
  and depth-sorts its geometry live at a cost of nothing. **A look and a
  technique can have opposite fates**, which is the confusion findings 20 and
  21 were both circling.
- **`art-shard`.** The claim here is the surprising one. GIF is run-length
  encoded, so a field of one flat colour costs almost nothing while a
  photograph costs everything. Hard-edged flat polygons were **the cheapest
  ambitious picture available**, and the most recognisable graphic language of
  the period is partly downstream of how a 1987 image format stores rows of
  pixels. It takes SURVIVES, alone among the imagery studies, because it was
  vector-native before vectors were available.
- **`art-glitch`.** Two pretences pointing opposite ways. The smear is motion
  blur, a physical record of something moving past an open shutter, and
  nothing moved. The micro-type, the index counting nothing and the barcode
  that will not scan look like information and are texture. That second half
  is **exactly the charge finding 11 laid against our own `hud-console`**,
  reached independently from a different wing. Two unrelated routes to the
  same conclusion, which is the pattern finding 18 came from.
- **`art-wireframe`.** A wireframe shows the edges of a solid, and here there
  is no solid. Nothing was shaded, textured or lit, so nothing had to survive
  being shaded, and the picture cannot be caught out because it never claims
  to be finished. **The authority of engineering at the price of a diagram.**

The reduction panel on each page turned out to be worth more than expected. It
ranks the vocabularies by what survives being made small, and the order is not
the order of quality:

| vocabulary | at 64px |
| --- | --- |
| `art-shard` | loses essentially nothing, a hard edge stays a hard edge |
| `art-render` | degrades gracefully, a soft image has no detail to lose |
| `art-glitch` | reduces to texture, having always been texture |
| `art-wireframe` | collapses into grey, worst in the set |
| `art-pixel` | does not degrade, it **breaks**, being defined in pixels |

That is a real property and none of it is about taste. It also explains
something about the period the lab could not otherwise: the wings that shipped
gallery thumbnails are the wings whose work reduces, and the pixel wing never
did, because it could not.

**28. Three independent routes to one charge, and the imagery wing keeps
arriving at it.**
Finding 11 said that in the print vocabulary the numbers count something real,
and that in most pages which borrowed the look they are decoration. That was a
charge against our own `hud-console`. The imagery studies have now reached it
twice more, from two directions neither of which was looking for it:

| study | what pretends | what it pretends to be |
| --- | --- | --- |
| `hud-console` | product codes, circled indices | a catalogue |
| `art-glitch` | micro-type, an index counting nothing, an unscannable barcode | data |
| `art-chassis` | notches, a trademark on an internal panel name, a scrollbar with nothing to scroll | an instrument |

Three unrelated builds, one answer: **the era's visual language consistently
dressed as an instrument and consistently instrumented nothing.** That is the
same shape as finding 18, which also arrived twice by accident and is worth
more for it.

What survives the charge is the part that was doing a job. `art-chassis` takes
ABSORBED, only the second study to, and it splits exactly along that line: the
module, a rectangle with a titled header holding one kind of content, won so
completely that nobody designs it any more, while every ornament fastened to
it died.

Three more imagery studies close the set out, and two of them break the
CONTESTED monotony finding 26 complained about:

- **`art-duotone` takes SURVIVES.** Collapsing a photograph to a ramp between
  two colours did two jobs at once: it cut the palette enough to make a
  photograph affordable in a 216 colour world, and it forced the picture to
  obey the site palette instead of fighting it. The byte economy is finished.
  The discipline is not, because it was never really about bytes:
  subordinating imagery to the palette is in every brand system now. **The era
  reached a design rule through a compression problem, and the rule outlived
  the problem.** That is the cleanest possible answer to this lab's question.
- **`art-scene` takes CONTESTED**, and the distinction from `art-render` is
  reference rather than style. An abstract form refers to nothing and so
  nothing can fail to be true of it. A scene implies a place, therefore a
  time, therefore a moment before and after, none of which a page has. The
  letterbox bars are the tell, exactly as "skip intro" is the tell in
  `splash-gate`: a page cropping itself to cinema is saying what it wishes it
  were. It dated harder and faster than the abstraction beside it, because a
  rendered city is a specific year's idea of the future, which is the least
  forgiving thing an image can be. Underneath the ambition: four depth layers
  and one light. **Aerial perspective, a painting convention centuries older
  than photography, running the most futuristic vocabulary in the lab.**

The four earlier imagery studies now carry the same reduction panel as the
rest, so finding 27's ranking is demonstrated on every page rather than
asserted here. `art-pixel` in particular now shows what "breaks" means: at 64
pixels it is a smear, while every other vocabulary merely thins.

**29. The fifth verdict, and the rule that produced it worked exactly as
written.**
Finding 24 found two techniques that died a death this lab had no name for:
the chromeless popup and `self.status`. Neither was outgrown, neither
constraint lifted, and both were switched off on purpose by browser vendors
because both let a page lie about where you were. It refused to invent a
verdict for two cases and set a test instead: **a third turns up, the
vocabulary changes**.

`ambient-audio` is the third. Sound that starts on its own still parses, still
decodes and still runs. Its live probe reports `AudioContext` coming back
`suspended` and `audio.play()` rejecting with `NotAllowedError`, and then both
working instantly on a gesture. Nothing broke. The permission was withdrawn,
by four vendors, between roughly 2017 and 2019.

So **REVOKED** exists, `popup-window` is refiled into it from ARTIFACT, and the
verdict spread is now 12 SURVIVES / 15 CONTESTED / 3 ARTIFACT / 3 ABSORBED /
2 REVOKED.

Two things worth keeping about the refiling itself:

- **ARTIFACT was not a small mislabel.** It means the problem went away and
  took the solution with it, which is `splash-gate`: nobody waits ninety
  seconds for a movie, so nothing needs to fill the wait. None of these three
  had a problem left to go away. Filing them under ARTIFACT was quietly
  crediting progress with somebody's policy decision.
- **`rollover` stays ABSORBED and that is correct.** Its channel C is revoked,
  but a verdict names a study's dominant channel, and channel A absorbed
  completely. What finding 20 could not do was name what happened to channel
  C. Now it can.

**30. The lab's own charge needed narrowing, and the study that narrowed it is
the one the charge did not fit.**
Finding 28 said the era's visual language dressed as an instrument and
instrumented nothing, on three routes. `dot-notation` makes a fourth, and it is
the first that can be **run** rather than argued: the syntax makes a claim a
machine can check, so the page resolves each phrase as a real property path
against the global object. Three genuine paths resolve, as controls. Zero of
eight era phrases do.

The census behind it is the better half. Across both captures, wrapper
stripped, there are **241 occurrences of 126 distinct dot-joined tokens, and
every one is a filename, a hostname or a property access**. The count set as
display text is zero. The punctuation was in the building the whole time, load
bearing, in the parts nobody looked at, and the convention is a designer
reading the source of their own page and liking how it looked.

Then `organic-motif` was built and **the charge did not apply to it at all**,
which turned out to be worth more than a fifth confirmation. There are no
numbers in it, no index, no code, no notch, no scrollbar and no syntax. It is
not a diagram, a photograph, a control or a borrowed sign. There is nothing to
catch, because it never claimed anything. So the pattern finding 28 found is
not *the era decorated*:

> **The era decorated while claiming to measure.**

Take the claim away and the decoration is only decoration. The problem was
never ornament, and that distinction was invisible until a vocabulary turned up
wearing no costume at all. `organic-motif` is still CONTESTED, because honest
decoration is still decoration, and every row of its inventory declares that it
encodes nothing, in a third column no other study has.

**31. Removing a limit can subtract from a vocabulary.**
Every study here so far has been written against a constraint that softened.
`chevron-hatch` is written against one that **inverted**. In 1999 the diagonal
was the most expensive mark available: tables are rectangles, and with no
transform, clip path, SVG, canvas or alpha, every angle on every page was a
picture, exported, sliced, downloaded, and unable to move, reflow or recolour.
It is now the cheapest mark on the platform. The study builds the same hazard
band both ways so the two sit side by side, and the right one is a single
declaration.

The vocabulary did not survive that, and the reason runs against what this lab
keeps expecting. **When angles were expensive they were used sparingly, and
sparing use is what made them read.** A diagonal across an otherwise
rectangular page is an event. Once the cost went to nothing the restraint went
with it and the diagonal became wallpaper.

That is the exact counter-case to `art-duotone`, which is the cleanest SURVIVES
in the lab precisely because a compression problem produced a design rule that
outlived it. Here cheapness did not preserve the technique. **It dissolved the
discipline that made the technique good.**

**32. Finding 1 has never been tested, and the study built to test it found out
why: the lab only has one wing.**
Finding 1 says the era is not one look, that the studio wing is dark and the
chrome and vernacular wing is dark-on-light, and that nearly every technique
behaves differently across that split. It is the oldest claim here and the
`palette` field on every catalog entry depends on it. Nothing had ever checked
it.

`palette-split` computes every number on load from three measured captures, and
independently reproduces finding 8's table exactly: k10k 1999 at 5.9% and
1.31:1, 2Advanced 2002 at 3.0% and 1.13:1, 2Advanced 2000 at 31.9% and 4.43:1.

The result is a correction, arrived at in two stages, and the second stage is
the lab's own method catching the lab:

- **The split is not a split in lightness.** Sorted by ground luminance, the
  three do not fall into two camps. One sits alone at 31.9% and the other two
  sit 2.9 points apart at the bottom.
- **The first draft of that study then claimed the real axis was separation**,
  the distance between panel and ground, and added a live check to back it up.
  The check killed it. Separation puts the same single capture alone, with a
  wider gap. The prose was rewritten and the number was kept, which is finding
  9's rule doing its job for the third time.

What was actually wrong is the evidence, not the claim. **All three measured
captures are studio sites.** k10k and 2Advanced are filed as different palettes
and are the same wing of the same scene. Of the ten light-palette studies, only
two carry measured colour and both of those measurements came out of a studio
site as well. So every measured colour in this repository comes from two
sites, and finding 1's other half has no evidence here whatsoever.

Finding 1 is therefore **neither confirmed nor refuted. It is untested**, and
this is the study that establishes that, which is worth more than a third
reading of the same two sources. It also gives the parked corpus run something
specific to fetch that it did not have before: not more captures, but captures
from **the other wing**. A portal, a personal page, a board, a shop.

One genuine result fell out on the way. The mid grey ground puts white text on
`#999999` and measures at 2.85:1, well under the modern 4.5:1 floor, while both
dark grounds clear it comfortably at 9.67:1 and 13.17:1. **By a current
accessibility standard the era's dark pages are the legible ones**, which is
the opposite of the way this is usually told.

**32b. A fourth studio site turned up the same day, and it is nearly white.**
Finding 32 concluded that the studio wing has three measured captures and the
other wing has none. Erin then supplied GreyscaleNet, a Golden Age site the
museum files under Design & Art and Techno, captured **2001-04-20** at
`greyscale.net`.

Sampled off the museum's own screenshot, its ground is a flat `#e6e5df`, and it
is flat: four widely separated points return the identical value and 67% of the
image falls in that one histogram bin. That is roughly **80% luminance**.

| capture | ground | luminance |
| --- | --- | --- |
| 2Advanced, 2002 | `#253048` | 3.0% |
| k10k, 1999 | `#404642` | 5.9% |
| 2Advanced, 2000 | `#999999` | 31.9% |
| **GreyscaleNet, 2001** | **`#e6e5df`** | **~80%** |

If that holds, the studio wing spans nearly the entire lightness range, and
finding 1's "the studio wing is dark" is not merely untested, it is **wrong**.
Finding 22 already showed one studio moving from 31.9% to 3.0% in two years;
this puts a third studio at the far end in between them.

**It is deliberately not being written up as a finding, and the reason is the
method.** Every number in the table above except the last comes from markup
somebody typed. The last one comes from a picture. A screenshot cannot say
whether `#e6e5df` was declared or painted, and finding 25 is the standing
warning about precisely that error: k10k appeared to declare a dark ground for
six studies and turned out to be running a randomly tiled GIF, with the two
quoted values sitting on cells on top of it. A pale flat rectangle in a
screenshot is exactly as ambiguous.

So `greyscale.net` is now in `corpus.py` as the highest-priority row in the
dict, and the census settles two things at once: whether that ground is a
declared value, and whether the overlapping translucent shards shipped as GIF,
which cannot do partial alpha, or as JPEG, which mangles the hard edges
`art-shard` argues were the cheap option. Either answer complicates a study.

**33. A negative is a measurement, and three of these studies are built on
one.**
The lab's provenance rule has always been measured versus designed. This batch
adds a third category that turns out to be stronger than designed and cheaper
than measured: **the measured negative**, a census that comes back empty and
bounds what the lab is entitled to say.

| study | looked for | found |
| --- | --- | --- |
| `ambient-audio` | `bgsound`, `embed`, `object`, `applet`, any audio file, `autostart` | 0 |
| `form-controls` | `form`, `input`, `select`, `option`, `textarea` | 0 |
| `dot-notation` | dot-joined tokens set as display text | 0 of 241 |

Each of those studies prints the terms it searched for, because a negative is
worth nothing if you cannot see what was looked for. And each says plainly that
its specimen is therefore designed, instead of quietly building on an absence,
which is how a lab talks itself into believing its own reconstruction.

`form-controls`'s negative is the one that carries an argument rather than a
caveat. Faced with a control they could not paint, these studio fronts **simply
did not have forms**. The search box and the mailing list lived on a separate,
plainer page, because the moment a form appeared the operating system appeared
with it. Avoidance is a real design decision and it is the one the era made.

**34. One study in this lab is about a constraint that never lifted, and it
changes what the question means.**
`form-controls` is the only page here whose limit is live. You still cannot
style the inside of a `select`, for exactly the reason you could not in 1998:
the open list is an OS window on top of your page, and CSS has no reach into
another window. The study probes eight declared styles against an observable
that would move if they took, and reports three outcomes. Four TOOK, zero were
IGNORED, and four are **UNOBSERVABLE**: there is no measurement to take,
because the pixels are not in the document. That third category is the study.
It is not a browser refusing you, it is a browser having nothing to refuse.

Nothing in it reads `getComputedStyle`, which will hand back every colour you
set on an `option`, forever, and paint none of them. That is finding 21's
failure mode one channel over, and a page that reported computed values here
would have produced a full table of confident nonsense.

The rebuild is counted rather than asserted, and the count corrected the
argument. The custom listbox came out at **1.3x** the nodes of the native
control, near enough to parity that the paragraph claiming it was expensive had
to be rewritten. What the numbers actually say is that **the cost did not grow,
it moved**: 8 listeners, 5 ARIA attributes, 7 key names and about a kilobyte of
handler source where the element needed none of any of it, and it still does
not do five things the original did for free. The markup got no heavier. The
obligations did.

**35. The most absorbed pattern in the lab is the one that stayed hand-built
longest, and those are one fact.**
`thumbnail-strip` is the third ABSORBED and it splits the way `art-chassis`
splits. THE PATTERN, a row of small things where pointing at one changes a
bigger region elsewhere, won so completely it has no name worth using: every
gallery, carousel, related rail, file picker and swatch row. THE MACHINERY died
entirely, and the measurements are what it cost: six icons at 19, 19, 17, 19,
19 and 21 pixels, all 21 tall, **114 pixels of pictures driving one 219 x 47
panel**, built from 19 files of which 12 are never seen. The resting caption is
a file called `ikontekst_tom.gif`, Danish for empty, downloaded so a slot could
be blank. Every caption is a picture of a sentence, which is finding 7 arriving
in the interface copy.

The cross-link is the part worth keeping. This is `rollover`'s channel B, the
one finding 20 said CSS could not express at all for twenty years, and the
geometry says exactly why: **the panel sits above the strip**, and everything
CSS had ran forwards. So the most thoroughly absorbed layout pattern here is
also the one that stayed script-driven longest, and that is not two facts. It
is one fact from either end: the pattern won so completely that everybody kept
paying for it, by hand, every year, until `:has()` arrived.

The CSS version on that page works, and it is honest about its own price: the
caption has to be delivered as `content`, so the page selects it and gets an
empty string. In 1999 the words in that panel were not text because they were
a GIF. Today they are not text because they are a generated string. **Twenty
years apart, the same compromise.**

**36. The era's signature picture was the most expensive one available, and
the lab had been arguing the opposite.**
Erin supplied five references of the shard vocabulary: designgraphik,
Chapter3 and ThreeOh, Desktop Imperium, Cubadust, Submethod. `art-shard`
already names four of those five, and it makes an economy argument about them.
GIF run-length encodes, a field of flat colour is nearly free, so hard-edged
flat polygons were the cheapest ambitious picture the era had, and the most
recognisable graphic language of the period is downstream of a 1987 format.

**The references are not flat.** They are translucent slabs, thirty and forty
deep near the origin, crinkled, with notched silhouettes, internal facets and
dense sheaves of hairlines. `shard-burst` rebuilds that properly and then
rasterises its own specimen three times, counting distinct colours:

| condition | colours | vs GIF's 256 | mean run |
| --- | --- | --- | --- |
| as drawn: translucent, antialiased | ~17,000 | over | 2.8 px |
| alpha forced to 1, still antialiased | ~5,100 | over | 3.9 px |
| alpha forced to 1, antialiasing off | ~150 | **fits** | 8.0 px |

Only the third fits, and it is the version nobody shipped, because every edge
in it is a staircase. So the economy argument holds for a **flat dialect** that
some sites did ship, and not for the flagship pictures. `art-shard` now carries
a scope note; it is a narrowing rather than a retraction, and both studies
stand.

What replaces it is a better finding. The burst is close to the worst case for
**both** formats the era had: transparency and antialiasing each blow past
GIF's palette independently, and the hard edges JPEG would ring on are the only
thing the picture is made of. **They shipped it anyway, at full size, as the
whole page.** Almost every study in this lab starts from a limit and finds the
design that beat it. Here the designers priced the limit and paid it, which is
a different relationship to constraint than anything else here has.

There is a cross-link neither study went looking for. `type-ladder` is about
turning antialiasing **off**, because a bitmap face at 10px is only crisp
aliased, and finding 4 calls that the strongest argument against the whole
aesthetic. This is about antialiasing being unaffordable to leave **on**. Same
page, same year, same property, opposite decisions, one reason: smoothing costs
colours and colours were the budget.

**Method note, and it is the third instance.** The prose under that table was
hand-written three times and falsified three times by the study's own next
run, because changing the composition changes the counts. It is now
**generated**, including its verdict, branching on whether the aliased figure
clears 256. Finding 9 said to delete the number from the prose and let the page
compute it. The stronger version, learned here: when a claim depends on a
measurement that can move, **the sentence has to be computed too**, not just
the figure inside it.

**One more signal on the palette question.** Four of the five references are
light: pale ground, dark and saturated marks, one hot accent doing all the
shouting. Only Cubadust is dark. With finding 32b's GreyscaleNet that is a
second independent pointer that the studio wing was not a dark wing, and it is
the same weak evidence, pictures rather than markup. Recorded, not promoted.

**37. The soft picture is hard. The blur is on the light, not on the form.**
`light-shaft` is the pair to `shard-burst`, built from the one dark reference
in Erin's set, and it was drawn wrong first. Reading "soft vocabulary" as
"blurred geometry" produced olive fog, which is not what the reference is.

The reference is **hard**: crisp faceted blades running from near black to near
white in one frame, with the bright values sitting on the **edges** of facets
rather than across large areas. What is blurred is the shafts, the bloom at the
hot point, and the motion streaks. **Softness is applied to the lighting, not
to the form**, and correcting that turned a cloud back into an object.

The distinction is what the measurement rewards, and it completes the pair:

| | high-frequency energy | colours |
| --- | --- | --- |
| `light-shaft`, as drawn | 2.67 | hopeless for GIF |
| same, blur removed | 3.75 | hopeless for GIF |
| same, moved to the light wing | 3.79 | hopeless for GIF |

High-frequency energy is the mean absolute difference between adjacent pixels:
what a DCT codec spends bits on and where ringing comes from. Hard facets put
that energy on a limited set of edges and leave smooth gradients between them,
which is the signal JPEG encodes nearly free. So this picture is **expensive in
colours and cheap in the currency JPEG actually charges in**, where the burst is
expensive in both. It is the one picture of the period the lab has found where
the ambition and the available format agreed.

**A hypothesis failed here and it is worth more than if it had held.** The
theory was that darkness is itself a compression strategy, and that the one dark
reference being the cheapest was partly the palette. The third row tests it by
moving the identical marks into the light wing with mark-to-ground contrast
preserved. It changes the cost by about **one percent**. Darkness is not doing
the work; the blur is. The lab had a tidy story available, which would have tied
neatly to findings 32b and 36, and the instrument declined it.

That third row also had to be built twice, and the first attempt is the clearer
lesson. It changed only the ground, which put pale marks on a pale field and
measured them **vanishing** rather than measuring the wing, and it came back
*lower* than the dark condition: the opposite of the hypothesis, and a broken
control rather than a result. A control that changes two things at once answers
neither. Same family as findings 3, 14 and 19: the number was well formed,
internally consistent, and measuring the wrong thing, and the tell was that it
was backwards rather than that it looked wrong.

**38. The corruption is snapped to the layout grid, and that is measurable.**
`bar-glitch` is the last distinct vocabulary in Erin's five references, from
the Submethod screenshot: displacement bars, packed stripe bands, cross
scatter, solid accent bars, on a bone ground with one hot red.

`art-glitch` already charges this family with looking like information and
being texture. That charge is about **pretending to be data**. This is a
different and more specific failure: **pretending to be damage**, and unlike
the first it can be tested rather than argued.

The page draws the same field twice from the same seed, once quantised to the
layout module and once with free offsets, and counts where the edges land:

| field | edges | on the module | share |
| --- | --- | --- | --- |
| snapped, as the era drew it | 302 | 287 | **95%** |
| free offsets | 292 | 10 | 3% |
| chance | | | ~4% |

Real corruption does not do that. A truncated GIF, a dropped scanline, a bad
packet: none of them know where your table cells are, and the damage lands
wherever the bytes ran out. **The disorder is total at the level of the mark
and near zero at the level of the composition**, and no real failure produces
that combination. Put the two panels side by side and the free one reads as a
broken image while the snapped one reads as a designed element, which is not a
matter of taste: it is what an arbitrary offset actually looks like.

So it is a picture of chaos that obeys the page, drawn by somebody who wanted
the layout to look survived without letting it interfere with the layout.

**A bug worth recording, because the instrument caught the drawing.** The
alignment first came back at **0%** for the snapped field. The measurement was
right and the drawing was wrong: offsets were being quantised and then added to
an un-snapped origin, so no absolute edge ever sat on the module. The claim was
true of the code's intent and false of its output, and only a measurement
taken from the geometry rather than from the picture could have told the
difference. Findings 3, 14 and 19 from a fourth angle, and the first time in
this lab that the instrument has corrected the artwork rather than the prose.

**39. Finding 7 replicates on a document it was never taken from.**
Finding 7's census, the largest single claim in this lab, came off the k10k
front page of `2000-03-01`, which is not in `corpus/cache` and cannot be
re-read offline. `text-as-gif` censuses the `1999-05-06` capture instead, which
is: a different document, ten months earlier, in the cache the whole time.

| | 2000-03-01, finding 7 | 1999-05-06, this study |
| --- | --- | --- |
| images | 40 | 39 |
| spacer GIFs | 8 | 8 |
| `<font>` tags | 2 | 2 |

The spacer count and the font count are identical and the image count differs
by one. The same page, built twice, ten months apart, with the ratio of
pictures to type unmoved. That is the strongest form this claim has been in,
and it cost nothing: the second document was already on disk.

What the study adds beyond the replication is the number that needs no
inference at all. **One image of the thirty-nine declares alt text**, and it is
not any of the three navigation buttons, not the logo, and neither of the two
files whose own names contain `typo`. The constraint forced the pictures. It
did not force the silence, and the silence is what decides the verdict.

The study also splits measured from inferred on its own face, because it has
to: markup can tell you an image is 134x16, but nothing in the markup can tell
you whether it is a picture of a word. So it states its headline twice, once
counting every row it infers to be text and once counting only rows whose own
filename says so, and the claim survives its weaker reading. Filenames are
readable because the site was Danish: `ikontekst` is icon text, `tom` is empty,
`baggrunde` is backgrounds.

**39b. And the drawing lesson repeated in prose form, for the third time in
one session.** The study's first run asserted that the text images "occupy a
band with a ceiling", which sounded right and was half true. The generated
numbers said ten of nineteen non-text images cleared the top of the band, which
is barely more than chance and does not support the word band. The real signal
is one-directional and is now what the page says: **no text image is taller
than 54px and everything above 54px is a picture, but the converse fails**,
because icons at 21px and rules at 4px live inside the same range. Finding 9's
rule is not enough on its own. Generating the number does not stop you writing
a sentence the number does not support, and the sentence has to be generated
too.

**40. One handler, three channels, three different endings. The lab had been
averaging them.**
Findings 20 and 23 counted k10k's sixteen rollover handlers and named their
three feedback channels. `status-caption` measures the shape of channel C and
finds the counts were hiding the interesting part:

| | channel A + B, the image swaps | channel C, the status text |
| --- | --- | --- |
| handlers that use it | 14 of 16 | 16 of 16 |
| `onmouseout` handlers that undo it | 14 | **0** |
| verdict | ABSORBED | **REVOKED** |

Two things follow. First, **two links on that page had the caption as their
only hover feedback**, so when the capability was withdrawn they did not
degrade, they went silent. Second, and this is the part no count could see, the
picture channel was a **state** and reverted, while the text channel was a
**message** and persisted. Nothing ever cleared the bar. Described plainly, in
period terms and with no hindsight, that is: a page-controlled line of text in
browser chrome that outlives the pointer. Which is exactly why it is not
available now.

There is a typographic finding sitting inside it too. All sixteen strings open
with the same two characters, `-> `. The status bar was drawn by the operating
system in a font, size, colour and position no author could reach, so the only
typographic decision available in it was the content of the string, and
somebody made it. On a page where nearly every other word was a hand-set
picture, these sixteen were the only type the designer could not touch, and
they are also the warmest writing on the site: "Little Computer People
Webcam!", "Job offers in Denmark?!". The voice went into the cheapest channel,
and the cheapest channel is the one that no longer exists.

This is the third REVOKED study and the first that is not sound, which was the
reason for building it. Findings 24 and 29 built that verdict on a chromeless
popup and on unmuted autoplay, which left it looking like a verdict about
media. It is not. It is a verdict about capabilities a vendor withdrew, and the
clearest case in the cache is text.

**Finding 23 applies at full force and the study says so on its face.** The
2000 2Advanced capture writes to `self.status` zero times, because
Dreamweaver's stock rollover had no idiom for it. So this is craft at one site.
The study's claim is narrower than the era and is not about k10k at all: the
capability existed, it was used well at least once, and browsers removed it on
purpose.

**41. The web-safe 216 could not draw the era's dark palette, and the lab has
been citing it as that palette's constraint.**
`NOTES` has described the colour area as being about the 216 since the first
draft, and the palette is invoked five more times across the lab as the reason
for a duotone, an ordered dither, a flat plate. Nobody had ever checked the
lab's own measured colours against it. Doing that turns the description over.

**Five of the nine measured colours in this repository are outside the
palette**, and the earliest document in the set, 1999, is already outside it.
This is not a drift that happened as displays improved.

The reason is arithmetic, not attitude, and it is sharper than expected:

- k10k's measured ground `#404642` and panel `#2f3431` **both snap to
  `#333333`**. Inside the palette they are not two colours. The separation
  falls from 1.31:1 to 1.00:1 and the panel stops existing.
- Of 216 colours, 41 sit below 10% luminance and **only two of those are
  neutral**, `#000000` and `#333333`. Inside finding 8's measured 3% to 6%
  band there is **exactly one** neutral, `#333333`. The other thirteen in that
  band are saturated blues, purples and browns.
- A ground and a panel need two. There was one.

So the two-value dark system that `two-value-dark`, `dark-studio` and
`palette-split` are all built on is **undrawable in the web-safe palette**, and
leaving the palette was not a stylistic choice, it was the only way to get a
second dark neutral. `websafe-216` demonstrates this by drawing the same layout
twice, and by quantising a ramp between the two measured values: 34 distinct
colours as designed, **one** snapped to the palette, five with an ordered
dither. That last number is why the dither existed, and why the flat panel won
instead, since the dither costs a visible grain at exactly the size a 10px
caption sits on.

A constraint abandoned while it is still technically binding, because it cannot
express the work, is the cleanest ARTIFACT case the lab has.

**The limit is finding 32 again.** Three of the four documents are studio
sites, and the vernacular wing has no measured colour anywhere in this
repository and may well have obeyed the palette completely. The one document
here that does obey it exactly, 2Advanced in 2000, is also the lightest thing
the lab has measured, which hints that compliance and lightness travelled
together. That is a hypothesis resting on one document, and it is labelled as
one. The arithmetic half does not depend on the sample: the palette contains
what it contains.

**42. The image map's death is demonstrable, and the first instrument built to
demonstrate it was lying.**
`image-map` fills a hole: the technique had zero mentions anywhere in this repo
despite being, for several years, how a Photoshop comp became navigation. An
`<area>` is specified in absolute pixels counted from the image corner, and
nothing in the mechanism knows the image has a display size.

The study ships a real `<img usemap>` with a real `<map>` and five real
`<area>` elements, plus a scale slider, and hit-tests the result. At 75%:

| link | its coords cover artwork px | what answers on its own label |
| --- | --- | --- |
| Products | 11 to 168 | Products |
| Services | 179 to 336 | Services |
| Support | 347 to 504 | **Services** |
| Company | 515 to 672 | **Support** |
| Contact | 683 to 843 | **Company**, and Contact is off the image entirely |

Two of five labels still answer to their own link, the hotspots cascade off by
one, and the last one walks off the edge into space that is no longer part of
the image and cannot be reached at any point. The failure is silent in both
directions: neither the picture nor the coordinate list can detect that they
have stopped agreeing.

**The first instrument was wrong and would have printed a confident invented
number.** It asked each `<area>` for its own `getBoundingClientRect`, which
returns all zeroes in Chrome, so every row would have reported a drift exactly
equal to its own position. That was caught by testing the instrument against a
known case before trusting it, rather than by looking at the output, which
looked entirely plausible. It is the same lesson as finding 38 from the other
side: there, the measurement was right and the drawing was wrong; here, the
drawing was right and the measurement was empty. Both look like a working page.

**PROVENANCE is DESIGNED and the measured negative limits it rather than
supporting it.** Zero `<map>`, `<area>` or `usemap` across both captures. That
is a real negative, but it is two studio pages, and image maps were a
vernacular idiom. Unlike the negatives in `ambient-audio` and `form-controls`,
this one is a gap in the corpus and must not be read as a fact about the era.
This is the first negative in the lab that argues against its own study, and
recording that distinction matters more than the study does: **a measured
negative is only evidence where the sample could have contained the thing.**

**43. The audit page exists, and its first requirement was the right one.**
`README` has carried this line since the browser-traps section was written:
"No audit page exists yet; when one lands, this is its first requirement." The
requirement was that `incomplete` must be printed beside `violations`, because
axe returns incomplete rather than a pass when it cannot resolve a background,
and this era is built on background images.

`audit.html` meets it literally and goes one step further: **it prints no total
score at all.** The two columns are never summed, because a single number is
precisely what would let a study with many unresolved background checks outrank
one with a real failure. The page also honours the other two traps in that
section by construction. Frames are audited only once they have actually
intersected the viewport, and the counter reports done, failed and remaining
rather than done alone.

The third trap turned up during the build, in the audit page's own harness. In
a headless shell the `IntersectionObserver` never fires at all, so the page sat
at zero forever and looked exactly like an audit in progress. That is the
success-only counter failure wearing a new costume: **"nothing has happened
yet" and "this will never happen" rendered identically.** The page now gives up
after six seconds and says plainly that nothing was audited and why, and it
still refuses to audit frames that never drew, because a clean report from an
unrendered page is worse than no report. A per-study timeout went in for the
same reason, so no single page can stop the run without being counted as the
failure it is.

**43b. And then it ran, and the prediction was right about the mechanism and
wrong about the ranking.**
First full pass, 41 studies, axe 4.10.2: **178 violations against 22
incomplete, and not one study clean.** The README's worry was that the lab
would be flattered by the incomplete column. It is not, because it fails out
loud first.

What the run actually found, in descending order of how much it says:

- **`html-has-lang` fired on 41 of 41.** Every study in the lab opened
  `<!doctype html>` and went straight to `<meta charset>` with no `<html>`
  element at all. One attribute, forty-two files, forty-one violations, and
  nobody had noticed in twenty-three commits. **Fixed**, and the total fell to
  134 in the next run. This is the single most useful thing the page has done
  and it is exactly what a lab-wide audit can see that a per-page score cannot:
  a defect that universal is a property of how the studies are built.
- **`landmark-one-main` and `region` also fire on all 41** and are not fixed.
  Wrapping every study in `<main>` and landmark regions would be importing a
  2014 document model into pages whose whole argument is that the era had no
  such model. That is a real tension between the audit and the thesis and it is
  left standing rather than resolved quietly.
- **`color-contrast` appears in BOTH columns**, 406 nodes decided against and
  172 undecided. This is the case the two columns exist to separate: the same
  rule, part failed and part never resolved, so no study can be called clean on
  it in either direction.

**The contrast failures are in the lab's own furniture, not in the era
vocabulary it reproduces.** The failing nodes are captions and labels: 6pt
`.label` at 3.99:1, `.src` caption text at 4.03:1, both a hair under the 4.5
threshold, and both belonging to the study chrome this lab wrote rather than to
anything measured off a 1999 page. The four new studies were corrected. That
result is worth keeping because the opposite was assumed: the era's body text
was white on `#404642`, which is 9.7:1 and passes comfortably. **The dark
studio palette is not the accessibility problem. The tasteful grey caption
underneath it is.**

And one correction to the README's own prediction. It expected `incomplete` to
be caused by unresolvable **background images**. The actual reason axe gives
here is *"partially obscured by another element"*, which is overlap, not
backgrounds, because these studies draw with CSS rather than shipping sliced
GIFs. The requirement was right, the column earns its place, and the mechanism
behind it is not the one that was written down.

---

## Open questions

- **Does `SURVIVES` for `dark-studio` hold, or is it flattered by novelty?**
  The claim is that stripping the era markers leaves a disciplined two-color
  system. That needs a control: the same content laid out with no era markers
  at all, to see whether the discipline or the nostalgia is doing the work.
- ~~**Is there a fourth verdict?**~~ **Answered.** ABSORBED exists and
  `rollover` is filed under it. What the build added to the idea is that
  absorption is not all-or-nothing: one technique split into three channels
  absorbed at three different rates, and the verdict now has to name *which
  part* was absorbed. The fold is still an unfiled candidate.
- **Does SURVIVES need splitting?** Finding 18 says a technique can survive
  because it was excellent or because it was self-effacing enough to have no
  surface to age, and those still share one label. Eight studies carry
  SURVIVES; `foundry-store` and `community-board` clearly took the second
  route, `two-value-dark` and `tdr-numeric` the first. This is the next
  vocabulary decision and it is deliberately not made yet, because the two
  candidates so far were both reached by accident and a third case would say
  more than a taxonomy invented at a desk. `palette-split` is now a possible
  third for the self-effacing route: a ground and one step off it has no
  surface to age because it expresses nothing. But it was built on purpose to
  examine a palette claim, not reached by accident, so it is a weaker witness
  than the first two and the question stays open.
- ~~**Is `revoked` a fifth verdict?**~~ **Answered, and the last candidate on
  the watch list is what answered it.** That list named autoplaying sound, and
  `ambient-audio` is now built and measured. REVOKED is in `VERDICTS`,
  `popup-window` is refiled from ARTIFACT, and finding 29 records what the
  refiling actually bought. **`status-caption` is now the third**, and it was
  built specifically because the first two were both sound and left the verdict
  looking like a claim about media. It is text, it is measured, and finding 40
  records it. The two candidates that remain from the original list, framesets
  and `document.write`, were looked at again and **neither qualifies**: both
  still work in every current browser, so the verdict's own test, that the
  technique would still work if the capability had not been withdrawn, has
  nothing to bite on. `document.write` is the better of the two and is at most
  CONTESTED, since what happened to it was an intervention against injected
  scripts rather than a withdrawal of the capability. Leaving them unbuilt
  rather than forcing them into the slot is the same discipline finding 24 used
  when it refused to add the verdict on two cases.
- ~~**GreyscaleNet: what year, and what is it?**~~ **Answered 2026-09-02.**
  `greyscale.net`, archived 2001-04-20, filed by the museum under Golden Age of
  Web Design / Design & Art / Techno. Read from a tab Erin already had open,
  which cost the museum nothing and is the arrangement finding 17 settled on.
  It raises a bigger question than it answers, recorded as finding 32b, and it
  is now the corpus run's first target.

- **How much Flash can this lab honestly cover without running Flash?** Ruffle
  is MIT/Apache and would run period `.swf` in the browser, but hosting other
  people's compiled work is a different decision from writing a study about a
  technique. Undecided, and deliberately not started.
- **Weight budget realism.** The harness compares against 60 KB at 5 KB/sec.
  That is a defensible average, not a measured fact, and the studies are all
  well under it because they carry no images. The budget only starts to mean
  something once a study is image-built.

---

## Deliberately not doing

- **Cloning real sites.** Every study isolates a technique, written clean. The
  study is of the technique, not of anyone's page.
- **A shared design layer.** See the house rules. The absence is the argument.
- **A component library.** This is a study lab. If it starts growing a
  publishable kit, it has stopped being one.

---

## Attribution

The designers and studios that made this era are named in prose, in this file
and in study comments, as historical reference. No study reproduces anyone's
work, and no vendored asset ships without its licence recorded in
`vendor/fonts/LICENSES.md`.