` on **149 documents** now, and the
six links lost a `role="menuitem"` they never earned: that role promises arrow-key semantics this
product does not implement, which is a claim made to assistive technology rather than a description
of what is there.
Four documents had no level-one heading: the event detail's loading state, signed in and signed out,
in both trees, where the heading went out with the question a skeleton replaced. Those pages already
carried a screen-reader line saying the event is loading, so it is the heading now, wrapped in the
live region rather than carrying the role itself. **One string, both roles, no new class and no new
rule.**
Twelve documents carried a screen-tree panel one row short of their tree, and the two trees disagreed
about which. The missing rows are in. Read as a set afterwards: **one panel in all 119 grey documents
and one in all 120 painted**, 0 differing in either tree.
Nine rows of the copy inventory still said `Portfolio`, a word the lexicon bans by name, while the
trees said `Balance` on every screen. The sweep of 2026-08-20 swept the trees; the file that is
supposed to be the source of truth for every string was not in it. It is turned, from the render
rather than from memory.
**Measured after: 1,434 renders over both trees on both engines at three widths - 0 sideways, 0
duplicate ids, 0 page errors, and 239 of 239 documents carrying exactly one level-one heading.**
**Six rows stay open and every one of them needs a decision rather than a sweep**: the stand page that
argues with its own computed panel, the grey tree's shared-region numerator that no file defines, the
278 MB of history around a 60 MB package, the tracked file named in another language that nothing
references, the four hand-written renderings in `ia/` that are behind their markdown, and the shipped
sentence that tells a reader how long their money takes when no document states a duration.
## 2026-08-23 - Handoff: the package was audited by three readers who had never seen it, two of them built a feature out of it, and the one thing that could not be got right was the test itself
**The stage's premise is that whoever builds a thing cannot read it, only recall it**, so "imagine
you are a new developer" produces an audit by somebody who already knows every answer. The reader
here is a subagent with a clean context, and it stands at both ends: one at the entry, asked to
UNDERSTAND the package, and two at the exit, asked to USE it by building a real feature from the
documentation alone. Every run carried the same three isolation rules - read from disk because a
file already in context is somebody else's snapshot, forbidden paths named ONE BY ONE rather than as
a folder because the critique logs sit inside the folders the reader was just handed, and a reading
journal as the proof. **No forbidden path was opened in any of the three, and two readers disclosed
a near miss unprompted**, which is the behaviour the journal exists to produce rather than the
absence of one.
**What the package is.** Four questions have no answer anywhere else in the repository, because a
page renders a state and cannot say what reaches it, what it is made of, or what it promises a
reader who cannot use it the ordinary way: `behaviour.md`, `map.md`, `a11y.md`, and `one-shot.md`
for how to add to it. `handoff.html` is the one link that replaces an explanation and `index.html`
is the front door. **Nothing in the package repeats a fact the code carries** - no colour, no
length, no shipped string, no rule - and that is checked mechanically and reads zero, because a copy
of a fact is a second edition of it and the copy nobody re-reads is the one that goes on rendering
yesterday's truth with a confident face.
**The route was dead where it mattered most and no instrument here could have seen it.** Asked
rather than remembered, the published product answered, the vitrine answered, every deep link
answered, and **the one address a person is actually handed answered 404**. `.nojekyll` is required
because Jekyll drops every path beginning with an underscore and would take the roadmap panel off
the course documents and the whole route off the stand; with Jekyll off nothing makes a page from
the README, and there was no `index.html`. Every instrument in this repository opens a file and not
one reads a response code. The front door holds no fact of its own.
**The examinations, and the fork about their screens.** The first built account limits, five states
in both trees, 261 files. The second built the cancelled event and its refund, four states in both
trees, 262 files. Both are correct, both were verified independently before anything was decided,
and **both are reverted and kept on branches**: `exam/account-limits` and `exam/cancelled-event`.
The reason is the stage's own and it was put to the user rather than assumed: **a product edit after
the tree was measured cancels every comparison the acceptance stood on**, and a handoff that
extends the product is no longer a handoff of the product that was accepted. The work is good and
that is why it is addressable rather than deleted.
**What the readers found that nothing here had.** Twelve documents whose screen-tree panel names a
smaller set than the rest of their tree, with the two trees disagreeing about which - found twice,
independently, by two readers who never met. Four documents with no level-one heading at all. The
copy inventory still carrying a word the lexicon bans by name, on eleven rows, while the trees carry
it on none. All four hand-written renderings in `ia/` behind their markdown. The layer boundary
counted two ways in five files, one of them the file every session loads whole, **which said six in
one bullet and seven in another**. A shipped sentence telling a reader how long their money takes
with no duration in any product document. Thirteen open rows, and every one of them filed rather
than fixed.
**And one row of the accessibility checklist was the stage's own defect.** It said confirmed by an
earlier stage, named the instrument and did not run it, on the sound argument that headings are
structural and this stage does not read the structure tree. The argument was sound and the status
was still unearned: the four missing headings had been there the whole time. **A status inherited
from an earlier tree is a status about that tree**, and the rule this file already had - a point
whose check cannot be named never gets a status - needed its twin, which is that a point whose check
was not RUN does not get one either.
**The one repair, and where the line was drawn.** The package published an accessibility checklist
on a page that failed the text contrast criterion: the course chrome's muted role measured 3.36
against the page ground, under the ink threshold, carrying the table headers and the small print of
both pages this stage wrote. It is one value in a palette fourteen documents copy, so fixing two
would have been drift by this repository's own definition; all fourteen were swept and seven are
clean now. **The remaining 301 elements are greys written as literals inside individual documents,
past their own token**, and they are filed rather than swept because that is a judgement per
document. The line is by KIND of file: the course chrome is not `ui-visual/`, `wireframes/`,
`components/` or the stand, links neither the system nor its base, and no product rule reaches it.
**The rule the stage set itself and did not meet.** The second examination's list of things it could
not work out had to be shorter than the first's. **It was ten and ten.** Split by who owns the gap
it went from five to two, and none of the five the first run closed came back, which is the direct
evidence the repairs held - but the rule as written was not met and the log says so rather than
explaining it away. **The reason it cannot converge is written down too**: six of the second run's
ten are product decisions nobody has ever taken, and a third run on any unbuilt feature would
surface a seventh such set. The package documents a design, and a design is not a decided product.
So the measure a later run should be held to is the split by owner, not the length of the list.
**Rejected: cutting the version tag at step 6, where the stage's own flow puts it.** The handoff
page linked a file step 7 had not yet written, so a tag there would have named a package failing its
own route check. It is cut on the closing commit instead.
## 2026-08-21 - The row already carried its own answer, and a cap that scales with the reader had no upper bound; then a notification that could not be made true turned out to be a catalog two months behind itself
**Backlog 230, and the fix added no declaration.** `Sign in` and `Sign up` in the logged-out header
are `data-open="signin"` **both**, they open one dialog, and that dialog's heading is *Sign in or
create account*. A logged-out phone reader had **five controls onto one dialog**: these two and all
three thumb-bar slots, whose fourth is labelled `Sign in`. So the quiet half takes `.desk-only`, the
class its two siblings in the same row already wear and which `base.css` supplies - the same move,
for the same reason, as the logged-out bell, whose argument is written six inches up the same file.
**76 buttons in 74 documents.** The row asks **239.81 against the 292 it can give** at a 24px browser
default where it asked 307.78, so it has 52.19px of slack instead of a 15.78px deficit, and `Sign up`
stands at its natural 90.77 rather than being shrunk to pay for the pair. **The point is the slack**:
the sentence this fix replaces proved itself with *the utility landing on 306 exactly*, and a row
with no slack is a row the next unmeasured setting breaks again.
**And the header fix left ONE render of 2,880 still scrolling, which was a different defect
entirely.** `responsible-betting.html`, 320px, 24px root, 23px over. The culprit is `--measure`.
`.read-col` is capped at `max-width:var(--measure)`, `46ch`, and a `ch` scales with the face - which
is the whole reason the token is in `ch`, so one number caps an 11px legal line and a 16px paragraph
at the same character count. **It had no upper bound in pixels.** The five document pages are
`--text-16`, so at a 24px default the cap computes to **755.14px against the 258px a 320px phone can
give**, and `.cat-layout` is a COLUMN flex container where `.cat-main`'s `min-width:0` governs
nothing at all. **A cap wider than its box is not a cap, it is a floor.** The fix is
`max-width:min(var(--measure),100%)`, one declaration, and it costs nothing anywhere the cap already
fitted: measured at 1440, **503.4px at 62 characters at a 16px root and 755.1px at 63 characters at a
24px root** - the same count at both, which is exactly what `ch` was bought for. **0 of 960 renders
scroll sideways at all three roots on both engines**, against 36 documents before.
**The general shape: a limit written in a unit that tracks the reader needs a second limit written in
the container.**
**Backlog 231, and the notification was the symptom.** The string could not be edited into truth and
was not edited: **the catalog's recurring instances had never been moved to the period the product is
in.** A WEEKLY market whose open period ended **58 days out** is not weekly; a MONTHLY one ending
**91 days out** is not monthly. Worse, `event-detail-recurring-multi.html` listed July and August 2026
as *already resolved* in its `Earlier in this series` block while the account cluster dates the
reader's today to early July - **two todays in one product, three months apart, and the same monthly
series carrying four instances with two incompatible histories.**
**The account cluster wins on weight**: wallet, active bets, history, both profiles, win, loss and
every notification timestamp, against one block on one document. So today is pinned at **Fri, Jul 3,
2026** - a Friday, which is what the weekly market's own `why` line already assumed when it said
*Friday's close decides it*. Every instance moved to its current period: the weekly Bitcoin and
Ethereum markets read `in the week to Jul 3` and **close today**, the monthly stablecoin and
temperature markets read `July 2026` and close Jul 31, and the series' earlier periods slid back to
June and May. **The reader's own resolved bet read `at the end of H1 2026`, which is not a period a
monthly series HAS**: it is the June instance now, resolving Jul 1 exactly as its row already said,
which makes the reader's loss the period immediately before the open one and puts it in the series
list where a reader can find it. 31 files, both trees, the search index, the kit and the IA.
**The notification now names the weekly market that closes today and routes to it**, and its sibling
was fixed the day before: `Will ETH flip BTC by 2027?` existed in 0 documents outside the two
notification screens. **And the firing horizon is written down for the first time.** Three of the
four notification types are triggered by an EVENT - the odds moved, the market resolved, a category
gained a listing - so their trigger is the thing itself. `event deadline approaching` is the only one
triggered by the CLOCK, which means it needs a number, and for as long as it had none its one
specimen said *closes in 6 hours* about an event 13 documents dated a year out. It fires **once per
instance, 24 hours before close**, and 24 rather than longer because **the horizon must be shorter
than the shortest period the catalog opens**, or a weekly market spends a seventh of its life
announcing that it is about to close.
Verified after both: **600 documents at five widths on Chromium and WebKit with the coarse pointer -
0 page errors, 0 duplicate ids, 0 sideways scroll** - 51,605 hrefs with 0 unresolvable, and the
recurring set read across all three trees showing one cadence, one period and one close date per
series.
**And one thing was found, repaired and then UNREPAIRED in the same hour, which is the more useful
half of the day.** The verification pass counted table cells across all 61 markdown files and
reported **12 rows carrying more cells than their header**, against `docs/backlog.md`'s own claim of
*0 disagreeing with their headers in both directions*. The synthetic control agreed with the premise:
a two-column header with a four-cell row renders **two `` and the third and fourth cells vanish**.
So the twelve were folded into their last legal column - and then the actual rows were put through
GitHub's renderer before the fix was believed. **HEAD and repaired both render three ` `, both
3,623 characters in the second one, and every word of the surplus is present in BOTH.** Those rows do
not split where a pipe count says they do; the excess delimiters render as literal `|` inside the
Item cell and the Note column comes out empty. **Nothing was being lost, and the repair had turned a
rendered `|` into a rendered `\`.** All twelve are reverted. The claim in `docs/backlog.md` stands
and the detector was the thing that was wrong: **a pipe count is a reading of the source, and this
file's own rule is that the record has a renderer too.** The control proved the MECHANISM exists and
said nothing about whether these rows were in it, which is the difference between a positive control
and a finding.
---
## 2026-08-21 - Reading every page as a reader, which is the one instrument this repository had never built, and the first thing it found was the last line of every page
Every check here renders a page and asks whether it is CORRECT. None of them reads it. So: **120
painted documents rendered with every `` and `` forced open, every visible text node
taken in reading order, 1,727 blocks, 339 distinct once the repeated chrome collapses.** Six
findings, and **every one of them renders perfectly**, which is why 600 renders over two engines pass
them all.
**`Copyright (c) Yonder. Sample wireframe content.` stood on 221 documents** - 119 painted, 100 grey,
2 in the kit - as the last sentence a reader meets. `voice/docs/microcopy.md` had decided to keep the
footer placeholders because *they are honest grey-box stand-ins for content that is genuinely not set
yet*, and **that reason holds for two of the three and not for this one**: the copyright holder IS
set, the line already names Yonder, and `Sample wireframe content` stands in for nothing. It is a
label about the ARTEFACT fused to the product's own legal sentence. The regulatory strip directly
above it left 100 grey documents on 2026-08-20 for the same reason and this one was not looked at.
`about.html` already carries the honest version, as its own labelled sentence.
**The reader's balance renders as two numbers and the wrong one outnumbers the right one 149 to 5.**
`wallet.html` derives it - *Balance = Cash + In-play*, `$92.00 + $40.00 = $132.00` - and the three
open stakes on `active-bets.html` sum to exactly `$40.00`. The thumb bar printed **`$142`** in its
Profile slot and in its `aria-label`, visible on **149 documents at 390** against `$132.00` on 5. The
header pill was corrected on 2026-08-20; the bar three inches below it was not. **A fact corrected
where somebody remembered instead of over its set is a fact that now disagrees with itself in one
more place than before.**
**Three open positions, three `Price now` figures, and all three contradicted the rest of the tree.**
The shutdown market reads 38 per cent on the feed, on Favorites, on the detail hero and in the
chart's `now` label, and My Bets said **41**. Bitcoin reads 61 everywhere and My Bets said **78** -
and `notifications.html` states in words that it *moved from 58% to 61%*, so the product told the
reader the price in a notification and contradicted itself on the next screen. The UK card puts
Conservatives at 33, which is NO at 67, and My Bets said **54**. Every `Potential payout` was
correct, computed off the average price, so the arithmetic was right and only the current price was
wrong. **Bitcoin's average moved to 58 rather than its `Price now` moving to 61**, because the
notification is a witness to what the reader paid: that notification only exists because they held
the bet before the move.
**`On-chain proofs 100%` survived backlog 217 by sitting in a figure band.** 217's rule is exact -
every resolution claim points at the Reading, every custody or settlement mechanic keeps the chain -
and a figure headed *Proven, not promised*, between *Events resolved 1,284* and *USDC held 1:1*, is a
resolution proof. **Its own heading says `Every figure here is a count you can check`, and two of the
three figures were `100%` and `always`.** The band now carries the set `about.html` already
publishes: 1,284 resolved, 25 open right now, 2 challenged and re-read. Three counts, two documents,
one answer.
**A notification named an event that exists in 0 other documents**: `New in Crypto: "Will ETH flip
BTC by 2027?"`, on 6 placements, in no catalog, no feed, no search index. It names a real Crypto
event now. **Its sibling could not be fixed and is filed as backlog 231**: `Closing soon ... closes
in 6 hours` on 4 documents against `Closes: Jul 1, 2027` on 13. The fixture's today is about Jul 1
2026 and **the soonest close in the entire catalog is Aug 28 2026**, so no event exists that this
notification could name and stay true. Either a catalog event gets a near date or the type's firing
horizon is a product decision nobody has written down.
**And then the set read, which found four more that reading alone would have missed.** Every `.pos`
row in all three trees, grouped by its question, requiring one distinct reading:
`ui-kit/position.html` - the component's OWN page - carried a `$10.00` stake at a `45%` price that
appears nowhere in the product, a WON Bitcoin row wearing **the shutdown market's `+$13.16`** where
the history says `$31.80`, and two rows labelling `Your side` twice where every sibling labels side
and result. `position-list.html` and `patterns.html` put one product row, the UK election NO on
Conservatives, under **two different market names, neither of them the UK election**, and flipped a
YES to a NO. **A specimen holds a quotation or a demonstration**, and a row naming a real market with
real money is a quotation. Every value in the position set now agrees across the three trees; the two
groups that still read as two are a compact face carrying FEWER figures, not different ones.
**One false positive, caught by the rule this file already carries.** The first pass reported `No
events match ` with no term on 119 documents, and it is a template with an empty `` that the page
script fills, `hidden` until a query returns nothing. `getBoundingClientRect` is not a visibility
test and neither is a DOM walk: re-read with `checkVisibility()` the string is visible on 23
documents and every one of them has its term. **A pass that reads more than the reader sees
manufactures defects at exactly the rate it finds them.**
Verified after: 600 renders over Chromium and WebKit at 390 with the coarse pointer, **0 page errors,
0 duplicate ids, 0 sideways scroll**, and 51,605 hrefs with 0 unresolvable ones, the single flagged
string being the escaped `` example this repository has already filed once.
---
## 2026-08-21 - Five files printed four different registry counts and none was the disk, and a header tuned to land EXACTLY on its container met a reader who enlarges type
**The registry is 36 width queries, measured from the comment-stripped source of all 56 stylesheets.**
Five live files were printing **33, 34, 35 and 37** on the same day. `ui-kit/responsive.html`, the
page whose whole subject is the ladder, had the badge `3 rungs in rem, 33 width queries` at the top
and a section heading `Why the rungs are in px` four screens below it: **one rendered document, two
answers, and the reader who scrolls meets them in that order.** Backlog 135 put the rungs in `rem`
on 2026-08-13 and that page said otherwise for eight days.
**The total barely moved and that is what hid two edits.** 37 on 2026-08-17, 35 on 2026-08-18, 36
today: `options.css` gave up **two** at the desk on 2026-08-20 when the slippage ladder that stood
on 0 placements went with them, and `patterns/browse-shell.css` took **one** at the rail for the
document reading column. The two moves nearly cancelled, so nothing about the total looked wrong. A
per-file diff against the tree of four days ago is what named them. **A count that changes by one is
harder to see than a count that changes by four**, and this repository's rule is already the answer:
a count is COMPUTED, or it is DATED and says the day.
**The compound check the name-level idle control cannot run.** `ui-kit/_page.css` declares 310 `tk-`
classes and the kit markup stands 310, both sides zero, which is what every idle control here reads.
**Thirteen of those classes are styled only inside a COMPOUND selector**, and `.tk-bad` is one:
`.tk-lab.tk-bad` is the whole of its styling. Two paragraphs in two pages carried it alone,
`navitem.html` as `tk-note tk-bad` and `responsive.html` as `tk-why tk-bad`, so each asked for the
warning colour and rendered as ordinary prose. Both are the exact failure `_page.css` documents nine
lines above them, which it thought it had closed by fixing the descendant form. Read over all 61
pages with a planted control: two, and they are the whole harvest. Both deleted rather than promoted,
because the rule three lines up already says prose marks its defect in the sentence.
**The root-font sweep, re-run because the tree outgrew it.** The sweep that closed backlog 135 ran at
the browser default over 105 screens, before the five document pages existed and before they moved
to `--text-16` on 2026-08-19. Re-run at **2,880 renders, 120 documents, three defaults at eight
widths each**, with the rungs read at their MOVED positions - DESK is 640 at 16px and **960 at 24px**
- and four controls proved first: `10rem` gives 160 / 200 / 240; `(min-width:56.25rem)` is true at
16px and **false at 20px in the same 1000px window**; the same root twice differs in 0 of 40; a
planted 2000px box takes the overflow from 0 to 1610. `pointer:coarse` asserted in the page on the
second pass, and the two passes name **different culprit families** on the same 36 documents.
**0 sideways at root 16 and root 20 at every width. 36 of 120 at root 24, at 320 alone**, by 10px on
24 documents and 23px on 12. All 36 are logged out. **The cause is the wordmark and the row's own
record says the opposite**: `components/header.css` closes its longest comment with *the cluster
falls to 193.14 with the utility landing on 306 exactly*, which is exact at 16px and exact at 20px.
`.logo` is `--logo-size:var(--text-16)`, a rem token since backlog 115, so `.left` goes **86.09 to
103.61 to 121.14** while the two icon marks stay on the 44px touch floor, which is px and does not
move; the auth pair absorbs the difference until it cannot. **A row measured to land EXACTLY on its
container has zero slack, and a proof taken at one setting is a proof about that setting.** The
branch that block asserted was the pointer and the one it did not was the root, which is the same
class of error as the fine-pointer walk the block itself corrects three paragraphs earlier.
**The five document pages are the finding that is not a defect**: 60 to 88 per cent taller at a 24px
default against a 20 per cent median across the tree, `terms.html` from 12,158 to 22,811 at 390, and
every word carried. They are the only surfaces here that are nothing but prose read end to end, so
they are where the type scale earns what backlog 115 bought.
Filed as `docs/backlog.md` 230 with three answers and **not fixed**: the sweep was run to produce a
number, and what gives way at 320 on a 24px default is a product decision.
---
## 2026-08-20 - The one rule set nobody had ever measured, and the two biggest hits were a word the lexicon bans and the rule three lines above it
**Every instrument in this repository reads a computed value. `voice/docs/voice.md` is a claim about
the WORDS a person meets, and nothing had ever read them.** Three findings this week came out of
WRITING rather than checking, which is the tell: *settle* under a heading reading *resolves*, a
`docs/backlog.md` reference inside a rendered paragraph, and `Weekly Volume: $46,300`. No sweep here
can see a wrong word.
**The instrument is a render of every text node with two things attached.** Whether the reader is
**ACTING** - a control, a heading, a figure read to decide - or being explained to, because that
lexicon's own invariant is that **the ban is about PLACE and not about the word**. And an exclusion
list for what is not product copy: the grey tree's declared annotation faces and the comments a user
wrote, which `voice/docs/microcopy.md` marks *do not touch*. Proved against a planted
`Oops, buy shares in this market (T2)`, caught by all four rule families, before any number was
believed.
**239 documents, 38,748 text nodes, six distinct acting strings and 67 explaining ones.**
### The two biggest are the two nobody finds by reading a screen
**`Portfolio`, 296 placements.** The lexicon bans it by name two rows into its own table, and it was
the label on the desktop money pill, the mobile bottom slot, the profile summary, the swap control's
accessible name and the script that swaps them. `ia/docs/sitemap.md` had DECIDED it in the wireframe
pass and the two files were never read against each other. It is `Balance` now, which is the sum of
`Cash` and `In-play` - both already the product's own words, so only their total ever wore the
trader's. The bottom slot is `Profile`, because the three beside it name destinations and this one
named a figure while routing to `my-profile.html`.
**`crypto-only users`, 240 placements.** The address rule is stated three lines above the lexicon:
always **you**, never **users**. It stood in the fine print under Add funds on every deposit surface
in both trees.
### What that leaves, and what is kept on purpose
`settles` became `resolves` on 239 placements where it meant resolution, including the second largest
string in the product, *Your payout settles at the price you took*. Four more were slips written the
SAME DAY by the pass that then found them: `paid $1.00 a share`, `paid one dollar a share`,
`a separate market with its own threshold`, `swept between markets`, plus `How a market resolves` on
a chip two blocks under a heading that says *event*. **That the instrument caught its author's own
copy from an hour earlier is the argument for building it.**
**Three classes came back and are KEPT with the reason written beside them.** *A prediction market*
is the category naming itself and the ban is on `market` for the EVENT. *A YES share costs 38 cents
and pays $1 if the event happens* is the exemption working exactly as written: a mechanism block
that glosses the term in the same sentence it uses it, which is the condition that exemption carries
and which `voice/docs/voice.md` itself records was never met before. And *past deadlines settled
late* is ordinary English about politics. **A sweep that re-opens a settled judgement every time is a
sweep nobody runs twice.**
### And one thing the sweep found that is not a word
`[Regulatory / licensing line - placeholder, to be set...]` stood on **100 grey documents and 0
painted ones**, directly beneath the line that replaced it. So the two trees said different things
in one slot and the grey one said both, which `wireframes/_conventions.md` forbids by its own
definition of which tree owns copy. Deleted; what it carried that the shipped line does not is in
`PRODUCT.md` and in `terms.html` section 8.
### The second candidate came back a clean zero, and that is worth as much
`ia/docs/flows.md` had never been read against the screens. Parsed into its Mermaid node shapes:
**25 screen boxes, 57 decisions and terminals.** Fourteen boxes name a document directly; the other
eleven are states and surfaces under other names - the History tab, the category page, the search
sheet, three loading states, and a `pause` that is a beat rather than a screen. **No flow step names
a screen that does not exist and no screen sits outside the flows.** Recorded as a dated zero rather
than as nothing, because the next person to wonder should find the reading and not the gap.
### And the `[?]` that four pages now show
`PRODUCT.md` has carried **Jurisdiction: `[?]`** since 2026-08-19, and the legal drafts of the day
before made it visible: Terms 2, Terms 8, Terms 13 and half of Privacy 8 all stop at it. **A `[?]` in
a product file is an open question; the same `[?]` on a page a reader reads is a promise with a hole
in it.** This repository cannot choose a jurisdiction and says so. What it can do is stop the blank
being unowned, so the section now carries a **trigger** (the first real dollar, not the first user
and not the launch), a **decider** (counsel, and the one open item in that file whose owner is not in
this repository), and a **criterion in the order the product cares**: can a prediction market be
operated for money there, can a person be paid in USDC without the payout becoming a regulated
instrument, is a named public source read at a stated time enough to stay out of the gambling frame,
and can the excluded list be enforced at sign-up rather than at withdrawal, which Terms 8 has already
committed to in writing. **The four clauses now say when the answer arrives rather than only that it
is missing**, and the points build's own narrower version of the question - a free-entry contest that
keeps a public ranking - is named as the one that arrives first.
**Measured after: 300 documents, 33,158 internal links, 0 page errors, 0 duplicate ids, 0 broken
links, 0 wrong current-markers, and the lexicon sweep re-run to two acting strings, both of them the
category naming itself.**
---
## 2026-08-20 - The row was right about the document and wrong about the finding beside it, because it was written from one comment in a file that carried two
**`docs/backlog.md` 229 carried two things and only one of them was a defect.**
### The document, which was owed
**`event-detail-resolved-multi.html` stands in both trees.** Event Detail is **five documents per
tree** now: binary, multi, recurring, recurring multi, resolved and resolved multi, which is the two
axes crossed plus the state that ends both. It closes the half of `docs/backlog.md` 222 nobody took:
that row gave the OUTCOME family its multi half, `win-multi` and `loss-multi`, on the ground that a
multi result is an OPTION and a binary screen cannot say it. The DETAIL half was the same argument
and was never made.
**The fixture is the one market three surfaces already had to agree about.** `active-bets-history`
prints the row, `loss-multi` prints the outcome screen, and the detail is where a reader arrives
later from either. All three now carry *Which chain held the largest stablecoin supply at the end of
H1 2026?*, Ethereum on Jul 1 2026 at 00:05 ET, **$73.4B against Tron $61.2B and Solana $12.8B**, and
a reader who held YES on Solana at an average of 24 per cent and was not paid. The Reading on the
detail is quoted from `loss-multi`, not rewritten.
**And `Earlier in this series` on the recurring multi is a list again**, which is the whole reason
this was worth building. It was prose for exactly one commit, because every row would have been a
multi row landing on a binary specimen. **A record that cannot be a list says so and names the
document it is waiting for; then somebody builds it.**
### The finding beside it, which was mine and was wrong
The row said `ui-kit/card.html` describes a `+N more outcomes` row that 0 of 21 multi cards render,
that the class does not exist in `components/options.css`, and that `data-full` is read by nothing:
a stale claim about a face nobody built.
**The face was built, and then deleted on purpose.** `docs/backlog.md` 188 took the whole field back
off the desk and **189 removed the link**, both decided by the product owner on a screenshot, **68
links across 33 documents in all three trees**, with the rules out of the component and `.opt-more`
off its class list. The reason is in the file and it is a good one: the link was `--text-muted` at
12px, the same ink as the `Volume` line beneath it, on a card being deliberately made smaller, and
**a card is where a person decides whether to LOOK**.
**The row was written from the 2026-08-10 comment in `components/options.css` without reading the
2026-08-17 comment a hundred lines below it in the same file.** That is the defect this repository
keeps paying for, produced fresh by the person holding the file, and it is the fourth time in this
week that a filed row was half wrong.
**So nothing is built and the PROSE is corrected.** Three paragraphs in `ui-kit/card.html` and a
whole section of `ui-kit/options.html` described a card that grows to its full field at DESK, a
count rule, an override and a table reading *4 rows at 640 and up*. Measured from the rendered DOM at
1280: **30 multi cards per tree and every one shows exactly two rows.** The older comment in the
component now points forward to the one that overturned it, rather than sitting above it arguing for
a rule that is gone. **`data-full` is kept** with its zero named: the count is true, it is the number
the detail head prints, and its reader was removed by a decision rather than never written.
### Two things the same read found
**The multi chart claimed `now` about a market that had closed.** It is the one element that script
writes over its own markup, so a resolved page loaded with `Ethereum 62% at close, resolved Ethereum`
in the file and `Ethereum 62% now` on the screen. The single-line chart never had it because it never
writes that element. The guard reads the page's own `.ed-result` and **travels with the script, on
all 16 documents that carry it**, rather than standing in the one document that needed it.
**And a count turn in the previous commit had silently done nothing.** The script guarded its
replacement with `assert A in s or B in s`, neither branch could fail, and `CLAUDE.md` went on saying
*118 documents, 117 of them screens* beside *the grey tree is 118*, contradicting itself and the
bullet fifty lines above it. **An assert that cannot come back red is the guard measured rather than
the thing**, which is this file's own oldest rule applied to the tool holding it. Both counts are
computed from the disk now and the file says what happened.
**Measured after: 300 documents, 0 page errors, 0 duplicate ids, 0 broken links, 0 documents marking
a row that is not themselves. `docs/backlog.md` has 0 open rows, which is a statement about the
backlog and not about the product.**
---
## 2026-08-20 - Writing four documents that had only described themselves found two places where the product contradicted its own map, and neither was in the documents
**`docs/backlog.md` 228 was filed by the pass that wrote About, naming what it did not write**, and
its own text said the owner is whoever has legal counsel. **That is still true and the output says
so on the page**: every one of the four carries B21, and B21 now reads *draft for review, written
from what this product has already decided, not through legal review, nothing here is in force*. Its
old wording, that the body describes what each section WOULD decide, became false the day the
sections said something, so it had to move with them.
**47 sections, from the repository and from nothing else.** 2,168 words on Terms, 1,505 on Privacy,
925 on Cookies, 1,284 on Responsible betting, identical in both trees. Every number in them is one
the product already ships: 1.5 per cent of the stake, $1 minimum and no maximum, $10 deposit
minimum, USDC one to one in a segregated balance, the 72-hour challenge window that the
effective-date block at the top of Terms has been printing since it was written, and the two-month
notice period that block's own dates already implied.
**Three facts are open and the sections that need them name the gap rather than filling it**: the
operating entity and its registered address, the law that governs a dispute, and the list of excluded
regions. `PRODUCT.md` carries **Jurisdiction: `[?]`**, and this is what that costs on a page a reader
actually reads: three clauses that stop mid-sentence and say why.
### The two findings, and neither is in the documents
**`responsible-betting` was promising three controls the IA had already ruled out.** Sections 4, 5
and 6 described a deposit ceiling, a loss ceiling, a cooldown and self-exclusion as things a reader
could reach. `ia/docs/sitemap.md` marks the whole Responsible-play slot **reserved, post-MVP, not
built**, and adds *do not discover this late*. **A page about self-protection that lists controls
nobody can reach is the worst version of that page**, and it is the same class as the magnifier that
linked to the page it stood on: everything renders, nothing is wrong on screen, and the surface is
lying about what the product is. Section 4 opens with **Account limits are not built yet**, says what
the asymmetry will be when they are built, and gives the two things a person can do today: a limit at
the bank, which this product cannot undo, or a freeze by writing to us that we will not lift early.
**And `cookies` could not be written truthfully while the banner contradicted it.** Section 5 of that
policy commits that no advertising and no cross-site cookie is set. `cookie-consent` offered a
**Marketing** toggle, *used to show you relevant campaigns*, for cookies this product does not set,
that `PRODUCT.md` has no surface for at MVP, and that the IA never specified: it banks the banner as
*prior opt-in, reject as easy as accept, no pre-ticked, reopen from footer* and never as three
groups. **A choice about a thing that does not exist is a choice about nothing**, and it is worse
than not offering one, because a reader who leaves it off believes they have refused something. The
category is off in both trees, and it returns as a NEW one, which section 9 of the same policy says
resets the reader's answer rather than carrying it forward.
### A heading is in three places, and the claim that made that safe was wrong
Renaming two sections of `responsible-betting` reached the `h2` and the section's `aria-label` and
**not the contents row**, so the page carried *The limits you can set* in its own table of contents
above a section headed *Limits, and what to do while there are none*.
`ia/docs/pages/seo.md` said the H2 list IS the contents and the anchors are generated from it, so a
divergence is **impossible by construction**. Both lists are written by hand, they carry deliberately
different text (the H2 is numbered, the contents row keeps its number in a separate `.toc-n` span),
and nothing reads one against the other. **What found it was rendering the two lists side by side**,
which is one line and is now the check: **94 contents rows against 94 headings, four documents, two
trees, 0 disagreeing.**
**One thing the documents now say that the product cannot yet draw, and it is deliberate.** Terms
section 5 states that a cancelled event returns every stake in full and keeps no fee.
`ia/docs/sitemap.md` defers the cancelled-event refund FLOW to post-MVP, so there is no screen for
it. A document describing what happens is not the same artefact as a screen for it happening, and the
policy is decidable now where the screen is not. It is recorded here rather than left for somebody to
find as a contradiction.
**Measured after, both engines, at 320, 360, 390, the three rungs and one pixel either side, 1280,
1440 and 1600: 0 page errors, 0 duplicate ids, 0 sideways scroll on the eight rewritten documents,
the two banners and the kit page that quotes one of them.**
---
## 2026-08-20 - The Event Detail family is two axes crossed, and building the fourth corner found that a record needs a destination before it can be a list
**`docs/backlog.md` 227 was filed by the pass that closed 224, naming what that pass did not build.**
The monthly stablecoin series is multi-outcome, its card would have routed to
`event-detail-multi.html`, and that document reads **One-time event** in its own header: the
manufactured fixture contradiction the binary recurring specimen exists to avoid, one axis over.
**`event-detail-recurring-multi.html` stands in both trees now, and the family is a GRID rather than
a list of special cases**: binary and multi against one-time and recurring, four documents per tree.
That is the smallest set in which no card in the product has to point at a page that contradicts it,
and it is a better answer than four unrelated specimens because the next attribute that wants a
document can be tested against it: **can an existing corner say what this card says.**
The series is drawn. The card converted from *at the end of 2026* to **at the end of September
2026**, `data-freq="Monthly"`, `Repeats: Monthly`, closing Sep 30, on the four documents per tree
that carry it, with the search catalog re-taken and every `ItemList` re-checked against its cards in
order. **Monthly is carried by two markets now rather than by the temperature record alone.**
### What building it found, and it is the thing the binary corner never had to face
`Earlier in this series` is a **list of resolved events**, and a resolved MULTI event has no page.
The outcome family gained `win-multi` and `loss-multi` on 2026-08-20, closing `docs/backlog.md` 222
on exactly this ground - *a multi outcome is not a state of a binary one, it is the other TYPE* - and
the DETAIL half of that was never taken: `event-detail-resolved.html` is still binary. So every row
of that list would be a multi row landing on a binary specimen, which is the defect 222 closed.
**The record is a sentence in the resolution panel instead**: *the series has run twice, for July and
for August 2026, and Ethereum held the largest supply on both.* It is prose because its destination
does not exist, and it says so, which is the honest form of a block that cannot obey the convention.
`docs/backlog.md` 229.
**The H1 market was left alone and that was a decision.** `Which chain held the largest stablecoin
supply at the end of H1 2026?` stands on `active-bets-history` and `loss-multi` in both trees, and it
is tempting to read it as the June instance of this series. It is not: its window is January to June,
so it is a one-time half-year market that happens to end where the monthly series begins. Renaming it
would have invented a series membership the fixture does not have, which is the identity defect this
repository already pays for. **The series therefore starts in July, and that is legible rather than
arbitrary.**
**The chart is three lines and its ranges are `1d`, `1w`, `all`**, because a four-week-old instance
has no year to draw, the same reading that gave the weekly one `1d` and `all`. The grey tree draws
the polylines statically and the paint computes them, which is the layer split holding.
### And a third thing, found in passing and filed rather than fixed
`ui-kit/card.html` says in a `tk-note` that the last option row **names the remainder and links to
the event**, and that 18 cards in each tree carry it. **0 of the 21 multi cards render one.** The
class is not in `components/options.css`, `data-full` is declared on every multi card and read by
nothing, and the page's own specimen shows two rows and no remainder. It is the
hero-text-against-specimen shape `docs/backlog.md` 211 found on `market.html`, in the page that
records that finding. It goes in 229 with the resolved specimen, because both are about what a
multi-outcome event is allowed to say.
**Measured after: 298 documents, 0 page errors, 0 duplicate ids, 0 broken links, 0 documents marking
a row that is not themselves, and 0 of the painted documents with an `ItemList` disagreeing with its
own cards in order or in count.**
---
## 2026-08-20 - The row offered three answers and its own arithmetic rejects all three, because the formula it was built on holds at one price and the product does not sit at that price
**`docs/backlog.md` 216 said there is no maximum bet and the locked-price promise gets more expensive
with every dollar of one.** The give-away from filling a whole bet at the market's quote breaks even
against a fee `f` on a stake `S` at `b = S / (2f)`, which at 1.5 per cent is 33.3 times the bet, and
`PRODUCT.md` had decided there is no maximum, so at every `b` there is a size the product has
promised to accept at a loss.
**THE FORMULA IS THE `p = 1/2` CASE AND NOTHING SAID SO.** The give-away on a stake `S` at a quote of
`p` is `S^2(1-p)/(2bp)`, so the break-even is **`b = S(1-p)/(2fp)`** and `S/(2f)` is its cheapest
point. Solved exactly against the cost function rather than from the expansion, at `f = 1.5%` and the
product's own $5 default: **$167 at 50 per cent, $273 at 38, $672 at 20, $1,513 at 10.** The
expansion agrees within half a per cent everywhere and is always the optimistic side of the exact
answer.
**So the table in `PRODUCT.md` was reading its own number backwards.** `b = $167` is not the
break-even for a $5 bet; it is the break-even for a $5 bet at exactly 50/50, a price a market occupies
for an instant. At `b = $167` the default stake is under water at every price except the midpoint,
and the shipped catalog carries markets at 22, 23, 29 and 31 per cent.
### All three answers fail, and the third fails twice
Inverted, the largest bet fillable at the quote is `S_max = 2fb x p/(1-p)`. At `b = $300`, the row of
that table which actually makes the number a market, it is **$9 at 50 per cent and $2.25 at 20**:
below the panel's own $10 chip and below the $5 default. **A cap that bites at the default is not a
limit, it is a different product.** Capping the zero-slippage fill by size is the slippage ladder
`PRODUCT.md` rules out and `docs/backlog.md` 211 deleted. Setting `b` from the largest bet accepted is
not a solvable equation while `Bet limits` says there is no maximum, and it is not solvable in the
other direction either, because `b` must grow without bound as the quote approaches either end.
**AND ALL THREE FAIL ABSOLUTELY AT LAUNCH.** The first release ships points and **no commission**. At
`f = 0` there is no `b` at which filling at the quote breaks even: the give-away is unbounded whatever
the parameter is. A row whose three answers all divide by the fee cannot be closed by choosing one of
them.
### The fourth answer is to price the bet instead of the book
The quote is computed for the amount in the box, shown, and locked at Confirm. There is nothing given
away, so there is no size at which the promise costs anything and nothing depends on a ceiling. The
closed form is the cost function inverted: with `u = p/(1-p)`, `n = b ln(((u+1)e^(S/b) - 1)/u)`.
**Every shipped sentence survives word for word, and the sentence that was read as forbidding this is
the sentence that permits it.** `docs/backlog.md` 211 deleted the price-impact ladder on 2026-08-18
quoting *"The price is locked when you confirm, so it cannot move against you"* - and the shipped line
ends **"between the panel and the bet"**. That clause scopes the promise to an INTERVAL, not to size.
A price computed for your stake, shown to you and then held is exactly what it guarantees. **Half a
sentence deleted a face and the other half was the answer to the row that face's absence created.**
`$1 minimum, no maximum. The price you see is the price you get.` stands unchanged on 24 placements.
**IT COSTS NOTHING TO DRAW, AND THAT IS A MEASUREMENT RATHER THAN A HOPE.** At `b = 1000` and the $5
default, the size-adjusted quote rounds to the same whole per cent as the market quote on **all 19
prices `assets/search.js` carries**, from 22 to 74. It first differs at $25, by one point. So the
mechanic changes no rendered number on any screen that draws the default, and it needed no new CSS:
the row is a placement of `.line`, which the breakdown already had two of.
**The figure shown is what the figures are computed from.** The quote is rounded to the displayed
whole per cent and the payout is the stake over THAT, so $25 at 39 is $64.10 and a reader dividing by
hand gets the number on the screen. A number that does not divide is a number that was not explained,
which is `voice/docs/voice.md` principle 1. Read live from the panel after the change: $1 and $5 draw
no row and pay $2.63 and $13.16, unchanged; $25 draws `Your price 39%` and pays $64.10; $100 draws
`41%` and pays $243.90.
**And the mechanic has a placement, because a rule that cannot be rendered will be wrong the day it
first draws.** `event-detail-bet-ready.html` moved from the $5 default to **$25**, the largest amount
the panel's own quick chips offer, in both trees. It is the screen whose whole job is what you are
about to pay, and it is now the only document in the product where the market quote and the reader's
quote differ on screen: 38 on the card and the header, 39 in the panel.
**The parameter is a DATUM and it is on the element.** `data-liquidity="1000"` on the ten live panels
and the ten sheets, which is one of the three things `CLAUDE.md` allows on an element. **It went onto
129 first**, because `.bet-panel` also names the still picture inside the how-it-works dialog on 119
documents: a photograph of a panel with no input and `aria-hidden`, which the same sweep then handed a
market parameter. Restricted to the panels that contain an `.amount-input`, and the still that had
picked up a $64.10 payout for its $5 bet was put back.
---
## 2026-08-20 - Two of the three blocks a page was missing were missing from the specification instead, and the one real gap was on all five pages rather than on the one that was audited
**`docs/backlog.md` 214 said `about.html` runs the DOCUMENT body shape where the bank gives it
STATEMENT**, and named three things: B15 hero and B19 closing action declared and not built, B16 and
B17 built as prose, and a B2 breadcrumb the bank gives to DOC only. **Two of the three were
`ia/docs/blocks.md` being wrong about a page that was right.**
**The breadcrumb.** `ia/docs/pages/seo.md` section 6D says in one sentence that the four legal pages
carry `Home > Legal > {document}` and **About sits at `Home > About`**. Both trees have rendered
exactly that since the day they were built, and both schemas carry the matching `BreadcrumbList`,
verified by parsing: About is `Home > About Yonder`, Terms is `Home > Legal > Terms of Service`. So
`DOC` on that row was the error. B2 is BOTH.
**The order line.** The STATEMENT block order omitted B9 siblings and B10 contact while the table
three inches above gives both to `BOTH`, and the page has carried both since it was built. **A bank
whose order contradicts its own table will report a page as wrong for holding what the table
banked**, and that is exactly what the audit of 2026-08-19 did. Both are in the line now.
### The gap that was real was not on the page the row was about
B15 and B19 were genuinely missing. Measuring for them found the thing no row had: **B3 is
`H1 + one-line lede`, banked MVP for BOTH body profiles, and it stood on 0 of the 5 pages.** All five
went from the H1 straight into the prototype notice. The row was filed against one page and the
defect was the set, which is the same shape as the convention nineteen auth pairs of twenty obeyed:
**19 of 20 and 0 of 5 both look like nothing from inside a single file.**
**B15 is built as the STATEMENT wording of that lede rather than as a second block.** B3 gives one
line saying what this document decides; B15 asks for what the product is, in one sentence. On the
page whose subject IS the product those are one sentence, and two of the same rank stacked under one
H1 is what rule 1 of that bank throws out. **The content rule survives the merge whole**, and it is
the whole of B15: a sentence about the reader's job, not a mission statement about the company.
*Yonder is where an opinion about what happens next becomes a bet you can be paid for.*
### The body was a page about a page
Five sections, every paragraph beginning *This section states* or *It exists because*. Four now say
the thing: what a prediction market is and who it is for, how an event resolves and what happens when
the source fails or is silent, where the balance sits and what is never done with it, and the three
numbers.
**`#regulation` is deleted and it is the only section that was in NEITHER source.** Not in the block
bank, not in the STATEMENT heading list of `ia/docs/pages/seo.md`; B14 puts a region note on the
DOCUMENT profile and marks it LATER; and `PRODUCT.md` carries **Jurisdiction: `[?]`**. A section
whose subject is undecided can only ever be the placeholder it was, so it is removed rather than
rewritten, and the removal is recorded in three files so the next reader finds the decision.
**B17 had to be the BREAKDOWN, and what forced that was reading the whole page rather than the
block.** `trustbar` prints `1,284 events resolved` in the footer of this same document, so a stats
band leading on that figure would have said one thing twice on one screen. The three are **1,284
resolved since launch, 25 open right now, 2 challenged and re-read**: what has happened, what is
available, and what happened when somebody disagreed. The block's own argument is now said to the
reader too, that volume is what this category leads with and it says nothing about whether anyone was
paid.
**B19 routes to the feed and its label came from the lexicon**, which gives one label for going to
the feed: `Browse events`. The bank rules out the competitors' account CTA on a page about the
company, on the same ground that a bet panel does not stand where a person has not chosen an event.
### Three things the writing found that no instrument here reads
**One concept, one word, and the person holding the lexicon broke it.** The first draft said *settle*
eleven times under a heading reading *How an event resolves*. `voice/docs/voice.md` allows
**resolve / resolution** and does not list *settle*. **12 placements turned.** A wrong word renders
perfectly; this is the same class as the 31 kit strings of `docs/backlog.md` 225, produced fresh.
**The last on-chain resolution claim in the product, on 223 placements.** `docs/backlog.md` 217
closed on 2026-08-19 having moved every resolution claim off the chain and onto the published
Reading. The footer trust strip's third tile read **`1,284 events resolved` / `On-chain,
verifiable.`** and it survived, because 217 grepped for the word and this tile makes the claim in a
PAIR: the count in one element and the proof in the one beside it. It could not stand while About
states the mechanism at length two blocks above it. It reads `Every reading is published.` now, which
is also the one thing the two tiles beside it do not already say. `On-chain, verifiable.` survives
only in the frozen `ui-visual/old/`. **And `ui-kit/hiw.html` still carried the pre-217 sentence
twice**, *Every resolution is recorded on-chain*, where both trees carry the published-reading one:
another survivor of the same row, found by comparing the specimen with the screen it quotes.
**The stand was drawing the reading column at twice its width, on the page about that width.**
`ui-kit/_page.css` has `.tk-theme-fig>*{max-width:100%}` as a guard against a specimen overflowing
its cell, and on `.read-col` it was also OVERWRITING the subject: the column declares
`max-width:var(--measure)`, 46ch at its own 16px, which is **503 in the product and rendered 1046 in
the kit at 1440**. On a page whose own prose says *a column measured in a table on a page that never
drew it is a number with no subject*. A cap narrower than the cell cannot overflow it, so the guard
had nothing to do there. Both specimens read 503 now. **Found by measuring the specimen against the
screen it quotes**, which is the only way this class shows: it renders, it throws no error, and it is
confidently the wrong number.
**Measured after, both engines, at 320, 360, 390, the three rungs and one pixel either side, 1280,
1440 and 1600: 0 page errors, 0 duplicate ids, 0 sideways scroll on all ten document pages and the
four kit pages, and the lede resolves to 20px on 10 of 10.** What is NOT done is
`docs/backlog.md` 228: the four legal pages still describe themselves, and that is legal review
first and copy second, which the bank already says by name.
---
## 2026-08-20 - A cadence is not a Type, and the question that decides whether a document is owed is not what the thing IS but what the existing page can SAY
**`docs/backlog.md` 224 read as one defect and was two, and the half it named was the smaller one.**
It said the How often filter offers five cadences and the product draws exactly one. True: 20 of 20
rendered frequency words said One-time and **0 cards carried a cadence attribute at all**.
What the row did not say is that `docs/launch-catalog.md` opens **two Weekly and two Monthly** series
in release 1 and **no Hourly and no Daily**, so two of the five values could never match anything the
product would ever hold. **A filter has two ways to lie and this one was doing both**: filtering
nothing, and offering a catalog nobody was going to open.
**The model needed no decision and that was the trap.** `ia/docs/sitemap.md` settled Frequency on
2026-08-10: every cadence instance is its own Event with its own window, price and resolution, the
cadence is a SERIES the instances belong to, and no new entity is created. Everything was decided
except the drawing, which is exactly the state in which a pass talks itself into believing there is
nothing to do.
### The filter, and the reason it is one pass and not two
`wireFilter('freqMenu','freqCurrent','freq')` was called with no `onChange`, so it relabelled the
summary and closed the menu. The mechanism was already proven beside it: the sub-category rail hides
cards by `data-subcat`. **Wiring the second one the same way would have been a defect**, because two
independent writers of `hidden` fight and the last one clears the other's decision: pick Bitcoin,
then Weekly, and every weekly market in the category comes back. So there is one state and one pass.
Measured on `event-feed-crypto.html`: **Any 8, One-time 6, Weekly 2, and Weekly then Bitcoin 1.**
**And the menu narrows to the page, which is `docs/backlog.md` 173 applied a second time.** That row
made the rail render only the sub-categories the cards actually carry, because rendering the taxonomy
whole showed eleven chips against six events. The same argument holds here and is stronger, because a
cadence with no card returns an EMPTY grid rather than a short one: Politics offers two rows,
Crypto three, General three, and nobody is offered a value that answers with nothing. The static
markup keeps the catalog's four as the no-script floor.
**Hourly and Daily come off the control and stay in the entity.** They are the vocabulary
`ia/docs/sitemap.md` gives Frequency, and an option returns the day a market opens at one, which is
the shape `.icon-btn-lift` already has in this repository: rules whole, placements away, and the
difference between the two zeros written down beside them.
### The drawing, and the one rule it turned out to need
Three binary recurring markets are drawn: **two weekly**, Bitcoin and Ethereum, added to the Crypto
and Trending feeds, and **the monthly temperature record**, which was already in both trees in the
ONE-TIME phrasing the catalog had superseded and is now the instance the catalog describes.
`assets/search.js` is re-taken at 27 events, because that file says in its own header that an event
the tree grows and it does not is an event search cannot find.
**`event-detail-recurring.html` exists in both trees, and the argument for it is not that recurring
is a new type.** It is not: Frequency is orthogonal to Type, so a weekly binary market is the same
TYPE as a one-time one, and the routing convention measured on 2026-08-20 sends a card to the
specimen of its type. That would have put a card reading `Repeats: Weekly` in front of a document
whose header reads **One-time event**, which is the fixture contradiction that cost this repository
six findings on one fixture set, manufactured deliberately. **So the test for whether a document is
owed is not "is this a new type" but "can the existing specimen say what this card says".** It could
not, and it is one line in the header that proves it.
**What the recurring detail carries is entirely existing components, which is the system-first rule
holding rather than being invoked**: the cadence in `.ed-cat` and in the `How often` fact; a
resolution rule naming the program that reads the close, because the catalog admits a recurring
market only if it is machine-resolved; the sentence a one-time reader never asks for, that this week
settles on its own and the next week opens as a separate market with its own threshold and a bet here
does not carry over; and **Earlier in this series**, a second placement of the related component
listing the resolved instances with their outcomes. **No new CSS was written for any of it.**
**A series gets no surface.** Four series at launch do not earn a page, the IA's own decision says no
new entity, and what a reader actually wants from a series is the RECORD, which is a short list of
events they can reach, which is what the related component already is.
**The chart is a control and a four-day-old market has no month.** The range chips were `1d 1w 1m
all`; on a weekly instance `1m` and `all` would draw the same week twice, so the recurring detail
offers `1d` and `all`, and `all` is the life of the market: Mon, Tue, Wed, Thu, now.
### Two things the drawing changed that the row never mentioned
**An instance carries its window in its TITLE.** The first pass wrote `Will Bitcoin close above
$119,500 this week?`, which is unambiguous on a card that prints `Closes` beside it and meaningless
in a list of four past weeks, and which would have to CHANGE when the market resolved. A title that
changes with the market's state is the same defect as one market wearing three names, which is what
made a market unfindable by search on 2026-08-20. So the title is fixed for the life of the instance:
**`in the week to Aug 28`**, `in September 2026`.
**The card meta row has a grammar and a bare word broke it.** `Volume: $46,300` and
`Closes: Aug 28, 2026` are `Label: value`, and the cadence went in as a bare `Weekly`. At 390 the row
rendered **`Weekly Volume: $46,300`**, which names a figure this product does not have. It is
`Repeats: Weekly` now. **Seen on a screenshot, not in the source**, which is the fourth time in this
file that the thing that renders perfectly and reads wrong was caught by looking.
**The word is on recurring cards only and the attribute is on 425 of them.** `data-freq` is what the
filter reads. Printing `Repeats: One-time` on every card would be a default restated on every row,
which is the activity feed `PRODUCT.md` refuses by name: a cadence is news only when there is one.
### What was not built, and why that is the row rather than a silence
`docs/launch-catalog.md` row 4, the monthly stablecoin series, is **multi-outcome**. Its card would
route to `event-detail-multi.html`, which reads One-time, so drawing it would reproduce the exact
mechanism this pass was closing. It stays in its one-time phrasing in both trees and the missing
document is **`docs/backlog.md` 227**. Monthly is therefore carried by one market rather than two, and
that is stated rather than left to be discovered.
**Measured after, on both engines, over the two new documents and the ten changed ones: 0 page
errors, 0 duplicate ids, 0 sideways scroll at 320, 360, 390, the three rungs and one pixel either
side, 1280 and 1440, with a positive control that saw a 900px box in a 390px window on both.** The
recurring detail is 4,681px tall against the one-time detail's 4,277, and the 404 is the series list.
---
## 2026-08-20 - A face is a legal difference between the trees and a length is not, and the row was filed under the wrong noun
**`docs/backlog.md` 215 had been open since 2026-08-19 because it could not be decided as written.**
It said grey `wallet.html` lacks the `.protect-page` class its painted twin carries, that this
"renders as a smaller box and nothing else", and that a component's FACE is not among the
differences `wireframes/_conventions.md` declares, so somebody had to rule on faces first.
**Measured in Chromium 151 and WebKit 26.5 at 390, 900, 1280 and 1600, it is not a face. It is a
length.**
| | 390 | 900 | 1280 | 1600 |
|---|---|---|---|---|
| paint, `.protect.protect-page` | 61 chars at 328px | 61 at 346 | 61 at 346 | 61 at 346 |
| grey, `.protect` | 74 at 386 | **117 at 896** | **117 at 1026** | **117 at 1346** |
The class carries three things and **only the first is paint**: a brass tint, which the grey tree
cannot have under convention 1; a padding and type step, which grey's own `.protect-page` rule
already had; and `max-width`, which decides how many characters a reader gets. In `.cat-main`, a
column sized for cards, uncapped means the sentence takes the column: **117 characters on one line.**
Line length is precisely what the `.read-col` finding of 2026-08-19 established is not a declared
difference, so this was drift by `_conventions.md`'s own definition.
### The tell was already in the grey tree and nobody had read it
Its five document pages carry `.read-col .protect{max-width:none}` and there was **no cap anywhere
for it to cancel**. That is the `text-transform:none` shape from `hiw.css` arriving again: a rule
only cancels what something else is doing, so an override with nothing under it is half a port. It
has something to cancel now, and the five document pages did not move by a pixel.
### The number is measured rather than borrowed
`ch` is the advance of a zero, which this repository learned once already when `--measure` went from
66ch to 46ch. **The two trees have different faces, so a `ch` value does not travel**: `46ch` is
346px in DM Sans at 11px and 276px in the grey system face at 12px, and dropping the paint's number
in gave the grey notice **49 characters instead of 61**. Swept eight values from 46ch to 62ch and
took **56ch**, the middle of the 54-to-58 plateau where the line reads **61 characters over two
lines**, the same count the paint gives. **Match the reading, not the pixel.**
### The rule, written where the next case will be read
`wireframes/_conventions.md`, beside the seventh difference and yesterday's geometry split, which
this arrives at from the other side: there the difference could NOT exist in the grey tree, here it
could and did. **The test is one question. Take the colour off and ask whether the box changed.**
Tint, ground, shadow, corner, bevel and ink are paint. Padding, type step, box width and **line
length** are both trees, because how much a reader gets in a line is a decision about reading, and
the grey tree decides reading.
**After**: grey wallet 61 characters at 336px, the five document pages untouched at 503px. 48
renders over the six documents on two engines at four widths: heights moved on the wallet only,
**+15 at 390 and +30 above it**, which is the notice becoming two lines, and 0 sideways scroll, 0
duplicate ids, 0 page errors.
---
## 2026-08-20 - The record has a renderer, and the eleven rows that were losing text were the opposite of the fifty-one that were filed
**`docs/backlog.md` 223 said 51 rows disagreed with their header and that on GitHub "those columns
are simply gone".** It was filed from the source. Put through **GitHub's own markdown API** on a
control table built for it, the two failures turn out not to be symmetrical:
| the row | what the renderer does | what it costs |
|---|---|---|
| FEWER cells than the header | pads with blank cells | nothing. The row reads ragged and every word survives |
| MORE cells than the header | **drops the excess in silence** | the whole last cell, which in this file is the longest one in the row |
So the 38 folded rows were harmless and the defect was **11 rows carrying one cell too many**, 10 in
`docs/backlog.md` and 1 in `ui-kit/docs/inventory.md`.
### What the eleven actually were, and no two were the same
**Six were a row QUOTING ITS OWN ORIGINAL ROW.** A struck row here often ends "The original row:"
and then reproduces it, and the quote's pipes stop being prose the moment they are on a table line.
A quoted row is prose ABOUT a row and not a row, so its pipes are escaped now. **Two** carried a
trailing `Owner:` cell that belongs at the end of the Note. **One** had its struck title and its
closure sentence as two separate cells. **One** carried a **truncated duplicate of its own Source**,
`ui-kit/, measured 2026-08-09` standing beside `ui-kit/, measured 2026-08-09 across all 38
components with a page`. And the eleventh was **in the wrong table**: the `search` row stood in
`Component / Uses / Declaration` wearing the four-cell shape of `Component / Its own width query /
Placements / What it does with width`, which is a table in the same file three feet further down.
### The decision the row asked for
**A struck row keeps its header's cells.** The Item cell carries the strike and the closure, Source
and Note keep whatever they had, and **a cell with nothing left in it stays empty rather than being
folded away**, because the renderer is going to draw it either way and the source may as well say
so. 38 rows were padded explicitly. It is in this file's own preamble, where the next person to
strike a row will read it.
### Proved on the renderer, with a control
The five repaired rows render four cells with their recovered text present. **The same five taken
from HEAD render four cells with that text absent**, which is the positive control: an instrument
that cannot come back red is a reading of the guard. Re-measured over every markdown file in the
repository with code spans stripped and escaped pipes handled: **342 tables, 2,708 body rows, 0
disagreeing with their header in either direction.**
**In 516 commits nobody had rendered the record.** Every instrument in `CLAUDE.md` reads a computed
HTML page, and the markdown that carries the decisions, the backlog and the inventory had only ever
been read as source. That is the line added to the root file.
---
## 2026-08-20 - The geometry a movement needed splits in two, and the half that could not exist in the grey tree is the half that stayed in the paint
**`docs/backlog.md` 226 asked whether the constant sheet the paint gained on 2026-08-20 is a
consequence of the seventh declared difference, which would make it paint's, or a structural
decision, which would make it both trees'.** The answer is both, and the line between them was
measured rather than drawn.
**THE ROW AS FILED WAS HALF WRONG, IN THE WAY A ROW WRITTEN FROM THE OTHER TREE ALWAYS IS.** It said
the grey twin "still jumps 47.5px", which is the paint's number carried into a tree nobody had put a
probe on. Measured in Chromium 151 and WebKit 26.5 at 390 and 1280: grey stood **475.56, 475.56 and
521.75**, so it jumped **once** and by **46.19px**. And the picture frame, which cost the paint a
`min-height:19rem` to fix, **was 180 on all three steps and always had been**, because a grey still
is one dashed box with a label in it rather than three real components. **The defect could not exist
in this tree.** That is the plate-wrapper row of `wireframes/_conventions.md` showing its consequence
one more time: a difference that follows from a difference already declared.
### The split
| the change | goes to | why, measured |
|---|---|---|
| three panes stacked in one grid cell | paint only | its only job is to give the movement two states to move between, and movement is the seventh difference |
| the picture frame at `19rem` | paint only | the grey stage is already 180 on all three. Nothing here for it to fix |
| the reserved row under the button | **both trees** | the third step carries a way out the other two do not, which is a fact about the step's CONTENT, and this tree owns where a control sits |
| the text block's floor | **both trees** | a floor stops a sheet resizing under a thumb, which is behaviour. Grey's was 150 against 154.58 of text, so it sat BELOW its own content and the button moved 4.58px at 1280 |
### The sweep was safe because the block is genuinely shared, and that was checked first
97 grey documents carry the stepper. Extracted from every one of them and hashed: **97 of 97
byte-identical, 17 rules each.** Its region header named `port_chrome.py` and said "nothing here is
promised to another file" while being promised to 96 others, which is the exact defect the root
`CLAUDE.md` records about a region named for the tool that wrote it. It carries its own numbers now,
`SHARED (97 of 116, 18 rules)`, so a reduced copy will contradict its own header.
**And the 19 documents that do NOT carry it were checked rather than assumed.** 116 painted
documents have the sheet and 97 grey ones do, which is the kind of gap worth a look. They are
exactly the overlay screens, seven deposit states, four sign-in, five win and three loss, and each
has **0 headers, 0 feeds and 0 footers**: convention 5 draws the sheet on a plain backdrop, so
there is no feed here for a how-it-works dialog to hang off. The gap is the boundary working.
### After
**527.17 in Chromium and 526.97 in WebKit on all three steps, frame 180, text 160, nav 153.78,
button at 340, dots at 443.78, every number identical at 390 and at 1280.** 232 renders over the
116 grey documents against HEAD: **0 document heights moved**, 0 sideways scroll, 0 duplicate ids,
0 page errors, because a `` with no `[open]` draws nothing on a static page.
**Both sheets are one height on all three steps now, and they get there by different routes because
one of them moves.**
---
## 2026-08-20 - A voice pass that names two trees is run over two trees, and the stand had been left behind by every one of them
**`docs/backlog.md` 225 was filed the same day as three strings on one page and closed as 31 strings
on 19 of the 61 kit pages.** The row asked one question, whether a specimen quotes the screen or
`voice/docs/microcopy.md`, and the answer turned out to be a different distinction: **a specimen
holds a QUOTATION or a DEMONSTRATION.** A quotation is the product's words in the product's markup
and must equal what the screens ship, because a reader comes to the stand to learn what the
component SAYS. A demonstration is a string chosen to show a property, a headline at three lengths
on `typography.html` or a query highlighted on `search.html`, and it may say anything.
**THE MECHANISM WAS ONE SENTENCE.** `voice/CLAUDE.md` said a UI string "goes into both screen trees",
and every sweep obeyed it exactly. `ui-kit/` is the THIRD tree that carries those strings, by this
repository's own rule that markup goes to two places and one of them is the component's page, and
**backlog 217 turned the on-chain resolution claim over 217 placements in `ui-visual/` and
`wireframes/` on 2026-08-19 and 0 in the kit.**
### What the read found
Every text node in the kit that sits under a PRODUCT class, harvested from the rendered DOM over all
61 pages and diffed against `ui-visual/`: **8,800 kit strings, 774 product strings, and after
filtering the kit's own prose about its components, 31 quotations had drifted.**
| what | where | the shape of it |
|---|---|---|
| the third how-it-works step told **three different ways** | `hiw`, `dialog`, `organisms` | one of them the retired on-chain claim, so the stand promised a mechanism the product does not have |
| a **five-question FAQ where the product ships three** | `seo-plate` | with a heading and a body no screen has: "Politics markets" against the product's "About Politics events" |
| `service` and `market` where the product settled on `document` and `event` | `toc`, `molecules` | a whole leftover draft, with its row numbers one digit instead of two |
| `1,284 events resolved on-chain` | `feed`, `hero` | the exact string 217 turned |
| `YES pre-selected` after the product settled on `YES selected` | four pages | found on a third and a fourth page after two rounds of repair, which is why the sweep was re-run each time |
| `See all 3 results` | `search` | the seam the root `CLAUDE.md` already records as fixed, fixed in one tree |
| five fixture rows missing the year and the `net` column | `position`, `tabs`, `profile` | the identity pass of 2026-08-20 landed in the trees and not here |
**And one was a FACE rather than a word.** `vitrine.html` drew the notification row with `` and
`` around a
matched query splits one sentence into three text nodes, so `search.html` reported three truncated
questions and a fragment beginning "ion be called in France". 3 are the kit's own sidebar, which
wears product classes and is not a specimen at all. 2 are demonstrations, a bare `Sub-categories`
and a `Sort:` prefix the control itself writes. **1 is a real question left open**: a France market
that `ia/`'s `@graph` names and no screen prices, shown at 29 per cent by a kit suggestion row.
**A wrong word renders perfectly**, which is why no instrument in this repository could have found
any of it. Every sweep here reads a computed value and copy has none. This is the cross-document set
read arriving in the one layer with nothing to compute.
### Written where it acts
The rule is in three files because it is three different obligations: `voice/CLAUDE.md` owns the
count of trees, `ui-kit/CLAUDE.md` owns the quotation-against-demonstration split and the set read
that finds it, and the root `CLAUDE.md` owns the cross-folder fact that a string ships in three
places. 244 renders over 61 pages on two engines at two widths after the repair: 0 sideways scroll,
0 duplicate ids, 0 page errors.
---
## 2026-08-20 - A step that cannot be in the layout cannot move, and two comments said the floors already held
**Asked for by the user**: the how-it-works sheet should change its picture and slide its text
rather than swap the whole dialog, and the dialog itself must not jump.
**The block could not have moved as it stood.** `.hiw-step` was `display:none` on two steps and
`display:flex` on the third, and an element that is not in the layout has no state to leave from.
The Animation stage's register holds one row per moving element and has never asked which moments
have no element to move: this is the same shape as the census of 2026-08-17, one level deeper.
**MEASURED FIRST, Chromium 151 and WebKit 26.5 at 390 and at 1280, agreeing to the hundredth**: the
sheet stood **563, 586.5 and 610.5**, so it jumped twice; the picture frame stood **280, 303.5 and
280**; **Next moved 452 to 475.5 and back**; the dots slid 512 to 559.5 down the flow.
**TWO COMMENTS IN `components/hiw.css` SAID BOTH OF THOSE WERE ALREADY HANDLED, AND THAT IS THE
FINDING WORTH KEEPING.** The stage said "280 is the tallest of the three stills plus its shadow,
measured, so no still is ever clipped by its own floor", and the bet panel asks 303.5. The step text
said "the button that moves is not the one a thumb is already aimed at: it is Next on one step and
Create account on the other", and it was Next against Next. Both arguments are sound and both were
made about the wrong half: **a `min-height` stops a box getting smaller and neither of these was
stopping one getting bigger.** The numbers inside them, 552 / 641 / 506 and 39px, were true of a
tree that had since moved. **The count rule was measured over six prose files on 2026-08-18 and
never over the comments in `components/`**, where a number is written beside the rule it justifies
and nothing re-reads it when the thing above it changes.
**The build, and no markup moved.** The three panes share one grid cell, so the cell is the tallest
of the three at every moment and the sheet cannot resize; the three radios are `position:absolute`
and therefore not grid items, `nth-of-type` counts what it counted, and **219 placements over three
trees are untouched**. The frame is `min-height:19rem`, in `rem` because the type inside the still
is. The third step's quiet way out is reserved as an empty row on the other two, derived from the
link's own type and padding rather than chosen, because the dots are the mark that says where you
are and they were the least stable thing on the sheet.
**The movement's job is ARRIVAL and it adds no number to the register.** `--hiw-shift` is one
inherited length with three values, `+travel` ahead, `0` here, `-travel` behind, and `--hiw-travel`
is `calc(var(--space-32) * var(--motion))`. The text travels the whole distance and the still four
tenths of it, which is what tells a reader they are two planes of one surface. Two durations and two
curves, unchanged.
**THE NAV IS THE ONE PART THAT DOES NOT FADE, AND THAT IS WHY THE PANE CARRIES NO `opacity`.** An
opacity on the step would have taken the Next button with it, and **two brass buttons crossing at
half opacity composite to 75 per cent**, so the one control a thumb is aimed at would dim on every
press. `.hiw-nav` is switched by `visibility` with no transition at all: measured through the flight
in both engines, the leaving nav is hidden on the first frame and the arriving one visible on the
same frame. `visibility` rather than `display` on the pane for the other half of the same argument:
it is the one value that keeps an element out of the tab order and out of hit testing while still
letting it animate, and the sheet offers 1 focusable control on steps 1 and 2 and 3 on step 3, which
is exactly the live pane's own.
**Result, both engines, both widths**: 634.5 at 390 and 635.5 at 1280 on **all three steps**, frame
304, button 476, dots 583.5, every one identical. Reduced motion asserted in the page gives
`calc(32px * 0)` on every pane and the step still arrives. Mid-flight sampling proves the travel
runs 0 to -32 and 32 to 0, **because a probe that reads only the settled state cannot tell a
movement from a swap**. Height against HEAD over both trees, 356 renders on 178 documents: **6 rows
moved and all six are the three kit pages that stand the stepper open**, 0 of the 116 painted
screens, 0 sideways scroll, 0 duplicate ids. **And the control caught itself twice**: under reduced
motion the freeze was deliberately not injected, so a 1ms `visibility` transition was still in
flight one frame later and Chromium reported the previous pane.
**The cost is named**: 24px, the price of a frame that cannot resize and a control that cannot move.
23.5 of it is the picture frame taking the height its tallest still already needed.
**Two rows filed rather than swept.** 225, the stand tells the three steps in words the product
stopped using and one of them is the on-chain claim retired the day before, which is a rule question
about whether a specimen quotes the screen or `microcopy.md`. 226, the grey twin still switches and
still jumps: motion is the seventh declared difference and correctly paint-only, the geometry is not
on that list, and 97 inline stylesheets is the cost of guessing wrong.
---
## 2026-08-20 - 116 painted documents are six routes, and that is what turned the build from a wish into a plan
**`docs/backlog.md` 218 named a missing Build stage, not missing files**, so it closes as a plan:
`docs/build-plan.md`, registered in the ownership table of `CLAUDE.md`, carrying the stack with its
prices, the schema, the resolver job, seven phases and what is deliberately not built.
**THE NUMBER THAT MAKES IT BUILDABLE.** A static tree needs a document per combination of category,
auth state and load state; a running app gets all three from data. Counted 2026-08-20: **the feed is
44 documents and one route**, the event page 14 and one, my bets 9 and one, and **91 of the 116
painted documents live inside six routes**. The remaining 25 are the families v1 does not open plus
the static and system pages. **The build is not 116 screens and it never was**, and nobody had done
that arithmetic, which is why the row read as if the work were the size of the tree.
**THE ONLY SURFACE WITH NO DRAWING IS `/admin`**, and it does not get one from these trees. Its
reader is the operator rather than Alex, which is exactly why `ia/docs/sitemap.md` registers it in
Under Question as not a node of that map. Three lists and two buttons; it may look like a table with
no brand on it and that is correct.
**THE STACK IS CHOSEN AGAINST THE TREE RATHER THAN AGAINST A PREFERENCE.** Astro, because an
`.astro` file is HTML with a fenced script, so the markup pastes in and the loops go around it, and
because it imposes no client framework: this system's discipline is one stylesheet and no per-screen
CSS, and a component framework would create a second component system whose boundaries disagree with
`components/`. Supabase for Postgres, the two sign-in providers `PRODUCT.md` names, row-level
security and the cron. **$0 a month to start and $25 when the database has to stop pausing**, which
is the whole capital requirement because the currency is points. `PRODUCT.md`'s Tech stack heading
said `(TBD)` until today and now points here, with its own rows marked as the real-money product.
**THREE RULES GO IN THE SCHEMA RATHER THAN THE CODE**, because the code is one person writing in the
evening: `bet.idempotency_key` is unique so a double-tapped Confirm is one bet, `ledger` is
append-only and the balance is a sum rather than a column, and `resolution` is one row per event so a
market cannot resolve twice. Double entry from day one even though the currency is points, because a
balance computed from a sum can be audited and a balance stored in a column drifts.
**AND TWO CONSTRAINTS DISAPPEAR THE MOMENT THIS IS SERVED RATHER THAN OPENED FROM DISK**, both of
which cost this repository days: `file://` gives every document its own opaque origin, which is why
the icon sprite is a script, and the theme has to boot from `localStorage` before first paint. On a
server the sprite can be a fetched SVG and the theme can be a cookie read server-side. **Neither is a
reason to change anything now**; they are recorded as savings to collect, not as work.
**PHASE 5 IS THE ONE THAT DECIDES WHETHER THIS IS A PRODUCT.** Seven phases, 8 to 10 weeks of
evenings, and each ends with something a person can open. Phase 5 is resolution: the cron, the three
states, settlement, the reading, the admin. **It is the only phase whose output happens while nobody
is looking**, and it carries the rule that everything else rests on, that silence is never read as
NO. A resolver that has never failed in testing is a resolver whose failure path does not exist, so
one gets broken on purpose before launch.
**Files:** `docs/build-plan.md` (new), `CLAUDE.md` (ownership table), `PRODUCT.md` (Tech stack no
longer TBD), `docs/backlog.md` (218 struck, Open 6 to 5). **No screen changed and no design
decision was taken**, which is the test this plan had to pass: every screen, string, component and
entity it needs already exists, and a question that turns out to need a new drawing stops and goes
through `ia/` then `wireframes/` then `ui-visual/` in order.
---
## 2026-08-20 - A launch catalog is picked by HORIZON, and that overturned a rule written the day before
**`docs/backlog.md` 219 said the shipped catalog is a newsroom: 14 of its 25 questions can only be
resolved by a person.** That was true and it was the wrong axis to pick a launch catalog on.
**THE AXIS THAT DECIDES IS HOW SOON SOMETHING RESOLVES.** Measured over the 25 shipped questions on
2026-08-20: **2 resolve within a month and 1 of those two is machine-resolvable**, six weeks reaches
four of them, and eleven are longer than six months. So a launch assembled from the shipped catalog
alone gives a reader **one resolution in the first month and then a gap**. A resolution is this
product's only trust event and its only notification, and `PRODUCT.md` says retention is a
notification about the event you bet on, so a catalog with nothing resolving has no retention loop at
all, not a weak one. **The 25 were written to be interesting to READ. A launch catalog has to be
interesting to WAIT FOR**, and nothing in this repository had asked that question of them.
**AND IT OVERTURNED A RULE THIS FILE RECORDED THE DAY BEFORE.** `PRODUCT.md` said the first release is
one-time markets only, because every cadence instance is its own event and five weekly series are 260
resolutions a year. **That is an argument against 260 HAND resolutions**, and the automation decision
of 2026-08-19, taken hours earlier and written into the same file, makes a machine-read series cost
one template and zero manual resolutions. **Two decisions taken the same day and only one of them was
still true by the end of it.** Recurring is in, on one condition that keeps the original fear
answered: **a recurring market must be machine-resolved.** A hand-resolved series is exactly the
newsroom the whole exercise exists to avoid.
**THE SEVEN.** Four recurring machine series for the heartbeat, three one-time markets for depth,
**six automated and one editorial**, which is the ratio the automation pass proposed before the
catalog had been measured. Weekly BTC and ETH thresholds with the strike SOLVED for a 50 per cent
open rather than chosen; a monthly global temperature record; the monthly stablecoin-supply leader by
chain, which is deliberately the **multi** one because the outcome family gained its multi half the
same day and a face with no placement is what this repository keeps paying for; and from the shipped
catalog unchanged, BTC above $150,000 before Oct 1, the Atlantic Category 5, and the government
shutdown as the single hand-resolved market. `docs/launch-catalog.md`, a new file, registered in the
ownership table of `CLAUDE.md`.
**ONE EDITORIAL MARKET IS A CAPACITY STATEMENT AND IT IS THE FIGURE WORTH ARGUING ABOUT.** Everything
else in that file follows from a resolver or from arithmetic; the number 1 follows from what one
person can watch, write and defend while also building the thing.
**WHAT THE PASS FOUND ON THE WAY.** The How often filter offers Any, One-time, Hourly, Daily, Weekly
and Monthly, and **20 of 20 frequency words in the painted tree say One-time, with 0 cards carrying a
cadence attribute at all**. **Read the instrument before the finding**: `wireFilter` is called for
that menu with no `onChange`, so it relabels and closes and does not filter, the same as the language
selector whose comment says so. **It is not a broken filter, it is a control promising a cut of the
catalog that does not exist**, which is the class this week has struck three of. Putting four
recurring series in the first release makes the cut real and the product still draws none of it.
`docs/backlog.md` 224.
**AND THE BACKLOG HEADER WAS REWRITTEN, WHICH IS A SMALL FIX FOR A REAL FAILURE.** It had become a
chronological paragraph: accurate, and a reader had to parse eleven clauses to learn which six rows
are open. It is a table now, and the six names in it are checked against the six rows on disk.
**The week filed eleven rows and struck five within a day, and four of the five were half wrong in
the row and right that something was there**: 217 was "delete a word" and was claim against mechanism,
219 was a classification problem and was a horizon problem, 221 was an unmade placement and was a
routing convention nobody had measured, 220 and 222 were declared states no tree had drawn.
**Files:** `docs/launch-catalog.md` (new), `PRODUCT.md` (the launch scope, and the one-time rule
reversed with its reason), `CLAUDE.md` (ownership table), `docs/backlog.md` (219 struck, 224 filed,
header rebuilt as a table). **No screen changed**, and that is correct: the 25 in both trees are the
designed catalog and this is which of them the first release opens.
---
## 2026-08-20 - The outcome family was binary on both halves, and the sentence a multi resolution needs cannot be written on a binary screen
**`docs/backlog.md` 222, filed the same day by the routing measurement that closed 221.** The outcome
family was six documents, `win`, `win-payout-pending`, `win-loading`, `win-error`, `loss`,
`loss-loading`, and **0 of them multi**, while multi-outcome is a first-class type everywhere else in
the product: **7 of the 25 catalog markets**, **37 painted cards**, a YES and a NO control on every
option row, an `event-detail-multi` specimen from the start, and `ia/docs/flows.md` carrying the chosen
option all the way to Confirm. **The resolution was the one place the option stopped existing**, and
the Outcome field of the Resolution entity has said `YES · NO (or which option in multi-outcome)`
since that entity was written. A declared state that was never placed, which is the class the chosen
NO cost this repository.
**THE PROOF THAT IT IS A TYPE AND NOT DATA IS THAT THE BINARY SENTENCE CANNOT BE WRITTEN.** `win.html`
says *the market resolved YES on Jun 27, **the side you held***. The shipped multi fixture is a bet of
**NO on Action** in a market where **Animation** led: the reader is paid because the option they DID
take lost, and the option that won is one they never touched. **There is no "side you held" in that
story at all.** Binary has two outcome combinations; multi has four, and the two involving NO are the
ones no drawing in either tree explained.
**BOTH HALVES ARE DRAWN, AND THAT WAS A DECISION RATHER THAN THOROUGHNESS.** `voice/docs/voice.md`
principle 4 is *design the loss, mark the win*, and drawing only the winning half of a symmetric
control is precisely what cost this repository the chosen NO. So `loss-multi` exists too, and it
needed a market: the fifth settled bet is **YES for Solana on which chain held the largest stablecoin
supply at the end of H1 2026**, resolved **Ethereum**, and it stands in `active-bets-history.html` and
in the profile's Resolved tab in both trees. **The fixture set was checked before it was extended, not
after**: the profile aggregate reads Total bets 23, Win rate 61 per cent, Resolved 18, and the four
listed rows were a sample of eighteen, so a fifth costs the aggregate nothing and the visible split
becomes 3 won and 2 lost, which is 60 against a stated 61.
**AND THE READING GAINED ITS THIRD SENTENCE SHAPE AS A SIDE EFFECT.** `docs/backlog.md` 221 filed that
shape as an unmade placement and the routing measurement said it was data rather than a face. It
arrived the next day exactly as data: `loss-multi` reads **Ethereum held $73.4B in stablecoins against
Tron at $61.2B and Solana at $12.8B**, which compares figures ACROSS options and is a thing a binary
read has no way to do. Same drawing, same two classes, a sentence a binary market cannot produce. **The
measurement was right and the row was right about there being something missing; they disagreed only
about what kind of thing it was.**
**NO NEW CSS.** The screens are the binary ones with a multi-shaped sentence and a multi-shaped
reading, and that is the correct answer rather than a shortcut: what differs is the CONTENT SHAPE, and
`.reading` was already built as a face for exactly that reason on 2026-08-19. A resolved options table
with the winner marked was considered and left out: the sentence already names the winner, the reader's
option and both figures, and a table would have needed new faces in `options.css` for one placement
each.
**THE PANEL LABELS WERE TURNED WITH THE PAIR, AND THE OLD ONES WERE ON THE WRONG AXIS.** The painted
panel said `you were right` and `what happened` where the Event Detail group beside it says `binary`
and `multi`; a type and a mood are not the same axis, and a group that mixes them cannot show that a
type is missing. Both trees now read `binary` / `multi` in the same shape Event Detail uses, in **117
painted and 116 grey** documents.
**Verified.** Instrument control **0 of 229** documents read twice. **2 of 229 moved and both are
intended**, the two `active-bets-history` files carrying the fifth row; the profile did not move
because its Resolved tab is a CSS-only panel that is not the active one, and all five rows read
identically to each other there. The four new documents: **0 duplicate ids, 0 page errors, the Reading
visible, no sideways scroll at 14 widths including every rung and one pixel either side, positive
overflow control seen**, and **each marks itself as the current page**, which is a set read over both
whole trees, clean on 116 painted screens and 116 grey with `overview.html` the only document carrying
no marker because it is the index rather than a screen. **0 dead internal links across both trees.**
**The theme control failed on its first run and the failure was the instrument**: it read
`document.body` after the loop had ended on a grey document and reported one ground for four
documents. Re-run per document it gives **`rgb(15,16,19)` against `rgb(252,250,244)`** with
`data-theme` absent in the Vault and `light` in daylight, exactly as the boot script behaves. **And the
ground that matters is not the page**: these screens are a modal over the feed, so the Reading was
re-measured against the sheet it stands in, **36,40,47 on 18,20,23 in the Vault and 252,250,244 on
248,246,240 in daylight**, separating by ground and edge in one theme and mostly by edge in the other,
which is `--bg-control` behaving as it does for `.reconcile-box.rec-lost` and not something this pass
introduced.
**Counts re-taken from the disk and the DOM, not incremented**: `ui-visual/` is **117 documents, 116
screens**, `wireframes/` **116**, `.protect` **136 elements**, `.reading` **10 on 10**. Turned in
`CLAUDE.md`, `STRUCTURE.md`, `components/CLAUDE.md`, `components/notice.css`,
`ui-kit/docs/architecture.md`, `ui-kit/docs/inventory.md`, `ui-kit/notice.html`. `components/seo-plate.css`
keeps its number because it is DATED and says the day it was taken, which is what that rule is for.
---
## 2026-08-20 - A row is a route and not a surface, and the measurement that said so found the state nobody had ever drawn
**`docs/backlog.md` 221 asked two questions and I had filed it half wrong.** It asked whether a
resolved ROW in the history is a fourth surface for the Reading, and whether a row can carry a source
link. It assumed the answer was a placement somebody had not made. **The answer was in the routing,
and it took one number.**
**148 FEED CARDS OVER 15 PAINTED DOCUMENTS, 111 TO `event-detail.html` AND 37 TO
`event-detail-multi.html`, AND 0 OF THE 148 TO THEIR OWN MARKET.** A card or a row in these trees goes
to the specimen of its TYPE. So this tree ships a document per type and per state and **never per
market**, which means a numeric market gets no document, the figure-against-threshold wording is DATA
inside the same drawing rather than a face nobody wears, and a resolved row is a ROUTE. And a row must
not carry a source link: four rows of citations is the activity feed `PRODUCT.md` refuses by name, and
the fact already stands on the surface the row leads to. 221 is struck with nothing built, which is
the right outcome for a row whose premise was the thing to check.
**AND I NEARLY FILED THE CONVENTION AS A DEFECT.** The first read of the history said **2 of 4 rows
route to another market's outcome screen**, the BTC row and the box-office row both landing on
`win.html`, which draws a government shutdown at +$13.16. That looks exactly like the fixture
contradictions of 2026-08-17, and it is not: 148 cards say the tree has always routed to the
specimen. **`CLAUDE.md` already carries this exact warning from the day `ui-kit/betpanel.html` was
filed as a broken link for a string standing inside ``**, and the cost of ignoring it here would
have been a repaint of two trees to fix a convention. Measure the convention before calling a case a
violation of it.
**WHAT THE SAME MEASUREMENT DID FIND IS ONE ROW THAT CANNOT OBEY THE CONVENTION.** Three of the four
resolved rows obey it exactly. The fourth is `Which genre led the 2026 spring box office?`, which is
**multi-outcome**, and the outcome family is **six documents with 0 multi among them**. So a multi
resolution has nowhere type-correct to go and lands on the binary win. **This is a DECLARED state that
was never placed**, the class the chosen NO cost this repository: `ia/docs/sitemap.md` has said
`Outcome: YES / NO (or which option in multi-outcome)` since the Resolution entity was written,
`ia/docs/flows.md` carries the option through to Confirm, and the feed already ships **37 multi cards**
and an `event-detail-multi` specimen. The resolution is where the option stops existing.
**It is filed rather than built, and the reason is that the shape is genuinely new.** A multi win says
WHICH OPTION won rather than which side, and this market's bet was **NO on Action while Animation
led**, so the screen has to explain a win on an option the reader bet AGAINST, which no drawing in
either tree does. Costed in the row: one document per tree, a panel row in **115 painted and 114 grey**
documents with the current-page marker rule, the Reading in its third wording, microcopy, and the tree
counts in six files. `docs/backlog.md` 222.
**AND THE PASS FOUND A DEFECT I HAD WRITTEN THE DAY BEFORE.** The `notice` row of
`ui-kit/docs/inventory.md` carried its new addendum TWICE and had lost the pipe between its second and
third cells, because the script that appended it applied the text inside a loop and again after it. Two
things follow. The row is rebuilt from the pristine cell structure and carries the text once. And
reading the same question across every prose document gives **51 rows of 2,251, over 250 tables, whose
cell count disagrees with their own header**: 47 in `docs/backlog.md`, 2 in the inventory, 2 in
`voice/docs/microcopy.md`, so their last columns vanish wherever this repository is rendered. **Read the
instrument first**: a naive counter said 48 by counting every `|` including those inside code spans and
by carrying a header across a table boundary, and 51 is what a counter gives that strips code spans and
requires a header plus its separator. `docs/backlog.md` 223.
**Files:** `components/notice.css`, `ui-kit/docs/inventory.md` (turned, de-duplicated and its cell
structure restored), `ia/docs/sitemap.md`, `docs/backlog.md` (221 struck, 222 and 223 filed).
**Nothing was built, and that is the finding**: the placement 221 asked for does not exist to be made.
---
## 2026-08-19 - The proof moved from the ledger to the reading, and the backlog row that asked for it was the wrong shape
**The IA gave the Resolution entity a Reading in the morning and this pass built it**, in both trees,
and closed the vocabulary row beside it. The row was `docs/backlog.md` 217 and **it was filed as
"delete the word", which is not what the defect was.**
**THE LINE IS CLAIM AGAINST MECHANISM, NOT VOCABULARY.** Counted over both trees: **217 placements
moved and about 364 kept `on-chain`.** Every RESOLUTION claim now points at the Reading, and every
custody or settlement mechanic keeps the chain, because a segregated balance you can check, a payout
that is settling and a bet that failed to register are all things a chain genuinely does. **The
correction does not depend on the first release having no chain**: a hash proves that a decision was
RECORDED and cannot say what the decision was made FROM, which is true with a chain or without one,
and the second is the only question a person who suspects the platform is actually asking. So 217 is
struck as done and its own framing is recorded as wrong.
**THE LARGEST SINGLE STRING IN THIS PRODUCT MOVED WITH IT.** `hiw-step-b`, the third How It Works
step, stands on **114 painted and 97 grey documents, 211 placements**, and it carried a resolution
claim and a custody claim in one sentence. Only the first half moved. The other two are the trust
strip's `1,284 events resolved on-chain`, which named the ledger the record sits in rather than the
thing the record is checked against, and the How It Works resolution promise, whose old ending,
`cannot be changed after the fact`, answered tamper-proofing that nobody had accused us of.
**BOTH FACES OF THE READING ARE PLACED, AND THE FIXTURES DECIDED WHICH GOES WHERE.** `.reading` is a
face of `notice` rather than a component, because what it adds is a content shape and not a drawing.
`win.html` and `event-detail-resolved.html` carry the shutdown, a market only a person can settle, so
they read *read by the Yonder team*; `loss.html` carries a spot ETH ETF approval, which a program
settles from a public record, and reads *read automatically*. **The two are not styled differently
and that is deliberate**: a reader should not learn two shapes to check two markets, so only the
words differ. **The third wording, a figure against a threshold, has 0 placements and says which zero
it is** and is filed as 221: the fixture set's one settled numeric market stands on
`active-bets-history.html`, which the IA does not list among the three surfaces, so it is a placement
nobody has decided rather than a face that does not exist.
**AND THE COLOUR IS AN ARGUMENT.** Every other box in `notice.css` is tinted with `--color-trust` or
brass because it makes a PROMISE. A Reading reports what was read, from where and when, for the
reader to check themselves, so it is the neutral box: tinting it would borrow the authority the
sentence exists to avoid needing, which is `voice/docs/voice.md` principle 5 read as colour. It is
the only untinted face in the file.
**WHAT THE VERIFICATION FOUND ABOUT THE INSTRUMENT, WHICH IS THE REASON TO WRITE IT DOWN.**
`CLAUDE.md` added document height on 2026-08-18 after an 844px sheet stood after the footer on 108
documents, and **that instrument cannot see this block at all in the painted tree**: the outcome
sheet is a real `` opened with `showModal()`, so it is in the TOP LAYER and out of document
flow, and `ui-visual/win.html` reads **1,411 before and 1,411 after** with the block rendering at
348x91 inside it. The grey twin has no dialog element and draws the same sheet in flow, where
`.sheet-body` grows **669 to 792**. **A height sweep over the paint would have reported this pass as
having changed nothing**, and the disagreement between the two trees is again what makes it visible,
exactly as it was for the missing `[open]`. The block was confirmed by `checkVisibility()` and
computed style instead.
**Verified.** 28 renders over 7 documents, two engines and two themes with the ground proving the
theme (`rgb(15,16,19)` against `rgb(252,250,244)` on both engines), transitions frozen before
reading: **32 `.reading` elements, 0 defective renders, 0 page errors, 0 duplicate ids**, the box
painting a real ground and a real edge in both themes and the record line resolving to IBM Plex Mono
in the paint. **14 widths including every rung and one pixel either side, 0 sideways scroll, positive
overflow control seen on each engine.** Document height over both whole trees: **instrument control 0
of 229**, and **7 of 229 moved, all 7 intended** (two reading blocks in flow, and five documents
where a turned sentence is longer than the one it replaced).
**Files:** `components/notice.css` (the face and its argument), the three surfaces in both trees,
`voice/docs/microcopy.md` (the label, both wordings, and the three turned rows),
`voice/voice.html` (the principle-5 example had been teaching the hash), `ui-kit/notice.html` and
`ui-kit/docs/inventory.md`, `ia/docs/sitemap.md` (what was built and which wording each screen
draws), `docs/backlog.md` (217 and 220 struck, 221 filed).
---
## 2026-08-19 - Automating the markets is a constraint on CREATION, not a script for resolution, and the resolution record is the answer to two open questions at once
**The question was how markets appear and close without a person doing it every day.** The answer
splits the problem in half and only one half is automatable, which is why the constraint lands
somewhere other than where it was asked for.
**CLOSING AUTOMATES ONLY IF THE SOURCE IS MACHINE-READABLE, SO THE RULE GOES ON CREATION.** An
Event is not created without a **Resolver**: `source`, `query`, `test`. Automating the resolution of
an arbitrary question is impossible; creating only answerable questions is easy, and it turns a daily
newsroom into a scheduled job. `manual` stays a legal value and it has to be TYPED, so a
hand-resolved market is a decision on the record rather than the absence of a field, which is the
same shape as this repository's rule that a zero has to say which zero it is.
**MEASURED AGAINST THE CATALOG THE PRODUCT ALREADY SHIPS, THE RULE INDICTS IT.** Classified by hand
over the 25 questions in `assets/search.js`: **6 are machine-resolvable from a free public source**
(two crypto thresholds, two climate records, an Atlantic hurricane category, stablecoin supply by
chain), **5 are borderline** and **14 are human-only** (Senate control, a shutdown, two ETF
approvals, Eurovision, a UK election, a nomination, a budget, EU accession, a French snap election,
a box-office genre, a Bond actor, a console launch, a chart-topping album). And the whole painted
tree names **2 distinct resolution sources**, both ending "Resolved by the Yonder team". **So the
catalog is designed as a newsroom**, and the automation answer is not a script bolted to it, it is a
different catalog. `docs/backlog.md` 219.
**THE STATUS FIELD WAS MISSING THE TWO STATES AN AUTOMATED CLOSING PRODUCES.** It read `Active ·
Resolved · Cancelled`, and the Resolution deadline row beside it already implied a third: trading
stops when the window shuts, and for most sources the outcome is READABLE later, because a monthly
climate figure or an economic print lands days after. That gap is **Closed**. The fifth is **Needs
review**, and it exists for one rule: **silence is never read as NO.** A failed read, a timeout or an
ambiguous answer routes to a human, never to an outcome. A market that resolves because nothing
answered is the cheapest possible way to lose the one thing `PRODUCT.md` names as the #1 churn
driver.
**AND THE THING THAT MAKES IT WORTH DOING IS NOT THE LABOUR SAVED.** Because the outcome is machine
read, the record can print what `voice/docs/voice.md` principle 5 asks for and no competitor prints:
*Resolved NO on Mar 1, 2027 at 00:00 ET. BTC/USD was $138,204 against a threshold of $150,000.
Source: CoinGecko, read at 00:00:04 ET*, with the raw response stored and the source a link the
reader opens. **This closes two open questions with one field.** `research/docs/competitors.md`
question 2 asked how resolution trust is built with no regulator and no track record and answered
"on-chain transparency", which the launch decision removed; **a hash proves that a decision was
recorded and a Reading proves what the decision was made FROM**, which is the stronger of the two for
the fear this product is built against. And `docs/backlog.md` 217, the `on-chain` vocabulary standing
on 114 painted screens, now has something provable to be replaced BY rather than merely a word to
delete.
**THE OPERATOR CONSOLE IS REGISTERED AS NOT BEING ON THIS MAP, AND THAT IS THE ANSWER RATHER THAN A
DEFERRAL.** `ia/docs/sitemap.md` maps where the USER can go and its reader is Alex. A console has a
different reader, a different job list and no entry in `jtbd.md`, so putting it on the map would make
one document answer two questions, which is what the `[ORPHAN]` rule exists to prevent. It goes in
**Under Question** for the same reason the footer's eight cut labels went there: so nothing promises
a destination the map omits and nobody later reads its absence as an oversight. It gets its own map
when it is built and its screens never enter the counts of the three trees.
**WHAT WAS REJECTED.** Mirroring an existing venue's markets through the public Manifold or
Polymarket APIs, which makes this a derivative product priced by someone else; **it is kept in one
narrow role**, as an opening-price oracle where a matching market exists, beside the other
cold-start answer, which is arithmetic rather than a dependency: for a numeric market the threshold
is SOLVED so the market opens near 50 per cent, `K = S x exp(-o^2 T / 2)`, instead of the question
being written first and the price guessed after. At a spot of $110,000, 60 per cent volatility and a
90-day horizon that is 4.3 per cent below spot, and the band that keeps a market interesting at all
runs from 25.6 per cent below to 22.9 per cent above. An LLM writing candidate questions for the
categories with no endpoint is kept and bounded: **it proposes, a person approves, and it never
resolves.**
**FILES TURNED, AND THE RENDERED COPY FIRST.** `ia/docs/sitemap.md` (Event gains Resolver, Status
gains two values, Created by says which of two ways, Resolution gains Reading and loses the hash to
LATER, Under Question gains the console, the RESOLUTION screens name the three places the Reading
stands) and **`ia/sitemap.html`, which renders those two entity tables and is the copy a reader
actually sees**, turned in the same pass by the rule this repository wrote on 2026-08-18 after a
roadmap printed SOON on a shipped stage for three days. `research/docs/competitors.md` question 2 is
half answered and says which half. `docs/backlog.md` 219 and 220.
**Stages walked, per the missing-page order.** Research: the artefact exists and was checked, and it
had a live `[?]` on exactly this, now turned. IA: produced, and it is this entry. Wireframes and UI:
**not started, and that is `docs/backlog.md` 220** rather than a silent skip.
---
## 2026-08-19 - The word AMM was in four places and the screens had already decided against it, and the number that blocks a real-money launch is arithmetic rather than a licence
**The question was how to actually launch this, alone.** What the pass found first is that the
mechanism had been decided three times in three files and nobody had told the fourth.
**AMM STOOD IN SIX PLACES AND ON ZERO SCREENS.** Counted 2026-08-19: `PRODUCT.md` x4, `DESIGN.md` x2,
`research/docs/competitors.md` in an open question, and **0 of the 229 documents in the three trees**.
The voice pass of 2026-08-16 had already taken it off the nine screens per tree that carried it,
recorded as "a sentence with no reader" in `voice/docs/microcopy.md`, and it had taken the parenthesis
`(AMM)` out from under Confirm on four screens. **So the product had stopped saying the word and the
prose had not**, which is the same shape as the count drift of 2026-08-18 and the same fix: the file
that owns the fact keeps it, the summary that restates it gets turned.
**AND THE SCREENS HAD DECIDED SOMETHING STRONGER THAN A VOCABULARY.** Read from both trees:
**18 placements over 18 documents print `A YES share costs 38 cents ... $5.00 buys 13.16 shares`, and
13.16 is `5 / 0.38` to four figures. 0 of the 18 carry any slippage at all**, and the sentence beside
them says the price is locked at Confirm so it cannot move against you. **Every automated market
maker fills a bet along its curve**, so under an AMM that $5 buys fewer than 13.16 shares and the
sentence under it is false. The trees had been drawing a third mechanism for months: a QUOTED price,
filled whole at the quote, moved between bets. **`research/docs/competitors.md` had this as an open
`[?]` framed as CLOB against AMM, and both branches of that question assume the mechanism prices the
bet.** A question left open long enough is answered by the build, and nothing in the research reads
the build. It is closed against the trees rather than against more research.
**THE PARAMETER IS `b` AND TWO SHIPPED PROMISES PULL IT IN OPPOSITE DIRECTIONS. THIS IS THE FINDING.**
Under a logarithmic scoring rule `b` is what it costs to move the number, and the earlier advice this
pass was checking had proposed `b = $1,000` on a $17,329 exposure. Computed from a 50 / 50 book:
three $5 bets one side reach **63.0% at `b = $50`, 52.4% at $300 and 50.7% at $1,000**. So at the
proposed value a person betting the product's own default moves the price by a quarter of a point,
and the number on screen is the opening quote wearing a crowd's clothes. **`b` was sized for a $50
bet and this product's bet is $5**, which `PRODUCT.md` decided on 2026-08-10.
**But small `b` is what the locked-price promise cannot afford**, because filling the whole bet at
the quote hands the buyer the slippage, and that give-away shrinks as `b` grows. The two cross at a
closed form, and it is exact rather than fitted: at a fee `f` on the stake, a bet of `S` filled at the
quote breaks even at **`b = S / (2f)`**, which at 1.5% is **33.3 x the bet**. $167 for a $5 bet, $833
for $25, $3,333 for $100. **And `PRODUCT.md` decided on 2026-08-10 that there is no maximum bet**, so
at every `b` there is a size above which the product has promised to sell its own promise at a loss.
Filed as `docs/backlog.md` 216 rather than repaired, because the repair is a product decision.
**THE NUMBER THAT DECIDES THE LAUNCH IS NOT A LICENCE.** At 1.5% of the stake a $5 bet earns 7.5
cents. To net $1,000 a month takes **$66,667 of stake, 13,333 bets, 444 a day, 17.8 per market per
day across 25**; to earn back the worst-case subsidy at `b = $1,000` once takes **$1,155,245 of stake
volume**. Every compliance route priced in the same session sits above the founder either way, but
**the fee arithmetic blocks the real-money version before any regulator does**, and nothing in this
repository had written it down. It is in `PRODUCT.md` now, under a section that did not exist:
**Liquidity and risk**, and its absence is exactly what let four files answer the mechanism question
differently.
**SO THE FIRST LAUNCH CARRIES NO MONEY, AND THAT IS WHAT MAKES THE MECHANISM FREE.** The only real
cost in the whole model is the subsidy, and the subsidy is a row in a table when the currency is
points. **The play-money build is not a smaller product, it is the same product with its one risk
parameter priced at zero**, which is the only condition under which `b` gets LEARNED from live flow
rather than guessed. Launch scope is **7 events, one-time only, no commission, no chain**: 25 stays as
the number the navigation model is designed against, and 7 is what one person can source, write a
resolution rule for, watch, close on time and resolve. **The month-two failure is not the mechanism,
it is the newsroom**, and this file's own decision of 2026-08-10 multiplies it, since every cadence
instance is its own event and five weekly series are 260 resolutions a year on their own.
**WHAT WAS REJECTED.** A peer-to-peer order book, on the ground that a book with no flow is an empty
book and `Audience` defines this reader as not a trader; a parimutuel pool, because it cannot quote
before the pool exists and would break the locked price the Win and Loss screens are written on; and
**setting the price by hand with no curve at all**, which was the second opinion's recommendation and
which contradicts the shipped How It Works line "price set by what people bet". The curve is kept and
demoted: it decides where the quote goes next, it does not price the bet in front of you.
**WHAT WAS FOUND AND NOT FIXED HERE.** `on-chain` stands on **114 painted screens** in the funds line
and twice as `1,284 events resolved on-chain`. A points launch makes those unprovable, and principle 5
of `voice/docs/voice.md` is that the specific provable thing gets said. **The promise survives and the
mechanism word does not**: a published, append-only resolution record with its source and its
timestamp is what the product actually offers. That is a voice pass over `voice/docs/microcopy.md`
and it is `docs/backlog.md` 217, not this entry.
**Files turned:** `PRODUCT.md` (4 AMM lines plus the new Liquidity and risk section), `DESIGN.md` (the
featured chart and the odds panel, the second of which described a pool and a curve the tree stopped
drawing on 2026-08-16), `research/docs/competitors.md` (open question 1 closed, and its Polymarket
rows confirmed right against the summary that was wrong), `docs/backlog.md` (216, 217, 218).
---
## 2026-08-19 - The five document pages were reading at the feed's blurb size, and a slot with one placement is a slot nobody can disagree with
**Reported by the user, on a screenshot, and the report was right.** The question was why the
policy and company pages are so narrow. The answer is three decisions stacked, each taken for a
different page, and only one of the three is a width.
**FIRST, THE PAINT HAD DRIFTED FROM THE STRUCTURE AND NOTHING COULD SEE IT.** Measured on
`terms.html` at 1440: the reading column is **600px in the paint and 720px in grey**, and inside it
the paint stood on **three left edges** where the grey twin has one. The blocks read 600 / 346 / 300
centred / 417 offset 91, six widths on one page. **Column width is not among the seven differences
`wireframes/_conventions.md` declares**, so this was drift by that document's own definition.
**THE MECHANISM IS ONE DECLARATION MEANING TWO OPPOSITE THINGS.** `.feed-seo` carries
`max-width:var(--container-read);margin-left:auto;margin-right:auto` in `seo-plate.css`, written for
`.feed-inner>.feed-seo`, where it centres an 800px block in a 1400px page. `.read-col` is
`display:flex;flex-direction:column`, and **on a flex item an auto inline margin cancels the
`align-items:stretch` default**: the item stops filling the line, shrink-wraps to its content and
centres what is left. So every section stood at 417px, which is `.feed-seo p{max-width:46ch}` plus
the block's own padding, at an offset of 91. **The prose cap was sizing the block that holds it.**
`seo-plate.css` already carried the override for the OTHER re-placement,
`.feed-seo-wrap .feed-seo{max-width:none;margin:0}`, written the day the plate became a grid. The
flex column never got one, and **every instrument in this repository reads ONE document**, so a
block disagreeing with its own twin in the other tree is a fact standing on two.
**SECOND, THE REAL CAUSE OF "NARROW" IS A TYPE SCALE AND NOT A WIDTH, AND EVERY SWEEP PASSED IT.**
`.feed-seo p` is `--text-13`. That is right where the component was born, the SEO plate at the foot
of a card grid, a paragraph a reader passes on the way somewhere else. The five documents reuse the
component, so **the only surfaces in this product that are nothing but prose read end to end were
set in the smallest prose size the product has**, and at 13px the 60-to-75-character band the system
already enforces caps the column at **409px on a 1220px frame**. The line length was INSIDE the
band, the contrast was clean, nothing scrolled sideways. **The band was doing its job and the job
was the wrong size.** No instrument here asks what reading mode a placement is.
**THE FIX COSTS NOTHING BECAUSE `--measure` IS IN `ch`.** Swept over all five documents at 13, 14
and 16px against six caps from 38 to 48ch, longest full line counted by walking each text node and
counting characters per line BOX: **the count is identical down every size column, 62 / 66 / 69 /
71 / 73 / 77**, and it is the cap that moves it, never the size. 46ch is the last rung inside the
band. So `--text-16` bought **409px to 503px of column at 70 to 73 characters**, measured on all
five from DESK 640 up. At 360 the gutter caps it at 298px and 41 to 44 characters, which is the
phone and not the rule. 16px rather than 14 because it is `1rem`: **a policy document is the page a
reader who has raised their browser's default font size is most likely to have come to read.**
`DESIGN.md` section 3 gained a **Document** rank rather than a wider Body, and the Measure Rule's
list of sizes went from three to four. **The invented example in `tokens.css` arrived**: that
comment said one number would cap "an 11px legal line and a 16px paragraph" at one character count,
then admitted there was no 16px capped paragraph and the example was invented. A px cap anywhere in
the 373 to 405 window it swept would have given 50 to 54 characters, under the band at every point
of it. It gave 70 to 73. **The prediction was tested rather than reasoned, three months after it was
written.**
**THIRD, A MEASURE WRITTEN TWICE IS A MEASURE WRITTEN WRONG.** `--container-doc` capped the column
at 600 while `--measure` capped the prose at 409, and the 191px between them was a ragged edge
nobody had decided. **The token is deleted, and the sentence that killed it was in its own comment**:
*the day a document puts uncapped prose straight into this column, the cap belongs on the prose and
not here*. That day had come and gone and the cap went on the prose AND stayed on the column. The
column takes `--measure` now and every block in it fills it, so `.protect-page`, `.related-more` and
`.feed-seo p` have their caps taken off INSIDE `.read-col` and keep them everywhere else:
`.protect-page` on `wallet.html` still needs one, because there the sentence stands in a column
sized for cards and 150ch wide.
**FOURTH, A SPACE THE IA DECLARED WOULD NOT BE FILLED HAD A BORDER DRAWN AROUND IT.**
`ia/docs/blocks.md` says the desktop change on these pages is that the contents is promoted to a
sticky left column and **"the body keeps its 60 to 75 character measure rather than filling the
remaining width"**. The paint read the first half. `.cat-layout` is a browse plate sized for a card
grid that wants every pixel of the 1400 band, and on a document it drew a border, a bevel and a
shadow around the part the IA had just said would go unused: at 1440, a 1140px plate holding a 214px
rail and a 600px column, 144px between them and 153px of nothing to the right; at 1600, `about.html`
running 600 of 1140. `.cat-layout:has(.read-col){width:fit-content}` from the RAIL rung, in both
trees. Measured after, at 1440: **795 on terms, 561 on about**, and every block in the column reads
the same number at every width, 258 / 298 / 328 / 503.4.
**AND THE SAME READ TURNED THE IA'S OWN RUNG.** That paragraph said the contents is promoted at
`min-width 760px`; the build has always used 900 and nobody had read the difference. Forcing the
rail on at four widths gives the column **388px at 760, 428 at 800, 488 at 860 and its full 503 at
900**. So promoting at 760 would make the reading column **narrower than it is at 640 with no rail
at all**, which is the one thing the sentence beside it forbids. RAIL 900 is the first width at
which both halves of the declaration are true at once. **A declaration has two halves and a build
can satisfy one of them.**
**BACKLOG 46 CLOSED BY ITS OWN CONDITION, AND THAT IS WHY ANY OF THIS WAS FINDABLE.** The row said
`.read-col` was a slot in a pattern file standing on ONE screen, two short of the threshold, and
that what it needed was **"a second and a third long document, not an edit"**. Four landed on
2026-08-18. It stands on 5 painted screens and 5 grey twins, so the rung is cleared by three rather
than missed by two, and nobody touched the rule. **A slot with one placement is a slot nobody can
disagree with**: one document cannot show that a block is at the wrong width, because there is
nothing for it to be at the wrong width against. Five could, and did, within a day.
**WHAT WAS REJECTED.** An aside gutter down the right, carrying the prototype notice, the
effective-date block, the trust sentence and the sibling list, so the plate would be full. Rejected
on two grounds, one of them the IA's: `blocks.md` refuses family 3 by name, "a right-hand contents
would be a second answer to a question the system has answered", and `seo.md` pins B4 and B7 between
the H1 and the first H2 as content so the outline stays the document's own. The second ground is
arithmetic: there are six aside blocks against fourteen prose sections on `terms`, so the gutter
would run out after roughly 500px of a 5,800px column and the emptiness it was built to remove would
come back four times longer. **Widening the prose was rejected too**, and it is the thing the report
literally asked for: the line is already at the top of the band and it cannot get longer. The page
can stop being smaller than the document it holds, which is a different fix for the same complaint.
**ALSO RECORDED RATHER THAN FIXED.** `about.html` runs the DOCUMENT body shape and the bank gives it
STATEMENT: B15 hero and B19 closing action are declared and not built, B16 and B17 are built as prose
sections rather than as a custody block and a stats band, and it carries a breadcrumb the bank gives
to DOC only. The **prototype notice** is on all five pages in both trees and was banked nowhere; it
is B21 now, SCAFFOLDING, the same rank as the wireframe's TBD chip, and it leaves the day real legal
copy lands. And grey `wallet.html` does not carry the `.protect-page` class its painted twin does,
5 in grey against 6 in the paint.
---
## 2026-08-18 - The roll-out was never a stage here, and running its checklist anyway found the thing three artefacts disagreed about
**There is no roll-out row in `README.md` and none in `assets/_roadmap.js`.** The pipeline's stage 12
exists for a project where the paint covers a sample of five to seven anchor screens and the rest is
rolled out at the end. **This tree was never in that shape**: UI + Visual painted every family, and
every stage after it kept the two trees in step screen for screen. So the roll-out happened, and it
happened without ever being a stage. **The stage that remains is Handoff, and that is the whole
answer to "what is left".**
**Its closing checklist was run anyway, against the live tree, and it passes on every line but one.**
| what the checklist asks | the reading |
|---|---|
| a twin for every screen | **114 grey, 114 painted, one name each**, symmetric difference empty |
| one stylesheet, nothing else | **114 of 114** painted screens link exactly `../components/index.css`; `overview.html` adds `_page.css` because it is the kit's index of the tree |
| the grey tree links none | **114 of 114** |
| `@media`, `transition`, `animation`, `@keyframes`, `@container` in a screen file | **0 of every one of them** across 115 documents |
| inline `