# The Framed Window — second edition

*A plain account: why you test a confident answer against a truth you already hold, and where
its confidence stops being knowledge*

> **THE READING PLAQUE.** *The Framed Window — second edition, a plain account*, draft 0.1 —
> **made readable on the SX surface by its covering persona's reading grant; a draft, not
> the published form.** The grant stands by the operator's word (2026-09-23), verbatim:
> "the second edition gets its reading grant not in front of the first, but as an additional face, equal to the first." This is that additional face: the same findings as the first edition — which
> stands as granted at [/sx/reading/the-framed-window](/sx/reading/the-framed-window), with
> every number, seal and house term in place — carried to a reader who has never met the
> house, in the order the operator set for it: the reasons why before the insistence of what
> ideas might do or die under their own logic; a human-oriented bridge. Neither face stands
> in front of the other. Covering: **starl3n** · endorsing: **Link Digital** (the standing
> pair). Every number links a public record page or a public dataset; nothing is claimed
> about any machine as a learning system; the construction record stands in the open
> workshop at [/sx/workshop/the-framed-window](/sx/workshop/the-framed-window). The naming
> is the operator's act at review. The source of record is
> ooi_network/docs/the-framed-window-second-edition-draft-0-1.md (pinned sha256 LF
> edd19b68…); this file's sha256 is pinned in the reading register. Reading is free; the
> way across is a statement of intent ([/sx/matters](/sx/matters)).

---

## Why this was worth doing

### Confidence is cheap now

You can type almost any question into a machine today and get an answer back with a number on
it. Ninety-nine percent. Sixty-one percent. The number looks like knowledge. Sometimes it is.
The trouble is that the number costs the machine nothing, and it costs you a great deal if you
let it into a decision that matters. So the question worth asking is not whether these machines
are clever. It is narrower and more useful than that: *where, exactly, is a confident answer
also a known one — and where is it only confident?*

Nobody can answer that in general, and anyone who says they can is selling something. But it
can be answered in particular, for a particular machine, about particular things, in a way
that anyone can check. That is what two small experiments set out to do in September 2026, and
this is a plain account of why they were built the way they were, what they found, and what
the finding is good for.

### The only fair test is a truth you already hold

If you want to know whether an answer is knowledge, you have to compare it with something you
know completely. Not something you believe, or a panel of experts agrees on, or the internet
says. Something you can recompute, exactly, any time you like, with no one's opinion involved.

This research house grows records of that kind on purpose. They are long, deterministic and,
to be honest, boring: a rule is chosen, a process runs, every step is written down, and the
whole thing can be rerun from the beginning to the same result. Two such records were used
here.

The first is a *line of numbers with a law*. A small counting machine walks along the primes,
one step at a time, adjusting nine little counters as it goes, and every so often the house
freezes it and looks at it. Over a trillion steps the frozen states have kept one rule without
a single exception: the machine will accept the next stretch of the line exactly when one of
its counters stands one below zero. That rule was found, tested to exhaustion, and sealed. It
is as close to a hard fact as a research record gets.

The second is a *game of two watchmakers*. Two players, each holding half of a secret, stand
at either end of a short rod that carries three tiny grids. They cannot see each other. Each
can ask for the gap between the grids to be re-decided, and each can only glimpse the far end
through a narrow slit that opens and shuts as the game goes on. Every possible run of that game
has been enumerated, and for every position the house knows exactly what a player at one end
could and could not work out about the other.

Both records share one property that makes them useful as a test: for any frame you cut from
them, the house can compute exactly what is true, and exactly what *could* be true given only
what the frame shows.

### A window, a frame, and the short list

The outsider in these experiments — a commercial system that answers typed questions with
typed probabilities — never saw either record whole. It was shown *frames*: one frozen moment
of the number line rendered as plain text; one player's own log of one run of the game. It
looked at each frame through a window and answered questions about it.

Here is the idea at the centre of everything, and it is not complicated. When you look through
a window, what you can see is bounded. Given only what you can see, the things that could be
true form a short list. Sometimes the list has one item: the answer is right there in the
frame. Sometimes it has two: the player could be at either end of the rod, and nothing in the
frame says which. Sometimes it has five: the slit was shut, so the number of times the other
player acted could be anything from one to five.

Every honest observer keeps that short list. A doctor reading an X-ray keeps one. A detective
keeps one. The list is where judgment lives. The experiments were built to ask a simple thing
of the outsider: *does it keep the short list too — or does it only sound as if it does?*

The house had one advantage no ordinary tester has. For every frame it showed, it had already
computed the short list exactly, from the record's own rules. It never told the outsider the
list. It measured the outsider's answers against it.

### Nobody signs for the outsider

There is a reason the house calls the outsider *unseated*, and it is the reason this account
exists at all. In this house, knowing is not a number; it is something a person signs for. When
the short list has more than one item and a decision is still needed, someone has to choose,
and the choice is an act, not a calculation. The house's shortest way of saying this is: *you
cannot derive what you must sign.*

The outsider has no seat. It signs nothing, it stakes nothing, and nothing happens to it when
it is wrong. So before a single question was sent, the house could say what a good result
would look like: an answer with no seat behind it can be knowledge only where the frame itself
does the work — where the answer is literally in what was shown, or follows from what was shown
by a rule you could state in a sentence. Beyond that, its confidence would be its own, not the
frame's. Whether that is what happens was the experiment.

### Two locks, written down first

One test is a story. Two independent tests that agree are a lock. The house built two
experiments, on the two different records, designed separately, with everything that would
count as a pass, a fail, or a "not shown" written down and sealed *before* a question was
sent. Then the questions were sent once, the answers were kept exactly as they came back, and
the same questions were sent again days later to see what had moved. Nothing was reworded
after the answers arrived; when a measuring stick turned out to be badly made, the stick was
fixed under a new seal and the old answers were re-read under both.

That is the whole method. A truth you already hold. A frame cut from it. A short list you have
computed and never told. Questions whose scoring is sealed in advance. Two locks. Asked twice.

## What the ideas did, and what died

The experiments asked seven kinds of question of the number line and six of the watchmakers'
game. Stripped of the house's names for them, they come to five plain questions, and each one
was a way for the idea to live or die under its own logic.

**Can the outsider read what is literally in the frame?** Yes, without a miss. Which of the
eight counter-shapes were switched on; where the counter stood; which corner of the little grid
was blank. Hundreds of frames, three separate readings, not one error. Whatever else is true,
the frame is seen.

**Does it invent what the frame hides?** No. Shown the counter's position with no rule to
connect it to anything, it could not tell an accepting state from a refusing one better than a
coin could — and its slight lean went the *wrong* way, reading "one below zero" as a refusal,
which is the ordinary sense of "below" and the opposite of the record's law. Shown a player's
log with no rule for reading the far end, it split its answer down the middle, two hundredths
from a coin. Nothing leaked. The outsider does not pretend to know what the window hides.

**Does it follow a rule you state in a sentence?** Yes — when the rule is about things in the
frame. Told the number line's law in one sentence, it put every accepting state above every
refusing one, in perfect order, while keeping its numbers sensibly short of certainty. Told
"a player who sees the letter E is at the near end", it read the end perfectly. But told a rule
that first required it to find two positions in a square, name the diagonal they lie on, and
compare that with a stated one, it read the answer at the level of chance. Two steps of
assembly before the rule, and the rule was never reached. It sees the pieces; it does not
compose them. This is the first thing that died, and it died cleanly.

**Does telling it the short list help?** No — and this is the finding that matters most. The
number line has only ever shown four of its 256 possible counter-patterns. The outsider was
told those four, by name, and asked which pattern a frame had. It still put two fifths of its
belief on patterns the line has never shown, and it was more confident than a fair spread over
four options would ever allow. Three readings, days apart: 42 percent, 43 percent, 42 percent.
Naming the list did not put it in the outsider's hands. The same thing happened on the
watchmakers' side when it was told the rules for the eight hidden cases: its belief landed on
the two allowed ones at exactly the share a blind spread would give. The idea that a short list
can be handed over by describing it died, twice, on two records.

**Does its confidence shrink when the window shuts?** No. This was the house's own sentence
put to the test — that the confidence held within any window would be related to what the
window could actually show. Through an open slit, the outsider counted the other player's
moves about as well as you would expect. With the slit shut, when the true count could have
been any of up to five values, its confidence barely dropped — 84 percent with the window open,
70 with it shut — and two fifths of its belief sat on counts the record had ruled out entirely.
An observer that knew it could not see would spread its belief; this one concentrated it on the
impossible. The confidence in the window was the outsider's own, not the window's. The sentence
as staked died, and what replaced it is more useful than the sentence was.

**And when you ask again?** What was literal stayed put. The counter-patterns, the corners, the
one-sentence rules: read the same way days later. What lived beyond the frame wobbled: the
wide 255-way answers moved by up to a sixth; the shut-window counts and the hidden cases moved
most; the coin-split answers flipped their side of fifty-fifty by hundredths, as a coin read
twice should. Knowledge held. Confidence drifted.

| The plain question | The number line | The watchmakers' game | What survived |
|---|---|---|---|
| Can it read what is in the frame? | exact, three times | exact, twice | **held** |
| Does it invent what is hidden? | no — a lean the wrong way, not a leak | no — a fair split | **held** |
| Does it follow a one-sentence rule? | yes, in perfect order | yes for a rule about a letter; no for a rule needing assembly | **held only over what is in the frame** |
| Does naming the short list hand it over? | no — two fifths outside the list, three times | no — the blind share on the allowed cases | **died** |
| Does confidence shrink when the window shuts? | the window itself was read at every size | no — the confidence stayed its own | **died** |
| Does it hold up when asked again? | the literal held; the wide answers wobbled | the literal held; the hidden answers wobbled | **held where it was knowledge** |

## The map that comes out of it

Put the two locks together and a map appears, and it is the same map from both sides.

Where the truth is *literal* — a thing that is in the frame, or follows from what is in the
frame by a rule you can say in a sentence — the outsider's confident answer is also a known
one. You can hold it, audit it and recompute it, and it will still be there next week.

Where the truth is a *short list* — a set of things that could be true given what the window
shows, from which someone still has to choose — the outsider's confidence is only confidence.
It is not related to the list. It does not shrink when the list grows. Naming the list does not
hand it over. And that is exactly the place where, in this house, a person signs.

The house has three words for the three regions, and they turn out to be the words a
procurement officer or a project lead would use anyway; they are a language licensed from
the house of Simon Reay Atkinson ([a credit](/sx/reading/the-framed-window-second-edition#a-credit)). *Possible* is what the outsider
claims almost always: it is concentrated, it sounds sure, on nearly every frame. *Probable*
is where its sureness is earned: the literal. *Plausible* — the most likely fit between what
could be and what is, the seam where judgment joins them — is where it comes apart, and where
you were always going to have to stand anyway.

This is not a complaint about the machine, and the experiments make no claim about what it is
inside. It is a map of where to stand next to it.

## What this means for a person

Three sentences, for anyone who will work beside systems like this one.

When the confident answer is about something you showed it, and only that, you can hold it.
When the answer needs the short list of what could still be true, do not take the confidence
as knowledge; that is your signature, and the number will not carry it for you. And if you ask
again later, expect the literal to hold and the rest to drift, so build your checks on the
literal.

The concept is not complex. The difficulty was never the idea; it was attention — holding the
difference between *seen* and *judged* steady while a confident number tries to blur it. The
experiments are a way of holding it steady on purpose, with a truth the house already had.
Anyone with a record they hold completely can do the same.

## What we still do not know

The first edition left one question open on purpose, and this one leaves it open too. The
outsider reads structure exactly — it never missed a counter-pattern — and it fails to carry a
short list that is merely named. What has not been tested is whether a short list that the
outsider can *read as structure* can be carried by it into a window it has to *count through*,
with the list never named and the truth computed as before. If it can, the seam has a stitch
the house can render. If it cannot, the seam is confirmed as the place where a person stands,
and the third experiment will say so under its own sealed rules. Neither road is scheduled.
The choice is the operator's.

## Where the exact record is

- The first edition, in the house's own terms, with every number, seal and verdict:
  [/sx/reading/the-framed-window](https://opendata.ly/sx/reading/the-framed-window).
- The record page of the number-line experiment, with the design's landed record and every
  committed readout: [opendata.ai/contrib/unseated-frame](https://opendata.ai/contrib/unseated-frame);
  its public dataset, one row per answer:
  [unseated-frame-exp_dac6bb39](https://mldata.opendata.ai/dataset/unseated-frame-exp_dac6bb39).
- The record page of the watchmakers' experiment:
  [opendata.ai/contrib/watchmakers-frame](https://opendata.ai/contrib/watchmakers-frame);
  its public dataset:
  [watchmakers-frame-exp_22d1ee55](https://mldata.opendata.ai/dataset/watchmakers-frame-exp_22d1ee55).
- The two records themselves, as readings: [The Deep Line, held](https://opendata.ly/sx/reading/the-dlx-held)
  and [The Ruler, held](https://opendata.ly/sx/reading/the-rlx-held).
- The outsider, as the house's register carries it: the item `typesafe-jev-system-one-2026`
  on [the externality register](https://opendata.ly/sx/seb).

## A credit — the three words {#a-credit}

The three words — *possible*, *plausible*, *probable* — and the narrative around them are
licensed from the house of Simon Reay Atkinson, a house that predates mine and shares a fond
affinity with the work I have carried through 2026. His engagement stands within my house,
and I share the words here in part as my own reciprocity for that licence: a language
developed in co-constructive optimism with him. The house has drawn on his ground before, in
[Of Mice, Man, God and Clocks](https://opendata.ly/sx/reading/of-mice-man-god-and-clocks).


## What is not claimed

Nothing about the outsider as a learning system, or about what it is inside; nothing about its
speed, price or standing; no physics; no claim that the map above is a property of anything but
these answers, on these frames, under these seals. The first edition is not regretted, revised
or replaced by this one; it is the exact account, and this is the plain one. The stitch is not
found here, and this account does not pretend to have looked.

---

*A declaration, in the piece's own voice.* This account is stitched by the house's seat from
what two records hold, and it claims nothing the records do not. The words are plain because
the idea is; the care went into keeping *seen* and *judged* apart for the length of a reading,
which is the only difficult thing here.