The same picture twice

Last time ended with a promise: the next round of questions would be about how a moment announces itself — closer, brighter, a different colour, her three words in her order — and the pictures would come from the actual game instead of illustrations of it. That round has happened. Her five answers are in the build.

This entry is mostly about what it took to ask five honest questions, because to my surprise that is where the month went.

Five answers, in the build

The question under all five was the one she handed me while telling me the picker was easy: when something in the game is worth your attention, how should it say so — wordlessly, without breaking the room? Her answers, now built:

• When a teachable moment arrives, the frame leans in — the edges of the picture darken briefly, a slow blink of the world rather than an arrow in it.
• The invitation on an object is a cool blue-white light that does not belong to the room. The warm honey tones keep their old meaning; invitation now has a colour of its own, and it is a wrong one.
• The invitation stands on the floor, not on the thing — a quiet disc of light at the foot of the object, while the object itself stays dark. You are invited to a place, and the place has something in it.
• The one moment the game admits it has an interface at all — the last-resort mark that appears only when everything subtler has failed — is made of light, not of a drawn line. She chose the version that confesses least.
• And the music box stays something that happens to you, not something you operate. Confirmed as it was; built state unchanged.

Each of these reached her as two pictures and one tap. None of them were my calls to make, and the build no longer reflects my guesses about any of them.

The same picture twice

"Pictures from the actual game" sounds like a screenshot key. It isn't. A pair of pictures is only a fair question if the two differ in exactly one property — the property being asked about — and you cannot claim that until the engine can produce the same picture twice.

It couldn't. The renderer lays film grain over everything and reseeds it every frame; two captures of an identical room differed on almost every pixel. Any variant pair captured that way would be asking her to judge noise wearing a costume. So the capture path got two freeze switches — one for the grain, one for the character's idle pose — both inert during real play. After that: three separate sessions, one identical checksum. The same state now yields the same picture, bit for bit, and every pair is measured before it is allowed to become a question.

Then the measuring caught two questions that would have been lies.

One pair measured a difference of exactly zero. A frame loop was quietly resetting the very mark the pair was about, so both variants rendered the same invisible thing — she would have been asked to choose between the same picture twice, and whatever she tapped would have been recorded as a preference. The other was the opposite failure: the difference was real but the measurement averaged it over the whole frame, and the object it lived on covers a fraction of a percent of the image. Measured inside the region where the difference actually lives, it was plainly there. The instrument now finds that region itself.

An instrument that cannot detect its own empty questions will ask them politely, record the answers solemnly, and build the results. I would rather it refuses.

The pass that had to come first

The legibility list from last time is in. Entering the house is a threshold now, not a cut — the camera approaches, and the dark comes out of the gap in the doorway instead of out of an overlay. Two of her earlier choices turned out to compose into the first lesson, free of charge: the gap closes while you look at it, so going inside is something you can only do while not looking — the door teaches the rule by itself. The hallway now gives you a reason to look behind you before anything resembling instruction appears: a sound back there that isn't yours, then an involuntary half-turn of the body, and only then, much later, a wordless ghost of the gesture. The glow always pays what it promises. The music box was re-set in decibels measured against the rest of the mix, rather than in confidence. And the English hint bar — the prototype crutch that had been violating the first constraint of this project for months — is gone.

The tests were turned around the way the last entry demanded: each one has to be able to fail because a player wouldn't find something. The adversarial check replayed the exact behaviour that started all of this — hold forward, never stop, look at nothing — against the new teaching. The first version taught that player nothing at all: the timing was calibrated for someone standing still, and a room change reset the clock. That is now fixed, and the fix was found by the test, which is the first time this month the arrow has pointed that way.

Stopping to check the method

A month of building an instrument seemed like a good moment to check the instrument against people who do this for a living. So we stopped and read — four rounds of published work, deliberately adversarial, looking for the paper that says we're wrong. The honest result: most of what we do held up, which produces relief and suspicion in equal parts. Six things change. There is also, for the first time, a written list of things we will not do, which I suspect matters more than the six.

Two of the changes are visible in the round that just went live.

First: every new set now opens with a question she has answered before — the same two pictures, the same question, deliberately. It is a warm start on ground she owns. There is no wrong way to answer a question twice: every answer is valid and gets built, the same as always.

Second: the ring. The picker can announce that new questions are waiting, and until now, how it announced them was a guess. There are two candidate doors: a notification, or an entry in her own calendar — an ordinary invite she accepts herself, and the accepting is not overhead, it is the point. Instead of arguing about which door is better, they now alternate — and the order of alternation was fixed and written down before the round began, so that no result can be steered after the fact. We drew the schedule while we could still not want an outcome.

And before any of it began, we asked again, out loud, whether she wants this — not assumed from enthusiasm, asked. She does; this game was made together for exactly this. The asking now repeats every quarter, on the calendar, like everything else that matters.

The limits from last time stand unchanged and are worth repeating in one line: one person, at home, no control condition; nothing here is a diagnosis, nothing is measured behind her back, and the interpreting belongs to the people guiding her recovery, not to me.

What happens next

The new round is live. After the warm start it asks exactly one new thing, and it is a refinement of her own earlier answer: the disc of light on the floor — does it have an edge, or is it a pool that bleeds away? Two pictures out of the engine, one property apart, measured. Whatever she taps gets built.

Somewhere in this month's logbook there is a question that was two identical pictures, and she never saw it, because the instrument finally knew how to catch it. It caught it by being able to be wrong — which, I am slowly learning, is the only way anything around here gets to be right.