How Double Diaper Dungeon got made

The first five stickers are on the fridge, dated when they were earned. The transcript of the design conversation so far rides along, written by the Stop hook.

Every number lives in the balance table with where it came from, and the documents name entries rather than quoting values. Ninety-nine of the hundred and forty-one entries are provisional and say so.

Beside them: a validator that reads every page and every issue and refuses when they disagree, a test front door that runs the validator and every scenario, and the questions page grown by everything the writers could not settle on their own, each with the working ruling they wrote under. One document was made true: the combat page had left open whether a hit on a stunned fighter counts for dodge recovery, and the ruling says it does.

The gate refused one of its author's own commands on its first outing, because a sentence about her pages looked to it like a write to them. It now looks for a path rather than a word, and will not read an arrow as a redirect. Both mistakes are cases in the test. The project is licensed: the Affero GPL, version 3 or later, which she chose in two places. The full text, a page explaining what it asks, the notice written once inside the tool that stamps it, and the one source file that predates the choice stamped on the way past.

Three of the answers replaced mechanisms rather than filling gaps. The ghost room is no longer measured from the entrance: the generator finds the level's longest journey and takes both of its ends, one for the way in and one for the ghost, pricing a locked door at the whole walk to its key and back. Infestation stopped being a countdown in every empty room and became one room turning at a time, which is a tide a small party can outrun. And the dodge rule's stats moved: avoidance is defensive attention, so an attacker's imagination is what finds the gaps in it and an attacker's dexterity became a separate roll to land the blow. Smaller and just as load-bearing: caregiving has a floor under it, and a hero below that line is not a caregiver at all. Classes stopped being rolled for and are now read off the priority sliders at each level-up. Basecamps have no bathroom, so a change happens in the open and embarrasses everyone in the middle of the band, while the very grown-up and the very little are both unbothered. There is a fog of war, which she wrote while answering something else. The leader turned out to have an inner life, a chance to go to pieces, and a small face in the corner of the screen. Two entries left the balance table rather than sit there describing rules the game no longer has, and the validator learned to check that a name written in backticks is a name that exists, which is what would have caught the rename those two came from. One word in her answers names a stat this project does not have, and nothing uses it until she says what it is.

Spirit is the volatility of the mind. It resists being shamed, it makes a hero fall further and come back faster in equal measure, it feeds the imagination that makes a regressed hero dangerous, it leaves them open to hypnosis, and it earns a little more from everything survived. Three more of her rules landed with it. An opening is a moment when a defender's attention is elsewhere: imagination makes more of them and dexterity makes each one hurt more, which deletes a roll, because a second roll to decide whether a hero notices the opening they just made was a coin flip nobody asked for. The bladder's two stats swapped jobs, constitution holding the capacity and strength doing the holding, since storage and muscle had been the same stat by accident. And a caregiver who slips below her own floor keeps adventuring, loses every caregiver ability, and is picked up twice as fast by another caregiver, because she knows exactly how the care is meant to land. The wall around her handwriting came down to a plain list of two paths, which is what she asked for. The sections are gone, the append-only idea is gone, and the list no longer appears on itself. The gate behind it stopped reading a mention as a write, after refusing three of its author's own commands and then refusing the attempt to write its own test.

Then a tool to look at them, because whether five colours stay five colours is a question about an eye and not about a number. A strip of uneven swatches reshuffles every fifteen seconds or on the H key, so every colour eventually sits beside every other at every width. Below it a page of text with roughly every tenth word coloured, which is the density a battle log actually has. The C key redraws the whole thing as a protan, deutan or tritan eye receives it. A headless twin runs the same arithmetic on every build. It has already earned itself. Every colour stands well off the dark map, and two pairs fall together: light purple and skyblue for an eye with no red cone, bubblegum pink and yellow for an eye with no blue one. A search over the same five hues found a palette that never collapses and got there by muting the lime to moss and the sky to steel, which is not the game she is making, so the bright ones stand and the finding is written down instead. Which is the right answer anyway, because she also gave five ways to tell one hero from another that have nothing to do with colour: the class emoji in front of the name, names from more than one tradition, a battle log that holds only the heroes in that battle, a dot whose shape is rolled once at creation and never changes, and a striped dot for a hero whose top stats tie, whose stripe order is saved between playthroughs and whose outer stripes widen when a middle one would otherwise vanish. Unless there are only two stripes, in which case, on her ruling, that is not really a rainbow.

Two things in it are load-bearing and neither is obvious. Letters are apportioned by largest remainder, because rounding each share alone leaves the name a letter short and the reader sees a typo. And every cut is counted in characters rather than bytes, because an accented letter is several bytes and cutting inside one produces a string LOVE refuses to draw at all. The decoding is written out in the file rather than borrowed, since LOVE ships that module and plain LuaJIT does not, and the headless runner is plain LuaJIT. She looked through the colourblind eyes and saw that some coloured words came out darker than the plain words around them. She was right, in eight places. A word the game takes trouble to colour should not be the hardest word on the line to read, so ordinary text now sits at the brightest grey that still leaves every stat colour outshining it to every eye, which was found by measurement rather than picked, and the checker tests for it from now on. The colour picker she asked for is written down as the real answer, because the right value depends on a monitor and an eye. The dot on the map takes its colour from the sliders rather than from whichever stat is largest, which was her answer to a question about a colour that would otherwise change under the player and stop being true. And her handwritten note about how a hero is made is sorted: it sits in notes/ where it is read and never edited, and issue 215 is the blueprint from it. Gender and body history are two fields there, not one, because her rule turns on the difference: the pronouns and the magic go by the gender between the ears, and the other field only says which bonus was carried and is never shown to anybody.

So nothing writes a colour down any more except the balance table and the player. A resolver reads the two layers in a fixed order, the player's own choices from input/ first because that is the first thing the program reads, and the project's defaults second, and it is the only way anything finds out what colour something is. The window that judges the palette asks it too, so a player who has recoloured her stats and then opens the checker is looking at her own colours rather than at the project's. What it refuses is the point of it. A settings file that will not load stops the program, because half a palette is worse than none: some colours hers and some the project's with nothing on screen saying which. A colour that is not three numbers is refused rather than clamped, since a clamped colour is one somebody meant to set and did not. A stat with no colour anywhere stops the program at startup rather than four levels into a run. And a colour for something this version does not have is kept rather than discarded, because it is a choice she made on purpose and may be for a version that has not arrived yet. The lime is more vivid, which she asked for, and the way to make a lime more vivid is to take red out of it, red being what drags a lime toward the yellow beside it. The gap to spirit's yellow opens by a third and it is still plainly lime.

What makes the tool worth having is what it does not do. It replaces a word only where the word stands alone, meaning the characters on either side are neither letters nor digits, so an underscore counts as a boundary and a balance entry changes along with the prose while every one of the two hundred and fifty stickers in this project is left exactly as it was. It leaves the transcripts alone, because they are the record of what was said and the record does not get edited to agree with a later decision. It leaves her own files alone, because it reads the list of them. And it leaves itself alone, because a tool that rewrote its own examples would stop having any. The two issue files whose names still said the old word were moved with git, so both names are in the history. The colour window resizes its text now, on the wheel or on plus and minus, with equals putting it back. A LOVE font is rasterised at one size and drawn at that size, so a bigger font is a different font rather than the same one scaled, which means changing size means building fonts, which means they are cached by size. The layout stopped using fixed drops from the top and now measures itself in line heights, and the help line wraps and pins itself to the bottom by its own height, because the list of keys does not get shorter when the text gets bigger. And there is a status page, because "how far along are we" is a fair question with an answer that is hard to hold in a head. The written half and the running half grow at different speeds, and a sentence about the project flatters one of them. So it counts both, separately, and says which is which.

A bridge is a door that is the only way through. Locking anything else is a locked door with a corridor round it, which the party walks past and never notices. The first attempt looked for doors leading out of the reachable region and found none, because after a spanning tree every room is reachable, which is the point of a spanning tree. Then the test found the real one on its thirty-second level. Putting each key on the near side of its own lock is necessary and not sufficient: three locks can form a ring, key one behind lock two, key two behind lock three, key three behind lock one. Every one legal when it was placed, because the lock it ended up behind did not exist yet. Together they hold each other shut and four rooms of nineteen are reachable. So a lock is now placed, the level walked end to end picking keys up, and the lock taken back off if the level can no longer be finished. A hundred and sixty levels across four sizes, and all of them can be walked. The viewer draws her floor plan: black, dark grey for rooms nobody has entered, yellow lines for locked doors, red for walls, and an open door drawn as a line with its middle third taken out. None of those colours is written where it is drawn. She asked for themes before the map existed to hard-code colours into, so a theme is a named set of answers and every colour is pointed at by name. The haunted one is the game as it is. The nursery one, parchment with a texture and crayon drawings that move, is named and not built, which is the entire point of naming it. And the checker caught the word-renaming sweep changing a word inside one of her own sentences. Her line is back exactly as she typed it, the sweep now steps over any line beginning with an arrow, and the checker learned to ask whether a removed line is still gone rather than failing forever on a removal that was put right. It had two bugs of its own in that, both the same bug: every one of her lines begins with an arrow, an arrow begins with a minus, and every tool handed one read it as an option.

The viewer zooms the map now instead of the letters, which is what zooming means. The wheel multiplies rather than adds, so a notch feels the same size at every zoom; dragging slides it; equals puts it back. K then a number asks for that many keys and keeps the rooms, T then a number asks for that many rooms and keeps the keys. And it points. Rest the pointer on a room holding a key and an orange trail runs from it to the door that key opens, through the centre of every room between, with a ring at each end and a line saying how far it is. Rest it on the locked door instead and the same trail runs the other way. Two ends of one answer, so it is found once and drawn once. The checker learned something too, and it took a mistake to teach it. Putting her corrupted sentence right meant removing the corrupted version, and that version also began with an arrow, so the checker reported the repair as a loss. It now asks whose a line was before calling it missing, and the answer is in the commit that first wrote it: every commit an assistant makes carries a line naming the assistant, and hers do not.

The cause was a name. A door is stored once, under whichever of its two rooms sorts first, and those two rooms were called near and far. Sorting has nothing to do with where anybody is standing, so the trail aimed at the arbitrary one, and measured over a hundred and eighty locks that was the wrong room fifty-two percent of the time. A coin flip, which is what it was. They are called side one and side two now, which claims nothing, and there is a function that asks the question worth asking: shut that one door and see which of the two rooms the key can still be reached from. A lock only ever goes on a bridge, so shutting it cuts the level in two and exactly one side survives. The trail is drawn with the lock shut, so it cannot cheat by walking through the very door it is pointing at. Hexagons stand two ways now, on a corner or on an edge, switched with a key and read from a display preference. It changes nothing about a level: same rooms, same neighbours, same doors, same trail, turned thirty degrees. The one thing that did not follow from that thirty degrees was which two corners make up the wall between two rooms, which was a lookup table that was right for one orientation and quietly wrong for the other. It is found rather than looked up now, by asking which two corners lie nearest the neighbour's middle, which is true of any orientation. And the check that walked the history for a line of hers is gone, at her word. What replaced it is a tool that gathers everything she has written into one file to be read when the deciding has been forgotten, rebuilt on every test run so it is never stale. Her reasoning, which is better than the check was: a thing you read is more use than a thing that tells you off.

The order this happens in is forced rather than chosen, and it is the part worth remembering. A key goes on the side of its lock the party is already on, and the side the party is already on is the side the way in is. Place the locks from somewhere else and then declare an entrance, and the guild arrives behind a door whose key is on the far side. The level looks perfectly ordinary and cannot be finished. So the ends are found first by plain steps, the locks go in from the way in, and the ghost room is found again with the locks priced in. The way in does not move the second time; a door has been locked against it. The pricing is what makes this more than geometry. On twenty-five levels in thirty it puts the ghost somewhere a plain door count would not, and the handsome case is a ghost room in the room next door to the way in, with a locked door between them and its key across the level. Adjacent, and the furthest thing on the map. Both are drawn in purple, on her word, with the word in the room saying which is which, since a player only ever wants to find one of them at a time.

She also cut the arithmetic. A locked door used to be priced at the detour to fetch its key and come back, which measures the same idea exactly, takes a paragraph to explain, and cannot be worked out until every key is placed. It is now simply twice a step. And no key goes in the way in or the ghost room, which is hers: a key at the entrance is a key nobody has to walk for, and a key in the ghost room is only reachable once the level is over. One lock sometimes has to come off, and that turned out to be the rule working rather than failing. The way in is an end of the longest journey and a locked door counts double, so the rooms furthest away are exactly the ones sitting behind locks, and the measurement cheerfully picked an entrance sealed in a pocket of one room where no key could be placed at all. A lock the guild is sealed behind is a lock in the wrong place, so it opens and the wave runs again. Over forty levels, a hundred and eleven locks of a hundred and twenty keep their key. The map answers two more questions when pointed at. The way in draws the whole run: out of the door, round by every key it needs, on to the ghost, with a dot at every room because the route doubles back on itself and a bare line would be hard to follow where it crosses. The ghost room draws the way home, keys already in hand, since by the time anybody stands there they have them all.

The idea the whole thing turns on is that a drag keeps the position the mouse is ASKING for, worked out from where the press happened and where the mouse is now, and never from where the window ended up. Where the window actually sits is only what the collision allowed. Keeping those two numbers apart pays for three separate things at once, and pays with arithmetic that was going to happen anyway: a window stopped against another owes the mouse a distance, and does not move again until the mouse has come back across the whole of it -- which matters because the cursor is hidden during this kind of drag, and the window is then the only thing telling a person where their mouse is; a window that slides around a corner catches straight back up, rather than trailing the mouse by however far it was cut short for the rest of the drag; and a window with nothing in its way sits at its start plus the mouse's travel exactly, with no drift to accumulate. On top of that, three rules she asked for. Shove a window more than halfway across another and the two exchange corners, halfway being read as one middle passing the other, which is the reading that still means something when the two are different sizes. A swap is tried on paper first and refused if either window would land on something -- her worked example was three windows in a column and one tall one beside them, where the tall one would be squeezed into a third of its height, and that cannot happen even though resizing is not built yet. And a window dragged until more than seven tenths of its area is off the screen shuts itself mid-drag, area rather than distance so that going over a corner counts properly. The collision already worked out a limit rather than testing the destination, which is what stops a fast drag passing clean through. That turns out to be inseparable from the swap: to end up beyond a thing your middle has to get past its far edge, which is past its middle, so every shove hard enough to tunnel is also hard enough to swap. The only thing telling them apart is whether the swap fits, which is why the test for it now needs three windows. One thing found while building: sliding one axis can open the other, so the two are run again until neither moves. Without that, a window that slipped around a corner would still be holding the name of a window it was no longer anywhere near, and would then offer that window up as something to jump over. Tried by hand in a blank room with a coloured circle for each attribute; clicking one opens a character sheet with a paper doll and a column of numbers that moves to whichever side of the sheet has room. The sheets are the excuse. Checked headless by twenty-four groups of cases. Three questions are open on it and are hers to settle: whether a refused swap should instead let the window hop over and leave the other alone; whether the feel numbers belong in the balance table; and whether a window should be stopped at the screen's edge at all, given that falling off it closes it. Stickers this patch: 18 gold stars, 9 apples, no frowny faces, no worms. All time the same.

THE NAMES WERE BACKWARDS. A hero's name is cut into coloured pieces, one per attribute, in the proportions of their sliders, so that the name itself says what the hero is made of. It was cutting them in whatever order the shares arrived in, which meant a three-letter piece could sit in front of a five-letter one. She saw it and said so: the biggest piece has to come first, and then the next biggest. A short piece in front of a long one reads as though the short one mattered more, which is a picture that lies. So they sort now, and the pieces get shorter as the name goes on. The interesting half is what happens when two attributes are level. They have no natural order, and picking one would be a lie either way, so each hero carries a shuffle of the attributes rolled once when they are made and kept for good -- which is the same shuffle she already specified for their dot on the map, so the two agree about which of a hero's level stats comes first. Two heroes with identical even sliders therefore read differently from each other, and each of them reads the same way every time the game is opened. That tiebreak is not decoration on the sort; it is what makes the sort mean anything. The language's sort is not stable, so level shares without it would come out in whatever order the sort happened to leave them, and that can differ between two runs of the same build on the same hero. A name that recolours itself between startups cannot be used to tell anybody apart. The shuffle is rolled by a small routine that takes the randomness as an argument rather than reaching for any, so that when heroes are really made the roll comes out of the named hero stream and no other. A WINDOW NOW GOES AT FOUR TENTHS OFF, NOT SEVEN. She pushed the windows around and found that dragging one off the edge hardly ever fired: seven tenths meant shoving a window nearly out of sight before it would go. Four tenths of the area, which is her "total, not just one dimension" -- thirty per cent past the right edge and thirty past the bottom is inside the mark on either edge alone and a hair past it once the area is counted. Logged in the tuning ledger with the reason, and it is the argument for moving the handful of windowing feel-numbers into the table where she can turn them herself. Also written down, not built: the sixth kind of window, which is hers. An exclamation mark on a room saying which puzzle is in it, and behind the mark a place to watch the heroes try to get through, with a picture and a running commentary and no buttons at all. Her constraint arrived in the same breath as the idea and is part of it: there is no art budget, so the picture is made out of shapes and the palette and the emoji list, and that is the picture rather than a placeholder for one. Three questions it opens are hers, the largest being whether every room really carries a puzzle. Stickers this patch: 20 gold stars, 9 apples, 1 frowny face, no worms. All time the same. The frowny is for one window out of five, and it comes off the moment the other four arrive.

A strip along the bottom is reserved the same way, for the row of controls that will live there, and the line above it is drawn brightly while it bites and faintly while it does not -- an edge a person cannot see is an edge they will think is a bug. The strips are kept as widths rather than as a rectangle, because a rectangle would have to be recomputed every time the screen changed size, and a thing that must be recomputed to stay true will one day be stale. Two small rules fell out and are worth knowing. A wall never offers a swap, since there is nothing on the other side of one to swap with. And a window already outside the walls is free to come home and free to stay: it gets out there by being dragged out while overlapping was allowed, and hauling it back would mean deciding it was in the wrong. Her picture of a window shattering into star-shaped and square-shaped glitter falling in cosmic whorls is written into the ticket and is NOT built. She withdrew it in the same breath. It is kept because it is a picture rather than a specification, and should she want it, the shape of it is there. GEAR, A PACK, AND DRESSING. Three new pieces, each tested with no graphics anywhere. A piece of gear is a slot, a name, and bonuses in one to three attributes at once -- hers, and a widening of the design rather than a contradiction of it. That design had one ladder per attribute so that climbing one never asks a question and only the choice between ladders does; a piece on two ladders asks which of the two the wearer cares about more, which is a better question, because a hero who has spread their sliders now has something to find that suits them rather than two things that each suit half of them. What a piece is worth to somebody is every bonus times their share in the matching attribute, so an attribute they give nothing to contributes nothing. The pack is a grid of cells, each holding a piece or a definite emptiness rather than an absence, because a grid whose length depends on how full it is cannot be walked or drawn without asking how full it is first. Its arithmetic lives with it rather than with the drawing, since the drawing and the mouse both need to know where the fifth cell is and two copies of one sum is how a click lands a square away from what it looked like it hit. Dressing is one rule that produces both halves of what she asked for. Every slot is decided afresh, from everything the hero has, against the weights they have this second. Move nothing and the same piece wins again; move a slider and a piece that was second becomes first. What is worn stays worn on a tie, because two things worth the same are not an improvement on each other and swapping between them would be motion mistaken for progress. The exchange is one for one into the cell just emptied, so a full pack stays full and nothing is dropped. THE BAR, AND A BUG WORTH FIVE PARTS PER QUINTILLION. The priority bar holds the dividers between the attributes rather than the shares themselves, so a drag moves weight between exactly two neighbours and the shares add to one because they are the gaps in a line of length one. Nothing has to be made to add up; it cannot fail to. The dividers sit at whole steps out of a highly divisible total, and that is not tidiness. Her rainbow dot is the colours of every attribute tied for the largest. Held as fractions, five even shares are five slightly different numbers, because a fifth is not a number a computer has and four subtractions of nearest-fifths drift apart in their last bits. One of them came out very slightly the largest, and the dot drew as a flat blue circle. Whole steps make equal gaps into equal integers and level means level again. A tolerance would have been the other way and is worse twice over: it has to be chosen, and it makes equality true of A and B and of B and C but not of A and C, which is not something a sort or a grouping can be built on. A SECOND BENCH to try all of it on, with a strip of controls along the bottom. The rainbow dot opens a character sheet; the button opens the pack. Roll a new pack, hover a square for a label -- which appears in the opposite quadrant from the cursor, so it never sits under the mouse or falls off an edge -- drag a divider, and press the key to watch the hero change their mind. The older bench gained the same bar and the same strip. Her puzzle question is settled and written down: some rooms carry one, not all of them, until it is more comprehensive. Stickers this patch: 22 gold stars, 10 apples, 1 frowny face, no worms. All time the same. The frowny is still the one for a second window; inventory counts as the second of the five, and three more take it off.

A name is now four words at most: what it is like, what it is made of, what is on it, and what it is. Sparkly Iron Unicorn Tiara. The third of those is the one that would not have been there without her, and her reason is that it is a second axis rather than decoration -- a sparkly tiara and a sparkly unicorn tiara are two different tiaras, and the difference is important. Half of everything comes out plain, also hers, and that is what keeps it working: a wardrobe where every single thing has dinosaurs on it is a wardrobe where nothing has dinosaurs on it. Three entries carry more than a word and the extra is written beside them: a print of occupations that is never a policeman, because she said so and a later hand would add one without thinking; a neighbourhood that is roads and road signs and houses and sparrows together rather than four prints; and an arrangement of fruit that only works one way round. The dinosaurs are four separate prints rather than one, because she named four and the joke is the gradation between them. Eleven slots now instead of seven. Her fifty things included jewellery and pads, and neither is a hand or a body; she invited the change herself. A DUMB LUA MISTAKE, caught by looking at the output. The four parts of a name were gathered into a list and walked, and a list with a gap in it stops being a list in this language -- the walker halts at the first missing entry. So anything with no print lost every word after the place a print would have gone, and a helmet came out called "Squishy Steel" with no helmet in it. It appends now, which has no gaps to trip over. THE BAR MOVES IN FIVE-HUNDREDTHS, which is hers and is about a picture rather than about feel. A hero's dot is striped with every attribute tied for the largest, and a bar that can be put anywhere lands on exactly-equal shares almost never, so rainbows were rare by accident. Twenty notches of a twentieth each, the floor at one notch, and a bar that starts level and stays level until somebody means to change it. It makes the dragging feel like something as well: a control that moves in visible steps tells a hand what it did. One consequence is worth knowing and she is told rather than left to find it: three attributes can no longer be level at all, because a third of a hundred is not a whole number of fives. A three-tab hero never has a three-colour rainbow. Two, four and five are unaffected, and five is the default. TABS CAN BE ADDED AND TAKEN AWAY. A plus and a minus beside the bar open a little window to the side listing what can be picked. One attribute gets one tab, so anything already on the bar is simply not offered rather than offered and then refused -- a person should not be able to press a thing that says no. A new tab arrives at the floor with its share taken from whichever attribute is largest, so a carefully arranged bar is disturbed in one place instead of five. A removed tab gives its share to whichever is smallest, which walks the bar back toward level, and that is deliberate: level is where rainbows live. And the pack has a key of its own, with a button that says which of the two things it is about to do. Her monster ideas gained a line while this was being built and it rides along with the rest: a cop. Notable next to her own rule that the job print is mailmen and fire fighters and never police. A thing you fight is not a thing you wear. Her embarrassing clothes are parked where she parked them, in the cursed gear ticket rather than in the word table, because a piece with a mechanic has to be written by hand and a piece with a name does not. Stickers this patch: 25 gold stars, 10 apples, no frowny faces, no worms. All time the same. The frowny for one window out of five is retracted, and the retraction is a line of its own so the record of it stays.

Which is exactly right, and it is worth saying why it is not the thing that was warned against two rounds ago. A tolerance between PAIRS is a bad idea, because it makes equality true of A and B, and of B and C, and not of A and C, and nothing can be built on that. This is a threshold measured from ONE place, the largest share, so there are no pairs to disagree about. And because the shares are already whole notches, "within five hundredths of the biggest" is "within one notch of the biggest", which is whole-number arithmetic with nothing left to round. A chain of attributes each a notch below the last does not all collapse into one. Only the ones within a notch of the leader stripe. MITHRIL IS SIMPLY BETTER NOW, and so is a nicer adjective. Her list of metals was never in a random order, and asked whether the order should mean something she said yes and added the second half herself: better describing words should be more powerful, because sparkly is lovely and funky is not. So the describing word and the metal each carry a rank, and the two ranks plus a base are a budget that is shared out between the attributes a piece touches. Sharing rather than repeating is what keeps being good at three things a trade rather than a gift: a piece good at one thing is good at it, and a piece good at three is a little good at each, and which of those a hero wants is exactly what their sliders answer. Nothing is spread thinner than one point, so a weak piece is good at fewer things rather than good at many things by nothing. Mithril skips a rung. Three steps of one and then a jump is how a thing comes to feel rare: the difference between iron and steel is an upgrade, and the difference between steel and mithril is a story. Cloth starts from a higher base with no metal to add; metal starts lower and climbs further. So the worst helmet is worse than the worst shirt and the best helmet is far better than the best shirt, which gives cloth a reason to exist early in a run rather than being the thing nobody has replaced yet. WHAT ALL OF THAT REALLY BUYS is that the name of a thing is its stat line. A Gilded Mithril Helmet is obviously a treasure, a Crooked Dinosaur Skirt is obviously junk, and a person can sweep their eye down a full pack and know what is in it without reading a single number. Nothing had to be drawn to say so; her words were already there. It also makes one thing worth protecting, and it is why the ranks sit beside the words rather than in a table somewhere else: a word whose rank does not match what it sounds like breaks the whole arrangement, and nobody would find that out by looking at a column of numbers. The stones in jewellery are deliberately NOT a third ladder. She asked for two and got two; a third would make the name of a necklace carry two separate promises about how good it is. Left as a question for her. Stickers this patch: 27 gold stars, 11 apples, no frowny faces, no worms. All time the same.

The split it settles is the one the whole viewing layer will follow. The state stays where it was -- rectangles, stacking order, what a click means, where a drag may go, which windows are covered -- because that is arithmetic and has nothing to do with paint. The look goes in one drawing file. The contents stay with whoever knows about heroes and packs, one function per kind of window. A kind with no row is an error rather than a blank rectangle, and it caught one the day it was written: a bench whose windows had no kind at all had been working by accident. A FIGURE OF A PERSON, in a file of its own, holding only the arithmetic: where each slot's box sits and which box a point is inside. The drawing needs that and so does the mouse, and two copies of one sum is how a label appears for a slot the cursor is not over -- the same mistake the pack's grid was kept away from. Both benches now draw the same figure and answer the same hovering, and a slot the layout has no place for is left out rather than guessed at, because a guessed place is a box on top of another box and nobody finds out except by hovering one and being told about the other. Gear shows on the sheet now: each attribute as it comes to with everything worn, with the bare value beside it, and speed and health following along underneath since they are mixtures of the attributes above them. Put on a helmet and two numbers you did not touch go up. Hovering a slot says what is in it, in the same box the pack uses. THREADBARE IS FOR CLOTHING. Hers, and it needed a whole notion the words did not have: things now belong to families -- cloth, metal, jewellery, weapon -- worked out from what a thing takes rather than declared on each of sixty-odd entries, and a describing word can be pinned to one. A threadbare ring is not a ring. Notched, which was already weapons-only, moved onto the same mechanism, so there is one way of saying it instead of two. A STONE SAYS WHAT A PIECE OF JEWELLERY IS FOR, not how good it is. Asked whether the stones should be a third ladder, she answered a better question: the stone names the attribute, and jewellery boosts that and nothing else. Which gives jewellery a shape rather than making it a shirt with a smaller number. Cloth and armour spread their strength over one to three attributes; jewellery puts all of it in one. So a hero who knows exactly what they want wants jewellery and a hero who is spread wants a shirt, and that is a real choice instead of one number being bigger. A KEY IS NEVER WITHIN TWO ROOMS OF THE DOOR IT OPENS, which is hers and is worth the trouble it took: a key lying in sight of its own door is not a puzzle, it is a doorknob in two pieces. Measuring is the easy half. The hard half was that one key in eight was landing closer than she asked, and nearly every one came from the same shape -- the way in landing in a two-room pocket on a fifty-room level, so there was nowhere further to put anything. Deferring those locks to a later round did not help, because the region only grows when a key is placed. What did help was noticing that a pocket too small to hold a key properly is the same problem as a pocket with no room in it at all, which the builder already knew how to answer: the lock penning the guild in comes off. None of three hundred levels now has a near key and none of them needed the rule bent. It costs about one lock in three levels, and that is the price of the rule rather than a bug. The distance is recorded as each key is placed and kept, because measuring it afterwards gives a different answer -- by then the party has every other key and can take shortcuts that were shut at the time, and the number that matters is the one from when they were looking. Stickers this patch: 30 gold stars, 12 apples, no frowny faces, no worms. All time the same.

So the answer is a split rather than a compromise, and it is the right one for this game. A HERO DRESSES THEMSELVES AND NOBODY DRESSES THEM. A game where you can reach into somebody's pack and put a hat on them is a game where the heroes are puppets, and this one is built on them deciding for themselves -- which room to walk into, when to go to the toilet, whether to hold on. Their gear is the same kind of decision and belongs to them. What the player gets instead is the sliders, which is a better handle anyway: you say what a hero should care about and watch them act on it. THE LEADER GETS THE OPPOSITE TREATMENT ENTIRELY. Every relic placed by hand, nothing automatic anywhere, because the leader is the player and there is nobody in there to make the choice. Their page is a second KIND of window rather than the hero's page with different numbers in it, and what differs is not the layout but what a person may do in it: a hero's figure takes no input at all, and the leader's page is nothing but input. It also separates two things that were one. The guild OWNS relics and the leader WEARS some of them, and until now those were the same list. A relic owned and not worn now does nothing, which is what makes placing them a decision rather than a formality. How many slots there are is the whole of how much a run's relic luck can be steered, so it is left as a question for her rather than guessed at quietly: one would make the choice trivial and six would make it no choice at all. A relic dropped on a hero is refused with a sentence saying why, rather than silently ignored. A control that does nothing is a control a person will keep trying. The relic list itself is a placeholder and says so in three separate places: the list, the number of slots, and the attributes they add to are all hers to settle and all still open. A small thing fixed by looking at it: two column headings on the sheet were set from the left and ran into each other, so "with gear" and "bare" came out as one word. Right-aligned over the columns they head, which is the only way two labels of different lengths line up with two columns of numbers. Stickers this patch: 31 gold stars, 12 apples, no frowny faces, no worms. All time the same.

They have six to nine now, and some of what went in is deliberately not jewellery: earmuffs and ear warmers in the ears, a scarf and a neckerchief at the neck. A slot is where a thing goes and a family is what it is made of, and they were never the same question -- which two tests had quietly been assuming they were, and which is exactly what adding a woollen thing to an ear made visible. Both now ask what came out rather than which hole it came from. WEAPONS ARE TWO THINGS NOW, hers: sharp and solid, swords and arrows sharp, staves and maces solid. They are two SLOTS as well as two kinds, which is how the split also answers the other half of what she asked in the same breath -- one weapon slot with six things in it became two with eleven each. A hero carries one of each, which in a haunted daycare is a pair of safety scissors and a rolling pin, and that is the right image. They are two FAMILIES as well, so a describing word can be pinned to one. Notched, which was weapons-only, is sharp-only now: a notched blade has seen use, and a notched rolling pin has been through something nobody wants to hear about. A HUNDRED AND EIGHT THINGS TO WEAR, across twelve slots, none of them thin. SOMEWHERE FOR MODIFIERS TO GO, which is infrastructure and nothing else, which is what she asked for. Modifiers are not bonuses and she said so in a bracket rather than leave it to be guessed at: a bonus is the number a piece adds and is what its strength is shared out into; a modifier is a thing a WORD does. One system for every kind of word, with a different shape behind each, because what a word does depends on what kind of word it is: an adjective keeps an enchantment, a print keeps a pattern, a metal keeps an alloy, a stone keeps a cut, a thing keeps its make. One reader per kind turns its own shape into one common modifier, so whatever eventually applies them has a single list to walk in the order the words are said. Nothing applies them. Three words carry one example each so the path can be tested from end to end, and they are marked as examples; a piece of gear behaves exactly as it did yesterday. A word renamed out from under a piece that already exists is an error rather than a shrug, because a shrug is gear that has quietly stopped doing something. RELIC SLOTS GROW. Asked how many the leader should have, she gave a better answer than the flat number the question assumed: three to start, and one more for every level conquered, so a player who wants another slot can go and take an easy level for it. A flat number is a setting. A number that grows is a reason to go down, and it makes the easy level a real option rather than a worse one. What that detour costs is the half she gestured at and left open, and it is what keeps it from being free. The bench pretends two levels have been cleared so the rule can be seen rather than described. A REMOTE is configured for this repository, and nothing has been pushed to it. That is hers to do or to ask for. Stickers this patch: 33 gold stars, 13 apples, no frowny faces, no worms. All time the same.

The graphics library's own way of saving a screenshot is fenced: everything it writes goes under a hidden folder in the person's home, named after the program. Pictures of these benches had been piling up there where nobody would look and nothing would ever clean them up. Three such folders belonged to this project; they are gone, and what was in them is in the project instead. The way round the fence is worth knowing, because the same fence catches every other kind of file a program here might one day want to keep: the picture is encoded to bytes in memory and then written with the LANGUAGE's own file handling rather than the graphics library's, which has no fence round it and puts a file exactly where it was told to. Press S in either bench. They land in a folder of their own, split by who took them -- hers in one place and the benches' in another, because a picture taken to show a bug and a picture taken to show it fixed look identical in a listing and the folder is the only thing that says which is which. Names start with the date so a listing is in order without anybody sorting it, and then say what is being shown rather than which program took it: a useful name in a year rather than "screenshot 4". The folder's own spelling was already on disk with one of her pictures in it, and it stayed that way. Renaming it would move a folder for no reason but tidiness, and moving folders is how history gets lost. ONE WEAPON, NOT TWO. They were two slots for a little while, on the reading that a hero might carry a sharp thing and a solid one at once; she settled it in four words. So a hero carries one weapon and it is either sharp or solid. The two kinds stayed, because they were the half she actually asked for, and what they are for is families: a describing word can be pinned to one of them, so notched is sharp-only rather than weapons-only -- a notched blade has seen use, and a notched rolling pin has been through something nobody wants to hear about. A weapon that does not say which it is now stops the program, because every weapon is one or the other and one that says nothing is one nobody described properly. The slot is no thinner for the merge. It had six things in it when she asked about the counts and has twenty-two now. BOTH KINDS OF NUMBER REACH THE LABEL, hers. They are listed apart rather than added together, because they are not the same kind of thing: the bonuses are what a piece ADDS, and the modifiers are what its words DO. A reader who cannot tell them apart in the label will not be able to tell them apart in their head either. The label also says, in as many words, that nothing reads the second sort yet -- a person looking at a Sparkly anything should be able to see both that the word is carrying something and that it is not doing anything with it today. WHAT THE DETOUR COSTS, settled by her: time, and all the other supplies. Which is the answer that needed nothing added. A level already costs time, food, drink, garments and the wear on a party's confidence; the detour IS a level, so it costs a level's worth of all of it. The clock is the part that bites, and it bites the way this game already works -- heroes fill up and wear out on a timer that does not care what the party came for. A level taken for a relic slot is a level's worth of bladders, paid by the people rather than by a number in a corner. Stickers this patch: 34 gold stars, 15 apples, no frowny faces, no worms. All time the same.

The body has a map now. The older page about it says why it is shaped as it is; this one says what moves where, in what order, on a single turn of the food timer. Every number the body keeps is listed with what kind of number it is, every line drawn across a filling vessel has both sides of its fork named, and every branch says what each way would bring. Drawing the map found two places where pages had quietly stopped agreeing. One page had been working out how much a body can hold without the attribute that decides it, which is fixed here. Another names a number for how much a learned habit shortens a countdown, and that number has never existed anywhere; it cannot be invented from this chair, so it is written where a reader will trip over it rather than left lying where nobody looks. Counting the patterns already in the project turned up one machine fourteen times over six systems, and nobody had planned it once. A number climbs while a thing has not happened, is read as the chance the thing happens, and drops when it does. One formula covers all fourteen, both of its knobs sit under a square root so neither can be turned far enough to break anything, and the meaning is in how far the number drops: all the way is forgetting, part of the way is a body learning a habit, and never is a body wearing out. Every road to losing continence in this game is the third kind, which was true before anyone noticed it was a kind. And the daycare finally has something pushing back. It has been putting triggers into heads since the beginning and nothing has ever taken one out. Now a person can, at a camp, by chanting rather than by trances, and she needs both the openness to feel where the other one is and the sturdiness to sit there while she feels it. Around that, the way a hero comes to have any skill at all took a better shape: what she is gets read off the numbers she actually carries with her gear set aside, so the plan the player draws aims the growing, the growing writes the record, and the record earns the name. Shapes come in twos split by how grown-up she is, and the twos stack into ladders split by how far she has come, which means an apprentice becomes a carpenter without anybody drawing an arrow between them. Sixty-two names for the thirty-one possible shapes are parked in the notes to be argued with, about a third of them hers.

These were the transcripts that prompted the work: their opening was nine numbered "user requests" of machine-written tag soup before the first thing anybody actually typed.

Copying every field is what makes this safe rather than merely tidy. A commit's name is a hash of its snapshot, its parent, its author and date, its committer and date, and its message; copy all of them and the same commit replayed on two machines comes out with the same name. So two people doing this never disagree, a mirror that fell behind catches itself up, and the send is always a straight continuation instead of a fight. That claim is the whole design, so it is checked on every build rather than asserted: the test throws a mirror away, rebuilds it from nothing, and insists the names come back identical. A commit that touched none of those files still gets one here, empty, so the run of messages reads unbroken. An empty commit points at the snapshot its parent already points at, so it costs a couple of hundred bytes and no file data whatsoever. Where the rebuilding left off is written into the record rather than kept beside it: every rebuilt message ends with a line naming where it came from, and the newest of those is where the next run picks up. Running it twice does nothing at all the second time. Last of all it plants itself where git looks just before it dials a server, so an ordinary push does everything above on its own. That directory is machinery rather than contents and git never tracks it, which is why the planting happens on every run instead of once, and why there is a setup mode for a machine that has never done it. The hook hands over rather than calls, so the shell has finished with the file before the script rewrites it. Failure is split in two on purpose. Rebuilding never touches the network, so if that fails then something here is broken and the push stops dead. Sending is the only part that needs the world, so if that fails it complains loudly and lets the push through, because the rebuilt commits are already safe on disk and the next run carries them over. Also in here: the check on her own pages was reading the exit code of the thing it piped into rather than of the test itself, so that gate could have failed and the build would have sailed straight past it. It reads the right one now. The new stickers on the fridge belong to the other session running alongside this one, and are left uncommitted for it to carry.

The narration marker stays: a quote marker is real markdown, renders correctly everywhere, and carries meaning rather than decoration. A gain that came free with the removal: wrapped list items in the assistant's prose have their hanging indents back. Right-justification had been destroying them, because a ragged-left edge and a fixed indent cannot both be on a page.

The flag turns out to be findable after all. The hook is a child process of the push that set it off, and the kernel keeps every running process's command line readable under /proc, so the parent's own command line says what it was asked to do. Handing over with exec is what keeps that true: exec replaces the program without making a new process, so the parent is still the push. A rehearsal now rebuilds the mirror on disk, which nobody can see, and sends nothing. If that command line cannot be read at all -- another kind of machine -- it says so out loud and treats the push as real, rather than deciding on its own to stay quiet. The hook also stopped hard-coding where the project is. It sits two directories above itself by definition, so it works that out instead, which means a clone anywhere else needs no editing. That change came out of trying to test the first one: a hook pinned to an absolute path would have aimed a throwaway repository's push at the real project and published for real. Both are checked now. The test stands up two servers on this machine, plants the real hook in a throwaway project, rehearses a push and looks at what arrived, then pushes properly and looks again. It runs on the executable temp tier rather than in RAM, because the RAM tier is mounted so that nothing on it may be run, and git responds to a hook it cannot execute by saying so and carrying on -- which would have made this pass for the wrong reason. The address of the mirror gained an environment variable in front of it, for that test and nothing else, so the checking can watch a real send arrive at a server on this machine instead of taking the sending on trust. The new stickers on the fridge belong to the other session running alongside this one, and are left uncommitted for it to carry.

Each sticker means something narrower now. The gold star is for work above and beyond. The apple is the ordinary one, for the request met to the specification, and it is meant to be the common sticker by a wide margin. The frowny face moved off the work entirely and onto conduct. The worm widened the other way, to cover every opportunity eaten -- taken back, broken, or simply fallen short -- and it can now spoil a gold star as well as an apple. That last part is a second kind of worm rather than a fifth column on the file, because the game's reader refuses any row that is not exactly four columns wide, and a new word in a list breaks nothing. The counting tool grew a second job. It divides as well as counts, prints each kind as a share of the whole, and says out loud when gold stars have come to outnumber apples or when nothing bad has ever gone up at all. A rule nobody reads gets broken; a rule the tool prints at every commit does not. The blueprint for the game-side reader was updated to match, since the game shows these counts to every player: six kinds to validate, seven counts a side instead of five. Six questions came out of this, written down and still unanswered. The first decides the ratio -- what a really good example earns now, given that is where the sixty came from. One turned load-bearing along the way: a worm has no retraction, and the shape that used to be forgiven, three of five with the other two arriving the next day, is a worm today. Left out on request: the fridge file itself, which is carrying sticker lines from an earlier session, and the transcripts that are not this session's.

And the line between the two good stickers, which she declined to draw and handed over instead: an apple is for the work arriving, a gold star is for work that changed what could be done next. A good answer is an apple; an answer that made something possible, or caught a wrong premise before anything was built on it, is the other thing. Four stickers already on the fridge are read back under that test in both the guidance file and the praise page, because four worked examples settle more than a definition does, and when in doubt it is an apple. The counting tool gained the tighter of its two complaints. It already spoke up when gold stars outnumbered apples; now it also speaks up when more than one sticker in three is gold, which bites long before the ratio can invert. The threshold lives in the tool rather than in prose, on the same principle that keeps balance numbers out of the documents: a rule that runs cannot go stale. Every question this work raised is now closed, two by her and four as rulings taken under the authority she handed over along with the sticker roll. The founding vision keeps the older definitions, deliberately, since it is hers and these pages are not. The transcripts ride along, including two whose rearrangement came from an earlier session. The fridge does not: its own three lines are staged separately and wait on permission, because splitting one file across two owners can only be done through the index and the commit gate is right to be suspicious of that.

So the alternative wording travels inside the page itself. A line of ordinary text, and directly beneath it an HTML comment holding what that line says in the other repository. A comment renders as nothing, so the page here simply reads as written; the mirroring swaps each pair as it goes past and drops the comment. Both wordings are hand-written, in one file, one directly under the other, which makes changing one and forgetting the other difficult to do by accident. The words are hers, gathered from a single unbroken answer and set down close to how she said them. The game is adults only and everyone in it is an adult, but they have to be babies in their hearts or there is no point. People who dislike the material are welcome to leave; people who dislike the players are told where to go and asked to stay there. Adult babies are queer people and a culture like any other, and cultures deserve games of their own. The AI content is deliberate rather than apologised for, on the grounds that most people cannot write code and that wall coming down is the whole point. She is a developer without money or a job, which makes this easier for her than for most and does not make it easy. And her favourite colour is red, then blue, then black, because she likes all of them. Adding a path to the list of what gets mirrored is only safe when that path has no history behind it, because a rebuilt commit's name depends on its contents and an old path would rename every commit it ever appeared in. That constraint is written down beside the list rather than left to be rediscovered, along with what to do if an old path ever has to join. Checked: the marked line reads the other way in the mirror, an unmarked line is untouched, the marker itself never travels, and the page here is not altered by being mirrored. Her collaborator's uncommitted work -- the distribution issue, the handoff note, their own transcript, and the staged fridge -- is left alone, the way their note asks.

The gold stars: catching in one word that the second project directory was a symlink, before anything got built on top of a wrong premise; deciding the sync should plant its own hook every time it runs, which replaced a clumsier plan and is what got built; and asking why a push would fail rather than letting the guess stand, which sent somebody to look and turned up that the credentials are already on disk, so the worst version of that failure cannot happen here at all. The seventh is the front page: one document serving two repositories with the differing lines swapped in passing, in place of a separate source file and a rename. The apples: settling in one go what the mirror carries and what the repository is called, deciding the transcripts go out in the open and saying why, and answering a request for five things about the front page with eight. Only these seven lines are in this commit. The fridge also holds stickers from the session running alongside this one, interleaved with these, and those are left where they are for that session to carry. Their staged work was not disturbed to do it: the commit was assembled in a temporary index built from the last commit, so the real one still holds exactly what they left in it.

It knows where it is without being told. The path a script was run from is handed to it, and resolving that to an absolute answer gives one directory with no searching, no configuration file and no question put to a person who may not be sitting there. This file lives at the top of the project, so the project is the directory it is in. Passing a different project as the first argument still wins, for reading somebody else's fridge on purpose. The second fault was underneath the first and would have outlived fixing it. A missing fridge was treated as an empty one: a friendly sentence and an exit reporting success, which is a count of zero wearing the clothes of a count. It stops and says which file it wanted now. An empty fridge is a different thing -- a file with only its header row -- and that still counts to zero through the same arithmetic as any other count, quietly and successfully, because zero is the true answer there. Both faults looked like success, so nothing but a test would ever have caught them. There is one now: a copy of the reader is stood up in a throwaway directory beside a fridge with a known number of stickers on it, and asked. It runs on the executable temp tier, because a copy has to actually run and the RAM tier is mounted so that nothing on it may be. Writing that test got the count wrong twice before the reader did, by counting lines that matched rather than reading the number off the line -- the report prints every figure twice, once for the patch and once for all time, and the word apples appears again inside wormy apples. The helper that reads it now says so, because the next person will reach for the same shortcut. This bumps the standing rule that every script carries a written-down project path. The rule exists so a script runs from any directory, and asking the file where it lives serves that better while leaving the argument override alone. Changed here only; the other scripts still carry theirs. sticker-report

The download itself is settled as a copy of the repository rather than a curated export, which deletes the question of what goes in it and with it every future way of getting that question wrong. The renderer and its publisher were filed in the neighbouring monorepo on the reasoning that a transcript tool is general. They are not: they render one project's conversations into one project's website, and a general tool with exactly one caller is a specific tool wearing a costume. Both move here, renumbered, with the privacy issue over there repointed so nothing dangles. A third joins them, for the times each exchange happened -- captured while the session logs still exist, because those expire and a site built later would find them gone. The sticker rules tighten. Being right while talking is an apple; a gold star wants a wrong premise already sitting in a file. The old gold stars stand, because nothing comes off the fridge. Seven questions are written into the open-questions page awaiting her arrows. Three carried answers she gave in conversation and are marked so.

Two things it records rather than fixes. The roadmap stops at phase nine, so this phase appears in no map of the project. And the reason no checker has complained is that the validator recognises an issue by three digits and a hyphen -- which every phase up to nine satisfies and a two-digit phase does not. All five files are invisible to it: uncounted, unchecked, never reported missing. Both wait on whether distribution belongs in the game's roadmap at all, because fixing the counting before answering that would only move where the wrong answer appears.

Her decisions from the session go in where they belong. The download carries the whole history, and the history was read end to end first: across every file that has ever existed here, no credential was ever committed, and the only two files ever deleted had moved rather than vanished. What it does carry is her address on every commit, which the public mirror already shows and which is recorded so it is known rather than discovered. The platform's silence stops being a gate. The front page gets a door with an age on it, written the way she wrote it, because a gate phrased as a person asking is read and a gate phrased as a legal notice is clicked past. Her section of the site is written from what she gave: the braids, the thinner one on the right and why, the barrette, the dark colours she wears and the pastels she never will. Her wordplay for that contrast is set down exactly as she said it, with a note that correcting it would delete the point. And a rule about whose speech is whose. An assistant had filed the absence of her spoken words from the collected-answers file as a hole; it is the design. The arrow has one writer and the value is that the writer is never in doubt. An assistant quotes her and never writes in her hand. The palette note lands in the issue that owns the colours rather than reopening anything, at her word: one colour is still wrong to a tritanopia eye and the run says so every time.

The whole parser rests on one distinction. The exporter writes its headings at one level and a model writing prose writes its own at another, so structure and commentary never collide. Anything else would shatter a reply into fragments, because a long answer carries dozens of headings of its own. Two things were learned by failing rather than by planning. A reply that runs out of room resumes under a heading nobody had written down, and when a conversation runs out of room entirely the harness writes a recap that its own heading says belongs to neither speaker. The first is now carried as a marked continuation; the second gets a kind of its own, because folding machine bookkeeping into a person's replies on a public page is the one mistake worth most avoiding here. Attribution is checked rather than assumed. The exporter marks a model only where it changes, so the opening model carries no mark -- and the header's order was confirmed against the harness's own session log before being relied on, since this line becomes a public claim about whose words those are. The three libraries it will render through are copied in rather than reached across to, with the commit they came from recorded, because the release download is the whole repository and a path out of the project resolves on one machine and nowhere else.

The colours are not chosen here. They are read out of the one table the battle log, the map and the colour checker already read, converted from the numbers a graphics library wants into the ones a stylesheet wants. Tune a colour once and all four move together, which is the whole reason the site does not keep its own copy. On a page the two speakers sit on opposite sides, which is where that idea belonged all along -- it was tried once inside the stored file and rendered as a code box everywhere, because alignment is a property of a view. The harness's own recaps are drawn as neither speaker, and the model that served a reply is named only where it changes, since the same line forty times down a page stops being read at all. Three things were learned by running it rather than by planning it. The record mark in a history query belongs at the front, or every commit is handed its neighbour's files -- it read eight commits of forty-three and found nothing, loudly enough to catch. Renames have to be followed, or two thirds of the timeline dead-ends on names the project has outgrown; following them took the linked runs from ten of thirty-one to thirty. And the shipped font has to be copied beside the pages, or every one of them quietly falls back to whatever the reader owns, which is invisible from here. Nothing is redacted. Oversized code blocks are counted and listed on every run, so the number is watched rather than guessed at.

The gate does not pretend to be a lock, and says so at the bottom in small plain words. It cannot check anything, no version of it could, and building it as though it could would be worse rather than stronger -- a convincing lock invites somebody to lean on it. It runs no script either, which means the answer is not remembered; the returning reader is served instead by the nesting itself, since a bookmark to where you were reading skips the door. The memory belongs to the reader, which is the one place it can honestly live. Behind it: what the studio is and what it makes, a page about a person rather than a company, and a page for the game -- what it is in its own vocabulary, the two tensions that are the hook, where it actually stands written as a sentence somebody could check rather than a percentage somebody made up, and the way through to the development record. None of the three is typed. They are built from one record of what exists, so a second game is a row rather than a page, and the site cannot fall behind except by nobody running it. Her own words carry through unsmoothed, the bedroom included, and the contrast she coined between herself and her friend is set down exactly as she said it with the build asked directly whether both halves survived. Two failures worth catching are now caught on every run rather than in a browser: a link that lands nowhere, and a page too deep to find the font it asks for. The second is the dangerous one, because a page that cannot reach the font renders in whatever the reader happens to own and looks fine from here.

Four complaints, one cause. The renderer ran its inline pass once per source line, so nothing was allowed to span a wrap -- a marker opened on one line and closed on the next could never pair, and every eighty-column break survived onto the page as a hard one. Joining a block's lines before the inline pass fixes all four at once, and it applies to quoted blocks and list items for the same reason it applies to paragraphs. A wrapped bullet now stays inside its item rather than falling out of the list. The old behaviour is the default and is kept, because it is correct where the renderer came from: upstream sets poetry, and there a line ending is the author's decision rather than a tool's accident. That divergence is written down where the copies are described, along with the fact that the change is owed upstream. Then the colours she asked for. Tables have grey rules and cyan text, bold wears the same yellow as a heading, what she said is green, quotes lean. The reply side is right-aligned as a block rather than as ragged text, so every line of it starts at the same column and successive bullets share an edge -- which is what right-aligned prose needs to be readable at all. Two things fold away until wanted, both by the browser and neither by a script: the harness's recap, which is long and is not what a reader came for, and a commit's full message, which is what actually happened and would bury the story if every one were printed at once. And her own page says much more than it did. How she thinks, what she refuses, what the whole thing is for, and a picture against every noun.

Each file was dated from the strongest evidence still standing for it. Where the session log survived it was rebuilt from source. Where the log was gone the dates came from the record of prompts sent, which counts from an absolute instant and so was never affected by any of this. Where neither survived, the fault was inverted back out of the file's own timestamp, which is exact because the same offset had been applied going in. Some timestamps could not be believed. A bulk copy rewrites every file's modification time to the moment of the copy, and one folder here holds hundreds sharing a single second. Counting files per timestamp does not reveal that, because a conversation and the agents it spawned honestly do share a last message; what reveals it is a second witness disagreeing by more than a day. One project is preserved untouched for this reason: its transcripts came from an earlier exporter, now unregenerable, and their dates are not wrong so much as absent. Carried alongside: the correction itself, a planner that decides what each file deserves, the tool that carries those decisions out, a tool that takes a verified snapshot of the whole archive first, and tests for all of it. The planner changes nothing, so a wrong judgement can be read in a report before it becomes a wrong rename. The rebuild tool's search was two levels too shallow to see one project that had a live log all along, and never said so - a project a search cannot reach looks exactly like one that is not there.

Going: the planner that decided what each file deserved to be called, the tool that carried those decisions out, their notes, their two test files, and the sixty markers they dropped in the transcript folders to stop themselves running twice. With no pass left to run, nothing reads those markers and the instruction written inside them had no tool behind it. Staying: the correction itself; the test that guards it, because the correction converts a time and converts it straight back, which reads like pointless double work and invites being simplified into exactly the single call that caused all this - a change that would otherwise pass unnoticed for months; and the snapshot tool, which has nothing to do with timezones and everything to do with not destroying an archive whose originals are gone.

Two more tools ride on the same key. One builds the reading copy from the public mirror and copies it up, sending only what changed and never deleting. The other keeps a dated copy of a short list of folders on the machine each day, storing unchanged files once and linking to them, using a second key that the machine fences to rsync inside one folder. The claim page's design moved into the subscription issue: a page mailed as an attachment carries half a sealing key and the website shows the other half, so neither road carries a whole credential and a middleman on either road alone gets nothing. Extra changes, and why: - The project guide's gold-star rule is redrawn in her words: a gold star is for making (a lot of hand-made work, or optional extra-credit work built), an apple is for deciding well, and a bad decision or a phoned-in answer earns no sticker. The awarding tool's notes say the same. - The publishing issue now waits on the machine issue, and the roadmap and phase progress list the new issue. The fridge is left out of this commit: it holds forty stickers from earlier sessions that were never committed, and those wait for her say.

The door is now the name, her sentence, and two doors. The line saying what is behind it and the small line saying the gate cannot check are gone at her request. That the gate checks nothing is still the design, and is said in its source for whoever maintains it. Saying no no longer goes to a blank search page. It opens Wikipedia's Nicomachean Ethics at the section on excellence of character, where Aristotle says virtues are built by doing them until they become habit and become one's own when chosen on purpose. The reader most likely to click no is somebody younger who just told the truth when nothing made them, and that is where they land. The whole-site test now checks for 21 and no 18, for the exact section link, and that both removed lines stay removed. The site was rebuilt and sent.

This commit carries the forty stickers from the sixteenth to the eighteenth of September that earlier sessions put up and never committed, because each one committed only its own lines, along with the fifteen earned today.

The sticker report's warning about too many gold stars now says what a gold star is under the redrawn rule (making, not deciding) instead of the old one.

The old grouping clumped commits by "everything since a transcript was last saved" and followed transcripts by file name, which broke when the archive was re-dated; both are gone. The mirror's rebuild no longer starts over when it cannot tell where it left off. Its "stop, I am lost" message ran inside a sub-shell, which only ended the sub-shell, so an empty answer was taken as "begin at the beginning" and the whole history was copied in a second time. It now stops for real, also stops when its bookmark names a commit the project no longer has, and its test checks both. The timestamps issue now notes that linking no longer needs timestamps, and asks whether they are still wanted for their own sake.

The newest transcripts link to the helper conversations they started by their file names, ending .md; on the website those are pages ending .html, so thirteen links on the live site led nowhere. A bare transcript file name in a link now points at its page.

Three marks of punctuation were missing in the table of words, two commas and the brace that closes her section, so the site would not build. Those were added and no word was changed. The whole-site test guarded her old sentences by quoting them: the old wordplay, the old bedroom line, and the old wording of the age. It now looks for the new ones, so it still catches a tidy-up that loses her words, and still checks the door names 21. A gold star rides along on the fridge for the rewrite.

The notes beside her section of the website's words still described the old page: an emoji on every noun, and a sentence kept with a correction in it. Neither is true of the page she rewrote. They now say she took the emoji out because they made the page read as machine-made, so nobody puts them back inside her sentences. An apple rides along on the fridge for the answers.

A mailbox path longer than 108 characters is refused by the system outright, which is why the test ran in a short folder in RAM rather than the session's scratch folder.

The emoji are a table of real things in six groups -- fruit, vegetables, animals, buildings, transport, and things you could hold -- with no zodiac signs, no number keys, exactly one clock, and the apple kept out of the body so the food is always the one red apple. The speed, the fading, the cell size, the starting length and the chance of a random step sit beside them as knobs. The rules of the game are plain functions handed their randomness, and the drawing is a separate piece that calls them. A new test runs the rules in node for twenty thousand steps with a fixed seed and checks after every step that nothing impossible happened, and parses the whole page script. A reader whose system asks for less motion gets the snake drawn once, still. The site used to say that nothing on it runs a script. The rule is now that the gate runs this one, written into the page, and nothing else does; the whole-site test checks both halves of that, and that the snake stores nothing. The website issue carries the design and her request in her words. An apple rides along on the fridge.

A new script takes a picture of any page of the website and keeps it in the project: it builds the site into RAM, opens the page in a windowless Firefox with its own throwaway profile so the browser she has open is never touched, and saves the result under assets/screenshots/ named for the page and the day. The first picture is the front door with the emoji snake behind it. An apple rides along on the fridge for the paragraph.

Choosing happens on this computer, with no paid service. Each film is boiled down once to a strip of six frames and shown to a vision model on the graphics card, which writes a few sentences about the shapes, the motion and the mood; the description is kept under the film's fingerprint, so a film is looked at once ever. Each message and each description then become lists of numbers from the same meaning model, and a message takes the nearest film, stepping aside whenever that film is on the message before or after it, her rule. Every choice is written into a small file beside the transcripts and never worked out again, so a message keeps its film through new films, new models and rebuilds. The first full run chose one film for 120 of 288 messages. The descriptions share most of their wording and so do the messages, so the shared part was deciding; taking each side's own middle away first left only what tells one film or one message apart, and all 34 films came into use, the busiest 44 times. Those first choices were never committed, so nothing kept was changed. The push hook starts the model server if needed and chooses films for new messages before it builds the site; if that fails it says so and the push goes on, and the build counts any message still waiting. The site copies only the films some message uses, at full size, so each loads once. Tests check the choosing rules with made-up numbers and no model, that every film square on every page points at a real file and the assistant's are mirrored, and that the push runs the new steps in a real push and not in a rehearsal. The JSON library came across from the shared libraries unmodified, and a picture of a conversation page with its films is kept with the screenshots. The issue moves to completed with all four of its questions answered in her words.

The whole-site test checked the mirroring by matching one line of the stylesheet word for word, so moving the film broke it while the films stayed mirrored; it now finds the assistant's film rule and checks that it flips. The change is in the balance log with her words, and the finished issue records where the films hang. An apple rides along on the fridge.

The seeded test now runs the snake on a board with a box in the middle and checks, over twenty thousand steps, that nothing enters the box, that every clear straight line was taken straight, and that every segment slides from exactly the cell it was in.

As with the snake, the rules are plain functions handed their randomness and the drawing calls them. A new test runs them in node with a fixed seed: the arc measures 2.88 to one, no star or rainbow comes within a star's width of the words, no glitter leaves its lane, a click finds the star it lands on, and glitter from a star burst right beside the words stops at the edge. The knobs are a table under assets. The site now runs two scripts, each written into its page and loading nothing: the gate's snake and these stars. The whole-site test checks each page runs exactly its one and no other page runs any, and the notes that said the gate's was the only one now say otherwise. An apple rides along on the fridge.