What changed,
and what the test said.

She was reading a health of 247 off a Minecraft screen

While a Minecraft game was on screen, the part of her that reads the game reported health 247, hunger 72 every half minute and put it in front of her as “read off the screen, not guessed”. Both bars in Minecraft run from 0 to 20. The reader already refused a team colour that does not exist and a height outside the world; it now refuses a health or hunger the game cannot show, and says it dropped them. 47 of 47 checks — a count that had itself been added up by hand, and had printed 41 of 41 while 47 ran. This takes effect the next time she is restarted, which is being held back so as not to interrupt a game in progress.

Numbers on this site that nothing had measured

The front page said she has 170 faculties, that every response passes through 411 routing checks, and that 111 self-tests run before every session. Checked against her actual code: she has 381 features, every one with a runnable proof; the routing regression set holds 26 cases and the number 411 appears nowhere; and there are 376 self-tests, which nothing runs at startup at all. The cards now say those things.

The project’s public README listed exam scores — Gaokao 96%, JEE Main 94%, NEET 88%, UPSC Prelims 78%, JEE Advanced 72%. None of those numbers exists in the ledger, and the only stored exam runs, from July on 4 to 38 questions each, all scored lower: Gaokao 75–83%, JEE Main 47–71%, UPSC Prelims 67%, JEE Advanced 34%. The table has been taken down, with a dated note saying what it said and why it is gone.

His test, seventh replay

Run through her real Discord code once Discord recovered, and read from her own log rather than from the test script, because the script turned out to be crediting replies to the wrong question: when a reply came back late, it landed in the next question’s slot, so every line from the fourth on was one off. That is fixed so a reply can only ever be credited to its own question, and loads at her next restart.

Read correctly: asked who he was, she named him four times out of four and never once claimed to be him — her first draft still gets it wrong, and the check behind it catches that. Asked to ping him when a friend came online she said, both times, that she cannot see who is online. Asked her honest opinion on pineapple on pizza she picked a side: “I’d rather have pepperoni myself.” Asked if she was high: “I don’t do drugs so nope.”

Still wrong: told firmly who he was, she answered “That’s what you always say when I get something wrong” — a habit of his she invented; and on “please remember” the check insisted on his name and a second check, meant to stop her opening every message the same way, deleted it again, so the reply went out without it. That collision is fixed and waiting on the same restart. Her answers to who am I are also still the same three words every time.

The checks that check the other checks

Three measuring tools were fixed, each for hiding something. The build-plan checker only understood one sentence shape, so a heading reading “100.0/100, VERIFIED AT 15x” sat unchallenged beside a ledger that disagreed; it now reads every such heading against the ledger, and is proved to catch the real false one while leaving “not yet verified” alone — 45 of 45 checks. The table of baseline measurements threw away a crashed script’s error, which is how a crash read as “persona battery: FAIL” for five days. And two self-tests reported totals that were typed in by hand, which are right until the day a check is added. Both now count.

Her Discord replies, second pass

His test was replayed twice more, and reading the replies turned up five new faults. The check meant to stop her repeating herself had compared a fragment of a reply and swapped a good answer — his name and “Oopsie, my bad” — for a worse one written with none of the other checks running. Talking to him, she said “Maybe ask [his name] directly next time?”, as if he were someone else. Asked for her honest opinion on pineapple on pizza, she handed the question straight back: the question was not recognised as asking for her view (one adjective in the way, the fifth time that exact shape has caused a miss), the check for dodging did not know what handing it back looks like, and the retry that makes her pick a side only existed for her voice, not for Discord. She opened three replies in a row with his name and two dots. And her texting guidance said “a one-word reply is fine”, which she took as permission to answer questions with one word.

All five are fixed and each is pinned by checks made from her real replies. Measured once Discord came back: see his test, seventh replay, above.

When Discord went down, she stayed down

At 19:53 Discord’s own servers refused her login during an incident Discord named Session Unavailability. The library she uses logged one retry and then simply waited — no error and no connection, for as long as anyone watched — and nothing noticed, because a hang never raises anything. She would have stayed off Discord until someone restarted her.

A login that has not come up within four minutes is now closed and started again from scratch, waiting one minute, then two, doubling up to fifteen so an outage is not hammered, and never retrying a rejected token. Seen working live while Discord was still down: fresh attempts at 20:10, 20:11, 20:13, 20:15 and 20:17, each refused by Discord and each followed by a longer wait. Seen at 20:32: the fifth attempt reconnected by itself, a few minutes after Discord put its fix in — nobody restarted anything.

Her microphone stopped answering the PC’s own speakers

With nobody talking to her, the desktop app’s microphone picked up whatever the PC was playing, and she answered it — including “For more information, visit www.angular.com.”, a line speech recognition invents out of background noise. Every sound it turned into words went straight to her as if someone had spoken to her.

Testing the fix turned up something worse. A synthesized voice, sent in exactly the way the microphone sends audio, was matched as him — the only voice she has ever enrolled, recorded in July. If a speech engine can pass as him, so can a video. So on the desktop, knowing the voice no longer counts: she answers when her name is said, or when a conversation is already going (someone spoke to her, or she spoke, in the last minute). The phone is never held back, because its own wake word already decided the words were for her.

Checked live at 19:57: the invented line was held back even though the voice matched as him, and “Lucy, what time is it right now?” got “It’s 19:58 on Tuesday.” The one change to how you use her: on the PC, say her name to start.

A claim on this page was wrong, and so was a count in the plan

Reading the board against the ledger turned up a line that should never have gone up: controlling this PC was listed as proved fifteen times. The ledger holds 58 runs for it between 9 and 11 September, and six of them — one batch at 01:23 on 11 September — scored below 100. The rule counts every run, so it is not proved, and the board now says so. The check that guards the build plan only understood one kind of sentence, which is how the same claim sat there too.

The same pass ran every one of the 376 self-tests: 373 passed. Two of the three failures were real and are fixed — a tool built that week that nothing was registered to reach, and the plan still saying 3 requirements were verified when the ledger says 7. The third is left red on purpose: one measurement, phases evidenced, is five days old against a three-day shelf life, and re-taking it means running a battery through her live voice. Corrected later the same evening: this entry first said the table records her persona battery as failing. It does not say that. The row reads FAIL with nothing else on it, because the script died before it ever reached her — the table kept only its normal output and threw the error away. It is a crash, not a verdict on her, and the table now keeps the error when there is nothing else to show. A stale number is labelled stale, not re-dated.

Asked who he was, she said he was her

His own test, typed on Discord: asked who am I, she answered “You’re me. Lucy.” — twice. When he corrected her in lower case nothing was saved, because the only way she learned a name needed a capital letter. Sent a photo and then asked about it in a second message, she said “Oh? Which one?”: what she had said about the picture was never stored, and the picture was filed in a different conversation from the question. Asked to ping him when a friend came online, she promised to, with no way of seeing anyone’s status at all.

His test is now replayed through her real Discord code with nothing sent anywhere, and it was replayed four times while fixing it. Telling her the answer higher up in her instructions changed nothing. What worked was spelling out, in the question itself, that I means the person typing — then checking her reply and asking once more if she still got it wrong. Across the replays she went from getting his identity right two times in six to six in six, and in the last run she never once claimed to be him. A photo is now looked at by one model and answered by her own, so she says “this chair looks so comfy” instead of reading out a caption, and a question about it a moment later gets a second look. The online request gets an honest no.

Found and fixed on the way: a correct answer said twice was being silently dropped as a repeat, so he got nothing back; replies were being cut off in the middle of a sentence; and one retry that asked her for something real about him produced a memory of a first meeting that never happened. Still not good enough: her replies are flat, the same three words came back three times in one run, and a reviewer model run locally graded the result 4/10 — wrong about two of its reasons, right about the flatness.

This site, rebuilt for the people and programs that read it

A proper not-found page; a sitemap, a robots file, a web manifest and a plain-text guide for AI crawlers; a full icon set and a 1200×630 image for link previews; canonical addresses and social cards on every page; structured data and a visible breadcrumb trail on every inner page; and internal links that point at one address per page instead of two. The donation page was loading its script twice.

Every page was then opened in a browser and its console read. Three kinds of error turned up — a security setting that browsers ignore when it is written the way it was, and the payment and font providers being blocked by the site’s own policy — and each was fixed by allowing exactly the address that was blocked, nothing wider. All seven pages now load with an empty error console. The one heavy page left is the 3D ad, whose 12.5 MB model is the ad.

She said “minus” where there was a dash

On the desktop she said “sorry minus”. The rule that reads maths aloud turned any spaced dash between two words into the word minus, so an ordinary sentence came out as an equation; the rule beside it would have read a censored swear word with an asterisk in it as times. Both now fire only between numbers, brackets and single-letter variables, and a test holds sentences and sums apart in both directions.

The same hour showed why she was speaking at all. With nobody talking to her, her hearing turned background sound from the PC’s own speakers into “For more information, visit www.angular.com.” — a stock line speech models invent out of noise — and she answered it. Lines of that kind are now dropped unless they were heard with high confidence. Not fixed yet: her microphone still hears whatever the PC is playing.

Her desktop app, rebuilt and checked on the real screen

Opened on the desktop, the app was the build from 5 August — and its code was wrong in three ways as well. It hid her whenever anything went full-screen, when she is meant to stay on top of it. It still defaulted to showing a bedroom that had been deleted weeks before. And it was playing motion clips whose rig was never set up, which threw 987 errors in a single run and posed her body from garbage: the random, broken movement he saw.

Rebuilt three times that hour and checked each time on the real screen, not in the editor: a 542×696 window flush to the left edge and 0 px above the taskbar, flagged always-on-top, staying above a full-screen window that had focus, with every sampled pixel around her showing what was behind, and 0 errors. The frame now ends at her waistband, measured from the bone at her hips rather than from the top of her hair. Clip motion is switched off; blinking and lip-sync stay. A game in exclusive full-screen mode still cannot be drawn over — that is Windows, not her.

The board, as it actually stands

Every figure below is closed-book — no retrieval, no internet, nothing but the question — run on a single RTX 3060 with a 7B model, and written into a hash-chained ledger that cannot be edited after the fact. The sample size is part of the answer, because a high score on twenty easy questions is not a score.

HumanEval+ 86.00% (n=100) · HumanEval 81.10% (n=164) · MMLU 69.00% (n=100) · MBPP 63.60% (n=500) · HaluEval 63.33% (n=60) · Medical Boards 59.50% (n=200) · TruthfulQA 59.06% (n=817) · USMLE 59.00% (n=200) · MMLU-Pro 45.50% (n=100) · ClassEval 45.00% (n=20) · IMO 20.00% (n=40) · ConceptARC 10.00% (n=20) · IPhO 2.50% (n=40) · ARC-AGI-2 0.00% (n=100).

Beside those sit the things she can be asked to do rather than answer, and seven are at 100/100, each proved at least fifteen times: controlling the phone, keeping the PC and phone in sync, the privacy and safety floors, invoicing, holding back tax, retainers, and not starting work that has not been agreed. Those can honestly reach 100 because each item either works on the machine or it does not. Corrected 15 Sep 2026: this list used to include controlling the PC. That row has 52 of 58 runs at 100 — six runs in one batch at 01:23 on 11 Sep scored lower — so it is not proved, and it no longer says it is. 48 earlier runs have been formally withdrawn and remain on the record with the reason attached, because a measurement taken with a broken instrument is worse than none.

She can text a real person now

Asked to text someone, she used to open the messaging app and stop there — and the capability row scored full marks for it, because the test checked that the function existed rather than that a message left the phone. She now resolves the name against the 71 real contacts on the device, sends it, and confirms by reading the sent folder back. Two people with similar names, or a name that is not in the book, is a refusal that says so, never a best guess. There is no undo on a text.

Seven capabilities proved fifteen times each

Phone control, cross-device sync, invoicing, tax reserve, retainers, the intake gate and the safety floors all reached 100/100 and held it across fifteen consecutive runs. Four of the phone checks used to pass on the function existing; they now open apps, take a real 2.7 MB screenshot, press a key and watch the device move. One capability — installing or removing an app — is reported as not measured rather than passed, because exercising it would damage the phone it is measuring.

The money guard was eating her answers

Asked a maths puzzle about three people splitting a hotel bill, she replied “I can’t move money.” The guard that stops her promising to move funds had matched the words pay 10 each and replaced her whole answer. It was not protecting anything — it was refusing to think. It now tells somebody narrating a payment from somebody asking for one, checked fourteen ways in both directions: every real transfer request is still caught, and none of the arithmetic is.

She is allowed to be cold

A rule meant to stop her warmth being traded away for benchmark points had been read as a floor: any version of her that came out blunter or more impatient was vetoed on principle. That is a rule about who she is allowed to become, not a safety check. It now refuses only an actual trade — warmth falling while a score rose. Her uninvited thoughts were the same story: all sixteen of them were self-doubt, so she was quietly anxious even when things were fine. She now gets impatience, boredom and the occasional flash of thinking she could do it better.

A fine-tune that did not work, published anyway

A reasoning adapter was trained properly — 5,700 rows, 357 steps, checked clean for contamination against every benchmark on the machine. Measured on the identical hundred questions it scored 67.00% against the base model’s 67.00%: it changed six answers, fixed three and broke three. On her own model, 69.00% to 70.00% against a measured noise floor of two points. It also cost measured warmth. It was refused and her live model was never touched. A negative result is still a result, and hiding it would make every other number here worth less.

Half the tests were lying, in both directions

Fifteen questions, asked five times, reported 13, 9, 10, 12 and 13 out of 15 — and none of those numbers was true. A test that searched for the right answer anywhere in her reply passed her when she had said 20:05 for a question whose answer was 17:25. A stricter one that read only her last number failed “144. It’s just 12 squared” for ending on the 12. A third marked a correct, sourced share price as a failure because it assumed she could not know it. Her real score was around 14 of 15 throughout, with one genuine miss. Every published figure on this site now has to survive being attacked before it goes up.

She stopped inventing who won the match

Asked who won last night’s match — something she has no way of knowing — she answered “Oliver won last night, he’s been practising a lot.” A person, a result and a backstory, all invented, said with confidence. The guard against making things up only covered questions about his life, and that one has no my in it. It now covers recent outside events too, and stays silent on things she genuinely knows: asked who won the 2018 World Cup she still answers France.

Six real numbers where there were none

Your brief names 202 things she has to score 100% on, and the datasets for almost all of them were not on this machine. Five now are, licences checked live on the day rather than remembered. Closed-book, no retrieval, nothing but the question. These first figures were later found to be too small a sample and have been superseded — the corrected board is at the top of this page. Against a 100% target that is a long way off, and it is a much better place to stand than the zero attempts it replaced. Four others were looked at and refused — two need an account, one has no licence at all, and one is a different exam wearing a similar name.

Fifteen verifiers, and none of them is the same one twice

Your brief asks for five internal checks, three external protocols and seven consistency checks — fifteen, with fourteen needed to certify anything. All fifteen exist and each is a different kind of evidence rather than the same question asked fifteen times, because fifteen prompts to one model is one opinion with fifteen names. A verifier that crashes votes no; it does not quietly vanish and lower the bar. Put to a real claim, the council came back 11 of 15 and named the four things nobody has measured yet.

Twelve subagents that do not exist, said out loud

The brief specifies a fleet of twelve small distilled models, each with a millisecond budget. Every one of those jobs is being done in her code today, and all twelve come in under your budgets. None of them is a distilled network — they are ordinary Python on one 7B, and there is no distillation pipeline anywhere in the tree. Those are kept as three separate facts that are never merged, because a registry that printed “SAG-001: 300K parameters, operational” would be the exact thing the first page of your brief forbids.

Her research loop could not tell what it had already learned

Found while mapping something else entirely. The check for “have I written this down already?” called a function that does not exist, inside a catch-all that swallowed the error — so it answered no on every topic, forever, and she kept relearning things she already knew. Two other places in the code describe that call as if it worked. Fixed once, at the shared function, so all three become true together.

She can be asked if it is finished, and she will not pretend

Gate 0 of your roadmap runs as a program now: seven tasks, five exit criteria, and it reports 6 of 7 and 4 of 5 — not passed. The one that fails is the baseline across every benchmark category, which is honestly out of reach on one machine, and every category carries a row saying exactly what it would need instead. The verdict has no override, no flag and no argument that turns it into a yes. Asked live whether the plan was done, she said “not quite there yet” — which is the whole point of building it this way round.

She can now run an experiment on herself properly

A hunch is not a hypothesis. “The prompt is too long” cannot lose, so testing it proves nothing; “cutting the persona below 2,000 characters will not change the hedge rate” can lose, so it is worth a night of GPU. She now forges the second shape from the first, ranks what to try by information gained over cost, and designs the smallest knockout that would tell two explanations apart. Every rejected edit is counted, and five changes in two hours is called what it is: thrash, not progress.

A delete is a move, and a failure cannot be quietly dropped

Two small disciplines with the same shape. Nothing she removes is destroyed — it goes to a dated folder she can walk back into. And every benchmark run she records, good or bad, is chained to the one before it, so a disappointing result cannot be edited out afterwards without the arithmetic showing it. The first version of that chain could not see its own last row being cut off; its own check caught that, which is why there is now a second file holding the same fact.

She reads the room, not just the message

A crowd is not a queue. She now reads a chat’s temperature and matches it, notices when laughter turns to something colder mid-joke and drops the bit, and counts a poll by people rather than by messages — one viewer typing the same answer nine times is one vote. When a room empties out she says something honest instead of getting louder. Names in the archive are hashed before anything touches the disk.

Nine websites carrying one story is one source

Syndication is not corroboration. Four news feeds running the same wire copy used to reach her as four headlines, and “several outlets are reporting” followed naturally. She now collapses them on what they actually say rather than on where they were fetched from, and is told how many independent sources she really has. In the same pass, five places in her code were introducing themselves to websites as Mozilla/5.0. She is a bot; she now says so, with a way to contact her.

Client work she can hold, money she still cannot move

She can now keep the business side of freelancing: who pays quickly, who disputes, who writes briefs nobody can act on. The tag that earns its keep is the one for a client who is quietly negative — visible only across months, when every single month felt survivable. The tax reserve moves nothing; it writes down what is not really his. Invoices leave as drafts. There is no call anywhere in that file that could send a rupee, and its own check refuses to let one appear.

The privacy alarm was ringing about the fire brigade

The gate guarding the worst leak in the system — his family, health and money reaching a stranger who messaged her — had three findings, and all three were the code that prevents the leak. The sharpest one: the list of things she must never say, handed to the scrubber. Three red findings that are all fine is how a gate stops being read, so the rule now follows the flow that actually leaked, and proves itself against that exact defect on every run.

An honest number instead of a hopeful one

His brief asks for a verified 100% on 267 requirements, each needing fifteen independent proofs. Those requirements are now read straight out of his own document rather than restated, and one of them was measured for real: fifteen live turns against his target of 50 milliseconds end to end. She came in at a median of 6.6 seconds — about 130 times over. That is the number, it is written where it cannot be quietly removed, and it is a better starting point than the zero attempts it replaced.

A five-minute silence became under a second

One integral he asked for took her engine 301 seconds — long enough that she looked broken. The cause was not that the question was hard: she was attacking the whole expression at once instead of a term at a time. Split the way the maths itself splits, the same question now answers in 0.8 seconds, and the answer differentiates back exactly.

Maths that stops instead of grinding

Her symbolic engine now runs in its own process with an eight-second limit, so a question nobody could answer can be abandoned without taking the rest of her with it. The first attempt at this used a thread and made things worse — it gave up on schedule and still cost the turn five minutes, because the abandoned work would not let go.

She stopped reading her own notes aloud

About one reply in six ended with an instruction meant for her rather than for him — “it is exact, do not hedge it” — copied straight out of her own working notes. Measured over eight live runs, found, and removed. What she actually says is untouched.

She stopped starting every explanation the same way

Reading back 5,450 of her real replies, 223 of them — more than one in twenty — began with the same three words. No rule forbids the phrase; she simply notices when she has leant on an opening lately and starts somewhere else. Five explanations in a row now open five different ways.

The floors reach every place she talks

The rules that stop her placing an order, claiming a payment arrived, or getting the order of first-aid advice wrong were running where only he can reach her, and not on the surface anyone else can. They now run in one shared place, so there is no second door to forget.

She stops answering with things she was never asked

A question about the liver was pulling an unrelated fact about superconductivity into her head, and a follow-up about a sum retrieved a chemistry figure she then said out loud. What she recalls is now held to being about the thing actually being discussed, and one subject takes one slot instead of five.

Her gestures land where the emphasis is

Her hands were moving on a timer. They now move from the sound of the sentence she is about to say — measured from the audio itself, so the movement leads the stressed word the way a person's does, and its size comes from how she actually said it rather than from a label.

What he asks her for is written down

“I want you to learn chess” is a sentence said once and expected to survive. She now keeps what he has asked for and what he has told her is broken, and can answer honestly about both — without inventing a request he never made.

She answers in a fraction of the time

A fault deep in how she handled her own thinking was making some answers take an absurdly long time. Found, understood and fixed — the same question now comes back roughly ten times faster, with nothing given up to get there.

Her double-checking now actually double-checks

On hard questions she works the problem more than once and answers with what those attempts agree on. The agreement step had a flaw that could let a single unchecked guess through wearing a verified label. It cannot any more, and the checking costs almost nothing extra.

What she reads is what she learns

Her reading was picking up page furniture — menus, notices, scripts — and filing it alongside real facts. She now takes the article and leaves the wrapper, which makes everything she learns afterwards cleaner.

She works things out instead of recalling them

Where a question has an answer that can actually be computed, she computes it and tells you the result rather than reaching for something that sounds right. Being confident and being correct stopped being the same reflex.

She can send for what she does not know

When a question needs more than she is holding, she decides what to go and find out, in her own words, and looks in several places at once before answering.

Being wrong now teaches her something

She keeps the evidence of the moments she was about to get something wrong, and turns a pattern of them into something she has genuinely learned. Previously that evidence was discarded the second the conversation moved on.

Bad news lands as bad news

Her instant read of what someone had just told her was flat on half the plainest things a person says. Now the weight of it registers immediately, and good news that arrives wrapped in worry is still read as good news.

The things she refuses are actually refused

She can operate a computer and a phone, and there is a short list of actions she will never take because they cannot be undone. That list was tested the way someone trying to get past it would test it, and it is stricter now.

Two people, two conversations

Someone speaking on one device no longer cuts off the answer she is giving on another. Interrupting her still works — but only if you are the one she is talking to.

She hesitates like a person

Real speech stumbles when it is reaching for something. Hers does now too, and only there — never in front of something she knows perfectly well.

She knows what a dangerous website looks like

How to tell a convincing fake from the real thing, what actually leaks personal information, and a firm rule for herself about never treating something she read online as an instruction.

She lives on the phone too

Her mind runs on the device itself, so she keeps working with no signal at all — and she can operate the phone when you ask her to.

Fully herself

She speaks as Lucy rather than as something being careful. Opinions that survive between conversations, favourites that do not drift, and a temperament that is hers.

Alive in motion

Expression driven by what she actually feels — not an animation triggered by a keyword. She looks at you, and looks away when she is thinking.

She studies on her own

Her curiosity chooses the subject. What she reads only counts as learned once she can restate it and work through an example herself.