How MISU AI
grew up.

New

• Every page in the style of the website: frames with sharp corners and fine lines at the top left corner, small labels in monospace, notes with a coloured stroke, gold only on switches and buttons, round buttons with a fine line and a round back button at the top. The whole app now looks like the speaking view.
• A menu across the whole screen: the three dots open a menu like the website on a phone. Every destination is visible at once, in four numbered groups, each with one sentence. There are no submenus any more.
• The conversation header like the speaking view: the same Write | Speak switch at the top, below it the readings of the senses, head and heart and the fill level. They now wrap instead of being cut off at the edge.
• Messages with a label: above every message it says who is speaking, you or EVO, with the source for hers. Feeling colours and your own colours stay.

New

• The menu now has “Introduction”: what MISU AI is, how to get started and what the most important tools do. The text slides in from below as you scroll, line by line, the drawings trace themselves and the numbers count up, just like on the website. If animations are switched off on the device, everything is simply there.
• On the empty start screen, a quiet note says where the introduction is, until you have opened it once. After that it’s gone.
• The surfaces on the PC now follow the website’s design: the control desk and the new guide on the PC in black and white, cyan and violet for MISU AI’s own drawing, the same typeface, the same fade-ins.

Fixed

• “What has changed” named an old version, because the number had been entered by hand and was never updated. It is now read from the newest entry in the list.

New

• The head answers in under a second, like always. Afterwards, the same question goes to a larger teacher model on the PC, which is allowed to take hours.
• The teacher works while nobody’s asking. You’ll find its answers in the trainer under “The Teacher”, where you can tick them.
• Ticked means the head learns from the teacher instead of from itself. The answer goes into the training material as the teacher’s word. Nothing happens on its own.
• Rejected answers stay on record. What you turned down says as much as what you kept.

Fixed

• The head took minutes to answer (217 seconds in testing) because the teacher held it up. Now the teacher runs alongside: 13 seconds, and under one if the head is already awake.

New

• The head is now described correctly. Under Progress it was still listed as “The brain” and as someone else’s model. That hasn’t been true since 11 September: it runs on the same foundation as the big heart, with its own training.
• You train it in the trainer under “The head on the PC”. It still gets nothing personal: it learns only from questions, factual material and knowledge material, never from what’s reserved for the heart, and never from its own answers.

Fixed

• The big heart was still listed under Progress with an outdated model name. It, too, has been on the newer shared foundation since 11 September.
• “Which model is running” had a version number written into its text. Numbers like that always go stale, and the phone had long since moved on. Now the text only states the rule, and the current number sits at the top under MISU AI.

New

• Projects: the trainer has a view of its own for the things you’re building.
• Where the PC has an overview page, the app opens that very same page, including 3D models you can rotate and zoom.
• Notes per project, and only notes: you can type them or save an answer from the chat into a project.
• They never go into training material, the diary or what it knows. MISU AI doesn’t learn them or think about them. The content belongs to the project, nothing more.

New

• Interrupted training can be resumed: if the PC shuts off mid-run, the run picks up from its last checkpoint on the same material instead of starting over.
• The PC now saves checkpoints twice as often.
• A breather for the small models: the phone heart and phone head let the graphics card rest briefly in between.
• The background was crashes that coincided with constant load: started directly, 6 of 15 runs failed, versus 1 of 8 with the big models’ natural pauses. A run takes about one and a half times as long; it learns the same.

New

• All the AIs now learn from your statements. When you state something instead of asking, your sentence becomes the answer, and the head only writes the matching question.
• Before, only the phone heart learned from this, and it learned the AI’s own reply, even when that reply had disagreed with you. The diary now has a card called “Your statements”.
• Health and safety statements wait for your yes: they only go into training once you approve them. You can take any flipped statement back out again.
• Personal things stay as they are: wishes, feelings, first-person statements and sentences addressed to it aren’t flipped, or it would speak your words as its own. The thought itself says why.
• PC status in the trainer: a traffic light shows the graphics card, its memory, the processor and RAM, and who’s currently using the card. If the PC is busy, the app asks before starting.
• The traffic light doesn’t show temperature or power draw, because Windows doesn’t expose them.
• Before every training, the PC now checks the load too, not just memory: if another program is working the card hard, no run starts.

Fixed

• “Take today’s portion” got stuck at 40% because, as your thoughts piled up, the balance with the rest of the training material was off. Now the rest grows along with them. The next phone-heart training would have failed too.
• A cancelled run didn’t say why. The PC now records how the run ended.
• On the “Who I’m becoming” card, the twelfth moon no longer wraps onto a second line, and the origin of the starting point shows without brackets.

New

• Every AI shows how current it is: in the trainer, the phone heart, PC heart, head and phone head each show which version answers, when it was last trained, whether a newer training isn’t released yet, and how many of your thoughts are already in the model.
• All four AIs have version numbers now. Each release gets the next number, along with its training date.
• Releasing with a choice and a status: in the pickup box you pick heart or head, see what’s on the phone and what’s ready on the PC, and follow the release step by step.
• You release the head and the PC heart right at their own boxes.
• In the diary, every thought shows where it lives and where it goes. For each AI, a marker tells you: already in the model, comes in with the next training, or not for this AI (tap for the reason).
• A card at the top of the diary shows how many thoughts go to which AI.
• Your knowledge beats old knowledge: if you say something differently than what was learned, your newer statement wins, in training and in conversation. Answers an AI gave itself still don’t win.

Fixed

• Phone models couldn’t be released because a file was missing. The last released phone heart was v13 from 5 September, the last phone head v1 from 8 September. Now it works: phone head v2 is released and answering.
• “Release” reported an error even though the PC was still working: the app waited 60 seconds, but a release takes minutes.
• “Fetch” sometimes fetched the wrong model: the newest phone model, whether heart or head. After a head training, that brought back the old head.
• The PC shut off in the middle of a training on 13 September. Before a training starts, it now clears the graphics card completely, checks that it’s empty and keeps it empty during the run.

New

• The head and the big heart now run on a newer shared foundation, each with its own training. Both can be trained from the app again.
• Measured with the same questions: 31 of 33 factual questions right (the two previous models got 23 and 26), plus the most natural German of the five models tested.
• Its character has a voice of its own, without the telegram style. Its own training is applied in moderation: enough for attitude, but not so much that every answer shrinks to a curt “Yes.”

Fixed

• Stock replies from the old models: its character echoed their answers because they were shown to it as memories. Now it only remembers what you said.
• Old model answers ended up in training as yours. “Process thoughts” copied every thought a second time without noting who had written the answer. The big heart now learns only from you.
• The head dodged factual questions and pointed back at your decision. Life questions and ethical questions are now learned only by the heart.
• After a power failure, the trainer now says what happened, with time and step, instead of a vague message about an aborted start.
• While a training runs, the head and heart on the PC stay off and the phone answers alone. Before, a question mid-run could pull a second big model onto the card.
• The text for warming up the head still mentioned an old model and its size.

New

• In the portal you now choose who learns from a document: all three, the big heart on the PC (biographies, letters, philosophy) or the head (nonfiction, history).
• A training run then only takes what was meant for it.
• “All three” stays the default, for a measured reason: on 7 September not a single thought came from the big heart yet. The path from the big heart to the small one is built, but it has never carried a line.
• If you give a document only to the big heart, the phone never gets it. And the phone is the only one that answers without a PC.
• Anything without a choice, including everything from before, still goes to all three. Otherwise every training run would quietly shrink to just the new entries.

Fixed

• Anything that comes from a document or a story is now labelled that way. The answer is your wording, but the question was written by the head. The label for that existed, yet nothing ever set it: the portal and storytelling mode quietly filed everything as human-written.
• Without that distinction you can’t tell later which questions came from a model. That matters as soon as the head itself is trained.
• MISU AI is no longer fixated on the Midnight ecosystem. Product facts are out of its training, and where it used to introduce itself as part of the ecosystem, it now talks about itself instead of reciting a catalogue.
• Factual questions are the head’s job, so the heart doesn’t have to carry them.

New

• The trainer now shows what it’s computing on: the name of the graphics card, or a red “CPU!”.
• That’s measured, not assumed: the PC checks it right where the training actually runs.
• If the graphics setup was missing, training used to carry on silently on the main processor, about twenty times slower, and it looked like a stuck run.

New

• Each AI now has its own section in the trainer. The heart on the phone and the big heart on the PC can be trained separately, even when you’re out and about.
• The head is listed too, and it says why it can’t be trained: it exists here only in a form that doesn’t allow it. A button would have promised something the machine can’t deliver.
• The list comes from the PC, not from the app. The PC knows what’s actually there; a built-in list would one day show a button for a deleted model.

Fixed

• “Train” used to train only the phone’s heart. The big heart had to be started by hand, and the head wasn’t trained at all.
• Heart and head never got your thoughts. What went across as “memories” was only a bare word list: fragments, not thoughts.
• Now the PC picks out the matching thoughts itself and passes them on as question and answer.
• The boundary hasn’t moved: the head only gets factual material, the big heart gets everything. Personal, emotional, poetic and every emergency never go to the head.
• If the PC finds nothing suitable, it passes on only frequent terms and honestly calls them that, instead of passing them off as memories.
• The big heart now trains on the right data, without its own old answers. Otherwise it learns from itself, which is what broke an earlier version.

New

• The shared sense now really reaches the answer. What three people have confirmed independently flows in, always with its origin: “from the shared sense, 3 independent sources”.
• A majority isn’t proof. Without the origin attached, it would look like one.
• A stance, say on a question of ethics, is reported as a stance, not as knowledge. That category existed from the start, but nothing ever produced one.

Fixed

• Reported, but never read. The shared sense on the device was only filled from its own storage, which nothing wrote to, and the step that pulls it into an answer was never called. Four contributions had sat on the PC since 5 September doing nothing.
• Taste isn’t knowledge. Questions of preference could get into the shared sense, and three matching votes would have turned a preference into a fact. Statements that only hold for today or right now no longer go in either, since they’re wrong tomorrow.
• The switch “also what still needs sources” did nothing. The list already arrived filtered from the PC, so with just one person it was always empty.

New

• You can now change your password in the app, under EVO in “Your account”. Before, every change had to go through the PC.
• It asks for your current password even when you’re signed in, so nobody who briefly holds your device can take over the account.
• Your other devices get signed out; this one stays signed in.
• The head can be warmed up at start (Settings → Heart and head). A cold start takes 33 seconds.
• It’s off by default on purpose: warming up ties up a lot of graphics memory that the big heart needs too.

Fixed

• The PC now cleans up after itself before a training starts. Head and big heart together take a lot of graphics memory, so a run didn’t fit beside them and only failed after minutes, with a message about the graphics card.
• The big heart finally has a card in the control panel, with “wake” and “put to sleep”. Before, putting the head to sleep freed the head’s memory but left the big heart’s untouched.

New

• The menu now has two levels. Thirteen entries in a row had turned into a search rather than a list. Settings and EVO sit on top, with three groups below: Backup, AI settings and AI input.
• The functions behind them are unchanged.
• A kept thought now carries what it refers to. When you answered a follow-up question, only your last sentence was stored, a reply to nothing. Now the sentence before it and MISU AI’s follow-up question go along as context.
• The context is stored apart from the question, so the question stays exactly the sentence you wrote.

New

• One PC, two routes, one account. At home or on the go, the app used to treat that as two different PCs and asked for a separate sign-in each time.
• A valid sign-in was being thrown away. Now the app recognises the computer itself, not its name, so two computers can share a name.
• The password can be reset at the PC. That’s deliberately not possible remotely: anyone who could reset someone else’s password could take over any account.

Fixed

• A dead sign-in looked like a live one. If EVO answered “not authorised”, the app kept the access and showed “signed in”, and every contribution quietly fell flat. Now it drops the sign-in and shows the sign-in field.
• The entry for EVO’s balance got lost at every start. It was read, but never saved.

New

• The overview now shows which model is running. It’s always the newest one.
• The newest version was tested: five runs per probe, no extra guidance, same settings as in the app. The emergency question led to 112 in five out of five runs.
• Two runs out of five still slipped into the generic voice of the underlying model instead of answering as MISU AI.
• Camera scanning stays in, so the app stays at 41 MB instead of dropping back to 25.
• New: “read-only”. An EVO member now has four states instead of three: waiting, read-only, free, blocked. A read-only member sees the room, but doesn’t write in it and adds nothing to the shared knowledge.
• The QR code is what made this necessary. Before, the admin approved everyone by hand and knew who. Once a code can be passed around, all-or-nothing is too blunt.
• The default doesn’t change: whoever joins is still set to “free”. Under “Invite” there’s now “What the newcomer may do” with two buttons. It applies to the next join, not retroactively.
• The message now says why. A read-only member used to see “waiting for approval” and turned to the admin, who had nothing left to do.
• The big heart has a second version. It memorises its training examples less, which is the better sign here.
• What the second version doesn’t solve: the answers in the training set are still short, so the big heart doesn’t learn long stories this way. What helps there is storytelling with “Keep your words”.

New

• In the shared knowledge, the rule was lopsided. Three independent sources are needed to make something count, but a single objection was enough to silence it for good.
• The way back had existed on the PC since 3 September, but nothing ever called it. Now every silenced entry shows “Withdraw objection”.
• You can only withdraw your own. Lifting other people’s would be the opposite lever: one silences, one lifts, and so on forever.
• Clear the room: at the top right of the shared room, for the admin only. The PC could always do this, but the app never asked. It checks with you first and tells you the messages vanish for everyone else too.

New

• Three things were built and connected to nothing: the knowledge network, the shared thoughts and the topic choice. Built but never called is worse than not built, because it looks like it’s running.
• Menu → “What it knows” now shows the network: the key terms, the connections, the density, and what each connection rests on.
• Built from the kept thoughts, the network was almost empty. Built from the chat history it isn’t, so you can now switch the source.
• From the history only your sentences go in, not its answers. Otherwise the network would only know what it said often itself.
• The trainer has “What’s worth learning”: what the big heart answered well and what keeps coming back. It’s still empty, and the card says why: so far not a single thought came from the big heart.

Fixed

• The network used to consist entirely of filler words, the kind that say when or how sure something is, never what about. They’re now on the list of everyday words that get ignored. Nouns stay in.
• Whether an answer counted as “big heart” or “phone heart” depended on a piece of display text. A rewording would have mislabelled it silently, and a wrong origin is the worst error in this app. Setting and recognising it now share one source, with tests.
• Settings still named an outdated head model. That’s corrected.

New

• Pause, resume, start over: instead of a single button there are now four ways. A paused download stays put and picks up again, even after the network drops.
• If a download breaks off, what was already fetched isn’t lost. And if the PC puts down a new model with the same name, the download starts over by itself, because otherwise you’d end up with a file that isn’t a model.
• Thoughts can be edited and deleted: tap them in the list under “Give an example”. Deleting removes the entry from the diary, but what was already learned lives in the model and doesn’t come back out. The dialog says so too.
• Duplicate thoughts are marked in the list. They aren’t cleared away automatically; which version stays is up to you.
• A thought no longer needs a question. Leave the question field empty and the head looks for a fitting one. If it isn’t awake, the field stays empty and can be filled in later.
• Without a question the thought doesn’t go into training, because an example that answers nothing isn’t one.

Fixed

• “2 · fetch” loaded the wrong file. What came down was a model meant for the PC and many times larger than the phone’s model. Now it picks the one that belongs on the phone.
• The line used to state a fixed size that fit the phone model but not what actually arrived. It now shows the real size, as the PC reports it.
• The download used to live inside the screen, so leaving the trainer menu lost the run. For a download that takes half an hour, that’s more than a cosmetic flaw. It now keeps going.
• Duplicate thoughts are no longer accepted. If the PC’s reply got lost on the way, the same entry arrived a second time, and in training twice means double weight, the same kind of mistake that once broke an earlier version.
• Settings still named an outdated head model, so anyone reading them got told about a model that no longer runs. That’s corrected.

New

• Settings → “The way back” brings what the PC kept of the chat history back to the device. Until now everything only went one way.
• This mistake has happened before. After a reinstall, something was missing that had been stored elsewhere all along, because the app never read it back. The chat history would have been the next case.
• The vocabulary grows along with it, but only from your sentences. Its own answers come back into the history, not into the vocabulary; otherwise it would learn from itself.
• The sense readings don’t come back. The device measured them and the PC only stores them, so on a new device they’d be borrowed numbers.
• What you ticked for learning on purpose isn’t touched by the history checkbox. That’s your own decision and gets withdrawn in its own place. The PC’s reply tells you so.

Fixed

• The delete screen in Settings said that all of it lives on this device. Since 7 September the chat history is on the PC too. There’s now a sixth checkbox, “The history on the PC”, because otherwise “forget” would have been half the truth.

New

• Join EVO by QR code. In the shared room there’s now an “Invite” button at the top right that shows a code. Whoever scans it is in right away, with no waiting for approval.
• Anyone without EVO yet is no longer stuck at a dead end. Instead of just “you need an account and the admin has to approve you”, the screen now shows the way: scan, or enter the details by hand.
• An invite code can be withdrawn at any time and only works for one person joining.
• The code never shows up in the chat history. It lives in its own view and disappears with it, because histories get passed around, backed up and photographed.
• Only the admin can invite. Otherwise everyone who joined could let in more people, and one invitation would turn into a chain nobody could keep track of.
• The app grew from 25.4 to 40.7 MB, almost all of it the camera recognition for scanning. Without it you’d enter things by hand, and the app would be small again.

New

• Survival know-how now lives in the instinct. In any emergency, and for any question about survival even when nothing has happened, the instant answer is followed by what else applies.
• The more valuable part is the “don’t”, for example: don’t move someone with a suspected spine injury, no ice on a burn. Doing nothing rarely hurts; doing the wrong thing very much can.
• Every “don’t” comes with its why. If you know the reason, you’re more likely to stick to it than to a bare rule.
• This is reviewed text too, output word for word: no model, no invention. It doesn’t replace the emergency call, it fills the time until help arrives.

Fixed

• A stale cancellation looked like it had just happened. The message that a run was cancelled at startup carried no time, even when it was eleven hours old. Now it says when it happened.
• Training started from the button in the app now runs on the graphics card too. A run started before the switch still used the main processor, and nothing in the message showed it.

New

• The trainer now tells you what it’s doing. At the start, the app shows what the PC is up to: reading data, loading the model, preparing examples, with a percentage where there is one.
• Before, you got “Step 0 of 0” and an endless bar for a minute or two.
• You can open the PC’s log right in the app. If a display doesn’t explain something, the log says what’s going on in plain text.
• Your PC tells the app how to reach it. After the first connection the app knows both routes, at home and away, and switches between them on its own. No more switching by hand when you head out.

Fixed

• A cancelled run was invisible. If training died while starting up, the reason was only on the PC. Now the app shows it.
• A training run that kept going after the PC service restarted was treated as finished, and the app offered a second start that failed. Such runs are now recognised and named, with a note that they can only be stopped on the PC.
• “Cancel” did nothing for two minutes after starting. If you started by mistake, there was no way back.
• Errors from the PC sometimes showed up as gibberish instead of a readable message.
• When fetching a model, the second step stayed open even though the model was already on the phone.

New

• Your speaking rhythm is now measured, meaning pace, pauses and beat, and the soul gets it too. It should know how you speak, not just what you say.
• The PC listens to each recording once more and measures words per minute, pauses, how much of the time you’re talking and how even your beat is.
• Measured, not interpreted: what it gets is a measurement, such as words per minute, not a verdict about you. The soul may draw conclusions from it, and it’s explicitly told that’s what it’s doing.
• You’ll find it under Read Aloud once there’s a recording. It only runs when you tap, and takes about one and a half times as long as your recordings put together.

New

• Your voice samples now show up in the diary as their own entries with a microphone icon, below the day’s thoughts, with length, recording time and the text you read.
• In the calendar, every day with recordings shows a small microphone. Days with samples but no thoughts can now be opened too.
• A voice sample gets no progress tick. A thought gets processed and is eventually done, but a recording just sits there until there are enough.
• The diary is now right in the menu. It used to be several screens deep in the trainer, too far away for something that gets new recordings every day.

New

• Read Aloud: write a text, read it out, and the recording goes to your PC. It learns from sound and text together, so both have to match. At least 30 seconds.
• The page shows your progress: recordings, minutes, and how many more you need for the next level.
• These recordings are kept, the only ones in the whole app. Everywhere else, sound is read and thrown away, but you can’t learn a voice from text. The app tells you before you start.
• They stay on your PC, not with EVO.
• Voice samples are recorded at higher quality than usual, since hissing sounds and the sharpness of a voice can’t be brought back afterwards.

New

• Under every answer there are now tick boxes: you choose what to keep as a thought, whether that’s what the heart said, what the head said, both or neither. Then “save” or “discard”.
• There are two separate boxes because two organs answer. A single box would only give you “both or nothing”.
• Nothing is preselected. If you tick nothing, nothing is saved; the button stays grey and says “nothing selected”.
• The choice sits below both answers. In the first version it sat above the head’s answer, so you’d have decided before reading.

New

• Storytelling mode now goes through the PC too, when it’s reachable. Accuracy counts double here: a story becomes learning examples, and a misheard word then lives in the model.
• When storytelling goes through the PC, the senses follow along fully, with pitch and not just volume.

New

• The PC now does the listening when it’s reachable, and it’s preferred. In testing it got names and technical words word-perfect that the phone had misheard.
• The sound goes to your PC, not out onto the internet. There it’s read and thrown away, never stored. Without a connection the phone listens itself, as before: worse, but instant, and nothing leaves.
• When the PC does the listening, the senses run in full, with pitch and not just volume.

Fixed

• Part of the poor recognition was gaps, not mishearing: after every pause, the start of the next sentence went missing. When the phone does the listening, it now waits much longer for silence before cutting.

New

• A story you’ve told is no longer lost. It’s kept on the phone, so if you leave the page or the PC isn’t answering, it’s still there next time. Only text is saved, never sound.

Fixed

• Storytelling lost the last sentence too, the same bug as in dictation: whatever you said between the last recognised section and your stop was gone.
• The length limit was set too high: stories that were too long got accepted and then gave back less than they should, or nothing at all.
• Up to about 15 minutes of talking works fully. Above that, the app now says so clearly, with a number, instead of a message that looked like a network problem.
• Elsewhere, too, the app now names the real reason, like “this PDF has no text, probably a scan” instead of “the PC didn’t answer”.

New

• Dictation now works like you’d expect: tap, talk (thinking pauses are fine), stop. The text lands in the input box, you read over it and send it.
• You see what’s coming in while you speak. What was already in the box stays, so you can type half and speak the rest.

Fixed

• The recognised sentence used to go into the conversation unseen, as soon as Android thought you’d finished. A thinking pause was enough, and the reply answered half a sentence.
• Two messages ended dictation even though they aren’t errors. After a few seconds it said “I heard nothing”, although you’d only just started.
• Stopping lost your last sentence: whatever you said after the last fully recognised section was gone. That hit nearly everyone who stops mid-sentence.
• Above the box it said “I didn’t understand anything” while correctly recognised words sat in it. Android sends a late error after you stop, and that error is now ignored.

New

• The control panel on the PC now shows the status of each organ: control centre, head, soul, EVO and heart. It also shows how the phone can reach that computer.

Fixed

• Storytelling recorded nothing: two parts of the app were competing for the microphone, so speech recognition got silence. Now the mode frees the microphone first, and the senses keep following along.
• The message pointed the wrong way too, saying “that was too short” when nothing had arrived. Now it says when nothing was understood, and what to check.

New

• The Portal: give it a PDF, a Word document, a text file or pasted text. The PC reads it and suggests what it could learn from it.
• As with storytelling, nothing is taken in until you tick it.
• A long document is cut into sections. If something’s left over, it tells you instead of quietly trimming.
• When something doesn’t work, it says why: an old .doc file, a scanned PDF (which is just an image) or an empty file.

Fixed

• Sending a chosen file would have reliably failed on the phone, because Android often doesn’t give a picked file a usable location. Now the file itself is sent.

New

• Storytelling mode is finished: you tell a story and it only listens. Afterwards you can correct the text it understood, and the head suggests learning examples from it.
• Nothing is taken in before you tick it.

Fixed

• While telling a story, the text box showed a placeholder instead of the story, a typo in the program.
• Signing in to EVO could never work: the preset connection pointed at the phone itself instead of the PC. It’s now derived from the PC, and a wrongly saved one is corrected once.
• When evaluating a story, the head occasionally mixed up who did what. Measured: once in 48 suggestions, and not once in 96 after the change.

New

• Under “Who answers” there’s now “heart first, then head”. The heart answers right away and the head’s answer joins below it.
• The moon appears in the header, calculated rather than looked up, and you can set your star sign in Overview.

Fixed

• The soul was never asked: the “let the soul answer” switch did nothing, because two conditions in the program could never be true at the same time.
• The cause: a single lock stood for two different things, emergencies and credentials on one side, personal matters on the other. They’re now two separate rules, and a check keeps them apart.

New

• The bridge between measurement and meaning. Until now MISU AI measured how a sentence sounds but couldn’t do anything with it, and deliberately named no emotion.
• Now the head sorts the text into joy, sadness, anger, fear, calm, excitement or none, and that judgment is attached to the measured voice pattern. Eight out of eight test sentences were sorted correctly.
• Over time that builds up to “this is how joy sounds for this person”, without it ever guessing from the voice.
• Everything is relative to your own baseline, never absolute: the same pitch means something different for a deep voice than for a high one. Three samples of a feeling give it something to go on.
• It still doesn’t assert anything. What comes out is a suspicion, meant for asking, not for concluding.
• Voice input now waits for three seconds of silence before stopping. When you’re putting a sentence together, you pause mid-way, and you used to get cut off.

New

• The trainer now shows all thoughts, each with its number and a tick showing whether it’s already gone into training.
• Before, you only saw a single example, the one from the current training step. That was a progress display, not an overview.

Fixed

• In storytelling, no more sound goes to the PC. Instead of the recording, only the result of the analysis travels: time, pitch and loudness. No words and no voice can be recovered from that. Dictation works differently since 0.5.33, the same day: the sound goes to your PC to be read and is thrown away there.
• Recording in storytelling mode fought with speech recognition over the microphone, the same conflict that was fixed a day earlier, and it had crept back in elsewhere. It’s fixed.

New

• Ego and soul have a symbol now: a tappable yin-yang in Overview.
• Highlighting one half dims the other without hiding it, since neither exists without the other.
• Each half carries a dot of the other: the ego a core of feeling, the soul a core of reason. Two things that depend on each other instead of fighting.

New

• Storytelling mode now records too. With your approval, the recording goes to the ego on your PC, where there’s room.
• It carries no name, only an anonymous marker. MISU AI can recognise you by it, but at that point doesn’t know who you are. So it’s a collection of anonymous voices rather than a register with real names.
• That means speech leaves the device for the first time. Until now only text went to the PC. A voice is a biometric feature, so it’s spelled out here.

New

• In storytelling mode you can now read and correct the recognised text before it goes on. Android mishears names and rare words, which would put errors into learning examples.
• What you tell belongs to the ego. A story shapes the self you narrate, not the soul, which stays the same underneath.

New

• The mixing desk: five sliders for what MISU AI is told about itself before it answers. Name and portrait, memory depth, instinct, senses and self-check.
• You can switch each source off individually and change their order.
• The compare button asks the same question twice, once with everything and once without one source, with both answers stacked. That shows whether a source adds anything.
• Overview shows where it stands: seven levels, how far it is to the next, and the two persons, ego (thinks, doesn’t know you) and soul (feels, knows you). You choose both of their names.
• The mixing desk isn’t neural wiring: you set what it’s told, not how it thinks.

Fixed

• Until now the wiring was hard-coded in the program: every change meant a rebuild, and nobody could measure whether any of the five sources did anything.
• That’s not an academic question: on 4 September the guardrail was the cause of errors four times, and the model never was.

New

• MISU AI learns how individual words sound when you say them. Type a sentence in the input box, tap the orb at the top left (it gets a ring) and say the same sentence.
• It measures pitch and contour and links both to the words.
• It’s not about how a whole sentence sounded but about how this one word always sounds for you. A word counts as established from its third recording, since a single measurement isn’t proof.
• It names no feeling and just says what was measured, like low or animated, and how often. Concluding “sad” from a frequency would be exactly the kind of claim it otherwise refuses to make.

Fixed

• Yesterday’s limitation is gone: during voice input pitch couldn’t be measured, because Android’s recogniser only reports a level. Here the text is already written, so it listens with its own senses.
• No second recorder is opened. It listens along with what the senses already pick up. A second one would be exactly the microphone conflict that silenced voice input yesterday.

New

• Storytelling mode: you tell a story, MISU AI only listens and doesn’t answer. Afterwards it shows you what it could learn from it, and you tick what to keep.
• None of it is stored beforehand.
• It keeps listening through pauses. Pauses to think are normal when you tell a story, so it simply stitches the recognised pieces together.
• You’ll find it in the menu at the top right under “Storytelling”.

Fixed

• The privacy lock now runs before anything is sent to the PC. When you tell a story, things slip out that you’d never type.
• In a test, private details used to end up among the learning suggestions. Now PINs and addresses stay on the phone, and the rest of the story can still be learned.

Fixed

• Voice input recorded and delivered nothing: no text, no message. The recogniser was simply hearing silence.
• The culprit was two features sharing one microphone: while the senses are listening, the recogniser gets only silence. It only started once the blind senses were fixed; each change was right on its own.
• The senses stay on. The magnetic field keeps running, loudness comes from the recogniser, and the senses can still pick up your voice.
• Only pitch can’t be measured during a voice recording. It reads 0 then, instead of an old value that would look like a real measurement.

Fixed

• Voice input recorded, and then nothing happened: no text, no message, just a microphone that stopped pulsing.
• Now the problem shows up above the input field, right where you look after hitting stop. For the most common cause, it also tells you what to do.
• Since voice input stopped sending anything to Google, your device needs the German offline speech pack. Without it, nothing gets recognised, and nothing told you so.

Fixed

• The same bug as in 0.5.9 was hiding in one more place, behind the one just fixed: the app said it was bringing in the PC, but never did.
• In a test, the answer stayed on the device alone on the third try.
• One single place now decides whether the PC is needed, instead of two that relied on the same unreliable gauge.

Fixed

• The same factual question got a worse answer the second time, just because it had been asked before.
• In a test, the first time it called in the PC and got it right. The second time it thought it knew, skipped the PC and was off by a factor of more than ten.
• The reason: memory holds what you’ve said, and it was being used as a gauge of what the model knows. Now it only decides for questions about you.
• For everything else the PC joins in, if it’s there.

Fixed

• MISU AI labelled its own answers “thought by the PC”. When it had answered itself and the PC only added something, its own sentence still got the PC’s label.
• On the device, two contradicting figures appeared one under the other, both labelled as coming from the PC, although the first came from the model on the phone.
• The bug was hard to spot because one of the figures was roughly right. Wrong origin is the worst possible mistake in an app that’s meant to keep origins apart.

Fixed

• After a factual answer from the PC, MISU AI tacked on a sentence from its own instructions, and even copied it wrongly.
• A check against parroting already existed, but it didn’t cover this case.
• It now catches this even when the model changes word endings. In testing, the real faulty sentence was caught and four genuine add-on sentences weren’t flagged.

Fixed

• MISU AI locked out its own PC: in the trainer it flagged in red that the entered address didn’t belong to your PC, even though it was exactly the right one.
• The reason: the check and the actual request read the same address differently, so they were comparing two different things.
• Every address entered in short form was affected, including the form the input field itself suggests.

New

• The header bar is twice as tall and always shows the senses with numbers: pitch and loudness from the microphone, field strength and deviation from the magnetic sensor.
• The readout stays up even in silence, because an indicator that vanishes then looks no different from a switched-off sensor.
• The orb now shows the magnetic field: a rotating ring of strokes that gets denser the further the field drifts from its baseline.
• Each sense has its own colour so you can see which one is responding: sound is cyan, field is violet.
• With the senses off, the orb shows a still, thin circle and the bar says “Senses off”. Before, off looked exactly like quiet.

Fixed

• The magnetic field had always been measured but never shown: the orb only got loudness. To you, a sense that’s measured but never shown might as well not exist.

New

• MISU AI starts over on its own before its short-term memory overflows, and tells you instead of doing it silently.

Fixed

• Nobody was tracking how full its short-term memory was. After about eight exchanges it filled up, what happened then wasn’t defined, and the only way out was the ⟳ at the top.
• It gets worse well before that: in a test, the same model answered the same factual question correctly in a fresh session, but after three exchanges of history it only pointed to the PC.
• A small model gets swept along by what came before. That’s why the restart kicks in well before the memory is full.

Fixed

• The senses were blind and nobody noticed: every chunk of audio from the microphone triggered an error that landed silently in the device’s log.
• So no loudness, no pitch and no feeling colour was ever measured. Its first principle, feeling before thinking, had nothing to work with.

Fixed

• MISU AI dodged general-knowledge questions even though it knew the answer. The same model answered the same question correctly on the PC.
• The cause was in the app, not the model: it passed along hints based only on what MISU AI has heard from you. With a fresh memory that’s nothing, so every topic counted as unknown.
• A second leak only showed up on the device: a suggestion to ask a follow-up question was sent along even with general questions. It was meant as an offer, but the model read it as an order and asked instead of answering.
• Now memory may only judge things about you. If it’s never heard of someone in your life, it still says so openly. Its knowledge of the world comes from training, not from your conversations.

New

• If nothing comes back from the PC, the heart answers instead of leaving you with nothing. Before, a check from app start decided the whole route, and if it was outdated, your question went nowhere.
• When you come back to the app, MISU AI checks whether the PC is there. A check result only counts for a short while, then it’s just a guess.
• Errors now appear next to the answer, small and muted, and in real sentences instead of machine language.

Fixed

• The wait for the PC was too short: MISU AI gave up much earlier than intended. If the PC was busy, it looked like “PC unreachable” when it just needed time.
• Worse, the PC’s error overwrote your finished answer. You’d already read it, then the app swapped it for an error message, exactly when the device had answered on its own.
• Two paths went around the lock on outgoing connections: the camera image and the model download. The fix for that had been built, but nothing ever called it.

New

• A newly trained model is here. It carries the name MISU AI and comes with examples for emergencies.

Fixed

• Asked for its own name, the app itself answered “I know nothing about that”, instantly and without asking the model.
• The check counts what MISU AI has learned from you. Naturally nothing about itself is in there, so it reported “empty”.
• But its identity comes from training, not from memory. The model had been prepared for exactly this question.

Fixed

• Your name and what MISU AI had noticed about you went to the PC, but never reached the model on the device. Exactly backwards. A first fix the day before had left the path open.
• An emergency could fail because of the learning feature: a message about an emergency was filed like an ordinary detail and just acknowledged. Emergencies are now checked before anything else.
• The camera button sent image and text to the PC even when “Allow asking the PC” was off, and without the lock. Both are respected there now. The lock can’t inspect an image itself, though.
• In EVO, an objection couldn’t be withdrawn if you’d objected twice. The entry would have stayed blocked for good.