260913202936.0.02026-09-13
Reviewed copy and reliable guidance
- German and English interface copy, emails and fixed speech were reviewed for natural language and accurate guidance.
- New or changed copy requires current review evidence. Download text, durations and email confirmations are aligned.
260913201213.0.02026-09-13
Three liquid gauges, including small screens
- A third jade privacy capsule identifies the selected processing location without invented privacy percentages.
- All three gauges now appear first; compact profile rows and dark-theme styling improve smaller screens.
260913192329.0.02026-09-13
Your AI settings: Simple or Pro
- A compact settings dialog offers Simple and Pro views on small screens.
- Animated liquid gauges display response speed and reasoning depth as relative preferences.
- Model version, reasoning effort and processing route stay in a draft until applied.
- A dedicated local text workspace connects directly to MTPLX on this device.
- European text processing and GPT Live are prepared behind deployment readiness controls.
260913123400.0.02026-09-13
Reliable status polling
- Status polling pauses without a session and resumes after session recovery.
- Pending call reconciliation respects session ownership and server retry delays.
260913115437.0.02026-09-13
Natural German conversation copy
- German greetings, concept updates and recovery messages use clearer, more natural wording.
- Conversation instructions reinforce idiomatic German and a consistent polite form of address.
260913112410.0.02026-09-13
Conversation menu release validation
- The release test inventory includes the new automatic conversation title checks.
260913105536.0.02026-09-13
Conversations with clear titles
- Conversations receive short titles from their first useful assessment and first concept; manually chosen names stay untouched.
- The conversation picker uses clear dates, an open-conversation label and consistent icons and spacing.
260913102816.0.02026-09-13
Stable transcripts and deliberate pauses
- Final transcripts replace live text without collapsing the message bubble.
- Deferred paused assessments no longer create false failures; real notices have a separate area.
- Attachment left, send right, dismissible account menu and aligned utility actions.
260913092323.0.02026-09-13
Clear chat and compact controls
- Opaque bubbles, matching send and attachment controls, and contextual typing guidance.
- Unseen concepts remain in the background and can be opened directly.
- Profile context beside the assessment and a clear personal contact action.
260913082801.0.02026-09-13
Conversation data immediately accessible
- Resumed conversations expose download and personal contact immediately, without waiting for email verification or assessment refresh.
260913080456.0.02026-09-13
Clear email access and direct personal contact
- Email verification has one primary action and understated download and personal-contact links.
- Locked assessment content no longer creates an invisible scroll area.
260913063947.0.02026-09-13
Credits and access in one place
- Credit recovery appears directly in the assessment panel.
- Locked sections replace the regular tab icon with a centered lock.
260912204712.0.02026-09-12
Reliable conversation resumption
- Resuming shows history, navigation and text input before a voice connection is started.
- Connection errors offer an actionable retry and retain unsent text.
- Browser history navigation reloads the current conversation state.
260912203643.0.02026-09-12
Shared, reliable conversation limits
- Conversation transcript limits now come from the pinned Aithema core, keeping transcript length and turn counts aligned with the shared requirements logic.
260912180951.0.02026-09-12
Clear conversation and pause status
- Service dots distinguish current activity, previous outcomes and stale observations. Background assessment no longer changes the conversation-model status.
- Delayed microphone-control notifications during pause no longer claim speech recognition failed. Actual failures and unconfirmed states remain distinct.
260912180044.0.02026-09-12
Reliable concept control toggling
- Repeated image clicks without pointer movement reliably toggle floating controls.
- Floating actions adapt their spacing to the actual footer height.
260912174816.0.02026-09-12
A clearer conversation library
- Conversation management prioritizes resume, search and contextual actions, with inline renaming and deliberate deletion.
- Email verification shows a stable thank-you message without an unreliable close countdown.
- Concept images use the full viewport height. Pointer movement reveals all controls; clicking the image toggles them together.
260912173424.0.02026-09-12
Immersive concepts and two-column assessment
- Concepts fit the full browser viewport without cropping. Floating controls fade after two idle seconds. Escape reveals controls; another Escape returns to the conversation.
- Assessment separates overview and details into two columns, with professional readiness copy and centered navigation icons across browsers.
260912164057.0.02026-09-12
Clearer icons and concept preview inside chat
- The unseen concept preview is centered inside chat at two-thirds size, 50% opacity and a subtle blur.
- Seven simple three-path SVG motifs replace the dense set. Size and stroke weight match the neighboring tabs; the previous five motifs remain excluded from the next change.
260912125447.0.02026-09-12
Reliable status transitions
- Rapid status changes handle cancelled animations cleanly. Conversation and attachment regressions are covered.
260912124840.0.02026-09-12
Waveform status and accessible attachments
- Current status floats in the waveform. The conversation menu contains the end action.
- Typing remains available while paused. Sending resumes text processing with audio held off.
- Attachments can be opened and removed from a labelled area. Progress sits directly below the clock.
260912123940.0.02026-09-12
Restore service dots beside the footer logo
- The three service dots retain their colors and states beside the logo at bottom left.
260912123340.0.02026-09-12
Concept background with a soft opacity gradient
- The unseen concept sits strictly below all page content. Its opacity fades from 0% on the left through 5% at the center to 10% on the right.
260912122109.0.02026-09-12
Immersive generation and subtle concept cues
- Image generation fills the canvas with calm motion and stable text. Timing details can be expanded.
- An unseen concept appears behind the page at 5% opacity and disappears when opened.
- Missing concept titles use “Your concept”. Navigation dividers span the full height with softly faded edges.
- Stronger display normalization lets very quiet speech use more waveform height.
- The borderless conversation rail sits at the top, with the logo and service dots at bottom left. Copyright moves to the imprint, and the AI orb again shows a speech waveform.
260912121305.0.02026-09-12
One conversation rail with a stronger live waveform
- The original-color AI orb, spoken output, pause and a single microphone share one rail above the chat.
- Adaptive display normalization makes the waveform use more height. Speech travels from the speaker’s side without changing capture or recognition.
260912114246.0.02026-09-12
Conversation control follow-up from QA
- The conversation control stays clear of navigation at intermediate desktop widths.
- During resume, assistive technology retains the paused state until acknowledgement.
260912113243.0.02026-09-12
Unified conversation control and varied greetings
- Microphone, pause and spoken output share one compact capsule. Typing stays available, and the waveform microphone controls the same input.
- Eight standard conversation messages each have ten variants. The five most recent choices are excluded from the next selection.
260912090723.0.02026-09-12
Calm status replies and clear mobile navigation
- Questions about a pending concept receive a brief status reply without further design questions. Requests to wait are respected.
- Overdue text uses the selected language. Mobile navigation remains icon-only during generation.
260912084624.0.02026-09-12
Contextual conversations while concepts render
- Ordinary answers continue the conversation. Only explicitly requested visual changes become separate notes.
- Silence and punctuation-only turns receive no reply. Finished concepts are announced without unsolicited revision prompts.
260912073512.0.02026-09-12
Equal navigation segments with animated concepts
- Conversation, Assessment and Concept stay equally wide and on one line, with icons on small screens.
- During generation the estimate replaces the title. Ten UI motifs morph smoothly alongside a subtle breathing light sweep.
260911193527.0.02026-09-11
Optional concepts with reliable guidance
- Visual concepts are requested explicitly; conversations can finish without image generation.
- While a concept renders, the assistant retains new input for later and discusses only independent operational topics.
- Ready announcements belong to the completed job. Accepted renders remain visible after the conversation ends.
260911135105.0.02026-09-11
Calmer design collaboration
- Preview, progress and remaining time live inside Concept. Its frame glows after completion until the first hover or keyboard focus.
- Settings use a wider layout with personal options in their own column and room to compare processing choices.
- The assistant receives authoritative design status. Automatic idle prompts are suppressed during rendering; useful discussion can continue before quiet waiting.
260911123241.0.02026-09-11
Designs follow the conversation
- The assistant bridges rendering time with relevant questions and announces completed concepts during a conversation pause. A clickable preview appears in chat.
- Operators sign in through Augmentoring Zitadel. Markus has unlimited credits; usage, concepts and estimated conversation costs in euros count upward.
- The small concept preview sits above the button, tilted left, and grows gently on hover.
- Download conversation data delivers a ZIP with JSON, readable Markdown and generated images.
- Clarified points expand individually or together. Added spacing and a subtle scrollbar improve the assessment.
- Email confirmation lives in the right panel. Locks pulse after submission, then snap open as the confirmed content slides into view.
260911105457.0.02026-09-11
Refine concepts in the conversation
- The preview beneath Concept expands to show the whole image on hover or keyboard focus. Only the button shimmers; the preview shows progress and approximate time remaining.
- Requested changes in the conversation can trigger a new image version. New concepts initially reserve about 50 credits and settle against reported image usage, replacing the previous five-image limit.
- Automatic drafts still follow readiness and quiet moments. Every image, including final concepts, requires available credits.
260911095914.0.02026-09-11
A concept in a quiet moment
- After about two minutes without new input, a first concept or a refresh with new information can appear even below the next readiness milestone.
- The small preview accompanies the conversation. No further images are made without new input; paused conversations, unfinished typing and hidden tabs wait, and a final image remains reserved.
260911092709.0.02026-09-11
The concept stays visible during conversation
- A small preview appears beneath Concept. Generation activity remains visible in the conversation view, retaining the previous image during an update.
- The conversation controls share one outline with straight separators. Conversation management is available from the first conversation, and assessment uses a chart icon.
- Reset completely is also available in Settings, with confirmation and a clear explanation of its effect. Cancelling returns to Settings.
v2.19.242026-09-10
Pauses stay with the right browser tab
- A second browser tab can no longer release an active conversation pause through its automatic recovery.
- The pausing tab keeps its pause with a short content-free lease; the same tab recovers immediately after a reload, while an abandoned tab is released automatically.
v2.19.232026-09-08
UI progress stays visible during updates
- During regeneration, the progress card now sits clearly above the dimmed existing concept while the previous glare steps back.
- Operator-authorized demo conversations bypass the normal five-concept UI limit and retain only the global safety boundary.
v2.19.222026-09-08
Long UI generations stay compact
- Progress cubes now fit the available width without horizontal overflow even when real measured generation time approaches two minutes.
v2.19.212026-09-08
The UI concept shows its work
- While a UI concept is being created, per-second progress cubes and more than twenty conversation-aware activities make continued generation visible.
- Successful generations are durably timed without conversation content; their average and observed maximum improve the next estimate, with 45 seconds used only before history exists.
- Visual approval and rejection remain visible as green or red conversation guidance, steer both agent and next image, and can be removed; UI concepts are triggered around 25 and 40 percent and no later than the conversation recommendation.
- Conversation history now uses an unmistakable chevron instead of a dot-like glyph, and the stray preview dot is gone.
v2.19.202026-09-08
New conversations connect again
- New typed and spoken conversations once again accept ElevenLabs' current signed response, where the conversation ID is securely carried in the WebSocket URL.
v2.19.192026-09-08
The UI concept renders again
- Visual concepts now use GPT Image 2 through the configured OpenRouter connection when no direct OpenAI key is present; specific failures are explained instead of collapsed into a generic message.
- Clicking “Conversation” now only changes the view. The separate arrow beside it opens conversation history.
v2.19.182026-09-08
One conversation, one clear control bar
- Conversation view and conversation history now share one entry point; assessment and UI sit flush beside it while settings remain predictably at the far right.
- Pause and view controls share the same height and remain calmly aligned on both large and small screens.
v2.19.172026-09-08
The UI preview closes reliably
- The enlarged UI preview now responds reliably to Escape even when other dialog controls remain active in the background.
v2.19.162026-09-08
The emerging UI evolves with the conversation
- A third UI view now sits beside conversation and assessment; a first concept can appear after two substantive answers or 20 percent readiness and evolve as new details arrive.
- Concepts can be refreshed, compared, enlarged, and downloaded; consented image references securely inform revisions and are deleted with the session.
v2.19.152026-09-08
Paused stays paused, causes become clear
- An intentionally paused conversation continues to appear paused even if a delayed transport failure arrives, and can be resumed normally.
- For a degraded or unavailable service, an accessible info control explains the sanitised cause and the next useful step.
v2.19.142026-09-08
Austria in mind, the conversation in view
- Unless told otherwise, the assistant now assumes Austria rather than Germany; explicit location information always takes precedence.
- New and growing conversation turns remain fully visible automatically, with enough reading space above the controls.
v2.19.132026-09-08
The conversation listens live again
- Spoken words appear again while you are speaking; brief opening vocalisations such as “mhm” also no longer trigger a reply or duplicate greeting.
v2.19.122026-09-05
Progress stays in view
- Assessment progress and key statements remain visible while scrolling; the privacy preview reaches the panel edge cleanly while keeping status and actions usable.
v2.19.112026-09-05
The assessment gets room
- The focus view now shows the complete assessment instead of an empty canvas; conversation state, input mode, and scroll position survive every switch.
v2.19.102026-09-05
The important journeys are tested
- Automated checks now focus on real conversation journeys and critical safeguards, so visible regressions surface earlier and the suite stays maintainable.
v2.19.92026-09-04
The assessment keeps up
- Answered questions and progress now react with the fast structural pass; the thorough synthesis follows independently without hiding the last useful result.
- Normal conversations no longer hit a multi-minute analysis countdown after a few answers; long wait windows now settle into a clear, manually retryable state.
v2.19.82026-09-04
The conversation stays central
- Pause, resume, connecting, and reconnecting now share one clear control in the centre of the header; role, email status, and credits form a compact footer status group.
- Notices, imprint, and data processing now sit together as concise links at the footer edge and remain orderly and reachable on small screens.
v2.19.72026-09-04
One shared assessment truth
- The assistant and understanding panel now use the same input revision: new answers immediately show as being checked, and a stale or still-open assessment can no longer be described as fully resolved.
v2.19.62026-09-04
Resume stays actionable
- After a normal voice resume, the pause control is fully re-enabled so repeated pause and resume cycles remain actionable.
v2.19.52026-09-04
Pause stays in the conversation
- If voice connectivity drops during a pause or shortly after resuming, the page now reliably retires the old connection and retries up to three times without parallel sessions or a stuck resume state.
v2.19.42026-09-03
A calmer start to the conversation
- The welcome screen now brings value, consent, and the start action together in one clear, compact composition across phones and large displays.
v2.19.32026-09-03
Consent stays cleanly aligned
- The disclosure and settings link now remain a calm, composed row in compact desktop windows as well.
v2.19.22026-09-03
Conversation and next step stay in flow
- Pause and resume now recover the conversation and live transcription with bounded retries, without duplicate sessions or billing.
- The assessment responds immediately to new answers, refines itself in the background, and keeps the last useful state visible.
- Consent and conversation controls are calmer, while the completion step now leads more directly and warmly to a personal next step.
v2.19.12026-09-02
Core journeys stay stable
- Pause and resume remain synchronized across delayed responses, preventing unnoticed live usage or billing.
- Demo access, dialogs, consent, and the masthead are more consistently aligned and keyboard-operable across mobile, tablet, desktop, and zoomed views.
- Activated demo conversations remain safely discoverable in conversation history without showing a misleading email status.
v2.19.02026-09-02
Demo access stays in the conversation
- Discreet operator access can now continue the current conversation as a demo without losing its transcript or assessment.
- The orb, primary action, credit display, and readable pause status now adapt cleanly from phones to wide desktops.
- A reconciled assessment pause can now recover in the same conversation without releasing a live manual pause or losing its billing adjustment.
v2.18.12026-09-02
The conversation can always be saved
- The CSV export now includes the complete conversation transcript and still works when the structured assessment is temporarily unavailable.
v2.18.02026-09-01
The conversation stays in flow
- Pausing, verification, and resuming now keep voice, history, and input mode reliably together.
- Temporary assessment failures can be retried without losing the last useful assessment.
- The entrance, credit notice, and generated view remain calm, compact, and fully usable on short and mobile screens.
- Provider notices no longer promise unverified EU processing and describe the global Mistral endpoint accordingly.
v2.17.02026-09-01
Voice gains a direction
- The live display now connects the listening microphone on the right with the speaking Augmentoring orb on the left, borderless and clearly responsive to each state.
- The view switch now stays anchored in the masthead, while older messages pass softly beneath a fine edge instead of being cut off abruptly.
v2.16.32026-08-31
Voice stays in the flow
- Voice and text input can now be switched at any time during a conversation without losing the active context.
- Selecting and uploading a PDF no longer interrupts voice input; the spoken conversation can continue immediately afterwards.
- Pause stays out of the way while entering an email address, and ending a paused call no longer leaves a stale error behind.
v2.16.22026-08-31
A calmer entrance at every width
- The headline, introduction, and consent line now form a deliberately composed, readable whole on tablets and around desktop breakpoints.
v2.16.12026-08-31
AI role stays in view
- The explicit AI-assistant role line now stays fully visible with comfortable spacing in short desktop windows.
v2.16.02026-08-31
Clarity, quietly built in
- The imprint and privacy information now explain more clearly who provides the service, why data is processed, and which routes are involved.
- A quiet but unmistakable label makes the AI role clear from first contact without taking attention away from the conversation.
- Machine-readable provenance now accompanies generated content, supported by internal evidence, risk, and response guidance for responsible operation.
v2.15.32026-08-25
Credits based on actual usage
- Spoken AI replies now consume their reported actual cost instead of a flat 25 credits per reply. Incomplete replies keep the conservative reservation.
- Technical requests without a readable conversation context now obey daily and monthly deployment ceilings and server-side model policy too.
- Call transitions, deployments, and the code-quality gate were hardened internally without changing the visible flow.
v2.15.22026-08-25
Silent transitions, visible signals
- Silent transitions between technical call segments now report their start, success or failure to the operational journal. These signals contain neither personal data nor conversation content.
v2.15.12026-08-25
Credits, not a time limit
- The visible seven-minute limit is gone. Technical session boundaries continue in the background while credits remain available.
- Only a silent one-hour safety guard ends a stuck conversation, with all existing context preserved.
v2.15.02026-08-25
Credits without a dead end
- The masthead now always shows the real remaining credits: 250 without a confirmed email, then 1,000 per conversation. A quality call starts from 120 free credits.
- When there are not enough credits, one fitting action leads forward: enter an email, resend the confirmation link, or request a personal continuation with the existing conversation context.
- Reset completely still removes the email and conversation list from this browser, but no longer creates another anonymous taster. A genuinely new conversation with its own credits remains available to confirmed addresses.
v2.14.22026-08-18
The last hand-drawn icons
- The microphone, the paperclip, send, both crosses that forget something, the gear in the settings window, its close cross and the tick beside “active” were all drawn by hand. They now come from the same library as everything else: one stroke weight, one geometry, one centre.
- The two crosses that appear during a live conversation are built from the library too now, rather than redrawn in the script. That is exactly where they had escaped the first round.
- What remains hand-made, deliberately, are the three small pictures on the processing tiers. Those are illustrations, not icons.
v2.14.02026-08-18
Every conversation gets its own allowance
- Anyone who has confirmed their email address now starts each new conversation with a full voice allowance. Until now every conversation of one person shared a single one, and the second began where the first had stopped.
- Without a confirmed address it stays one taster per browser. An allowance you can reset by clicking “new conversation” would not be one.
- There is a new per-person daily limit. Before, only the whole installation had one, which a single person could in theory have used up alone.
- Underneath is proper bookkeeping: every charge records which conversation and which person it belongs to, and every call carries its reference to the provider and back. When it is later settled against its measured duration, the credit returns exactly where it came from, even if you confirmed your address in the meantime or midnight fell in between.
v2.13.52026-08-18
Four faults in the header
- Two mail glyphs stood side by side. The glyph carries its own instruction to be visible, and it outranked the rule meant to hide one of them. Exactly one shows now.
- The conversation list was cut off at the right edge, because it still opened leftwards although its button has moved to the right.
- “Reset” on the entrance did nothing. Three places offer that button, and only the first one in the document was ever wired up.
- The role still had “inferred” spelled out beside it. The sparkle next to it already says that, and it is now a real icon rather than a text character.
v2.13.42026-08-18
Thinking happens quietly
- Some language models write their deliberation down before the actual answer. Until now that was spoken aloud, stood in the transcript and cost voice allowance. It is removed before the reply leaves the building, on both quality tiers.
- Deliberation already stored disappears from what the model sees as history too. Otherwise it learns from its own transcript that this is how it is done here, and every round reinforces it.
v2.13.32026-08-18
The header on one line
- Everything now sits on one horizontal line: the brand on the left, one sentence about you in the middle, conversations and settings as a single object on the right. The conversation picker used to drop onto a second line.
- The three dots beside the wordmark are the operating state. All well is three green dots; if a service fails, its dot keeps its size and place, becomes a red ring and pulses. Pointing at the dots or the wordmark explains them and offers the way to the Augmentoring site.
- Instead of the words “email confirmed” there is a mail glyph: with a tick when confirmed, with a question mark while the link is still waiting.
v2.13.22026-08-18
Real icons, and the sentence before the button
- Every icon was drawn by hand, which is why the gear never matched the keyboard and the keyboard matched nothing. They now come from a library (Lucide): one stroke weight, one geometry, one centre.
- On the entrance, the line saying what you agree to now sits above the button rather than below it. Read first, then agree.
v2.13.12026-08-18
Your address comes along to the next conversation
- A new conversation used to begin as though you had never left your email. That put you back on the small taster allowance, which the previous conversation had usually already spent, so the first thing you heard was that the budget was gone.
- The confirmed address now travels with you. Only the address: no transcript, no files. And an address you have not confirmed stays unconfirmed.
v2.13.02026-08-18
Saying what actually went wrong
- A denied microphone, a missing microphone, a dropped connection and a silent voice provider all said “unavailable”. That named neither the problem nor the remedy.
- Now each says its own thing: allow the microphone in your site settings; plug one in; the connection dropped, start again and nothing is lost. When the problem is the machine, the cursor moves to “Text input”, the way that does work.
- The Privacy tier also tells a recording it could not understand from a reply it could not read aloud. One costs you a sentence, the other only the sound.
v2.12.92026-08-18
The assessment waits for the confirmation
- Typing any address was enough to see the assessment in focus. Typed proves nothing, so the blur now lifts on the click in our email instead, and the CSV download follows the same rule.
- Anyone who already left an address is not asked for it again: the panel says the link is waiting in their inbox, and a button sends it once more.
v2.12.82026-08-18
Hearing the last sentence out
- When your voice allowance ran out, the voice was cut off mid-word and the email dialog appeared. The sentence explaining what was happening, and what the email is for, was the one thing lost.
- It now finishes that sentence before the conversation ends. If it gets stuck, the call ends after twenty seconds anyway.
v2.12.72026-08-17
The clarify action finally looks the way it was meant to
- Under “what is still missing”, every question trailed a grey box in the browser's own button styling, wedged mid-sentence. It is now a quiet ghost: just the words until you point at it, then a softly tinted pill.
- And it sits in the same place on every row instead of landing wherever the question happened to end.
v2.12.62026-08-17
Agreed to what, exactly
- The start button said “Agree and start”, and what you were agreeing to was only in the settings. It now sits under the button: AI conversation, and where the processing happens. The last word of the line opens the settings, where the full sentence is and where consent is withdrawn.
- The line always names the tier actually running: USA or EU.
v2.12.52026-08-17
A cough is not a turn
- A cough, a closing door or a passing car could leave an empty “…” bubble sitting in the transcript. Those now disappear again, including while you are in another tab, and they no longer reach the stored transcript.
- More importantly, the conversation no longer answers them. It used to hear a pause, take it for your turn, and fill it, usually with “are you still there?”. Now it stays quiet and keeps listening.
v2.12.42026-08-17
Around confirming your email
- The masthead said “email confirmed” while the right-hand column still asked you to confirm your email. That notice now goes the moment the confirmation lands, and the assessment that the spent allowance had blocked is asked for again straight away.
- While you are entering your address, the conversation knows what it is waiting for. Instead of “are you still there?” you hear at most one sentence about that step — otherwise it stays quiet.
- When you start speaking, the voice now drops by half as much. It used to sound switched off; now it sounds like it is giving way.
- The time an allowance resets is finally written in German rather than as “Aug 18, 2026, 7:02 PM”.
v2.12.32026-08-17
The language you chose stays chosen
- Change the language and it stays changed, including across “Reset completely”. Until now a reset deleted that choice along with the session, and the page came back in a language nobody had asked for.
- With no choice of your own, the page now opens in German whatever the browser is set to. English is one click away in the settings and is remembered once taken.
v2.12.22026-08-17
The consent lives in the settings
- The consent sentence stood on the entrance, with a link beside it to process elsewhere. Both are gone: the button says what you are agreeing to, and what that means in full is in the settings, right above the choice of processing location.
- That is also where you withdraw it, with the same tick box that records it.
v2.12.12026-08-17
The entrance down to four things
- The entrance now shows the orb, one line and the button. Below, set clearly apart, three quiet icons: text input, settings, start over.
- The explanatory lead and the welcome-back line are gone. The button already said what they said.
v2.12.02026-08-17
Product development too, and an off switch for the voice
- Anyone who wants to build a product rather than automate a process is no longer turned away. Both are our work; the assistant now understands what the project is instead of deciding whether it fits.
- The assessment recognises whether this is process improvement, product development or both, and passes that to the team.
- The consent step shows only the tick box and one button. Escape takes you back to the choice.
- New during a conversation: "Voice output" can be silenced with one click. The microphone keeps listening and replies still appear in the transcript; only the speaker goes quiet.
v2.11.32026-08-17
A tidy header on a phone
- On narrow screens the voice allowance bar sat on top of the conversation list. It has its own row now, as intended.
v2.11.22026-08-17
The way to a conversation no longer dead-ends
- "Arrange a conversation" sent anyone without an email on file back to the start page without a word. The click now asks for the address the proposal needs anyway, and carries on from there instead of starting over.
v2.11.12026-08-17
Honest about the allowance
- When the allowance ran out, the message said the budget of "this session" was used up. That was not true: the allowance belongs to you, not to the single conversation, and starting a new one does not reset it. That is what it says now, along with what actually helps.
v2.11.02026-08-17
Type in the middle of a conversation
- When the room is too loud or a word came out wrong, you can now write without hanging up: one click on "Or type" and the field is there.
- While you write, the microphone stays out of it and the assistant waits in silence instead of asking whether you are still there. Leaving the field hands the conversation back.
v2.10.62026-08-17
The conversation stops running away
- A longer reply pushed the transcript out of view for good: from then on every further reply stayed below the edge until it was scrolled by hand. The end stays visible now, and anyone who scrolls up themselves is no longer yanked back.
v2.10.52026-08-16
The assessment comes in again
- Anyone who spoke never saw the right-hand column at all: it only opened for typed turns. It now comes in as soon as you start talking, in both processing tiers.
v2.10.42026-08-16
Your own sentence is back on screen
- In the privacy tier, a placeholder was left standing where your spoken sentence belonged. It was understood correctly all along and the reply matched it, but you could not read it back. Now it says what was understood.
v2.10.32026-08-16
Measured apart, waiting cut short
- The wait until the first word is now measured separately from the provider's first sign of life. The two used to be one number under two names and could not disagree; since the split they sit a few milliseconds apart.
- A reply that never arrives no longer blocks the conversation for two minutes but for twenty seconds. The slowest measured reply so far took three.
v2.10.22026-08-16
Logs outlive the release
- Nothing on the surface, but it matters for everything after this: until now every release deleted the server's logs, because they belonged to the container being replaced. That is exactly why a fault reported the day before could no longer be traced. From now on they are kept.
v2.10.12026-08-16
A clock you can see coming
- In the last two minutes of a conversation, a quiet line says how much time is left. Before that it says nothing: a clock ticking for seven minutes rushes a conversation that has no reason to be rushed.
- When the conversation ends at that limit, it now says so instead of hanging up without a word: you can start talking again right away and everything said so far is kept.
- Ending it yourself still gets the plain goodbye, even seconds before the limit. The Privacy tier bills per turn and has no minute limit at all.
v2.10.02026-08-16
One orb, one sentence, one button
- The way in is the orb itself now: one sentence, one labelled button, nothing else. The empty assessment box waits until there is something to understand, then slides in from the right.
- The choice comes first and consent follows it. Until now both buttons were dead until consent was given — a first visitor's first click landed on nothing.
- Language, appearance and processing location now sit together behind the gear. The status dots only appear when something is actually wrong.
- During the conversation everything that does not support the sentence being spoken is gone: the headline, the duplicated button row, the keyboard hint before anyone types, and a footer with nothing in it.
- Everything that arrives or leaves moves while doing so — turns, files, notices, the understanding panel. Anyone who asked for calm interfaces gets the quiet state change instead.
v2.9.242026-08-16
Remove attached files again
- Every attached file can now be removed with one click. Its content leaves the conversation immediately, the assessment is rebuilt, and the assistant is told the file is no longer available.
v2.9.232026-08-15
Take back what was misheard
- Any of your own turns can now be removed. The assessment on the right is then rebuilt, and the assistant is told explicitly that the statement was withdrawn.
- What the assistant said stays: it is the record of what it actually answered.
v2.9.222026-08-15
The first pass really is quick
- The preliminary assessment now covers what you want to read right away: summary, signals and open questions. The detailed part follows with the thorough pass — and overwrites nothing you have already answered.
v2.9.212026-08-15
Answers without the detour
- Spoken replies no longer force a thinking pass that two short sentences never needed. Measured, that halves the wait again.
v2.9.202026-08-15
The assessment arrives in two passes
- Instead of waiting a minute for the assessment, you see a first pass within seconds. The page says openly that it is preliminary and replaces it shortly after with the thorough one.
v2.9.192026-08-15
The transcript stays tidy
- The transcript reliably follows along while you speak — and no longer yanks you back to the end while you are rereading something further up.
- While spoken words assemble, leftover word fragments no longer slip outside the bubble.
v2.9.182026-08-15
Speak faster, assess thoroughly
- The model that answers during the conversation is now separate from the one that writes the assessment. Replies arrive noticeably faster, while the assessment on the right still comes from the strongest model.
- If the fast model refuses once, the assistant quietly tries a second one instead of apologising straight away.
v2.9.172026-08-15
The engine picker responds again
- “Select” now works even before a conversation is running. Previously the button flickered and nothing happened, because a precondition was missing in the background and nobody was told.
- If saving does fail, the dialog now says so instead of staying silent.
v2.9.162026-08-15
Clarify now really asks
- Clicking “clarify now” genuinely steers the assistant to that point. In the first attempt the point arrived but the reply talked past it.
v2.9.152026-08-15
Raise an open point directly
- Next to every open point in the assessment, a “clarify now” link appears during the conversation. One click and the assistant takes up exactly that point next, without you having to phrase it yourself.
v2.9.142026-08-15
You can see again whether everything is reachable
- Three status dots are back at the top: language model, speech service and email confirmation. They now reflect the conversation you are actually having, not just a rarely used side path.
- If your address is still unconfirmed, a button to send the confirmation again sits right there.
v2.9.132026-08-15
The assessment now says what is going on
- When the assessment on the right takes longer or is unreachable, it now says so, together with a button to try again. Until now the panel simply stayed empty in those cases.
- If the conversation budget is spent, that is stated once honestly instead of being retried in the background over and over in vain.
v2.9.122026-08-15
Honest about the smaller budget
- The privacy page now states both halves: a conversation starts without confirmation, but the smaller spending limit can be reached. The conversation then ends with an audible note and can be continued after confirming, without losing what you told us.
v2.9.112026-08-15
Response times are finally measured
- Operationally: how long the assistant takes to answer is now recorded persistently instead of being lost with every rollout. That is the basis for switching to a faster model if needed. Nothing changes for visitors, and no conversation content is stored.
v2.9.102026-08-15
Unambiguous operating settings
- Operationally: settings defined twice are now detected and refused during rollout instead of quietly deciding which provider the privacy page names. Nothing changes for visitors.
v2.9.92026-08-15
The privacy page knows both routes
- The privacy page now also names Mistral AI as the provider behind the Privacy tier and explains that its global API route does not guarantee a processing location. It also states plainly that typing is no escape from processing: both paths run in the same session at the same provider.
v2.9.82026-08-15
No form without a limit
- Every input this site accepts is now hard-limited while it is read — even when a sender conceals its size. That protects availability for everyone else while somebody tries.
v2.9.72026-08-15
Your budget belongs to you
- The conversation budget now attaches to your email address only after confirmation, and to this browser alone before that. Entering somebody else's address reveals neither their usage nor lets anyone spend it.
- The internal counters for budget and confirmation emails are additionally protected with a server-side key, and the privacy notice now describes them as they actually work.
v2.9.62026-08-15
Call duration now has a hard limit
- A conversation reliably ends at its time limit, even when the browser stays silent about the ending: the limit is now also enforced on the server, and a drift in the voice platform fails the deployment instead of quietly causing costs.
v2.9.52026-08-15
Uploaded files cannot stall the server
- A maliciously built Word or Excel file could occupy the server for minutes and stall every conversation in progress. Reading such files is now fast in every case, and one request has a fixed overall time budget.
v2.9.42026-08-14
No asterisks in the ear
- When the language model slips into formatting characters, they are now removed before speaking and before display — spoken text stays spoken text.
v2.9.32026-08-14
Documents belong in the conversation
- Files can now go straight into the live conversation: attach them or simply drag them into the window. The assistant knows the content from the next sentence on, the assessment on the right includes it immediately, and an unreadable file is honestly marked as such.
v2.9.22026-08-14
A genuinely German voice
- The privacy tier now speaks with a natively German female voice (cloned from a public-domain LibriVox reading) instead of an English accent.
v2.9.12026-08-14
The Privacy voice speaks
- The privacy tier now has a voice: a female speaker who also answers in German. Without a chosen voice, no audio came out at all before.
v2.9.02026-08-14
The privacy tier becomes real
- The engine picker can now genuinely switch to Privacy: speech recognition, reasoning and voice then run directly through Mistral AI instead of the Quality tier's conversation providers. The configured global API route does not guarantee a processing location.
- The tier only becomes selectable once Mistral AI is actually configured on the server; until then it honestly keeps saying “coming soon”.
v2.8.82026-08-14
No break mid-sentence
- If the connection to the language model breaks mid-reply, the assistant now finishes the sentence itself with a brief apology and the conversation continues — previously the whole realtime connection dropped at that moment.
v2.8.72026-08-14
A tidy restart dialog
- The new-conversation dialog is reset: one-line buttons with room to breathe, the complete reset cleanly separated with its hint right beneath it — no more button squashed into a circle.
v2.8.62026-08-14
Every refund finds its account
- The per-minute refund always lands on the account charged at call start, even if the email was entered in between; multiple open refunds no longer overwrite each other, a closed tab files its claims on the next visit, and network errors are retried patiently too.
- Under the hood the assistant now measures its response latency, as the basis for the upcoming model decision.
v2.8.52026-08-14
Polish and purpose-bound keys
- The engine-choice bars carry their color by value: full percentages are teal, low ones shift to amber, both bars by the same rule.
- No orphaned three-dot line remains after a call ends, the email dialog keeps keyboard focus inside itself, and the masthead meter can no longer be repainted by a late response.
- The conversation context for the voice platform now carries its own purpose-bound signature and can never pass as a session credential.
v2.8.42026-08-14
The assessment always keeps pace
- The assessment now follows the persisted conversation state and never lags one answer behind.
- The blurred preview for anonymous conversations survives a page reload, a withdrawn rating clears its old display, only one empty state shows at a time, and the CSV download appears only with an email on file.
v2.8.32026-08-14
The email card never lies
- If sending the confirmation link fails, the dialog now says so honestly and offers to send again, instead of claiming a delivery; an already-confirmed address is greeted as such.
- The assessment pane shows which address the assessment goes to, with a change option right beside it.
- When the voice allowance runs out, the email dialog opens even if you had dismissed it earlier with later.
v2.8.22026-08-14
The link belongs to the human
- The confirmation link now leads to a small interstitial with a confirm button. Mail scanners that open links automatically can no longer verify your address unintentionally.
- The confirmation pages lead back to the conversation, the click no longer changes the conversation language, and a late-clicked link is still recognized when you return to the tab.
v2.8.12026-08-14
Every limit tells the truth
- The budget announcements now know all five states: a verified user's session limit and the monthly limit speak their own truthful line instead of wrongly talking about today.
- After two stumbles in a row the assistant reports a disruption honestly in exactly one sentence and ends the call, instead of endlessly asking you to repeat.
- The outdated reference to the classic version is gone from the error message.
v2.8.02026-08-14
Engine choice and a harder foundation
- New: the gear at the top right opens the engine choice. The Quality tier runs today (ElevenLabs + Opus, US); the Privacy tier and self-hosting are visible and honestly marked as coming soon.
- Hardening round under the hood: conversation turns are bound to an active call, the allowance meter no longer double-counts reservations, orphaned calls end themselves after 15 minutes, and the privacy page now tells the full truth about US processing for typing too.
v2.7.42026-08-14
Order, continuity, honesty
- A late assistant reply now lands above your words in progress, not below them: bubble order is right again.
- A follow-up call in the same tab keeps the visible history and greets you as a continuation instead of starting over.
- When you interrupt the assistant, its bubble now shows only what was actually said, not the unspoken remainder.
v2.7.32026-08-14
A stumble does not end a call
- If a single language-model reply fails, the assistant briefly apologizes and asks you to repeat, instead of the connection dropping.
- If a call dies while the email dialog is open, submitting now lands you back on the start card instead of an empty page.
- After email confirmation the card shows only the confirmation and keep-talking, without competing choices below.
v2.7.22026-08-14
The cost in parentheses
- The masthead bar shows in parentheses the platform credits this session's calls have actually consumed; the figure updates after each call ends.
v2.7.12026-08-14
Everyone pays only their own conversation
- Conversation turns were internally billed to a single account shared by all anonymous visitors. Once empty, the assistant reported an exhausted allowance to everyone, resets included. Every session now pays into its own account.
- Calls terminated by the platform also receive their per-minute refund.
v2.7.02026-08-14
The assessment behind frosted glass
- No dialog interrupts the conversation anymore. From your third sentence your real assessment forms behind frosted glass in the side pane, inviting you to sharpen it with your email.
- Entering the email lifts the blur in place.
- The email can now be entered in any conversation state, even late in the call.
v2.6.92026-08-13
Confirmed? Keep talking.
- When the email confirmation arrives while no call is running, the card immediately offers to keep talking: one click and the conversation continues with the large voice allowance.
v2.6.82026-08-13
The assessment keeps pace
- The assessment on the right now updates after every one of your statements instead of every second one — a clarification in your last sentence never reached the panel before, and answered questions stayed listed as open.
v2.6.72026-08-13
The refund actually lands
- The per-minute refund after a call did not wait long enough for the platform's record and was lost — it now retries patiently, if necessary on the next page visit.
v2.6.62026-08-13
Short calls cost short
- After every call the 0.70 $ flat charge is trued up against the real call duration (0.10 $ per minute) — three short calls no longer devour the whole voice allowance.
- In type mode Cmd/Ctrl + Enter now sends too; a subtle note below the box says so.
v2.6.52026-08-13
Honest announcements, certain ending
- The budget announcement is now true for every state: without an email on file it asks to leave one instead of claiming a link that was never sent; the daily limit says goodbye without demands.
- After the announcement the assistant ends the call itself (end_call) — on every limit path including the daily one; the browser-side safeguard now checks all ceilings server-side.
v2.6.42026-08-13
No more budget loop
- When the voice allowance is exhausted, the call now ends itself after the notice and the email dialog opens — instead of the same announcement looping forever.
- Long sentences in the bubbles wrap correctly again.
v2.6.32026-08-13
Email-first works
- The email can now be entered right after a complete reset or before the first conversation — previously that ended in an error (no-session).
v2.6.22026-08-13
Three rough edges fewer
- A call only starts when the voice allowance also carries a few replies — the limit notice comes before the call, not after your first sentence.
- The email dialog now also closes with Escape.
- The focus ring of the buttons under the conversation is no longer cut off on the left.
v2.6.12026-08-13
Word morph and correction glint
- Text changes in your live bubble now morph word by word: unchanged words stay put, corrected ones glide — on top of the light sweep and glow.
- When the final transcript genuinely corrected your live words, a glint flashes briefly in the bubble — only then.
v2.6.02026-08-13
Your words as you speak
- As you speak, your words appear live in your bubble — via a parallel transcription session at the same provider, with no new data sharing.
- When the conversation's final transcript arrives it replaces the live text with a light-sweep effect; your bubble stays put from the moment you start speaking.
- Your messages now carry their own color with a soft shadow, clearly separated from the assistant.
v2.5.82026-08-13
A second lane for the voice
- Voice calls now also work over the classic channel (?transport=websocket) — an escape hatch when a corporate network blocks WebRTC.
- Failed call starts now name their cause in the console; the groundwork for word-by-word live transcription is in place for the moment the platform delivers it.
v2.5.72026-08-13
Confirmation without switching tabs
- The confirmation link now opens only a short confirmation page — the conversation continues in the original window, which notices the confirmation on its own and shows it in the masthead.
- The assessment pane now explains that it needs an email address and offers the entry directly instead of staying empty.
v2.5.62026-08-13
The restart dialog everywhere
- New conversation opens the same dialog everywhere: masthead menu, conversations page, and right in the card after ending a call.
- Without a stored email the dialog offers the complete reset instead of silently doing nothing.
v2.5.52026-08-13
A clean restart for the stage
- New conversation now asks: continue normally, or reset completely (forgets email and conversation list in this browser).
- In presentation mode the masthead calmly shows unlimited voice allowance instead of the meter.
- Bottom button shadows are no longer clipped at the sides.
v2.5.42026-08-13
A goodbye is not an outage
- When the assistant hangs up after a goodbye, no error appears; if you ask for an appointment it points to the arrange button first.
- Your speech bubble stays put instead of vanishing, new turns glide in, the thread scrolls smoothly.
- The masthead meter is now called voice allowance.
v2.5.32026-08-12
Fewer words
- Test-version badge removed, opening copy tightened, duplicate privacy link removed from the card (the footer keeps it).
v2.5.22026-08-12
Lighter card, calm masthead
- The start card carries less text: one consent line with the privacy link, the microphone note lives inside the speak button.
- Masthead glitch fixed: switches sit firmly right, nothing stretches anymore.
v2.5.12026-08-12
Light, dark, or like the system
- New appearance switch top right: system, light or dark, remembered per browser.
v2.5.02026-08-12
The budget becomes visible
- New budget display in the masthead: remaining budget as a thin bar, your tier beside it (no email, email entered, confirmed) and the hint that confirming raises it.
- Masthead decluttered: step rail removed, language switch top right.
- Footer: notices as a plain link, subtle underlines that strengthen on hover.
v2.4.22026-08-12
Opaque panels, polite dismissal
- Version, notices and conversation panels are opaque again in dark mode.
- Incidental panels close on an outside click or Escape; the email dialog deliberately stays until you answer it.
v2.4.12026-08-12
Memory repaired
- Resumed conversations verifiably know their history again: the platform delivers our context key nested, the wiring read the wrong spot. Our own conversation rules now apply to every reply as well.
- Footer decluttered (nothing overlaps the privacy link), opening view of the left column recomposed.
- Note: the short-lived intermediate build still wore the 2.4.0 badge.
v2.4.02026-08-12
Quiet holds, visible version
- Pause and the email dialog hold the conversation through a real state machine: microphone verifiably muted, voice eased down, no artifacts in the transcript.
- Volume changes glide instead of jumping.
- This version badge with its history.
v2.3.02026-08-12
Proposal, understanding, cost ceilings
- The proposal step (arrange a conversation) is back, briefing mail included.
- The understanding pane shows the readiness meter, open and settled items, and the CSV export.
- Call and model spend count against the classic budgets; an exhausted budget says so in the conversation itself.
v2.2.02026-08-12
Own voice, own thinking
- Replies come through our own model wiring with this house's conversation rules.
- The email question is a calm dialog inside the conversation, resuming works from the confirmation link, hanging up works by saying so.
- Short consent wording with details one link away; AI transparency on the privacy page.
v2.0.02026-08-12
The conversation becomes real
- A real spoken conversation with barge-in, instant transcript, and typing as an equal path.
- The classic version is archived; every path leads to this surface.