FABLED FLOW FOR WINDOWS — early test build (win-v3.5)
=======================================================

Thank you for testing! This is an early Windows build of Fabled Flow,
a dictation tool where everything runs on YOUR computer — no cloud, no
audio ever leaves your machine.

WHAT YOU NEED
- Windows 10 or 11, 64-bit
- A microphone
- About 1.5 GB of free disk space (the app + its speech model)
- Internet for the first launch only (model download, ~640 MB)

HOW TO START
1. Unzip this whole folder anywhere (e.g. your Desktop).
2. Double-click FabledFlow.exe.
   * Windows will likely show a blue "Windows protected your PC"
     screen because this early build is not yet code-signed.
     Click "More info", then "Run anyway".
3. On first launch a small window downloads the speech model
   (one time, ~640 MB). When it says Ready, you're set.

CHINESE (中文) SUPPORT
- Right-click the Fabled Flow tray icon and choose
  "语言：中文（简体）/ Chinese". The Chinese speech model downloads
  once (~230 MB), then Chinese dictation just works.
- In Chinese mode the recognizer auto-detects Chinese AND English —
  you can mix both freely in one dictation.
- Switch back to "Language: English" any time from the same menu.

HOW TO DICTATE
- Click into any text box (Notepad, a browser, chat, email…).
- HOLD the RIGHT CTRL key, speak, and let go.
- Your words appear at the cursor a moment later.
- A small dark pill at the bottom of the screen shows a live
  waveform while you speak.

WHAT THIS TEST BUILD DOES
- English dictation with automatic punctuation and capitalization
- Filler-word removal ("um", "uh")
- Smart number formatting: phone numbers, thousands separators,
  currency ($100), counts ("one hundred lights" → "100 lights"),
  decimals ("one point oh" → "1.0"), percentages, times, fractions
- Dotted addresses and versions: "46 dot 17 dot one seventy five
  dot one twenty" → 46.17.175.120, "one point two point three" →
  1.2.3, "seven point x" → 7.x
- A built-in abbreviation lexicon: spelled-out letters glue into
  known terms ("h t t p" → HTTP), tech forms convert ("three D" →
  3D, "four K" → 4K), and famous casings are fixed (iPhone, GitHub)
- Spoken web and email addresses: "example dot com" → example.com,
  "john dot smith at gmail dot com" → john.smith@gmail.com;
  multi-word names join and lowercase ("Crown Faith Alliance dot
  com" → crownfaithalliance.com)
- Extra safeguards against invented words on very short recordings
- OPTIONAL sign-in (right-click the tray icon → "Sign in…"): links
  this PC to your account by email code. Dictation never requires it.
- Keeps a local history of everything you dictated
  (%APPDATA%\FabledFlow\history.json)

WHAT IT DOESN'T DO YET (coming in later builds)
- Languages beyond English and Chinese, custom dictionary, AI document formatting,
  settings, colorways — the Mac version has these; Windows will too.

KNOWN HONEST CAVEAT
This build was produced on a Mac (the app is built cross-platform)
and has passed its full automated test suite, but it is young on real
Windows machines. If ANYTHING misbehaves — no text appearing, wrong
key behavior, crashes — please tell the person who sent you this,
ideally with the file %APPDATA%\FabledFlow\debug.log attached.

PRIVACY
- Your voice is processed entirely on this computer.
- No audio and no text ever leaves your machine.
- The app makes exactly two kinds of network requests: the one-time
  model download, and an anonymous launch counter (app version +
  Windows version — nothing personal, no content).

To uninstall: quit Fabled Flow (right-click the tray icon), delete
this folder and %APPDATA%\FabledFlow. That's everything.


WHAT'S NEW IN 3.5
- Long dictations paste almost instantly now: the app transcribes IN
  THE BACKGROUND while you speak, so letting go of the key only has
  the last few seconds left to finish — a ten-minute dictation lands
  about as fast as a short one. During long recordings the pill shows
  a small timer and word count ("12:07 · 1,842 words"). Hours-long
  dictations no longer pile up in memory.
- Power-cut and crash protection: while you dictate, everything
  needed to rebuild your words is continuously saved to disk. If the
  PC loses power, the app is killed, or it closes mid-dictation, the
  NEXT launch recovers the dictation into History, honestly marked
  "Recovered". Closing the app normally cleans up after itself.
- A second listen where it matters: the places where long-dictation
  audio windows meet are re-heard by an independent second decode and
  the two readings are adjudicated — the seam-error class is caught
  automatically. Repeated-word sites get a second ear too, so a
  genuinely spoken "pace, paste" is never collapsed into one word.
- Mixed Chinese/English dictations: each part is now recognized and
  formatted under its own language's rules, including when a crash
  recovery rebuilds them.
- Number echo fix: when the engine writes one spoken number twice
  ("take 18 eighteen seconds"), the duplicate collapses ("take 18
  seconds"). Genuine number series are never touched.
- Spoken years come out as years: "eighteen eighteen" → 1818,
  "nineteen oh five" → 1905, "nineteen eighty-four" → 1984. Counting
  ("seventeen eighteen nineteen") and lists are never touched.
- Modern words like "standalone" are never split apart any more —
  the app now also consults Windows' own spell checker.

WHAT'S NEW IN 3.4
- The app now looks and works like the Mac version's dashboard:
  a "My account" button lives at the bottom of the sidebar and opens
  a proper account page — see which devices are signed in to your
  account, sign any of them out, or sign this PC out.
- Report a problem in one click: any History entry that still has its
  recording gets a "Report a problem" button next to Play. It opens a
  pre-filled report; the recording and text are sent ONLY when you
  press Send, and the entry then shows "Reported".
- History entries are never cut off any more — cards grow to fit the
  whole dictation (only truly enormous ones start folded, with a
  "Show the whole dictation" button).
- Pop-up messages now stay on screen long enough to read — longer
  messages get more time.
- Colors matched to the Mac's current look (the light "Default" style
  is now the same warm beige; the dark style's black pill is cleaner).

WHAT'S NEW IN 3.3
- Faster pastes in WhatsApp-class apps: some text boxes accept our
  direct text write but never report it back, and the app used to
  wait 1-2 seconds for that report on every long paste. It now
  remembers those boxes (per app and box type) and goes straight to
  the reliable paste path — your text lands about as fast as the
  words are recognized. The memory heals itself if an app starts
  reporting honestly again.
- Play your dictations: History now has a Play button on recent
  entries — hear exactly what the microphone recorded (the app keeps
  the last 30 recordings on this PC only; older entries simply have
  no button). Deleting an entry deletes its recording too.
- Correct Chinese word counts: a Chinese dictation used to count as
  "3 words" because Chinese isn't written with spaces. Every Chinese
  character now counts as one word (Japanese and Korean too), so
  your word totals and words-per-minute are finally honest. English
  counts are unchanged.
- The "Show what the engine heard" text in History is gone — the
  Play button tells that story better.

WHAT'S NEW IN 3.2
- Dictated text now lands in modern app text boxes (WhatsApp, the new
  Notepad) DIRECTLY through Windows' accessibility channel, with the
  result read back and verified — no more pressing Ctrl+V yourself.
  Where a box doesn't support the direct channel, the keystroke path
  remains, now paced like a human keypress.
- Chinese ID and phone numbers read digit-by-digit now come out as
  digits: "3210271995幺零零幺 2447" becomes 321027199510012447.
  Everyday Chinese phrases are never touched.

WHAT'S NEW IN 3.1
- Long dictations can no longer lose their endings: audio past 22
  seconds is heard in overlapping windows, stitched without losing a
  word.
- Invented trailing words can no longer ride a rescue: any recovered
  ending must be heard identically by two independent listens.
- Misheard non-words heal to the intended word ("mid-ditation" becomes
  mid-dictation); made-up brand words are left alone.
- A paste can never be silently marked delivered while you type: the
  app now verifies YOUR dictated text is actually in the box.
- "Log in in a terminal" style phrasing is no longer collapsed.
