1.2.0 — Public Beta

The tuning table, and a panel to use it: the numbers that decide how a
workout feels are now visible, adjustable, and you can watch the effect
while you change them. Plus four defects found during a real VR workout.

Added

• Tuning is a feature, not a debug hatch. It sits behind the Advanced
  button in the workout preview, follows the app language like everything else
  — all 25 value names, descriptions, block titles and provenance tags — and
  has its own guide: docs/tuning.md.

• Scroll the track with the mouse wheel over the preview. The step is
  musical, not a fixed time: one beat per notch, a quarter of a beat with
  Shift — because what you are looking at is the distance between two hits,
  and that distance is measured in beats everywhere else in the program.

• Space pauses the preview, as in any player — but never while you are
  typing in the panel, where a space is a space.

• Two numbering schemes. VERSION_WEB is what someone downloading sees and
  moves only when a release goes out; BUILD is the internal counter. In two
  days the public number had burned through five versions nobody ever saw.
  rilascia.py promotes an internal build to a public release.

• Tuning panel. Open a track's workout preview and press Advanced (or
  Ctrl+Shift+T): 25 adjustable values appear as sliders beside the preview.
  Move one and the choreography is rebuilt — 33 ms measured — so you see the
  effect instead of imagining it. Nothing is written to disk until you press
  Save. Full guide: docs/tuning.md.

• Every value carries where it came from — measured on the game,
  confirmed in VR, or chosen and never verified — and a range that refuses
  the absurd with a message saying why, rather than ignoring it in silence.
  This is what makes the panel safe to hand to anyone: the guardrails are in
  the values, not in hiding the panel.

• Values grouped by what they affect, in three collapsible blocks: on every
  kind of workout, only on what the tool generates, only on the hits you throw.
  The grouping was measured, not deduced from reading the code — each value
  was pushed to an extreme and the result compared across all three modes.

• A reset next to each slider, not only one for everything: while looking
  for one value you move five, and when one has gone too far you want that one
  back without losing the other four.

• Methods: a named setting, in a tuning/ folder next to the executable,
  picked from a dropdown. Import and export go through the system file dialog,
  and each method carries who made it, what for, when, and with which version —
  a bag of numbers from a stranger does not tell you whether it still applies
  to the build you are holding. On import you are told what your version does
  not recognise, rather than having it applied half-way in silence. A method
  that holds up in VR can ship with a release.

• confronta_tuning.py: regenerates the same track with and without your
  tuning and puts the measurements side by side, averaged over several
  generations. A table with no way to see the effect does not remove the
  guesswork — it moves it from whoever writes the code onto whoever uses it.

Fixed — what using the panel turned up

Every one of these was found by driving the sliders and watching the preview,
which is the whole reason the panel exists.

• The rule now starts from the positioning. Until now the only answer to
  "these two do not fit" was to shift the move or drop it. Now:

  is there room for two SWINGS?    yes → keep them    no ↓
  they become two STRAIGHTS
  is there room for two STRAIGHTS? yes → keep them    no → the second one goes
  

  Same principle already used for squat chains — the instant stays, the move
  changes — brought to where it mattered most. Measured with dense markers:
  0 violations across four combinations of breathing room and exemption,
  against 19 before.

• Raising the breathing room removed the wrong punches. With the full
  exemption on your markers, the breathing rules did not touch your own hits:
  raising the value deleted 12 automatic punches — which are not the problem —
  and left 13 close pairs, all of them yours. The exemption now defaults to
  applying to straight punches but not to hooks and uppercuts, which are wide
  movements you cannot throw back to back. Measured cost with realistic
  markers: none, 115 of 115 markers kept.

• The exemption only looked at the first of the two hits. A hook right
  after a straight got through, and no slider could push it away. Verified in
  runtime. It now looks at both: a wide movement counts whether it comes first
  or second.

• There was no breathing room before a swing. We only looked at the hit
  that came first, so hook → jab asked for the swing rule while
  jab → hook asked for the ordinary one. The official repertoire makes no
  such distinction: a full beat minimum in all four cases (before a swing, 8
  cases on the same arm and 141 on the other; after, 4 and 154). New value,
  breathing room before a swing.

• A separate breathing room for your own hits. 1.225 works well between
  two hits you marked, but imposing it on everything made two figures in the
  euclidean catalogue unplayable — tresillo and wide cinquillo sit at exactly
  1 beat, as the official repertoire does. The automatic layer builds on the
  game's own figures; your markers do not.

• The preview refused to rebuild on already-generated tracks. Those are
  exactly the ones the dropdown lists, and at startup none of them has an
  analysis in memory, so every slider did nothing. The analysis is now computed
  on demand when you move a slider.

• In the visualizer, every block was drawn with a squat under it. A block
  drew a bar — the same shape used for a squat — plus the shield icon just
  below it, so it read as squat + shield always, and the two separated as
  they approached. There were zero blocks simultaneous with a squat in the
  data: it was not a squat, it was its shape. A block on its own now draws only
  the shield; the bar stays for the squat+block combo, where it is real.

• The track name is gone from the playback bar. It repeated what the
  dropdown above already said and cost up to 260px — exactly what the position
  slider needs, since its precision is the one thing there that depends on
  length.

Fixed — found during a real VR workout

• Low punches did not exist. Not rare: zero, on every preset and in
  every mode, against 4.3% in the game itself. No hooks to the body, no low
  blocks. They now follow the proportions measured on the official repertoire,
  verified across twelve generations: low hooks at 8.8%.
• Punches came too soon after a hook or an uppercut. 44–50% of the moves
  that followed one landed under a beat away, with minimums of 0.00 — an
  uppercut and a jab at the very same instant. The 1.5-beat rule only applied
  to the same arm; on alternating sides half a beat was enough. It is now 0.75
  beats, i.e. +125 ms at 120 bpm, and moves under a full beat dropped to 6%.
  Deliberately less than the full beat the game uses: copying that would make
  our choreographies as sparse as its own.
• A high block arriving next to a squat. Reported twice, six weeks apart.
  The rule turns out to be deterministic, measured without a single
  exception: a block paired with a squat is low 12 times out of 12, a block on
  its own is high 44 out of 44. You crouch and block low, or you stand and
  block high — a high block while crouched is a position that does not exist.
  A first attempt used a random 21% quota: the right number with the wrong
  mechanism, since drawing at random produces exactly the two things the game
  never does.
• "No obstacles" only worked halfway. It reached your own markers but not
  the automatic layer, so its squats all survived: 11 measured in Harmonize and
  5 in "Markers only", with a chain of eight. They now go, and the block+squat
  combo goes with them.
• Squat chains had no limit. Measured up to eight in a row; the official
  repertoire reaches sixteen but with a median of one. Past the limit a squat
  becomes a punch: the instant stays, the move changes.
• Extend learned where you punch, not how much. Marking two hits per beat
  is 241 punches/min while the most intense preset targets 88 — a 2.7× gap,
  pulled back to the ceiling every time. That is why it "could not learn
  punches in quick succession". It now follows your cadence when you set a
  faster pace than the preset.
• An uppercut and a hook on the same arm, almost touching. Your own markers
  are exempt from the spacing rules — that is what makes "your marker always
  wins" true — but the exemption was all-or-nothing, so nothing could separate
  them. How wide it goes is now yours to choose, because the two requests in
  play are both legitimate and in conflict: let me pack my own hits closer
  against these two are too close together.

Fixed — the same track now gives the same choreography

Generating a track twice, with the same markers, produced two different
workouts: the seed came from the clock. The purely automatic layer never had
this problem — it derives its seed from title|artist on purpose — but the path
that uses your markers did not honour it. Nobody had noticed, because there is
no reason to generate the same track twice, until a panel arrives that does it
twenty times in a row. Moving a slider now changes only what that slider
governs: 236 of 278 instants stay untouched instead of everything
reshuffling.

Fixed — in the tools themselves

• The panel never regenerated anything. A cache one level below the reque...

1.1.2 — Public Beta

The first public build. Download the .exe below, double-click, and it runs — no installer, and it changes nothing until you ask it to.

What you need

• Windows and the Microsoft Edge WebView2 Runtime (normally already there on Windows 10/11; if it is missing the app says so and points you to it).
• BoxVR installed, if you want to install what you make straight into the game.
• Your own MP3s. The toolkit ships no music.

What it does

tool	what it does	needs the patch?
Fix	re-reads the audio of a song the game already imported and repairs the wrong intensity levels	no
Generate	builds a brand new track from an MP3, with full coverage and no silent gaps	yes
Playlist manager	rename, reorder and clean up already installed playlists	no

Inside Generate there is also Sidecar: tap out the punches you want yourself while the track plays, and the tool builds the workout around yours — keeping only yours, weaving them into the generated ones, or learning your cadence and rhythm and carrying them through the rest of the track.

Read this before the patch

Generate writes explicit choreography, and the game only reads it with a small modification to one game file. The patch is reversible, takes a backup, and Steam's "Verify integrity of game files" undoes it too. It touches nothing to do with anti-cheat, networking or purchases. A game update overwrites it and it has to be reapplied. See docs/patch.md for exactly which bytes change and why.

The move repertoire the generator uses is FitXR content and is not distributed here: it is extracted from your own copy of the game when you apply the patch.

Where things are written

Nothing is written into the game library without an explicit confirmation, and existing files are backed up first. Generated tracks live in a working folder and are cleared when the tool closes — once you have installed them in BoxVR, the copy that counts is the one in the game.

Known, and honest

• Three tests that open a desktop window are flaky, and one choreography test is red. They are in the repository, red, rather than hidden.
• Sidecar is marked beta for a reason: it works, it has been played in VR, and it is the newest part.
• This is a beta. It writes to your game library, with backups, but it is a beta.

Licence, and how it was written

GPL-3.0. Unofficial modding project, not affiliated with FitXR.

Most of this code was written by a language model, working with the author over several months. docs/ai.md says what that means in practice — including what went wrong and how it was caught.