Anneal 5.2.0

5.2.0 — 2026-10-20 14:34

New

• The auto EQ now updates 300 times a second in every latency mode (you asked: "make it 300/s
  instead of 94/s"). Before, Normal and Precise updated 94 times a second and Low 188. The EQ follows
  the music in finer steps; how fast it reacts is set in seconds, so its timing and character are
  the same, only smoother in time. 5.1.1 is untouched; this is a new version.

• At every sample rate the step is the sample rate / 300 (160 samples at 48 kHz, 147 at 44.1 kHz).
  160 doesn't divide the frame sizes, so the overlap-add now scales each position in a step on its
  own; with neutral settings the output is still the input exactly (within 0.0000002), in every
  mode and block size.

• Latency is unchanged in every mode (it depends on the frame size, not the step).

• CPU: more updates cost more. At 48 kHz, against 5.1.1:

  • Stereo: Low 1.1 → 2.0 %, Normal 0.8 → 2.8 %, Precise 1.3 → 4.5 % of one core;
  • Mid/Side: Low 2.2 → 3.9 %, Normal 1.5 → 5.2 %, Precise 2.2 → 7.8 %;
  • Zero latency: Stereo 2.4 → 4.1 %, Mid/Side 3.9 → 7.1 %.

  The slowest 0.1 % of 64-sample blocks take at most 21 % of the deadline (Mid/Side Precise), so it
  stays safe in real time even at tiny buffers.

Fixes

• None (a direct request, not a daily run).

Tests: 156/156 acceptance checks, 3 new: 300 updates/s in Low, Normal and Precise at 44.1 / 48 /
88.2 / 96 / 192 kHz; neutral output equal to the input with a 160-sample step in every mode; Precise
at 48 kHz running 300 frames/s (was 93.75). 23/23 host checks on the built VST3, setup checks pass.

Anneal 5.1.1

New

• Faster engine, part 2 (the CPU item on the roadmap): the hearing model now converts its levels
  to and from dB in batches, four values at a time, with Anneal's own fast log and exp (no tables, no
  new libraries). Before, it called the standard log10 / pow one value at a time, about 5,000 times
  per frame. The sound is unchanged: over 5,031 frames of music in Low and Precise, with Reveal,
  Smooth, Tilt, Sidechain (Make room and Unmask), Reference match, Guard and Punch all on, the curve
  stays within 0.000004 dB of 5.1.0. The display levels stay within 0.00001 dB.

• In full Mid/Side the stereo harmonic generator no longer runs. Its harmonics were always replaced
  by the Mid and Side ones, so it was wasted work; the sound is the same.

• Measured in the same session on this machine (AnnealTests profile, per STFT frame):

  • the model: Low 30.9 → 21.6 µs, Normal 33.3 → 22.0 µs, Precise 38.5 → 24.4 µs;
  • the display levels: 4.7 → 2.4 µs (three per frame);
  • the whole stereo engine: Low 73 → 57 µs, Precise 163 → 141 µs.

• Mid/Side is now under the 3 % budget in every latency mode. At 48 kHz, 64 samples (Normal at
  512), against 5.1.0's run:

  • Mid/Side: Low 3.7 → 2.2 %, Normal 2.5 → 1.5 %, Precise 3.6 → 2.2 %;
  • Stereo: Low 1.7 → 1.1 %, Normal 1.2 → 0.8 %, Precise 1.8 → 1.3 %;
  • Zero latency: Stereo 3.3 → 2.4 %, Mid/Side 5.6 → 3.9 %.

  The slowest 0.1 % of 64-sample blocks: Stereo Low 5 %, Precise 14 %; Mid/Side Low 10 %, Precise
  20 % of the deadline (5.1.0: 7 / 15 / 15 / 29 %).

Fixes

• None found. The bug pass looked at the engine, the zero-latency filters (5.1.0's newest code), the
  processor's host paths (prepare, reset, latency and Zero switches, state), yesterday's build log
  and benchmark, and the interface snapshots.

Tests: 153/153 acceptance checks, 3 new:

• the fast dB conversion against 10·log10 from 1e-20 to 1e30 (largest error 0.00003 dB), with SIMD
  and as plain loops (what non-x86 builds run), including 0, negative, NaN and infinite inputs;
• the fast gain conversion against 10^(dB/20) from −200 to +200 dB (largest error 0.000013 dB);
• the model's curve within 0.001 dB of 5.1.0's math (measured: 0.000004 dB).

23/23 host checks on the built VST3 and 25/25 setup checks pass.

Anneal 5.1.0

5.1.0 — 2026-10-18 20:34

New

• Zero latency (Latency menu → Zero 0.0 ms): Anneal with no delay at all, for tracking,
  monitoring and live playing. The hearing model still analyses the sound in Low-latency frames, but
  its curve is applied through minimum-phase filters, so nothing waits for a frame. A new filter
  is designed from the curve every ~5 ms and crossfaded in over ~1.3 ms.
  • Works at zero latency: Reveal, Smooth, Tilt, Response, Range, Guard, Freeze, Sidechain,
    Reference, Paint, Match, Mix, Delta, Width and Mono bass. Width and Mono bass are built into the
    Side filter.
  • Rests at zero latency: Heat writes its harmonics into the FFT frames, Punch has to boost an attack
    as it happens, and the Listen solos are made from the frames, so these three wait for Low or
    more. The Heat knob says "off at zero latency", and the display shows a ZERO LATENCY tag.
  • The curve follows the music about one frame (~5 ms) later than in the latency modes. A
    minimum-phase EQ also shifts phase where it boosts or cuts, so at Mix below 100 % the dry blend
    can colour the sound a little.
  • The Ceiling has no look-ahead here: it holds sample peaks exactly but can't catch peaks between
    samples.
  • Measured:
    • reported and applied latency 0 (also through the built VST3);
    • neutral settings pass the input within 0.0000001;
    • Smooth cuts a sustained 3150 Hz ring by 8.9 dB, the same as Low;
    • Heat writes nothing (2nd harmonic −191 dB);
    • Width 0 % is mono (sides −183 dB), Width 200 % lifts the sides by 6.02 dB, and Mono bass at
      200 Hz takes 60 Hz sides to −34 dB while 1 kHz stays within 0.01 dB;
    • the filter design matches its target curve within 0.01 dB, with all its energy in the first
      quarter of the response, as a minimum-phase filter should;
    • the Ceiling at −1 dB holds a signal 6 dB over full scale at −1.00 dB.
  • Off by default. It's a separate setting, "Zero Latency", so the Latency menu's other choices and
    every saved project are unchanged; projects from earlier versions open with it off. Not
    automatable: the host re-aligns tracks when latency changes.

Tests: 150/150 acceptance checks, 8 new:

• the filter design and minimum phase;
• latency 0 with the input passed through;
• Smooth cuts as in Low;
• Heat rests;
• Width and Mono bass;
• switching Zero and the latency modes while playing, including Mid/Side;
• the Ceiling without look-ahead;
• older projects open with it off.

23/23 host checks on the built VST3, 1 new: Zero latency through the host, with latency 0 and the
impulse in place at Mix 0 % and 100 %, then back to 1091 samples. Setup checks pass.

CPU at 48 kHz / 64 samples: Zero latency Stereo 3.2 %, Mid/Side 5.6 % of one core, against Low
1.7 % / 3.7 %. The filters cost about 1.5 % of a core. With Zero latency off, everything is as in 5.0.0.

Anneal 5.0.0

5.0.0 — 2026-10-18 20:10

The third major update: a SHAPE group for the punch and the stereo image of the sound, and your own
presets. With the new controls left alone, everything sounds as in 4.0.2.

New

• SHAPE (new group, rose, between REFERENCE and OUTPUT):
  • Punch −100 … +100 % (Off by default): a spectral transient shaper. In the bands where an
    attack happens it makes the attack louder (+) or softer (−), up to 8 dB, and leaves the body of
    the sound alone. It uses Guard's per-band attack detector but follows only the attack itself:
    held while the detector holds the attack, gone within about 15 ms after. It works inside Range and
    Paint like the rest of the auto EQ, and in Mid/Side it follows the EQ routing. It's sharpest in Low
    latency; in Precise, the 85 ms frames hold an attack longer. Measured on 2 kHz plucks in Low:
    • attacks (first 15 ms): +7.9 dB at +100 %, −7.9 dB at −100 %;
    • the body (80–160 ms): moves 0.01 dB;
    • with Range at 5–20 kHz the 2 kHz plucks are untouched.
  • Width 0–200 % (100 % by default): the stereo width of what Anneal outputs, set per frequency
    bin, so it's linear phase with no crossover filters. 0 % is mono; 200 % doubles the sides
    (+6.0 dB). The centre stays where it was (within 0.01 dB), and mono tracks are left alone.
  • Mono bass Off / 20–400 Hz (Off by default): everything below the frequency becomes mono,
    fading in over a quarter octave either side, so the low end stays centred and mono-safe while the
    rest keeps its width. Measured:
    • sides at 60 Hz: −85 dB with Mono bass at 200 Hz (Low), −122 dB at 120 Hz (Precise);
    • sides at 1 kHz: unchanged (0.00 dB).
  • Width and Mono bass shape everything you hear, harmonics and Listen solos included. Mix blends
    them with the dry signal like the rest.
• Attack marks on the display (roadmap item 11): with Guard or Punch on, short rose bars at the
  foot of the EQ lane show the bands where an attack is happening right now (taller = stronger).
• Your own presets. The preset menu lists them under YOUR PRESETS, after the factory ones (the
  arrows step through both), and has three new commands: Save current settings as…,
  Delete “…” and Open the presets folder.
  • A preset holds what the factory presets set: Reveal, Smooth, Tilt, Response, Range, Heat, Color,
    Focus, Mix, Output, Flavour, Punch, Width and Mono bass.
  • Each is a small XML file, %APPDATA%\Qulic\Anneal\Presets\<name>.annealpreset, shared by every
    project and every instance. Saving under an existing name replaces that preset; deleting sends
    the file to the Recycle Bin. Characters Windows doesn't allow in names (\ / : * ? " < > |)
    become spaces.
  • The bar shows the preset's name, with * once you change something. Loading one is one undo
    step. A project remembers which of your presets it was set from, as long as the file is still there.
  • The DAW's own preset list (host programs) still holds the 8 factory presets.

Changed

• Wider window: 1464 × 640 (was 1312 × 640) for the SHAPE group, resizable from 1098 × 480 to
  2196 × 960.
• Factory presets also set Punch off, Width 100 % and Mono bass off, so each sounds as before.
  Projects from earlier versions open with those settings too.

Tests: 142/142 acceptance checks, 22 of them new:

• Punch: louder and softer, body untouched, inside Range only;
• Width: 100 %, 0 % and 200 %, with Heat working, through the whole processor, and mono tracks
  untouched;
• Mono bass in Low and Precise;
• typed values and read-outs, older projects, factory presets;
• user presets: save, load and undo, remembered by a project, broken and older files, safe names,
  the preset bar, delete;
• the SHAPE layout.

22/22 host checks on the built VST3, 1 new: Punch, Width and Mono bass through VST3. Width 0 %
leaves the sides 107 dB under the signal, and all three were swept while playing. 25/25 setup
checks.

CPU at 48 kHz with the new controls off, as in 4.0.2:

• Stereo: Low 1.7 %, Normal 1.2 %, Precise 1.8 %;
• Mid/Side: Low 3.7 %, Normal 2.5 %, Precise 3.6 %.

With Punch, Width and Mono bass on, Low and Precise at 64 samples measured 1.4 % and 1.6 %, against
1.8 % and 1.6 % without them in the same run: no measurable cost. Linux was not rebuilt on this PC;
the Linux workflow runs on GitHub for the push.

Anneal 4.0.2

4.0.2 — 2026-10-18 19:39

Lighter on the CPU: Anneal now runs its spectral transforms on its own FFT, about four times faster
than the one it used, so every mode costs less. The sound, the latency and your projects are
unchanged.

New

• Faster engine (the first part of the CPU item on the roadmap). Anneal's own FFT replaces
  JUCE's portable one, which Windows and Linux builds get: 6.6 µs instead of 24.7 µs per 4096-point
  transform, and as accurate (within −133 dB of the old one, as close to an exact calculation).
  • CPU at 48 kHz on the build machine, of one core: Stereo Low 2.0 → 1.7 %, Normal 1.8 → 1.2 %,
    Precise 2.7 → 1.8 %. Mid/Side Low 3.7 → 3.6 %, Normal 2.8 → 2.4 %, Precise 4.3 → 3.6 %. With
    Reference Match and Paint on: Low 2.2 → 2.0 %, Normal 1.9 → 1.3 %, Precise 2.8 → 2.0 %.
  • Slowest 0.1 % of 64-sample blocks: Stereo Precise 25 → 14 % of the deadline, Mid/Side Precise
    36 → 29 %, Mid/Side Low 19 → 15 %; Stereo Low stays at 8 %.
  • Reference files are analysed faster too.
  • Mid/Side Precise is still above 3 %: the perceptual model is now the largest cost there (three
    models per frame). That is the next part of the item.
  • For development: AnnealTests profile shows where one frame's time goes.

Fixes

• Freeze survives the host preparing Anneal again. Hosts do this when you change the buffer
  size or the sample rate, and many when an export starts (a VST3 plug-in is prepared every time
  it's activated). Until now a frozen curve was thrown away there and re-made from the first moment
  of audio afterwards, so an export could sound different from what you had frozen. Now it holds,
  also when the rate changes (44.1 kHz and up). Measured: a −9.0 dB cut stays −9.0 dB after a new
  buffer size and after 48 → 44.1 kHz; through the built VST3, −8.9 dB before and after the host
  deactivates and activates it (4.0.0: back to 0.0 dB).
• Loudness Match keeps its gain when the host prepares again. Before, it started over at 0 dB
  and swelled to the matched level over half a second, for example at the start of an export. Now it
  carries on from what it had measured (−0.65 dB stays −0.65 dB).

Tests: 120/120 acceptance checks (new: the FFT against JUCE's and an exact DFT, and both fixes,
which failed before them), 21/21 built-VST3 host checks (new: Freeze through deactivate/activate,
which fails on 4.0.0), 25/25 setup checks. The Linux build was not re-run for this update; the new
code is plain portable C++ with no new libraries.

Anneal 4.0.1 — Linux support

Anneal 4.0.1 adds verified Linux x86-64 builds and fixes Linux installation and native editor testing failures.

• Linux downloads include VST3, CLAP, LV2 and standalone, plus install.sh. Extract the ZIP and run bash install.sh from the extracted folder. Built on Ubuntu 24.04; requires compatible glibc 2.39/libstdc++, ALSA, FreeType/fontconfig and X11 or XWayland.
• Fixed headless Linux editor initialization (BadAtom), native editor teardown (BadWindow), and incomplete Linux workflow inputs.
• Linux upgrades now stage replacements and preserve the installed plugin if copying fails.
• Windows VST3 and standalone setup is included.

Verification: 112 acceptance checks, 20 actual VST3 host checks, 10 CLAP checks and 23 packaging checks passed on native Ubuntu. LV2 discovery/metadata and runtime dependencies passed; LV2 audio in a DAW is not covered. 24 Linux editor snapshots were rendered and representative layouts inspected. Windows passed 112 acceptance, 20 VST3 host and 25 isolated installer checks. macOS remains unverified.

Free public downloads and screenshots: https://www.kvraudio.com/product/anneal-by-qulic/downloads

The Linux ZIP contains the tested native build from run 38837427218. The final source revision adds documentation only after that build. The release workflow will additionally attach a Linux tar.gz and checksum after rebuilding this tag.

Anneal 4.0.0

The second major update. Anneal can now match the tone of a reference, and you can draw where it
works. With the new controls left alone, everything sounds exactly as in 3.1 (tested: bit-identical).

New

• Reference match (new REFERENCE group, teal). Give Anneal a reference and it steers its auto EQ
  toward that reference's long-term tone:
  • Drop an audio file (wav, aiff, flac, mp3, ogg) on the display, or use Load a reference
    file… in the Reference menu. It's analysed in the background, never in the audio thread, and
    Reference switches to Match. The display outlines itself in teal while you drag a file over it.
  • Learn key: learns the tone of whatever plays into the sidechain input (another mix, a
    bounce, a reference track on its own channel). The live line under Amount counts the seconds
    learned. Choosing it with nothing routed keeps the reference you had.
  • Match compares the reference with the last ~3 s of this track, both at the same overall
    level, and adds the smoothed difference (about 1.5 ERB either side, at most ±6 dB × Amount) to
    the auto EQ. Range, Freeze and Paint treat it like the rest of the curve. Measured: a mix with a
    +6 dB shelf at 4 kHz (5.5 dB brighter than a flat reference) comes out 5.3 dB closer to it at
    Amount 100 % (8 kHz −3.6 dB, 500 Hz +2.0 dB), in Stereo and in Mid/Side. A reference learned from
    the key is within 0.28 dB of the same audio analysed as a file (200 Hz – 16 kHz).
  • Amount 0–100 % (default 50 %): how far Match steers. The live line says how far it is
    steering now ("matching −2.4 dB").
  • The reference is a teal line on the display at the level of the output, with a REF tag naming it.
    It's saved with the project. Clear the reference is in the menu.
  • Off is the default; projects from before 4.0 open Off with no reference. Off, Learn key and
    Amount 0 % are bit-identical to 3.1 (max difference 0).
• Paint (pill in the header): draw on the display how much of the auto EQ and of Heat acts in
  each region, beyond the two Range handles.
  • Drag in the EQ lane or the Heat lane: the top of the lane is 100 %, the bottom none. Alt-drag or
    right-drag erases back to 100 %, a double-click resets that lane, Escape leaves paint mode. Each
    stroke is one undo step.
  • The EQ amount covers everything the auto EQ does there (Reveal, Smooth, Tilt, sidechain and
    reference). Measured: painting the EQ to 0 % over 2–5 kHz takes a −9.0 dB Smooth cut at 3150 Hz to
    0.0 dB and leaves the −7.7 dB cut at 800 Hz alone. Painting Heat to 0 % there takes a 1.5 kHz
    tone's 2nd and 3rd harmonics from −6 / −10 dB to below −170 dB and leaves its 4th (6 kHz) as it was.
  • Painted-out regions stay dimmed, like outside the Range, and the hover read-out shows the amounts.
    Saved with the project (64 points per lane); 100 % everywhere is the default and bit-identical to 3.1.
• Keyboard shortcuts: Ctrl+Z undo, Ctrl+Shift+Z or Ctrl+Y redo (Cmd on macOS), once you've
  clicked in Anneal's window and if the DAW passes keys to plug-ins. Keys Anneal doesn't use go back
  to the DAW.

Changed

• Wider window: 1312 × 640 (was 1160 × 640) for the REFERENCE group, resizable from 984 × 480 to
  1968 × 960. Everything else is where it was.
• The project format is version 2: it adds the reference and the paint. 3.x projects open as they
  were; projects saved by 4.0 don't open their reference or paint in 3.x (the knobs do).
• The Windows CLAP is on the release again: Anneal-4.0.0-Windows-CLAP.zip, with the two new
  parameters (27 in all).

Fixes

• Freeze lost the held curve when the DAW stopped and started playback (the curve was re-learned
  from the first moment after the restart) or when the latency mode changed. A frozen curve now
  stays exactly where it was (−9.0 dB before and after both).
• Switching the Ceiling back on could pull quiet audio down for about 80 ms: its detector still
  remembered loud audio from the last time it was on. It now follows the audio while off.

Tests: 112/112 acceptance checks (19 new: the two fixes above, Reference match (identical when off,
the shelf moved 5.3 dB, Learn key, files analysed off the audio thread, Freeze holds the match), Paint
(identical at 100 %, EQ and Heat paint, a stroke is one undo step), reference and paint in saved projects
and older projects, the shortcuts, paint strokes and a double-click reset on the display, a dropped file,
the display following Learn key, and the REFERENCE group in the layout). 20/20 host checks on the built
VST3 (new: the Reference parameters, Learn key through the sidechain bus, and the reference and paint
travelling in the VST3 state into a new instance: Heat painted to 0 % there writes nothing, −300 dB,
against −15.3 dB unpainted). 25/25 isolated installer checks, ClapCheck 10/10 on the Windows CLAP, file
version 4.0.0 on the VST3 and the app. Editor rendered at 100/150/200 % and
at the new minimum and maximum sizes, plus Reference, Learn key, Paint, painted and file-drop snapshots.
CPU at 48 kHz on this machine: Stereo Low 2.0 %, Normal 1.8 %, Precise 2.7 % of one core; Mid/Side
Low 3.7 % (64), Normal 2.8 % (512), Precise 4.3 % (64), the same as 3.1.0. With Reference Match at
100 % and Paint on: Low 2.2 %, Normal 1.9 %, Precise 2.8 %. Slowest 0.1 % of 64-sample blocks: Stereo
Low 8 %, Precise 25 %; Mid/Side Low 19 %, Precise 36 % of the deadline.

Anneal 3.1.0

3.1.0 — 2026-10-13 23:06

New

• CLAP for Windows: Anneal-3.1.0-Windows-CLAP.zip on the release, for Bitwig, Reaper,
  FL Studio 2024+ and other CLAP hosts. Same sound, controls, sidechain input and latency as the
  VST3. Copy Anneal.clap into C:\Program Files\Common Files\CLAP (or
  %LOCALAPPDATA%\Programs\Common\CLAP just for you). Checked with a small CLAP host
  (tools/ClapCheck.py, 10/10):

  • ports: Input and Sidechain in, stereo Output;
  • 25 parameters;
  • latency 1091 samples at 48 kHz, the same as the VST3;
  • 2 s of audio processed cleanly.

  Not tried inside a DAW yet.

• macOS (Apple Silicon and Intel) and Linux, prepared, not built yet. The project now targets:

  • on macOS, one universal binary (arm64 + x86_64) with AU (Logic Pro, GarageBand, MainStage),
    VST3, CLAP and the standalone app, for macOS 10.15 or later;
  • on Linux, VST3, LV2, CLAP and the standalone app.

  The interface uses Helvetica Neue on macOS and the system sans on Linux. A GitHub Actions
  workflow builds them, tests them (acceptance tests, host check, CLAP check, auval -strict,
  a both-architectures check) and packages them with an install script. It isn't running yet:
  GitHub needs one extra permission on the build PC's login first (see "Turning on the macOS and
  Linux builds" in docs/DISTRIBUTION.md). Until then there are no macOS or Linux downloads, and
  that code hasn't been compiled for those systems.

• docs/DISTRIBUTION.md:

  • which DAW uses which format;
  • what Pro Tools (AAX) and notarized macOS builds need;
  • a ready-to-paste KVR listing;
  • notes on Plugin Alliance and on licences before a public release.

The Windows VST3 is unchanged: same sound and look (all editor snapshots are byte-identical to
3.0.1's).

Tests: 93/93 acceptance checks, 17/17 host checks on the built VST3, 25/25 isolated installer
checks, ClapCheck 10/10 on the Windows CLAP, and file version 3.1.0 on the VST3 and the app.
CPU at 48 kHz on this machine: Stereo Low 2.0 %, Normal 1.8 %, Precise 2.7 % of one core; Mid/Side
Low 3.7 % (64), Normal 3.1 % (512), Precise 4.4 % (64). The DSP code is the same as 3.0.0's, so
3.0.0's higher figures (2.5 / 2.1 / 3.4 %) were the machine's load at the time, not a slowdown.
Slowest 0.1 % of 64-sample blocks: Stereo Low 10 %, Precise 24 %; Mid/Side Low 17 %, Precise 37 %
of the deadline.

Anneal 3.0.1

3.0.1 — 2026-10-13 22:54

Fixes

• The plug-in file's Windows version (Explorer → Properties → Details, and what some plug-in
  managers show) said 0.2.1 in every release since 0.2.1, including 3.0.0. The plug-in itself
  always reported the right version to the DAW. JUCE only rewrote that version stamp when the icon
  changed. The build now refreshes it whenever the version changes, and build.ps1 refuses to
  continue if the VST3's or the standalone app's file version differs from the release version.
  Sound, controls and projects are unchanged.

Tests: 93/93 acceptance checks, 17/17 host checks on the built VST3, 25/25 isolated installer
checks, plus the new file-version check (failed on the 3.0.0 build: 0.2.1; passes on 3.0.1).

Anneal 3.0.0

3.0.0 — 2026-10-13 19:05

The first major update. Anneal can now listen to another track, hold its curve, and keep its
output under a true-peak ceiling. Everything from 0.2.6 sounds the same with the new controls off.

New

• Sidechain (new SIDECHAIN group, violet). Route another track into Anneal's sidechain input
  and pick what it does:
  • Make room: dips the main only where the key is loud, in the key's own bands. A vocal key
    over a music bus pulls the bus down at the vocal's formants and leaves the rest. Measured with
    a 2 kHz key at Amount 100 %: −9.0 dB at 2 kHz, 0.0 dB at 300 Hz and 8 kHz.
  • Unmask: lifts the main where the key would mask it (+5.6 dB at 2 kHz in the test).
  • Amount 0–100 % (default 50 %). The ring's live line shows how far it is ducking or
    lifting right now. The key's spectrum is drawn on the display as a violet line, with a KEY tag.
  • With no key routed, or a silent one, the output is identical (max difference 0). The key's FFT
    runs only while Sidechain is on. Off is the default, and projects from earlier versions open Off.
  • Routing: in Cubase, enable the plug-in's side-chain input and send the key to it. In FL Studio,
    send the key's mixer track to Anneal's track with Sidechain to this track, then pick that
    track as Anneal's sidechain input in the wrapper's Processing tab. Both were tested here
    through JUCE's VST3 host, not inside the DAWs.
• Freeze (AUTO EQ title row): holds the auto-EQ curve where it is, so Anneal works as a static
  EQ shaped by the music it just heard. Heat, Tilt and Range stay live. The sidechain dip is part of
  the curve, so it holds too. The display shows "AUTO EQ FROZEN". Measured: a −9.0 dB cut stays at
  −9.0 dB after the music changes (without Freeze it returns to 0.0 dB).
• Ceiling Off / 0.0 / −0.3 / −1.0 dBTP (OUTPUT): a true-peak look-ahead limiter at the very end,
  after Mix and Output, so Heat and an Output boost can't clip. Peaks are found 4× between the
  samples (16-tap windowed sinc), with 1.25 ms look-ahead and 80 ms release. With a sine at fs/4 and
  music driven 12 dB over, the true peak (measured with an independent 8× meter) lands from
  0.01 dB under to 0.10 dB over the ceiling. The output meter shows the reduction in orange under
  the peak, holding the deepest value for 1.5 s. Off is a pure delay (bit-exact) and the default.
• Undo / Redo: two arrows in the header next to Latency. Each knob drag, typed value or menu
  choice is one step, and loading a preset is one step. Host automation isn't recorded. Undo
  history starts fresh when a project loads.
• Wider layout: 1160 × 640 (was 1000 × 620), resizable from 870 × 480 to 1740 × 960, with a
  SIDECHAIN group between HEAT and OUTPUT.

Changed

• Latency is 1.4 ms longer (67 samples at 48 kHz) in every mode: the Ceiling's look-ahead.
  It is reported whether the Ceiling is on or off, so turning it on never shifts timing. At 48 kHz:
  Low 22.7 ms, Normal 44.1 ms, Precise 86.7 ms. Hosts compensate automatically.
• The plug-in has a second, optional input bus (the sidechain). Hosts that route nothing to it
  see a stereo effect as before.

Fixes

• Hover read-outs with whole decibels could show four decimals ("heat −24.8106 dB"); they now round.

Compatibility: every parameter ID is unchanged. Four parameters are added (Freeze, Sidechain,
Sidechain Amount, Ceiling); projects from earlier versions open with all of them off.

Tests: 93/93 acceptance checks (10 new: a silent key changes nothing, Make room ducks only the
key's band, Unmask lifts, the processor's sidechain bus, Freeze holds, true peak under each
ceiling, Ceiling Off is a pure delay, undo/redo with a preset as one step, older projects open with
the new controls off, and the 3.0 layout with every control inside and none overlapping). 17/17 host
checks on the built VST3 (new: the sidechain bus through VST3, a routed key ducking the main −9.0 dB;
latency checks now include the look-ahead). 25/25 isolated installer checks. Editor rendered at
100/150/200 % and at minimum/maximum size, plus new Sidechain and Freeze snapshots.
CPU at 48 kHz on this machine (Ceiling and Sidechain off): Stereo Low 2.5 %, Normal 2.1 %, Precise
3.4 % of one core; Mid/Side Low 4.6 % (64), Normal 3.5 % (512), Precise 5.3 % (64). That is above
0.2.6's run (1.9 / 1.7 / 2.7 %) and close to 0.2.5's (2.4 / 2.1 / 3.4 %); the runs weren't side by
side and the machine's load varies, so treat it as unconfirmed. Slowest 0.1 % of 64-sample blocks:
Stereo Low 10 %, Precise 28 %; Mid/Side Low 20 %, Precise 43 % of the deadline. Ceiling-on and
Sidechain-on CPU weren't benchmarked.