Bias-X changelog

• v1.0.1windowsJuly 31, 2026

  Grit no longer turns the volume down as you turn it up. It used to lose 4.4 dB across the knob at mixing level and 13.3 dB on hot material, so turning it up sounded like turning the plugin down. It is now within 0.3 dB across its whole range, and it is roughly three times stronger at the same knob position — 0.10–0.30 is a normal working range now. Output is no longer a hidden drive control. And every factory preset is far gentler on delicate material: vocal presets measure 0.37–1.08 % distortion at mixing level where they used to be 2.40–2.97 %.

  All of that shipped in 1.0.0. It is repeated here because 1.0.0's notes opened with a paragraph about test tooling and buried it three paragraphs down — and the first paragraph is the part most people read. The 1.0.0 notes are left as they were published rather than quietly rewritten.

  > 1.0.1 changes no audio. Same DSP, same presets, same behaviour — the plugin binaries > are functionally identical to 1.0.0. If you are already on 1.0.0 there is nothing here for > you but a better manual. This is a documentation release.

  Fixed

  • •BX-066: the manual published the wrong figure for the plugin's own noise floor, in three places and a chart. It said ~2.9 % THD at −18 dBFS with every control at minimum; the real figure is 1.924 %. The cause was in the measurement harness rather than the prose: BiasMeasure's Setting had no field for stage_gain2 and never set it, so the sweep row labelled *"everything at minimum"* was quietly running the second stage at its 6 dB default. The gap is exactly Stage 2 — the authoring grid puts Drive 0 / Stage 2 0 at 1.884 % and Drive 0 / Stage 2 6 at 2.759 %. A probe whose name does not match what it sets is worse than no probe. Setting now carries the field, defaulted to 6 dB so every canned setting measures exactly what it always has (verified unchanged: clean 20.69, driven 37.39, gritty 40.31, full 28.23), with the "minimum" row explicitly at zero. The manual's chart is redrawn from the corrected sweep.
  • •BX-067: the manual could not be exported to PDF usably, which is why no PDF had ever shipped. Two faults, both invisible until something actually rendered one. Every bold word printed white on white — strong hardcodes its colour, so it survived the print stylesheet entirely and whole phrases were simply missing from the page. And section{page-break-inside:avoid} cost six near-blank pages out of 41, because a section longer than a page cannot avoid breaking and gets pushed whole onto a fresh one.

  Added

  • •A PDF manual, docs/manual/Bias-X-User-Manual.pdf — 29 pages, and it keeps the dark theme rather than inverting to a light document, because that look is the manual's identity and a PDF of it gets read on a screen far more often than it gets printed. Making that survive an export needs print-color-adjust:exact (Chrome drops every background otherwise) and @page{margin:0} with the inset moved inside, because Chrome leaves the page-margin band unpainted whatever the canvas background says and framed every page in white. Generated by build-manual.py --pdf via headless Chrome or Edge, so it is reproducible rather than hand-made.
• v1.0.0windowsJuly 30, 2026

  Bias-X 1.0. The version number is the point of this release as much as anything in it: every knob has been measured to do what its label says at the level people actually mix at, every factory preset is voiced to a distortion budget a test can fail, and the whole thing is gated by six harnesses plus pluginval at strictness 10, two sanitizers and clang-tidy.

  Upgrading from 0.7.0, which is the last public build. Everything below landed after it.

  ---

  A full audit, thirty-four fixes, and the whole factory library rebuilt — twice. The second rebuild is the one that matters: the library is now voiced to a declared distortion budget per category, enforced by a gate that fails, instead of to taste with the measurement taken afterwards.

  If you only read one thing: presets for delicate material are far gentler than any previous version. Vocal now measures 0.37–1.08 % THD at mixing level where it used to be 2.40–2.97 % — about 2.7× quieter — and the same is true of the mix-bus (Vintage) and texture/pad (FX) categories. Drums went the other way, on request.

  Two of the fixes change how a control behaves, and if you have used an earlier build you will notice: Grit is roughly three times as strong at the same knob position, and no longer changes the output level as you turn it, and the Output control is no longer a hidden drive control. Both are corrections, not re-voicings — the details are below. Every factory preset was re-solved against them.

  User presets are untouched: their stored parameter values recall exactly as before.

  Why the library was rebuilt

  Every measurement in this project until now used a 0.4–0.5 amplitude test tone (−8 to −6 dBFS). That answers *"how far can this be pushed"* but not *"what does it do to my mix"*, and for Bias-X those have very different answers: TriodeCurve.h's kSwing = 1.2 V of grid swing per unit of input, on a −1.2 V operating point, means a full-scale signal drives the modelled 12AX7 to its rated maximum. THD is therefore steeply level-dependent.

  Measured at a realistic −18 dBFS, the v0.7.0 library ran 5–40 % THD. "Silk Polish", commented in-source as *"barely-there polish… the lightest preset in the library"*, measured 15.7 %. There was no subtle option anywhere in it. The batch was authored honestly against the only ruler the project had; the ruler was pointed at the wrong question.

  That same blind spot is why two v0.7.0 fixes had to be reopened — see below.

  Changed — the factory library, second rebuild (BX-063/BX-064)

  The first rebuild in this release aimed at "much more conservative" and still came out too hot for voices, pads, textures and whole mixes. The cause was a method, not a mis-hearing: both this release's first pass and v0.6.0's chose parameter values, measured afterwards, and documented whatever the measurement said — so no preset could ever fail. What changed:

  • •Every category now has a THD ceiling declared before any value was chosen, and BiasMeasure exits non-zero if a preset exceeds it. Vocal 1.10 %, FX 1.10 %, Vintage 1.30 %, Bass 1.80 %, Guitar 2.40 %, Drums 7.00 %, Drive 9.00 %, Extreme 24.00 %.
  • •Three measurements made the rewrite solvable. A new control-interaction probe found that the engine has a floor of 1.924 % with every control at minimum, and that Mix is the only control that reaches below it — so a preset aimed at "a valve stage idling" *must* run Mix under 100 %, which is arithmetic rather than taste. It also found that the controls split into expensive (Drive 6 dB +2.01 points, Grit 0.15 +0.96) and free (both Iron shelves +0.001 across their whole travel, Age −0.021, Bias −0.336), and that Drive and Grit compound at 1.21× the sum of their parts. Iron is a linear low shelf and Age is a lowpass; they change the sound at no distortion cost at all. The library now buys character from those and treats Drive and Grit as a budget.
  • •Where it landed, THD at −18 dBFS, before → after:
  • •| Category | Before | After | | |---|---|---|---| | Vocal | 2.40 – 2.97 % | 0.37 – 1.08 % | ~2.7× quieter | | Vintage (mix bus) | 2.47 – 3.18 % | 0.75 – 1.30 % | ~2.7× quieter | | FX (textures, pads) | 2.06 – 3.08 % | 0.72 – 1.09 % | ~2.8× quieter | | Bass | 1.29 – 4.01 % | 0.59 – 1.75 % | | | Guitar | 1.27 – 5.22 % | 0.62 – 2.39 % | | | Drums | 1.99 – 5.00 % | 1.56 – 6.39 % | wider, hotter top | | Drive | 5.02 – 10.47 % | 3.17 – 6.91 % | | | Extreme | 17.67 – 22.00 % | 8.64 – 22.94 % | wider spread |
  • •Presets are now proved to be distinguishable from each other, which is the risk a conservative library creates — pull everything back far enough and five presets become five names for one sound. BiasMeasure renders a six-tone probe through each preset and compares tonal shape, distortion and modulation pairwise within a category. On its first run it found near-duplicates in six of the eight categories: Lead Sustain and Crunch Rhythm scored 0.04, i.e. the same preset twice. Every pair was separated using the free controls, so none of them got hotter in the process.
  • •Level matching is tighter than before as a side effect: the whole library now sits within −0.95 to +0.77 dB of unity, against −1.40 to +1.26 dB.

  Changed — the factory library, first rebuild

  • •All 40 presets re-voiced, measured at −18 dBFS. Six of the eight categories now sit entirely inside "audible warmth": Vocal 1.17–2.97 % · Bass 1.29–3.50 % · Guitar 1.27–4.85 % · Drums 1.99–3.72 % · Vintage 2.47–3.18 % · FX 2.06–3.08 %. Only Drive (5.02–9.55 %) and Extreme (16.4–21.6 %) go past it, deliberately. The gentlest preset went from 15.7 % to 1.17 %.
  • •Auto Gain is off on every preset (it was on for 29 of 40). Each is instead level-matched statically with stage_output to within ±1.5 dB of unity at −18 dBFS — measured, and now automatically flagged if it drifts. 0 of 40 flagged.
  • •Extreme presets pulled back from the ceiling. v0.7.0's peaked at 0.95–0.99 on a hot tone, effectively at clip; they now peak 0.13–0.19 and still measure 16–22 % at program level.
  • •Grit values are far lower throughout — a direct consequence of BX-023 below, not a change of taste.
  • •Age capped at 0.26 for anything mix-ready. Measured, Age at only 0.50 removes 8.6 dB at 8 kHz; several v0.7.0 presets ran 0.40 silently. It goes higher only where darkening *is* the preset, and those say so in their names.

  Fixed — DSP

  • •BX-060: Grit turned the volume DOWN as you turned it up. The one fix in this release reported by a user rather than found by the audit. Measured across the macro's full travel: −4.4 dB at −18 dBFS, and −13.3 dB on a hot −8 dBFS signal, while THD only moved 27.5 % → 45.7 %. Near the top of the knob Grit was mostly a volume control that got slightly dirtier, and since the ear weighs 13 dB of level far above 18 points of added harmonics, turning it up read as doing *less*.
  • •The cause was BX-023's own fix, below. Making the pentode unity-gain for *small* signals meant dividing by drive — but that cancels the shaper's small-signal prediction, which is only exact while the shaper is not compressing, and saturation *is* the output falling below that prediction. So the correction over-shoots, and over-shoots harder the further the knob goes: at drive 10 a −8 dBFS input reaches the shaper at 4.0 and comes back at about 1.0, so dividing by 10 over-corrects roughly fourfold.
  • •This is the same fault BX-022 fixed for the two Iron macros in v0.7.0, and TransformerSaturation.cpp's comment already spelled it out — *"exactly right at low drive and progressively wrong as drive rises, because saturation is precisely the actual response falling below that prediction"* — while the identical mistake was being written into the pentode. It is fixed the same way: the stage now measures its own output against its own input over ~80 ms and equalises them, so drive shapes harmonic content and mix carries amount. Worst level deviation across the whole macro is now 0.30 dB, at every input level from −30 to −3 dBFS, and THD stayed monotonic while getting steeper through the middle (9.85 % vs 8.54 % at 0.50).
  • •Two things worth knowing. The gate that should have caught this had been *printing* the level on every Grit step since v1.0.0 and asserting only on THD — the numbers were in its own output, unread — so BiasKnobAudit now asserts level at six input levels, and that assertion was verified able to fail before its clean result was believed. And the stage_output-with-Grit span widens from 1.38× to 1.93× against the 2.0× ceiling: that is the physically correct direction, since feeding a saturator 18 dB less signal *should* cost it harmonics and the old figure was partly the broken normalisation masking it, but the margin is now thin and the ceiling was left at 2.0 rather than raised to fit.
  • •BX-023 reopened and re-fixed: Grit was still non-monotonic at program level. The v0.7.0 fix was verified only at hot test-tone level, where the fault does not appear. At −18 dBFS the macro measured 3.78 % THD at 0, 1.56 % at 0.30, and 13.72 % at 1.0 — a control captioned "extra edge" made the signal measurably *cleaner* over its lower third. Two causes, both now fixed: the pentode's inherent asymmetry offset had the opposite sign to the triode's, so their even harmonics cancelled rather than adding; and the pentode's small-signal gain scaled with its own drive, so crossfading it in raised the fundamental faster than the harmonics. Now monotonic end to end (3.78 / 4.86 / 6.22 / 8.54 / 12.52 / 18.51 %) and about 3× more potent at the same knob position.
  • •BX-026 scope corrected: the Output trim was a drive control for the back half of the chain. The v0.7.0 record says "identical THD at every setting from −24 to +6 dB", and that is true *of TubePreamp* — but grit, ironOut and age sit downstream of it, so through the whole engine an Output sweep moved THD 15.84 % → 3.45 % at Grit 0.6, a 4.6× swing. The BX-023 fix incidentally reduced that to 1.38×, which is ordinary level-dependence rather than a hidden control. Both halves are now asserted.
  • •BX-028: Auto Gain collapsed to its −12 dB floor during silence, so the first ~277 ms after any transport stop of more than ~1.2 s swelled up from too quiet. 29 of 40 presets shipped with it on.
  • •BX-029: Auto Gain's time constants were per-block, making it 68 ms at a 64-sample buffer and 872 ms at 2048 — the same preset levelled differently monitored and bounced. Now derived from wall-clock time.
  • •BX-030: Mix was the one unsmoothed parameter, so auditioning presets over playing audio clicked and automating it zippered. Now ramped.
  • •BX-031: a block larger than the prepared `maxBlockSize` overran the oversampler's buffer — heap corruption on the audio thread. Oversized blocks are now split into prepared-size chunks.
  • •BX-032: the Quality switch injected ~20 ms of a patch you do not have. Re-preparing the modules slammed every smoother to its module default rather than keeping the live value; worst case, with Grit at 0 the switch produced a burst of 100 %-wet pentode.
  • •BX-033: Sag stepped rather than ramping — an ~8 dB discontinuity in one sample on a preset recall.
  • •BX-034/BX-035: one non-finite sample could crash or permanently silence the plugin. std::clamp does not sanitise NaN (both its comparisons are false), so a NaN reached CubicLut's (int) cast and read ~2 GB out of bounds. Separately, the transformer's Jiles-Atherton state is a feedback integrator, so a NaN latched there and silenced the plugin for the rest of the session. Hardening those two individually was not enough — the new regression test proved it, since this chain has feedback state in a dozen places including JUCE's own oversampling filters. Non-finite input is now scrubbed at the engine boundary and the chain resets.
  • •BX-046: the host's buffer size was the modulation sample rate. At 1024 and 2048 samples the fast tempo-synced LFO divisions aliased into a slow, arbitrary wobble and Square degenerated into near-random flips — the same session sounded different monitored and bounced. Processing is now subdivided when modulation is active, and costs nothing when it is not.
  • •BX-048: the capacitor stage's "even-harmonic" term was an odd function. The one stage whose stated job is warm 2nd-order colour was producing 3rd and 5th. Replaced with a genuinely even, DC-free term.

  Fixed — presets, UI and app

  • •BX-043: a truncated or corrupt `.biaspreset` loaded as a success, silently leaving a hybrid of the previous preset and whatever fragment survived. Files now carry a parameter count and a short one is refused with a message.
  • •BX-044: parameters absent from a preset file kept their current value instead of returning to their default, so the same file sounded different depending on load order.
  • •BX-045: the session forgot which preset it was on — the LCD read INIT after every reopen, and Save silently became Save-As and duplicated the preset.
  • •BX-047: the Standalone could launch a second copy. moreThanOneInstanceAllowed() returned true, which made the already-correct restore handler dead code; relaunching the shortcut started a second process competing for the exclusive ASIO device.
  • •BX-049: the scope and meter timer guards were dead code — visibilityChanged() is never delivered for an ancestor's visibility change, so both kept running at 30 Hz with the standalone window hidden to the tray.
  • •BX-050: the operating-point dot's animation speed depended on the repaint rate, since paint() was advancing its filter.
  • •BX-051: three of four ComboBoxes had no accessible name, and the entire preset transport was keyboard-unreachable while the knobs beside it were fully accessible.
  • •BX-052: the scope and meters kept repainting at 30 Hz underneath a full-screen overlay — invisible work, exactly while the user was reading the help panel.
  • •BX-054/BX-055: `SignalPathPanel` shaped its glossary text twice per frame, and the two paths disagreed about the scroll range by 28 design px. One shared cache fixes both.

  Fixed — release pipeline

  • •BX-036: `-SkipTests` shipped the previous build's binary under the new version number. --target Bias builds JUCE's shared-code *static library*, not the plugin — the .vst3 and .exe are separate targets that depend on it. The stale-bundle guard then passed against the previous build, and the pipeline signed, packaged, tagged, pushed and announced the old binary as the new version. Now uses Bias_All, plus a timestamp assertion that refuses to package an artefact the build did not touch.
  • •BX-037: no git command in the push step was checked. A rejected push still printed "Done. Pushed v$new" while nothing reached GitHub. Every call is now fatal, the branch is asserted to be main, and the push is explicit.
  • •BX-038: `BiasKnobAudit` could not fail — it printed MISMATCH and returned 0, sat outside the gate, and went unmentioned in TESTING.md, while six BUGS.md entries cited it as their regression test. It now asserts every measured value and is in the gate. It immediately caught BX-023.
  • •BX-039: pluginval was fetched unpinned and unverified and executed on the machine holding the live Azure signing session. Now pinned and checksum-verified.
  • •BX-040: nothing verified a signature actually attached. signtool verify is now a post-condition on both artefacts, and -SkipSigning renames the installer so it cannot be mistaken for a signed build.
  • •BX-041: JUCE was an unpinned absolute path in an unrelated project's folder with nothing recording which version shipped. CMake now asserts the version and fails loudly.
  • •BX-042: VirusTotal could not block a release — a "malicious" verdict was printed into the public notes and the installer published anyway. It now fails the job, and the notes only link a report when a scan actually ran.

  Added

  • •`docs/manual/Bias-X-User-Manual.html` — a self-contained, shippable user manual: annotated interface, signal-path and measured-response diagrams, a full control reference with the measured numbers, all 40 presets, recipes and troubleshooting. Rebuild with python docs/manual/build-manual.py.
  • •`docs/PRESETS.md` — every preset documented: what it does, what it is for, its full control settings and its measured numbers.
  • •`docs/AUDIT-2026-07-29.md` and `docs/WORKPLAN-2026-07-29.md` — the six-lens audit (45 candidate findings → 32 confirmed after adversarial verification, 11 refuted) and the sequenced plan that produced everything above.
  • •`BiasMeasure` program-level suite — a THD-vs-input-level sweep, a preset-authoring calibration grid over the (Drive, Stage 2, Mix) space, an stage_output cleanliness probe, and a per-octave tone-response probe for Iron and Age. The factory audit gained a THD@-18 column and a gain dB level-match column that flags anything more than ±1.5 dB off unity.
  • •New release-gating checks: factory recall for all 40 presets (with a createInit() assertion), a silence-recovery check for Auto Gain, NaN/Inf injection, oversized-block processing, and 0.5×/2× editor renders.

  Closed — the three items the first pass carried forward

  Two of the three turned out to be blocked on a wrong belief rather than on real work, which is the more useful finding.

  • •BX-001 closed with no DSP change at all. The row had said for three releases that the hot settings "await the 8x/16x HQ mode", and v0.7.0 added a guess that it might be "a model-order problem, not an OS-factor one". Neither had ever been tested, because BiasMeasure's canned settings only ever ran at Standard quality. Measured at both: driven −31.0 → −55.5 dB (−24.5 dB), gritty −24.9 → −38.0 dB (−13.1 dB), full −21.6 → −32.8 dB (−11.2 dB); clean was already at −98.5 dB. So the original framing was right, the mitigation has been shipping since v0.4.0, and what was actually missing was the measurement and the guidance. BiasMeasure now prints the Standard-vs-High delta on every run, the seven hottest factory presets select High, and the manual says which settings warrant it.
  • •BX-056: clang-tidy now runs, and needed nothing installed. It had been a documented "known gap" for four releases on the stated grounds that it required "a -G Ninja + clang-cl toolchain". Only the Ninja half was true — clang-tidy needs a compilation *database*, not clang as the compiler, and Ninja + cl.exe produces a perfectly good one. Both Ninja and clang-tidy already ship with Visual Studio. Added scripts/tidy-configure.bat and scripts/run-clang-tidy.ps1 (with -Changed and -Filter for an edit-check loop, since a full sweep is 30–60 minutes). 0 diagnostics across all 30 of Bias-X's own translation units, after fixing four real misc-const-correctness findings in Palette.cpp. Deliberately not release-gating: a clang-tidy version bump adding a new check must not be able to block shipping a fix.
  • •BX-057: JUCE migrated to `FetchContent`, pinned at `GIT_TAG 8.0.13`. v1.0.0's BX-041 added a version assertion as a stopgap; this is the real fix. When a local checkout exists the configure resolves to it and does nothing but add_subdirectory, so local builds are byte-identical and no slower; a machine without one fetches the pinned tag. Both paths verified end to end — a clean configure with -DBIAS_JUCE_DIR= cloned JUCE 8.0.13 from upstream and built a working 8.6 MB Bias-X.vst3. The CI-rebuild prerequisite recorded in CLAUDE.md as a v1 blocker is therefore removed; adding a rebuild-verification job is now ordinary work.
  • •BX-058: UBSan now runs, closing the last of the three. This one *was* blocked for a real reason — MSVC has no -fsanitize=undefined, and Visual Studio's "C++ Clang tools" component ships clang-tidy.exe but no compiler, so BX-056's trick of reusing the VS toolchain does not extend to it. With LLVM installed (clang-cl 22.1.8) and a BIAS_UBSAN CMake option, 0 reports across five harnesses — including BiasEnginePhase5's NaN/Inf injection check, so the BX-034/035 hardening does not itself invoke UB. Three obstacles worth recording: clang-cl reports as Clang while MSVC is *also* true, so the existing if(MSVC) branch would have silently dropped UBSan; Windows ships UBSan only as a static-CRT runtime, so the build dir has to be /MT; and -fno-omit-frame-pointer is silently ignored by the cl driver (/Oy- instead). Driven by scripts/ubsan-run.ps1, and verified able to report before its clean result was believed.
  • •BX-059: the documented ASan workflow could never build `BiasUnit` — found while checking the UBSan work had not disturbed ASan. docs/TESTING.md told you to build four harnesses with -DBIAS_SANITIZE=ON; the first failed at link with LNK2038: mismatch detected for 'annotate_string', because MSVC's ASan annotates STL container internals and Catch2 — pulled in by FetchContent — is not instrumented. The other three do not link Catch2, so whoever ran ASan got three greens and one error that reads like a toolchain fault. Fixed with _DISABLE_STL_ANNOTATION, plus the second trap: every instrumented binary exits 127 before `main()` unless MSVC's clang_rt.asan_dynamic-x86_64.dll is on PATH. scripts/asan-run.ps1 handles both — 0 findings across all four.
  • •> A caution worth carrying forward. The clang-tidy harness reported "0 diagnostics, all > clean" before it worked at all: PowerShell's Start-Process -ArgumentList does not quote > array elements, so the -p path — which contains a space in this repo — was truncated, > clang-tidy silently found no compilation database, and reported nothing. That is > indistinguishable from a clean result. It was caught only by deliberately injecting a > defect and checking the harness noticed, the same discipline Wave 0 applied to > BiasKnobAudit. Do it for every new gate.

  Fixed — documentation

  • •BX-065: the manual's library table can no longer go stale. BX-061 below corrected the manual's figures; this removes the mechanism that let them rot. build-manual.py held all 40 presets as a hardcoded string its own docstring described as "a paste of BiasMeasure's factory preset audit" — and a paste cannot know it is out of date. It now reads the audit at build time and fails with instructions if the file is missing, rather than emitting a manual from remembered numbers. docs/PRESETS.md is likewise emitted from the same data as the C++ table by scripts/presets/, so the two cannot disagree. Both of these were found while updating the docs for BX-060, and both are the same class of defect as BX-060 itself: a number that was already measured, already written down, and never checked against what it sat next to.
  • •BX-061: the user manual's "Measured performance" table shipped with v0.7.0 figures. clean was documented at −77.0 dB alias / −72.1 dB DC; the v1.0.0 build measures −98.5 / −77.3. clean uses no Grit at all, so BX-060 could not have changed it — which is what gave it away. The numbers trace back to the v0.7.0 changelog's own "Known trade-off" section, so they were copied from the previous release's notes rather than from a measurement run. All four rows regenerated. The manual's other five measured claims were then re-verified rather than assumed — Drive, Sag, Output Iron and Age all reproduce exactly; Stage 2's "20–45 % per 3 dB" actually measures 17–23 % and was corrected.
  • •BX-062: `docs/PRESETS.md` claimed "six of the eight categories never leave the warmth band", which was never true. The table printed directly above that sentence already showed five of the six crossing 3 %. Only Vocal was ever inside 1–3 %. Replaced with the measured statement, and every number in that document is now regenerated mechanically from BiasMeasure's factory audit — 40 headings, 6 trims, 40 table rows, 8 category cells — so a missed one fails loudly instead of silently.

  Documented, not changed

  • •Bias-X has no clean floor. With every control at minimum it still produces 2.89 % THD at −18 dBFS; at Init, 10.08 %. This is structural and deliberate — kSwing puts unity input at a 12AX7's rated maximum — and it was simply never written down. The manual now leads with it.
  • •The Iron macros add essentially no saturation — 0.03 percentage points across their entire travel at −18 dBFS. What they do is a clean low shelf: +4.2 dB at 60 Hz, nothing above 250 Hz. That is exactly what BX-022 deliberately built; the mismatch was between the behaviour and the name, so it is fixed by documentation rather than by code.
• v0.7.0windowsJuly 28, 2026

  A user reported that some controls seemed to do something other than what they claimed. A full audit — every knob swept across its whole range, measured for THD, level and spectral tilt — found four that measurably contradicted their own labels, plus two calibration faults. All six are fixed, and the factory library was re-solved against the result.

  This changes the sound. Every preset was re-voiced; user presets will also sound different, most obviously brighter (see Age) and, at high gain settings, more saturated (see Output).

  Fixed

  • •Both Iron macros did the opposite of their label (BX-022). Swept from Init, "Input transformer saturation" measurably *reduced* total harmonic content (25.58% → 22.20%), and Output Iron was a −2.1 dB volume trim with +0.6 points of THD. Two causes, both left behind by v0.4.1's BX-017 fix. The level compensation normalized by the model's *small-signal* prediction, which is exactly right at low drive and progressively wrong as drive rises — saturation *is* the response falling below that prediction — so the residual was a level drop that grew with the knob (−2.2 dB against only 2.3% THD gained). For Input Iron, which sits before the tube, that drop also starved the tube's drive, and the tube makes ~25% THD to the transformer's ~2%, so total harmonics fell. Second, the macros never touched the transformer's low-frequency shelf at all, leaving the magnetizing-inductance lift — the most recognisable part of a transformer's sound — permanently switched off. Both knobs now hold level while drive shapes harmonics, and both deliver a monotonic ~4 dB low-end lift across their travel.
  • •Grit's bottom third made the sound cleaner and quieter (BX-023), on a control that promises "extra edge". Its drive range (0.1–10, skewed) did not reach unity until 30% of the way up the knob, so the first third *attenuated* the signal before the shaper and blended that quieter, barely-distorted copy in. Every increment is now additive.
  • •Age was not neutral at zero (BX-024). Fully counter-clockwise, on a control captioned "capacitor aging", 6 kHz already sat 9.8 dB below 200 Hz — the aging lowpass was parked at the capacitor's own corner with the aging set to none, and it had no bypass at any setting. The plugin could not be made bright. Age 0 is now transparent (−0.08 dB at 6 kHz) and sweeps down to exactly the same fully-aged character as before.
  • •"Output trim" was a drive control (BX-026). It was applied *into* the tube stage's soft-clip ceiling, so it set how hard the stage squared off — THD moved 25.3% → 31.3% across its range. It is now applied after the ceiling and measures identical THD at every setting from −24 to +6 dB: level, and nothing else.
  • •The Standalone couldn't be quit from the tray (BX-027). Right-clicking the tray icon did nothing but toggle the window — the Show/Exit menu never opened. Since the window's close box only hides to tray by design, that made Exit unreachable and Task Manager the only way out. The cause is in JUCE's Windows tray dispatch, which strips every mouse-button modifier before delivering an up event, so the isPopupMenu() test in the mouse-up handler could never be true. The menu is now opened from mouse-down, where the button is still present — which is also the more correct place on Windows, since that is the only branch where JUCE gives the tray window foreground ownership.

  Changed

  • •Drive and Stage 2 no longer waste the top of their travel (BX-025). The triode cascade stops responding past ~32 dB and ~24 dB, so on a linear 0–40 dB range the top 20% and 40% of those knobs did nothing. Both are now skewed; measured dead travel falls to 0% and 10%. Note for existing sessions: this changes what a normalised value means, so a host automation lane already recorded on either knob will land on a different dB value. Presets are unaffected — they store plain dB.
  • •All 40 factory presets re-voiced. The Output fix means gain1/gain2 alone now decide how hard the tube squares off, and without re-voicing thirteen presets collapsed onto ~40% THD — the whole library bunched into one crushed sound. Each preset's gains were re-solved to restore its original character, then levels re-trimmed. Result: all 35 non-Extreme presets sit at rms 0.076–0.095 against the probe's 0.087, peaks 0.35–0.66; all five Extreme presets are back under full scale (0.87–0.99).

  Installer

  • •The brandmark is now on the wizard itself — the tall banner on the Welcome and Finished pages, and the header mark on every other page. Both are generated by BiasIconGen from the *same* live-vector draw code as the plugin's header glyph, so the installer art cannot drift from the product. Inno accepts BMP only for these and JUCE ships no BMP encoder, so the generator gained one.
  • •The uninstaller now removes the VST3 bundle wherever you put it. It only ever deleted the individual files it had logged, which left an empty Bias-X.vst3 folder tree (plus the desktop.ini/Plugin.ico that mark it as a system folder) behind in the user's VST3 path. The chosen VST3 location is now recorded in the registry at install time and read back at uninstall — the uninstaller runs no wizard, so it had no other way to know where a redirected bundle had gone.
  • •Add/Remove Programs now uses the multi-resolution .ico rather than the Standalone's embedded icon, so it resolves even if the .exe is gone and Windows can pick its own size. The Start Menu's Uninstall shortcut carries the brandmark too.
  • •Separate install locations for the VST3 and the Standalone, and the optional Start Menu / Desktop shortcuts, were already present (v0.5.0) and are unchanged.

  Known trade-off

  The clean canned setting's alias@11k went from −95.9 dB to −77.0 dB. Nothing got noisier — the old number was Age's permanent lowpass filtering the aliasing away, and that lowpass was the BX-024 bug. −77 dB is far below the noise floor of any real material. The hot settings moved the other way and genuinely improved: driven −27.6 → −31.0 dB, gritty −19.1 → −24.2 dB, full −19.5 → −20.0 dB.

  Full regression gate (BiasUnit 4296 assertions, BiasRenderSmoke, BiasPresetRoundtrip, BiasEnginePhase5, BiasUiSnapshot, BiasPresetRender 40/40, BiasKnobAudit, BiasMeasure 0/40 presets flagged) and pluginval strictness 10 → SUCCESS all pass.

• v0.6.0windowsJuly 28, 2026

  A rebuilt factory preset library — 40 presets, 5 per category, every one of them measured through the live engine rather than eyeballed from a parameter table. Measuring them is also what surfaced the one DSP fix in this release: the Mix knob has been comb-filtering since Mix was introduced in v0.4.0.

  Fixed

  • •The Mix knob comb-filtered the high end at every setting below 100% (BX-021). The wet chain runs through the oversampler and is delayed by its latency; the dry copy for the blend was taken straight off the input, so the two were misaligned and cancelled each other by a frequency-dependent amount — negligible in the bass, deepening upward until the offset approached half a period. Measured on the v0.5.0 engine: a −9 dB hole at 3.8 kHz at Mix 70%, and a −5.3 dB hole at 4 kHz at Mix 50%. The dry path is now delay-aligned to the active oversampler before blending, which also makes the plugin's reported latency honest at Mix < 100% (the dry component previously arrived early relative to what the host was compensating for). Anything using Mix — parallel saturation especially — is materially clearer through the top octaves.
  • •Found by measurement, not by ear: the existing Mix test probed a 220 Hz tone, where the misalignment is a negligible fraction of a period and the blend looks correct. That is precisely how the bug survived two releases. BiasEnginePhase5 now checks the blend at 4 kHz as well as 220 Hz, and the low-frequency probe is kept deliberately — it passes on the *broken* engine, documenting the blind spot in the test itself.
  • •Other than that fix, no DSP, UI, or installer changes; this release is the library.

  Changed

  • •Factory preset library rebuilt: 128 → 40 presets, 5 per category across Vocal, Bass, Guitar, Drums, Drive, Vintage, FX and Extreme. The old library's problem was breadth without a level standard — sixteen variations per category, many of them near-duplicates, with no consistent output level between them. The new set is small enough that every entry earns its slot and can be auditioned in one pass.
  • •Extreme now means Extreme. Only the Extreme category is allowed to square off, gate, or self-oscillate; it holds the five deliberate sound-design presets (Total Meltdown, Gated Cutoff, Siren Bias, Screaming Envelope, Bit Rot). Every preset in the other seven categories is a mix-ready starting point — none of them clips, gates, or needs a trim before use.
  • •Every preset is level-matched. Measured through the live BiasEngine on a musical probe (BiasPresetRender), all 35 non-Extreme presets land within roughly ±1.5 dB of unity (rms 0.074–0.103 against the probe's 0.087) and peak between 0.31 and 0.57 — so switching presets compares tone, not loudness. Auto Gain carries that for anything that saturates; the parallel-Mix and tremolo/chop presets, where Auto Gain would fight the blend or the modulation, are hand-trimmed to the same target instead. The five Extreme presets are deliberately hot but still peak below full scale (0.82–0.99).

  Added

  • •Authoring conventions documented in `BiasPresetManager.cpp`'s factory table, so a future batch doesn't have to rediscover them from the DSP. Three that materially change how a preset is written: stage_output is applied *into* TubePreamp's final fastTanh ceiling (so it sets how hard the stage squares off, and is not a clean makeup trim); age_capacitance is effectively the brightness/weight control, because CapacitorColorizer's LP/HP pair never bypasses (8k/80 Hz Small, 4k/30 Hz Medium, 2k/10 Hz Large); and age_coloration, grit_*, ironIn_* and ironOut_* must never appear in a preset at all, because BiasEngine drives them from age_age and the three macros every block and would silently override any stored value.
  • •Full regression gate (BiasUnit, BiasRenderSmoke, BiasPresetRoundtrip, BiasEnginePhase5, BiasUiSnapshot, BiasPresetRender, BiasMeasure) and pluginval strictness 10 → SUCCESS all pass. BiasMeasure's factory audit flags 0 of 40 presets. No CPU change from the BX-021 fix — the dry delay is one block copy at base rate, and the measured difference is inside the benchmark's own run-to-run spread on this machine (full engine, Standard: 7.6% / 10.6% / 8.0% across three consecutive runs).
• v0.5.0windowsJuly 27, 2026

  Three UX improvements, all user-requested after reviewing v0.4.1's installer: the Standalone gains a real system tray icon, and the installer gains real choices it didn't offer before. No DSP changes.

  Added

  • •System tray icon for the Standalone. Left-click shows/hides the main window; right-click gives a Show/Exit menu. The window's own close box now hides to tray instead of quitting — only the tray's Exit (or an OS shutdown/log-off) actually ends the process, so Bias-X can run continuously as an always-on processor without a window parked in the taskbar.
  • •User-choosable install locations. The installer no longer hardcodes where the VST3 and Standalone go — a standard directory page controls the Standalone's location (defaults to Program Files), and a second, dedicated page controls the VST3's (defaults to the shared system VST3 folder, with a note that this is what lets DAWs auto-discover it without extra configuration).
  • •Start Menu shortcut is now optional, matching the Desktop shortcut's existing checkbox — defaults on, so the out-of-the-box experience is unchanged unless a user actively opts out.

  Fixed

  • •The release pipeline's own pluginval safety gate could silently pass a real validation failure through — confirmed once (releasing v0.4.1): a per-test failure was logged, but $LASTEXITCODE came back 0 and the release proceeded to packaging/signing/publishing anyway (manual re-validation afterward showed the actual shipped binary was fine, but the gate itself didn't gate). Added a defense-in-depth check that scans pluginval's own output for its unambiguous failure markers, not just the exit code. Release tooling only — no plugin change.
• v0.4.1windowsJuly 27, 2026

  Five fixes from real-world use and a follow-up knob-correctness audit of v0.4.0 — full detail in the "Post-release follow-up" entries under [0.4.0] below and in docs/BUGS.md BX-016 through BX-020. No new features; this release exists to get already-fixed bugs into users' hands.

  Fixed

  • •HQ mode caused complete audio glitching (BX-016) — a host-callback reentrancy hazard plus an infinite re-prepare loop triggered by the plugin's own latency notifications. Standard mode also glitched under heavy Age/Coloration settings from the same root cause.
  • •Input Iron / Output Iron worked backwards (BX-017) — turning either UP reduced saturation instead of adding it. The underlying model's field-strength scaling was also recalibrated so the controls now produce genuinely audible character, not just a corrected gain stage.
  • •Age could collapse a signal to near-total silence at high settings (BX-018) — a phase-cancellation notch, now capped to a mild, genuinely "subtle" dip.
  • •Preset/bank storage could end up inside OneDrive on machines where Windows' Documents folder is cloud-redirected (BX-019) — moved to AppData.
  • •SignalScope's transfer curve and operating-point dot didn't animate with Env/LFO modulation (BX-020) — cosmetic; the audio was always correct.

  Added

  • •BiasKnobAudit (new tool) and expanded BiasMeasure coverage — every exposed knob is now verified for real, audible magnitude (not just direction of change), and all 128 factory presets are measured on every run.
  • •Full regression gate (BiasUnit, BiasRenderSmoke, BiasPresetRoundtrip, BiasEnginePhase5, BiasUiSnapshot, BiasMeasure, BiasKnobAudit) and pluginval strictness 10 → SUCCESS all pass.
• v0.4.0windowsJuly 25, 2026

  DSP was byte-stable from v1 through 0.3.0; this release deliberately breaks that streak — a full quality assessment (DSP algorithms, performance, UI, completeness) led to a ground-up pass on the nonlinear models, full-chain anti-aliasing, a real performance pass, a reference-grade hero visual, and a session's worth of workflow features. Every phase below was independently gated (unit + integration harnesses + pluginval strictness 10) before moving to the next.

  Changed — DSP core (the voice is intentionally different now)

  • •Full-chain oversampling. BiasEngine oversamples the entire nonlinear chain once (4× Standard, 8× High — see Quality below) instead of only TubePreamp's old internal 2×. Measured alias-to-signal ratio at 11 kHz: driven +3.9 dB → −26.7 dB, full-chain setting −4.0 dB → −32.1 dB, clean −22.2 dB → −99.5 dB. Latency (near-zero, minimum-phase IIR) is now reported to the host via setLatencySamples().
  • •Koren triode model (TriodeCurve.h) replaces the old 3rd-order polynomial "lookalike" — a real Norman-Koren 12AX7 plate-current model with smooth grid-conduction limiting (no hard corner) and an authentic bias-dependent harmonic balance. THD@1kHz fell from a cartoonish 120% (driven) to a musical ~21%; the plugin is genuinely more tube-like, not just cleaner.
  • •C¹-continuous pentode saturator (PentodeDistortion) — a single smooth asymmetric tanh shaper riding on a bias-dependent operating point, replacing the old tanh/atan split (which had a slope discontinuity at the zero crossing) and its discontinuous "add-a-square-term" 2nd-harmonic bolt-on.
  • •Rigorous Jiles-Atherton transformer (TransformerSaturation, both irons) — proper implicit dM/dH integrated with RK2 midpoint (verified numerically identical to RK4 at this oversampling factor, half the cost), a numerically-stable Langevin function + derivative, and the dead dMdt_prev state removed. No more clamp-as-stability-crutch.
  • •Cheap correctness/perf fixes folded in: capacitor-stage filter coefficients now recompute at control rate instead of every sample; the transformer's langevin() is called once per sample instead of twice; TubePreamp's Bias modulation is now smoothed (was a raw per-block jump).
  • •The 7 factory presets were re-verified against the new models (round-trip, factory recall) rather than re-tuned from scratch — the new models were deliberately voiced to read as "the same instrument, more real," not a different plugin.

  Added — Mix, metering & workflow

  • •Global dry/wet Mix, real IN/OUT peak+RMS meters (on-brand heat-ramp bars flanking the hero scope), and Auto Gain (RMS-ratio level-match between input and the driven signal, computed before the Mix blend so it doesn't couple with Mix%) — all in a new MASTER utility strip.
  • •Tempo-synced LFO — a corner toggle on the LFO Rate knob swaps it for a 9-way note- division combo (incl. dotted/triplet), reading the host's live BPM.
  • •Oversampling quality selector (Standard 4× / High 8×) — both oversamplers are pre-built at prepareToPlay, so switching is instant and allocation-free; switching also re-prepares the five DSP modules at the new effective rate and forwards the latency change to the host.
  • •A/B compare — swaps between two in-memory parameter snapshots (not persisted; presets remain the durable comparison mechanism).

  Added — Hero visual

  • •The SignalScope's context lanes (IN/STAGE/GRIT) are now trigger-aligned (locked to a rising zero-crossing) instead of sliding per block.
  • •The OUT lane is a phosphor-persistence trace — a 7-frame decaying afterglow fed from a new lock-free sample ring in BiasEngine, reading like a real analog scope.
  • •A live operating-point dot rides the transfer-curve plot at the actual input drive level.
  • •The header tube glyph and hero underglow now breathe with the tube's conduction level (a 20 Hz header-band repaint) instead of a static bloom.

  Added — Foundation polish

  • •Inter (Regular/SemiBold) and JetBrains Mono (Regular/Medium) are now embedded via BinaryData (OFL-licensed, pulled from the projects' own GitHub releases) instead of referenced by name — this actually fixed the real render on the primary dev machine, where neither font turned out to be installed.
  • •Tooltips on all ten panel knobs and every combo/button, in plain language.
  • •Knob accessibility names — VectorKnob now sets its slider's accessible title from its caption, covering every knob (present and future) from one shared code path.
  • •Dark-glass popup menus — ComboBox dropdowns now use a glass-gradient background, rounded accent-highlighted rows, and a small blaze dot in place of the system checkmark, instead of default JUCE chrome.

  Performance

  • •Full-engine CPU: 1.14% → 11.3% of one core at 48 kHz/512 (was 27% at a naive first pass with RK4 + no caching) — the 4× oversampling and reference models cost real CPU, and a dedicated pass bought most of it back. The Koren model's quiescent plate-current is now cached at control rate instead of recomputed every sample (halved the tube stage's cost, exactly — not an approximation). SIMD and further fast-math were evaluated and deliberately deferred (see docs/AUDIT-LOG.md) — the chain's recursive, branchy per-sample state makes a branchless rewrite higher-risk than the CPU headroom currently justifies.

  Testing

  • •New harnesses: BiasMeasure (THD / alias-to-signal / DC, the objective before/after ruler for the DSP work), BiasBench (per-module + full-engine CPU timing), BiasUnit (Catch2 unit layer — DSP-primitive invariants + parameter-layout checks), BiasEnginePhase5 (functional tests proving Mix/Auto-Gain/Quality-switch actually work, not just that they're wired), BiasPresetRender (renders every factory preset to WAV for audition). An opt-in BIAS_SANITIZE CMake option builds any target with AddressSanitizer (-DBIAS_SANITIZE=ON on a fresh build dir).
  • •Two real UI bugs were caught by inspecting actual renders, not just reading passing test output — see docs/BUGS.md BX-009 / BX-010.

  Verified (full gate)

  BiasUnit (9 cases / 4290 assertions), BiasEnginePhase5 (7/7 functional checks), BiasRenderSmoke, BiasPresetRoundtrip (32/32 params round-trip + factory + init), BiasUiSnapshot regression, and pluginval strictness 10 → SUCCESS all pass on the final build. Warning-clean.

  Post-release follow-up (same day) — user testing pass

  • •Fixed: Standalone opened to a solid white content area on the primary dev machine (native window chrome fine, JUCE-drawn content blank) — a hardware-accelerated (Direct2D) rendering context failing silently. Forced the classic software renderer (JUCE_DIRECT2D=0).
  • •Fixed: every standard button (help, PresetBrowser's Load/Delete/Save-As, MASTER strip's Auto Gain/A-B) showed a visible square halo around its rounded corners — drawButtonBackground's ambient drop-shadow was larger than the button and got hard-clipped into a square by JUCE's component-bounds clipping. Disabled the ambient shadow for buttons (kept for the larger group/master panels, which have real margin).
  • •Changed: the MASTER strip grew 56px → 104px tall (canvas 900×680 → 900×728) — the Mix knob and MASTER's buttons/combo were too small to read at the original height.
  • •Added: a "How it works" overlay (new header icon, next to "?") — a two-page Signal Flow + Tips & Tricks pop-over in the same idiom as the Preset Browser. Signal Flow is a live-vector, hover-interactive diagram of Bias's actual chain (Input → Input Iron → Stage → Grit → Output Iron → Age → out) plus the envelope/LFO modulation loop, each stage populating a description panel on hover/click. Tips & Tricks is a static two-column list (6 creative + 6 utility pointers, each naming a real control). Pattern adapted from Black Ion 76's SignalPathPanel, restyled for Bias's single-accent Blaze system (the modulation loop uses the same accent family dimmed to accentDeep, not a second hue).

  Post-release follow-up (2026-07-23, continued) — layout polish & signal-flow detail

  • •Changed: the MASTER label now sits top-left of its panel, matching every other group panel's convention, instead of vertically centered.
  • •Added: a "CAP SIZE" caption above the capacitance dropdown — it previously had no label at all.
  • •Changed: the Signal Flow overlay's bottom half is now two distinct bordered containers (Stage Detail + Control Glossary) instead of text floating on one flat field, plus a grid-textured diagram well and a "SIGNAL PATH" corner label.
  • •Fixed: tooltips had stray Unicode/ASCII artifacts from em-dash encoding and unnecessary bold formatting — scrubbed to plain " - " hyphens and regular weight (BiasLookAndFeel gained getTooltipBounds/drawTooltip overrides for this).
  • •Changed: the MASTER strip trimmed 104px → 89px tall (canvas 900×728 → 900×713) — still comfortably legible, reclaims the extra vertical space it didn't need.

  Post-release follow-up (2026-07-24) — layout restructure, dynamic glossary, category presets

  • •Changed — layout. STAGE / CHARACTER / AGE now share one row (tightened spacing: 5px panel padding, 3px knob padding, 10px inter-panel gaps). MOD and MASTER share the second row, with MASTER — now including the OUTPUT trim knob, folded in from its own former group — seated bottom-right where MOD used to sit, and MOD moved left. Canvas shrank 900×713 → 900×610.
  • •Fixed: the envelope/LFO modulation loop's return arrow in the Signal Flow diagram pierced 14px into the STAGE box, cutting straight through the "2× TRIODE" sublabel. The return leg now approaches and stops cleanly at the box's own edge, matching the diagram's other loop arrow.
  • •Fixed: MasterStrip's AUTO and A/B buttons were centered 6px off from their true divider-bounded cell — the divider lines were drawn at a flat -6px offset from each zone's edge that didn't match the zones' own symmetric content-centering. Dividers now draw exactly on the zone boundary, and both buttons were also narrowed (AUTO 84→64px, A/B 50→36px) to read as tighter, more proportionate controls.
  • •Changed: the Control Glossary panel in the Signal Flow overlay is now dynamic — hovering or clicking a stage populates it with that stage's actual on-screen controls, each with its own detailed sonic-effect description, instead of one static list for the whole diagram. A stage with no dedicated controls (INPUT) shows an explanatory note instead of an empty box.
  • •Changed: the header preset dropdown now groups presets as category-first submenus (was a flat, section-headered list) — hovering a category (VOCAL, BASS, …) opens a submenu of just that category's presets, with a right-pointing arrow indicating there's more inside. BiasLookAndFeel::drawPopupMenuItem previously discarded the hasSubMenu flag entirely; it now draws the arrow in the same hand-drawn triangle language as the combo box's own dropdown chevron.
  • •Added: the factory preset library grew 7 → 128 presets, 16 per category across Vocal / Bass / Guitar / Drums / Drive / Vintage / FX / Extreme (Drive is a new category — a graduated intensity ladder from "Whisper" to "Redline"). Each category blends bread-and-butter standard presets, creative ones drawn directly from the Signal Flow panel's own Tips & Tricks (negative Bias snarl, stacked Irons, envelope/LFO modulation, parallel blend via Mix), and utility presets (clean gain-stage references, Auto-Gain-on A/B pairs, High-quality final-bounce variants). All 958 parameter assignments across the library were validated against their registered ranges; the 128 names are unique.

  Post-release follow-up (2026-07-25) — CPU performance & legibility

  • •Fixed: user-reported ~40% CPU in Ableton (vs. 1-2% for comparable plugins). Root cause: per-sample transcendental math — the Koren triode's pow/softplus, the Jiles-Atherton transformer's double-precision tanh/sinh in langevin(), the pentode's per-sample tanh — running inside the 4×/8× oversampled chain, up to ~1.5M pow calls/sec. Replaced with Catmull-Rom lookup tables (FastMath.h, new) for the triode plate curve, langevin(), and the pentode/output-ceiling tanh (<1e-6 error vs. the closed form, unit-tested); age (CapacitorColorizer) moved outside the oversampled island — it's mild/near-linear, verified inaudible rather than a real timbral change; the transformer low-shelf now only redesigns on an actual value change (was every 32 samples unconditionally); BiasEngine's per-sample meter/mix/scope loops replaced with vectorised AudioBuffer/FloatVectorOperations calls. Full engine (Standard, 48 kHz/512): 17.29% → 7.19% of a core (2.4×); High: 14.25%. Real-world validated (Standalone + Ableton, via Task Manager): 2-3% — matches the benchmark and lands in the range of comparable plugins.
  • •Investigated and ruled out two hypotheses raised while chasing a stubborn in-DAW CPU-reading mismatch, both closed with direct evidence rather than assumption: the editor's SignalScope repaint cost (user A/B-tested the UI open vs. closed in Ableton — zero CPU difference) and VST3 host sub-block splitting for sample-accurate automation (checked the actual JUCE VST3 wrapper source — it steps each parameter once per host call, no splitting). The real gap traced to Ableton Live's own CPU meter reading pessimistically for this plugin's shape relative to real OS-level cost (confirmed via Task Manager on the same process) — a metering characteristic, not a code defect. Full evidence trail in docs/BUGS.md BX-015.
  • •Fixed: tooltips still rendered bold, despite BiasLookAndFeel explicitly overriding tooltip drawing to a regular weight. Root cause: the editor's TooltipWindow was constructed with a nullptr parent, making it a desktop-level window sitting outside the editor's component tree entirely — its getLookAndFeel() could never actually find lnf, so it silently fell back to JUCE's stock styling no matter what the override said. Now parented to the editor, per JUCE's own documented guidance for plugin editors.
  • •Changed: hero meter and scope text was too small to read comfortably. Increased size and kerning across MeterBar's IN/OUT labels and the SignalScope's IN/STAGE/GRIT/OUT/TRANSFER/GRID labels.
  • •Changed: the Signal Flow overlay's Stage Detail and Stage Controls panes were hard to read, and long control descriptions clipped silently. Both panes bumped to a genuinely legible size (title 13.5→16 pt, body 12→13.5 pt, control name 9.5→11.5 pt, detail 9→10.5 pt) — the old drawFittedText auto-shrink-to-fit was quietly compressing text below its stated size whenever copy ran long. Both now word-wrap normally and grow a vertical scrollbar only when content genuinely doesn't fit; the control glossary changed from fixed equal-height row slices (which silently clipped detail text on any stage with more than 2-3 controls) to a real scrollable stacked list.
  • •Added: BiasUiBench (new CMake target) — a headless GUI repaint-cost benchmark isolating SignalScope/MeterBar paint cost from DSP cost, motivated by the CPU investigation above. BiasBench gained CLI-overridable sample rate / block size and a Standard+High dual-pass in one run, so DSP cost can be checked under realistic DAW conditions instead of one fixed config. BiasUnit gained two regression cases (FastMath::fastTanh / CubicLut accuracy vs. the closed forms).

  Verified (full gate, re-run 2026-07-25)

  BiasUnit (11 cases / 4296 assertions), BiasRenderSmoke, BiasPresetRoundtrip (32/32 params round-trip + factory + init), BiasEnginePhase5 (7/7 functional checks), BiasUiSnapshot regression, and pluginval strictness 10 → SUCCESS (v1.0.4) all pass on the current build.

• v0.2.3windowsJuly 21, 2026

  Added an About pop-over with credits and a live link back to vst.tools.