Apotheosis

Zero overshoot, by construction. The lookahead attack has always been a sliding-window minimum — the future peak is already known, so no attack time constant exists. That minimum now runs through a cascaded-box FIR smoother, and the smoothed gain is provably never less attenuating than required. The test bypasses the safety clamp to prove the algorithm rather than the net: 10,080 randomised signals across four styles and three lookahead settings, zero violations. Running sums accumulate in double precision with exact re-summation every 2^14 samples — float running sums drift, and drift in a gain envelope is audible.

A ceiling you can measure. ITU-R BS.1770-4 Annex 2, implemented directly: a 4x polyphase FIR, 12 taps per phase, the standard's 48 coefficients with phase symmetry asserted at compile time. With True Peak Guard on, it runs on the base-rate output — the signal a delivery meter actually sees — and ducks by exactly the measured excess for exactly its duration. At a −1.0 dBTP ceiling, an independent compliant 4x meter reads −1.128, −1.287, −1.293 and −1.733 dBTP across the test corpus: every item at or below the ceiling, including a near-Nyquist inter-sample-peak case. A compliant 4x meter can itself under-read the true peak by up to 0.554 dB near that frequency — a property of the BS.1770-4 standard, not of this implementation, so the honest claim is "a compliant 4x meter reads at or below your ceiling", not "no sample anywhere exceeds it".

Oversampling that spends its budget where you can hear it. Alias energy folds back in the decimator, so the custom filter stages spend −106 dB on the way down against −86 dB on the way up, relaxing 10 dB per stage. Measured alias products sit at or below −70 dBc at 4x, −90 dBc at 8x and −100 dBc at 16x Linear, with pass-through magnitude flat to within 0.0004 dB out to 20 kHz across all six factor/phase combinations. The 4x minimum-phase default deliberately keeps the original filters, because its job is bit-exact compatibility with older sessions.

Dither designed, not borrowed. A 9th-order weighted error-feedback requantiser with per-channel TPDF, its coefficients least-squares-fitted offline to an inverse threshold-in-quiet audibility curve for 44.1 and 48 kHz. Measured 20.05 dB of F-weighted noise-floor improvement over flat TPDF at 16-bit, against a binding 15 dB floor — and a chi-square test confirming the dither generator is genuinely triangular, not an approximation.

Delivery-grade metering, O(1) memory. Integrated LUFS the specified way: 400 ms blocks at 75% overlap, −70 LUFS absolute and −10 LU relative gates, two passes over a 0.1 LU histogram instead of a growing block list, so a spec-compliant meter stays real-time-safe over an unbounded session. Verified against the EBU Tech 3341 test vectors to 0.1 LU, with Loudness Range per Tech 3342 to 1 LU.

Audition that actually nulls. Delta taps its dry reference through a second, filter-identical oversampler, so subtracting it leaves no filter-difference residue — 6 dB under the ceiling, the delta sits below −100 dBFS RMS. Unity Gain trims by exactly minus the Input Gain, so switching A/B compares character instead of level.

Latency, reported in full. Apotheosis is not a zero-latency plugin: reported latency is lookahead plus the oversampler's round-trip plus a 6-sample guard delay that sits in the signal path unconditionally below 176.4 kHz, whether or not True Peak Guard is engaged — deliberately, because a delay that appeared and disappeared with the toggle would force the host to re-run delay compensation mid-mix. At the default 5 ms lookahead, in samples at the project rate:

Minimum Phase is flat across factors because the IIR allpass chain's own latency is effectively zero at every factor; Linear Phase is where the oversampling factor shows up, and even there 4x → 8x costs only 3 samples.

Compatibility proven by corpus. Before any new feature landed, the previous engine was rendered to committed float32 fixtures with settings, checksums and generating platform recorded. Every commit since is compared against them bit-exact, after compensating that constant 6-sample guard delay — a pure integer shift, no sample value changes. Zero audio-thread allocations with 16x Linear Phase, the guard, weighted dither, the Safe style, Delta and Unity Gain all engaged simultaneously, counted under a replaced allocator.