V0A (2026)

A monophonic pitch corrector built as three staged deliverables over one shared C++20 DSP library. V0A is an offline command-line renderer that takes a mono WAV, a manually specified period and pitch ratio, and writes a resampled WAV plus a diagnostic CSV — it contains no pitch detection at all, existing only to isolate and audition the resynthesis primitive. V0B adds automatic detection, octave arbitration, chromatic targeting, and voiced/unvoiced gating, still offline and deterministic. V1-Lab wraps the same library in a JUCE AU/VST3 plugin with key/scale selection, retune speed, humanize, voice type, tracking mode, and a three-state note matrix. The plugin duplicates none of the DSP.

Correction runs through a period-synchronous variable-rate resampler. A fractional read head advances over a circular buffer; when read/write separation exceeds bounds, the read coordinate jumps by exactly one detected period — subtracting a period to repeat a cycle when shifting sharp, adding one to delete a cycle when shifting flat. Phase-vocoder and STFT correction, PSOLA, granular and WSOLA shifting, formant preservation, and neural pitch detection are all explicitly excluded, not merely absent.

The project reconstructs the behavior of Antares Auto-Tune EFX 3 from public patents and forensic technical research, and states plainly that it uses no proprietary binaries, extracted assets, or copied source. What makes that claim checkable is an evidence discipline maintained alongside the code. Implementation decisions follow a fixed authority order — the product directive first, then the architecture and evidence documents, then the supplied research reports, and only then new engineering inference — with an explicit rule that a report does not become authoritative by using confident language, equations, or concrete constants.

Every non-mechanical choice carries a provenance label: EVIDENCE-BASED for claims resting on the EFX 3 manual, the Hildebrand patent, or Hildebrand’s public statements; PLAUSIBLE RECONSTRUCTION for inference consistent with the evidence but historically unconfirmed; EXPERIMENTAL for testable hypotheses, disabled by default; and PRODUCT REQUIREMENT for things the project needs that make no historical claim at all. Confidence is tracked separately from provenance, on the principle that a high-confidence product requirement is still not evidence.

The evidence ledger holds fifteen numbered claims, each with its classification and the concrete engineering consequence that follows from it. A parallel registry holds sixteen hypotheses, each with a stable ID, a default state, a discriminating test designed to falsify it, and a promotion criterion it must clear before becoming default behavior. Where the two research reports disagree — on twelve topics, from target hysteresis and retune smoothing domain through humanize timing, tracking epsilon, and latency to resampler direction — the disagreement is preserved in a table recording each report’s position and the decision actually implemented, rather than silently resolved. Unresolved alternatives ship as disabled experiments instead of being deleted.

Measured plugin latency is 4,112 samples at 48kHz. The repository is explicit that this is its own figure and not a claim about the original, which remains one of several openly unresolved questions.