Premix AI Policy

1. The model we ship does not learn from your jobs

RestoreUNet — the denoise step we ship — was trained on an allowlisted benchmark (paired raw / reference). Customer jobs are not a corpus for that model. The training entrypoint refuses INCOMING, WORKING, EXPORTS, and object storage. We do not improve Premix for other customers from your tape.

Each job may fit a small adapter on this packet’s own noise — this HVAC, this lav, this room — so repair matches the input it was given. That adapter lives in the job folder. Cloud intro: deleted with the seven-day copy. Mixer License: stays in cart scratch, never uploaded. It is never written into the model we ship, and it is never used on anyone else’s job.

2. Mixer License — a client, not a hosting service

The downloadable app connects to folders you choose on a machine you control. Files never leave that machine. We do not receive a copy. We do not run analytics on the session. There is no Premix account required to process audio, and no middleman between the operator and the tape.

If the material cannot leave the set — confidential, unreleased, talent-sensitive — this is the path. Cloud intro is not a substitute for it.

3. Cloud intro — a processing job you send

Cloud intro is not “data not collected.” You upload production audio via MASV to Lemonwater’s worker so the same pipeline can run remotely. Processors: MASV (transfer), DigitalOcean (object storage), Stripe (payment), Vercel (this site), Resend (mail), Clerk (account at checkout). We do not set advertising or analytics cookies on the product.

Copies are deleted seven days after delivery. You may email info@lemonwater.io and we will delete sooner. Premix is not a storage service. Keep your own backups.

Who can touch the tape. Mixer License: no vendor. The file stays on the operator’s machine. Cloud intro: only the rows marked below. None of these is an AI company running inference on your voice. Payment, login, and the website never receive the audio.

No OpenAI, Anthropic, Google, Adobe, ElevenLabs, or other generative / voice-AI vendor is in this pipeline. If that list ever gains a name that can hear the tape, this table is updated first.

4. What we will not do with your tape — even if it would help us sell

Harvest and forum talk are consistent: people will not send unreleased work to a tool that silently trains, invents words, or keeps a copy for “research.” These are refusals, not slogans.

• No “improve our models” clause. There is no buried consent that your job becomes Lemonwater’s training set. The shipped model is not updated from customer tape. A job-local adapter is this packet only.
• No marketing use of recordings. We do not play, quote, or show your audio in ads, reels, demos, or Lemonwater Observer without a separate, written yes for that file.
• No sale or sharing of the tape. We do not sell recordings, transcripts, or derived embeddings. Processors listed in Privacy §9 see cloud jobs only to move or store them for the job. We will say so on this page before any new vendor is allowed to see audio.
• No third-party AI processor. Repair and captions do not POST your audio to OpenAI, Adobe, ElevenLabs, or anyone else’s inference or voiceprint endpoint. Transfer and storage for cloud intro are only the named vendors in §3.
• No confidential work on cloud by default. Unreleased, talent-sensitive, or grant-restricted tape belongs on the Mixer License. Cloud intro is only for files you are willing to let leave the set.
• A written log, not a black box. Every job writes a sound report: what was found, what ran, what fell back, channel by channel. That report is yours. It is how you show what happened — the opposite of Enhance Speech with no receipt.

5. What the pipeline will not do

• Invent dialogue. Assistive repair of a recording that already exists. It does not generate speech that was not spoken, and it does not clone voices.
• Replace the people already hired. It does not replace the mixer, the utility, the editor, or the re-recording mixer. It prepares a packet for them to work from.
• Phone home from the repair models. Denoise and captions run as local models on the worker or on the operator’s machine (DeepFilterNet, RestoreUNet, Whisper weights loaded on that machine). They do not POST your audio to an outside inference API.

6. Captions — local, optional, still a draft

Optional captions (English, off by default) run Whisper on the same machine as the job — the operator’s computer, or our worker for cloud intro. Audio is not sent to OpenAI or any other caption vendor. The model can still omit, mis-hear, or invent words, especially on noisy field audio. Treat a caption as a draft that needs a human read before it is trusted or published.

7. No voiceprint. No identity profile.

This is the product. Premix does not identify who is speaking. It does not enroll a voice, extract a speaker embedding, match a voice to a person, or store a biometric template. There is no voiceprint technology in the pipeline — not for customers, not for talent, not for “quality.”

The paying customer is usually the mixer or editor. The voices on the tape are often someone else. Repair consent from the operator is not talent consent, and it is not a license to profile the people on the recording or to prompt an LLM with their performance.

• No speaker ID, clone, or embedding library. Spectral repair is not biometric identification. We do not build a reusable print of anyone on the tape.
• No third-party voice processor. Audio is not sent to a voice-AI vendor (cloning, verification, emotion, diarization-as-a-service). Captions, if you turn them on, run Whisper on the same machine as the job.
• No LLM guidance on the tape today. We do not send your audio, captions, or sound report to ChatGPT, Claude, Grok, or any other LLM so the model can advise on the cut or the people speaking.
• No always-on listening. Premix is pointed at files. It is not a microphone that “helps you work.”
• You still need the rights. By sending a cloud job you represent that you may process those recordings. We do not obtain talent, union, or source consent for you. Production AI riders that forbid training, cloning, and likeness harvest are compatible with this page — that is why it is written this way.

8. Participation preferences — now, and when voice meets an LLM

Clients already state how they will and will not participate: files on the machine when the tape cannot leave; no training; no clone; no invented speech; a log of what ran. Premix defaults match those preferences. We do not wait for a settings panel to refuse the uses people already reject.

• Today’s defaults: captions off; Mixer License for confidential work; no shipped-model training; no voiceprint; no gen-API; job-local adapter dies with the job; cloud copies gone in seven days.
• If we add voice-driven LLM guidance later, participation is a preference, default off. We will not treat a repair job as consent to talk to a model about the people on the tape. The path, if any: you opt in; we prefer typed or pasted context over raw voice; a transcript used as context is not used to train that model; identity and voiceprint stay out. This page changes first.
• Guidance is a draft. An LLM does not become the mixer, the editor, or Lemonwater. You still own the cut. We still own advice we give in a consulting engagement.

9. Changes

If a job adapter is ever written into the shipped model, used on another customer’s job, if cloud copies are kept past the seven-day window, if a new vendor is allowed to see or hear the tape, if we add voiceprint or speaker-ID, if we add an “improve our models” option, or if we send voice or transcripts to an LLM for guidance, this page and the provider table are updated before that change ships — not in a changelog after.