# Questions

## What does it cost to run?

The software is free and MIT licensed. Your costs are hardware, electricity, and
hosting.

Hosting a published archive is the cheap part: static files on Cloudflare Pages
fit comfortably in the free tier, and bulk download archives too large for it can
go behind a free worker. The reference deployment publishes tens of thousands of
transcripts this way.

The expensive part is transcription, and it is paid in machine time rather than
money. See below.

## How long does transcribing take?

It depends entirely on your hardware and which backend you use, and any number
quoted without both is meaningless.

What is safe to say: it is the bottleneck, it wants a GPU, and it scales with
audio hours rather than video count. A back catalogue of several thousand hours
is a project measured in days of machine time, not minutes. Channels that already
publish captions skip this step entirely and cost almost nothing.

One trap worth knowing about: running too many jobs at once can push a GPU
backend into falling back to the CPU, at which point everything gets several
times slower simultaneously. More parallelism is not reliably faster.

## How much disk does it take?

Transcripts are small — text compresses well and a whole corpus of them is
megabytes. Media is not. Whether you keep the audio after transcribing is a
per-channel decision, and the software tracks which recordings are safe to clean
up and which are not.

Audio for recordings that have disappeared upstream is never cleaned up
automatically, on the grounds that it is the one thing you cannot re-fetch.

## Do I need a GPU?

No, but transcription without one is slow enough to change what is practical. A
CPU-only machine is fine for a channel that publishes captions, or for a small
back catalogue you are content to work through over weeks.

## Is any of this sent anywhere?

No. Downloads go from the platform to your machine, transcription runs on your
machine, and the build writes files on your machine. Nothing is sent to us —
there is no "us" in the runtime path at all.

The two exceptions are both yours to make: publishing sends the built site to
whatever host you choose, and the optional in-browser chat on a published site
sends queries to an AI provider using an API key the *visitor* supplies.

## Is this legal?

Downloading and keeping a copy of publicly published video is a question with
different answers in different jurisdictions, and it is not one a piece of
software can answer for you. The tool does what you tell it to; deciding what to
tell it is your responsibility, as is complying with whatever terms and laws
apply where you are.

What the software does do is make the record accurate: it says what it archived,
when, and whether the original is still there — rather than presenting an
archive as if it were the platform.

## What happens when a video is deleted?

Your copy stays. The archive re-checks whether recordings still exist at their
source and records the answer, so a deleted recording is marked as deleted,
remains searchable, and keeps its transcript.

The reverse claim is one the software is careful never to make. A recording
counts as available until something re-checks it and finds otherwise, and nothing
re-checks a whole corpus continuously. "Available" here means *not known to be
gone* — never "confirmed still there".

## Where is the git repository?

On this site, read-only: `git clone https://archilyzer.pages.dev/source/archilyzer.git`.
It is the main branch with its whole history, regenerated from the private
repository with every deploy — machine paths scrubbed on the way out, which is
why its commit ids differ. There is no GitHub, and nothing takes a push or a
pull request. [Source](/source/) also has every file to browse and the history
with every commit's diff, and [Downloads](/downloads/) the same tree as a tarball.

## Can I use it for one video?

You can, but nothing about it is designed for that. It is built around back
catalogues — thousands of recordings, kept current on a schedule — and almost
every design decision in it, from the paginated index to the incremental build,
exists because of scale.

## What this is not

- **Not a hosted service.** There is no account, no subscription, no upload.
- **Not a forge.** A read-only mirror: clone and pull, but no pushes, issues or
  pull requests.
- **Not a downloader.** It supplies no downloader — you install `yt-dlp`,
  `ffmpeg` and a transcription backend yourself, and it drives them.
- **Not cloud transcription.** Your hardware, your electricity, your queue.