Does the AT-1 container make the audio smaller?

No, and we do not claim it does. Encoded audio is already at the entropy wall. On this exact episode, xz -9e gets 1.032x and zstd -19 gets 1.039x — three to four percent, which is nothing. The container itself adds 0.24% for its index and hashes. You can reproduce all of it with `at1 podcast stats`. What the container buys is verification, byte-range addressing and query, not bytes.

What does 'verified in the browser' actually mean?

Every segment carries a SHA-256 recorded when the episode was sealed, and the whole payload carries a global hash. Your browser recomputes the digest of each segment with WebCrypto after downloading it and before handing it to the audio device. The page never asks the server whether the bytes are good — it checks them. If one byte anywhere in the file has changed, the check fails and playback stops.

Can it tell me where the file was altered, or just that it was?

Where. Because the hash is per segment rather than only over the whole file, a single altered bit is located to the exact segment and therefore to a timestamp. The 'corrupt one byte' button on this page does exactly that against real data — it flips a genuine bit in the received bytes and lets the ordinary check catch it. Nothing about that demonstration is staged.

Why does playing from the middle not download the whole file?

The container starts with a directory that records the byte offset, length and hash of every segment. Reading 23 KB of directory tells the player where everything is, so jumping to 12:50 is an HTTP Range request for that segment alone — about 92 KB, under 1% of the file. That is the same 'address it without unpacking it' property AT-1 provides for tables, logs and archives, applied to audio.

What is the waveform showing, if not volume?

Encoded bytes per second. In a variable-bitrate stream the encoder spends bytes in proportion to how much is happening, so the curve is a measure of information density rather than loudness. It is read straight out of the container index, which means the shape of the episode — and the boundaries marked on it — are computed without decoding a single audio sample.

What is the episode bound to?

The SHA-256 of the source document it was generated from. The document itself is not in the container and is not published; only its digest is sealed in. That is enough to prove on demand exactly which source produced the episode, without publishing the source. It is the same primitive we sell for training-data attestation, pointed at our own artefact.

Queryable verified media container

An audio file that can prove it hasn’t been edited

Twenty-one minutes on the experiments that failed, served from a sealed container. Every segment is verified in your browser before it reaches the speaker, and any moment is addressable without downloading the rest.

Information density

00:00 / 00:00

Each bar is one segment, and its height is encoded bytes per second— not loudness. It comes out of the container’s index, so the shape of the episode is readable without decoding a single sample. Bars turn green as they verify. Click anywhere to jump.

Verified

of 0 segments

Bytes read

0 B

0.0% of 1 B

Tamper-evident, to the second

A hash per segment plus a global hash. One altered bit is not just detected, it is located — to the segment, and therefore to a timestamp.

Addressable without unpacking

A 23 KB directory locates every segment. Playing from the middle fetches the middle: about 92 KB, under 1% of the file.

Queryable without decoding

Segment boundaries are found from the encoded frame-size curve stored in the index. No audio is decoded to compute them.

It does not compress the audio, and here is the measurement

Encoded audio sits at the entropy wall. There is nothing left to take, and we are not going to pretend otherwise on our own website. On this exact file:

audio, as served	9.64 MB	—
xz −9e	9.34 MB	1.032×
zstd −19	9.28 MB	1.039×
AT-1 container	9.66 MB	+0.24%

The container costs a quarter of a percent, and buys knowing the file is intact, knowing where it stopped being intact, and reaching any moment in it without pulling the whole thing. That is the trade, stated plainly. Run at1 podcast stats and check it yourself.

The media container See it on data, not audio Attestation by hash

Common questions

Does the AT-1 container make the audio smaller?: No, and we do not claim it does. Encoded audio is already at the entropy wall. On this exact episode, xz -9e gets 1.032x and zstd -19 gets 1.039x — three to four percent, which is nothing. The container itself adds 0.24% for its index and hashes. You can reproduce all of it with `at1 podcast stats`. What the container buys is verification, byte-range addressing and query, not bytes.
What does 'verified in the browser' actually mean?: Every segment carries a SHA-256 recorded when the episode was sealed, and the whole payload carries a global hash. Your browser recomputes the digest of each segment with WebCrypto after downloading it and before handing it to the audio device. The page never asks the server whether the bytes are good — it checks them. If one byte anywhere in the file has changed, the check fails and playback stops.
Can it tell me where the file was altered, or just that it was?: Where. Because the hash is per segment rather than only over the whole file, a single altered bit is located to the exact segment and therefore to a timestamp. The 'corrupt one byte' button on this page does exactly that against real data — it flips a genuine bit in the received bytes and lets the ordinary check catch it. Nothing about that demonstration is staged.
Why does playing from the middle not download the whole file?: The container starts with a directory that records the byte offset, length and hash of every segment. Reading 23 KB of directory tells the player where everything is, so jumping to 12:50 is an HTTP Range request for that segment alone — about 92 KB, under 1% of the file. That is the same 'address it without unpacking it' property AT-1 provides for tables, logs and archives, applied to audio.
What is the waveform showing, if not volume?: Encoded bytes per second. In a variable-bitrate stream the encoder spends bytes in proportion to how much is happening, so the curve is a measure of information density rather than loudness. It is read straight out of the container index, which means the shape of the episode — and the boundaries marked on it — are computed without decoding a single audio sample.
What is the episode bound to?: The SHA-256 of the source document it was generated from. The document itself is not in the container and is not published; only its digest is sealed in. That is enough to prove on demand exactly which source produced the episode, without publishing the source. It is the same primitive we sell for training-data attestation, pointed at our own artefact.