Let a scroll expand an uncompressed tarball - #14
Merged
Merged
Conversation
`assetArchives` accepted `zip` or `tar.gz`, so a publisher who ships a plain `.tar` could not be consumed at all — and that is the usual choice when every member inside is already compressed, where a second pass costs minutes to save nothing. The archive goes through the same listing, validation and extraction as `tar.gz`; the only thing missing was permission to say so. Compression is now left to the library on read, which is what it was doing anyway: node-tar reads the header and decompresses or not, and its `gzip` option applies to writing. Passing `gzip: true` on a read looked like a guarantee and was not one. The `tar` / `tar.gz` distinction records what the author pinned, and the asset's SHA-256 is what holds them to it. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Deploying scrollcase with
|
| Latest commit: |
64cb078
|
| Status: | ✅ Deploy successful! |
| Preview URL: | https://e40a0391.scrollcase.pages.dev |
| Branch Preview URL: | https://plain-tar-assets.scrollcase.pages.dev |
suffro
added a commit
to Liatir/liatir-app
that referenced
this pull request
Sep 11, 2026
Boltz-2 builds, signs, self-tests and now predicts. Given ubiquitin's 76
residues and nothing else, it reproduces the experimental structure of PDB 1UBQ
to a backbone RMSD of 1.99 A against a 3 A limit, at pLDDT 0.925. The largest
single deviation, 11.5 A, falls on the C-terminal tail, which is the part of
ubiquitin known to be mobile: physics, not a wrong fold. Two runs at the same
seed differ by exactly zero, and a separate validator process reproduced the
RMSD to the last decimal.
Every prediction here runs single-sequence, because an offline box has no
alignment server. That is the accuracy this product can actually promise, so it
is the accuracy that gets measured.
The evidence is bound to the signed box rather than to a directory: the
validator verifies the release signature and walks the payload digest before it
runs anything. Archive 609a3cc5, 12.69 GB, 20.3 GB installed. Development key,
dirty tree; a production build makes a different archive and must be measured
there.
Three things this turned up that only running it could:
- Boltz's CUDA path imports cuequivariance_torch unconditionally, even though
upstream calls those kernels an optional extra and the reviewed decision was
to leave them out. The first prediction died on ModuleNotFoundError. Boltz
keeps a pure-PyTorch implementation of the same triangular multiplication
behind --no_kernels, which the runner now passes: same mathematics, slower,
and no closed-source NVIDIA kernel shipped for an unmeasured speed-up.
- pandas 3.0.5 works with Boltz 2.2.1. That was the one loose upstream bound,
left for a real import to settle rather than pinned pre-emptively.
- mols.tar is an uncompressed tar, which Scrollcase did not accept at all.
Fixed upstream in suffro/scrollcase#14, which also drops a gzip: true that
read like a guarantee and was not one. Until that release lands, this scroll
cannot be built from a clean checkout.
The legal record is approved. mols.tar is a derivative of the wwPDB Chemical
Component Dictionary, which the PDB archive places under CC0 - traced through
upstream's generation script, the archive's own contents and the wwPDB policy,
and re-checked at build time: all 45,227 members are named by a PDB chemical
component identifier and nothing else.
One claim in that record was wrong and is corrected. frozendict was not this
project's first copyleft dependency. Every published box carries between 8 and
26 GPL-family conda packages, and native-tools, which ships inside the app, has
18 including bcftools and bwa under plain GPL-3.0. What was new is that it was
the first one anyone could see.
Protenix splits in two. v2 stays blocked on a checkpoint that is not ours to
unblock: HTTP 403 again, five days on, while every neighbouring object serves
200, with no official mirror anywhere. Mini now resolves, after the reviewed
manifest turned out to have no conda-first solution at all - 24 conda entries,
116 PyPI, none without a SHA-256, every upstream pin honoured. All 116 of its
licences are reviewed, and two of them should not be there: cuequivariance-ops
ships closed-source NVIDIA kernels that Protenix never imports.
test:verify 6/6. syngraphe check clean.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
This branch had an error being deployed
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
What
assetArchivesacceptedziportar.gzonly. A publisher who ships a plain.tarcould not be consumed at all — and that is the usual choice when every member inside is already compressed, where a second gzip pass costs minutes to save nothing.tarnow joins the enum and goes through the same listing, validation and extraction astar.gz. Nothing about the safety rules changes: entries are listed and checked before anything is written, and links and special entries are still refused outright.One thing that turned out not to be true
The read paths passed
gzip: trueto node-tar, which reads like a guarantee that the archive really is gzipped. It is not one — node-tar reads the header and decompresses or not on its own, and itsgzipoption applies to writing. Probed directly on all four combinations before deciding:So the option is dropped rather than threaded through, and the white paper now says what actually holds the author to their claim: the scroll's
tar/tar.gzdistinction records what they pinned, and the asset's SHA-256 is what makes it binding.Tests
Two added to
archive-security.test.mjs: a plain tarball expands withstripComponents, and the entry rules apply to it exactly as to a compressed one.Full suite on Linux: 566 passed, 36 files. On a Windows host 43 tests fail with
EPERM: operation not permitted, symlink— pre-existing, unrelated, and true onmaintoo.Where it came from
A project whose model bundles 1.86 GB of pickled molecule definitions as a plain
mols.tar. Third gap this line of work has found by real use, after PyPI licences (1.1.0) and download-then-expand (1.1.1).🤖 Generated with Claude Code