Skip to content
Gravlax

Reproducible demonstrations

Three Google Colab notebooks exercise the preprint’s central workflows. They contain no saved scientific output and no fallback data address. Their checked-in defaults pin the published demo-data-v1 demo-manifest.json and its SHA-256, 82c34aad442d478f1cb1243a6ccfe8ad9f937b81d9e1f946a8eb2cfc498214fd. Each notebook verifies the CLI, Python wheel, archives, annotations, and design files before executing them. Both installed programs must report the exact version declared by the manifest.

A collection commits source content identities rather than pathnames, so since 0.2.1 a capsule may publish a prebuilt, rooted .aicollection with shape routes plus a --locations manifest keyed by those identities. The demo-data-v1 manifest declares no collection, so the multi-archive notebooks download each rooted source archive and build a fresh collection inside Colab. When a manifest declares the optional collection section, they download that hash-pinned collection and its location manifest instead and resolve every committed source through --locations, whose relative paths are read beside the downloaded manifest. Nothing is rebuilt or rescanned. The declared collection must commit exactly the story’s archives with the same build options, its location manifest must resolve exactly the archives already verified by root, and collection inspect --locations ... --verify-routes must report the declared collection root and one shape route per archive; otherwise the notebook stops without falling back to a local rebuild.

  • Annotation reinterpretation compares two annotations against one rooted evidence archive and displays exact count deltas and a signed gene-level delta chart.
  • Multi-donor event discovery reverse-searches a collection for recurrent junction and splice-event evidence, applies unique-chain raw-UMI-class sample, donor, group, and annotation-gap filters, and charts the strongest event’s support across donor/group combinations.
  • Federated junction and co-occurrence queries one exact junction across a collection and drills into Boolean same-record region, junction, and terminal-tail predicates. It charts exact junction support by source archive and the resulting evidence patterns.

Every chart is generated from typed output produced during that notebook run. The notebooks use a small deterministic SVG renderer built from the Python standard library and Colab’s IPython display API, so there is no plotting package to resolve and no cached chart data.

The checked-in demo-manifest.template.json lists every required asset and story parameter. It is deliberately incomplete: null fields stop execution rather than silently selecting mutable or invented data. A publishable manifest must validate against demo-manifest.schema.json, use immutable release or repository-record URLs, and contain lowercase SHA-256 digests for every transport object. Archive assets should additionally declare their rooted aie-directory-root-v2 identities, and a published collection declares its aicollection-directory-root-v1 content root alongside the location manifest that resolves its committed sources.

The notebooks are now one-click demonstrations pinned to the immutable demo-data-v1 capsule and the v0.1.5 software it declares. Clearing a locator, using a mutable URL, or changing any downloaded byte still fails closed rather than selecting fallback or cached results.