Public, versioned evidence: results, hashes, and reproduction kits. Verify anything yourself.
Each package below is a self-contained bundle: the exact result we reported, a manifest of SHA-256 hashes over every file in the bundle, and a reproduction kit that regenerates the result from scratch. Nothing here depends on trusting us — pull the bundle, recompute the hashes, and re-run.
Every package carries a semantic version tag. Older versions stay published and immutable; corrections ship as new versions with a changelog, never as silent edits.
A MANIFEST.txt lists a SHA-256 digest for every file in the bundle. The manifest itself is hashed and that root digest is printed on each package card below.
A run.sh plus pinned dependencies and a fixed seed. On matching hardware the reproduction lands on the same numbers, bit-for-bit where the math allows.
Protected internal geometry is background IP and is stripped before publication. Public packages contain only inputs, outputs, and code that we are free to release.
The voxel-to-diagnostic pipeline reproduced on a single consumer RTX 3090, matching the frozen reference result within tolerance. Bundle includes inputs, the diagnostic output, and a run script.
Manifest root: sha256:4f1c…a09e
Ten independent runs from the same seed and pinned toolchain, showing output stability across restarts. Includes per-run logs and a diff report demonstrating zero drift on the reported metrics.
Manifest root: sha256:7b30…c14d
Side-by-side against a published open baseline on a shared task, with scoring code included. Labeled a demonstrated result on this dataset — not a claim of a verified public benchmark win.
Manifest root: sha256:e2a8…f6b1
An observation-stage package: what changes when public-side inputs are perturbed. Marked hypothesis, not result. Numbers are indicative and expected to move before a v1 tag.
Manifest root: sha256:1d55…8c27
Plain-language statement of exactly what the package shows, its evidence tier (observation, hypothesis, or demonstrated result), and its known limits.
One SHA-256 line per file. Recompute locally and compare against the root digest on the card to confirm nothing was altered in transit.
The scrubbed, public inputs and the reported outputs. Protected artifacts are excluded by construction and never enter the bundle.
Pinned dependencies, a fixed seed, and a single entry point. Run it to regenerate the outputs and re-derive the manifest yourself.
Download a bundle, then re-derive its manifest and compare the root digest to the value printed on the card. If they match, the files you hold are the files we published. Then run the reproduction kit to regenerate the outputs from the inputs.
Something we saw once or under narrow conditions. Not load-bearing. Published so the record is complete, clearly labeled as such.
A pattern we expect to hold and are testing. Numbers may move. Included to show the reasoning, not to assert a settled outcome.
Reproducible on stated hardware with the included kit. The strongest tier we publish, and still scoped to the dataset and setup described.
Scope note: everything in this room is development-stage research on public, scrubbed data. Protected internal geometry is background IP, is excluded from every package, and never appears here. No verified public benchmark win is claimed anywhere on this page.