Independent implementations

Verifiable conformance results from implementations maintained outside the 4dgs repository. The table deliberately has the same partial-support vocabulary as the SDK feature matrix: an honest partial result is useful, and a read path that was not submitted is visible rather than inferred. The source of this page is the public, machine-readable

results.json

catalog.

No independent results have been published yet. The route is ready for an implementation maintained outside this repository to submit the first verifiable result.

Corpus identity

Every result currently refers to the immutable 0.1.0 corpus:

Field	Value
Corpus version	0.1.0
Release tag	releases/corpus/v0.1.0
Peeled tag commit	d9a76528560496b06f4d9baa25d53c3d944537d8
Manifest schema	1
Archive	4dgs-conformance-corpus-0.1.0.tar.gz
Archive SHA-256	a147df71cb13eebb42997a5b3e6433b1e27e572a891bd1b28d9db7572b63ac3c
Harness variants	74
Indexed-eligible	73

The tag commit identifies the source that generated the corpus; the archive digest identifies the bytes that were actually tested. Both are recorded because either one alone leaves part of the run implicit.

How scoring works

The catalog records the harness's raw passed, skipped and failed totals, not a hand-computed percentage. Each submitted runner accounts for all 74 manifest variants. A streamed runner can answer all 74; an indexed runner can answer 73, because one valid variant deliberately carries no chunk index and is skipped. A complete two-path result is therefore 147/147.

This page derives each path's denominator and the combined score from the corpus metadata. A read path that was not submitted contributes zero to the combined 147 checks and is labelled Not submitted. A declared feature gap is a harness skip, not a pass; the runner's declines and refusals capability fields remain visible in its evidence details. Only runs with zero failures are published, but skips are welcome: 51/74 is an honest, useful result, not a failed attempt to claim 74/74.

The exactAggregates, canonicalStateOrder, aggregateDecodedBudget, lateFrontMatterRecords, and optionalIdentityDefaults and gaussianBirthChunkWindowIntersection capability gates are shown too. When either of the first two is false, the harness omits that transition field exactly as the runner protocol specifies; the score makes no claim for the omitted field. aggregateDecodedBudget: true records that the additional one-byte injected gate passed. lateFrontMatterRecords: true is meaningful only with refusals: true; it opts a streamed runner into the unreleased structured placement family. optionalIdentityDefaults: true opts the runner into both unreleased identity witnesses. gaussianBirthChunkWindowIntersection: true opts into the unreleased legal-overhang instant queries and direct path comparison. None of these unreleased families adds cases to the immutable 0.1.0 score shown here. These capabilities are displayed separately rather than added to the corpus score, whose denominator remains the 74 immutable manifest variants per path.

Evidence required

A catalog entry identifies one implementation version, language and permanent project URL, then records:

• the runner artifact's stable HTTPS URL and SHA-256;
• the corpus version;
• the UTC run date and platform;
• a stable HTTPS URL for the complete harness output and the SHA-256 of those exact bytes;
• a permanent reproduction document or script URL;
• for each submitted read path, the complete protocol-1 capabilities declaration, exact command, and raw passed, skipped, failed totals.

Evidence must be public without authentication and remain downloadable. Prefer a release asset or a file at an immutable source revision over an expiring CI artifact. The evidence bundle must include the capabilities response plus complete harness stdout and stderr, including per-variant failures and final totals, so somebody else can check the catalog without trusting a summary written for the pull request.

Verify and reproduce

First verify the exact corpus archive:

curl -fL \
  'https://github.com/avala-ai/4dgs/releases/download/releases%2Fcorpus%2Fv0.1.0/4dgs-conformance-corpus-0.1.0.tar.gz' \
  -o 4dgs-conformance-corpus-0.1.0.tar.gz
printf '%s  %s\n' \
  'a147df71cb13eebb42997a5b3e6433b1e27e572a891bd1b28d9db7572b63ac3c' \
  '4dgs-conformance-corpus-0.1.0.tar.gz' | sha256sum --check -
tar -xzf 4dgs-conformance-corpus-0.1.0.tar.gz
cd 4dgs-conformance-corpus-0.1.0/corpus
sha256sum --check CHECKSUMS.txt

Then download the result's runner artifact and run-evidence file, verify each digest from the table, and follow its reproduction URL using the recorded command and platform. The conformance suite page documents how to drive an external runner through this repository's harness. The archive's MANIFEST.json also has enough information to drive a runner without cloning this repository.

Publish a result

Publication is an ordinary documentation pull request under project governance, not a private certification process:

1. Capture the runner's --capabilities response, then run each read path separately so its final 74-variant totals can be copied without separating a combined log by hand. Preserve the complete stdout and stderr and compute the evidence bundle's SHA-256.
2. Add one entry, sorted by stable id, to website/static/conformance/results.json. The field contract is enforced by the dependency-free validator; its focused tests contain full and partial synthetic examples.
3. Run yarn workspace website test:conformance-results and yarn workspace website build, then open a pull request following CONTRIBUTING.md.

Review checks the catalog's structure, the cited digests and whether the evidence reproduces the numbers. Merging records that evidence; it does not ask a maintainer to vouch for the implementation.

What a result does not claim

A published score is agreement with one corpus version, not an endorsement, certification, security audit or general quality ranking. It says nothing about inputs the corpus does not contain.

The conformance harness materializes small canonical summaries, so a result does not prove that the implementation obeys the format's bounded-memory requirement on arbitrary input. That needs separate implementation tests and review; no score here turns an unbounded decoder into a conforming one.

Decoding ends at reconstructed gaussian state at time t. Rendering, ordering, culling, level-of-detail policy, GPU code, listener processing and image or audio quality are outside the format and outside this score. The catalog must not be used as a renderer comparison.