v0.5.0: Briefs, evidence and optional model review

Scene Acceptance 0.5.0

Scene Acceptance checks a saved USD delivery against general diagnostics and, when supplied, the requirements of a job. Version 0.5 adds a caller workflow for interpreting a text/image brief, approving its test scope, requesting views, evaluating a saved revision and checking a repair against that scope. The application supplies rendering and an optional model CLI.

Scripts return measurements and rule results. The optional judge returns opinions tied to the supplied evidence. Missing evidence stays visible; the judge does not turn an unmeasured requirement into a pass.

Caller situation	What 0.5 reports
A scope interpretation needs correction	NEEDS_REVIEW, exit 3 and review_specification, separately from scene findings
A repaired scene introduces motion/material areas outside the frozen judge scope	A coverage gap requiring preparation/review of a new scope before acceptance in judge/both mode
The renderer supplies an MDL library absent from the bundle	An optional caller attestation tied to the scene, exact reference and target environment; original resolver evidence remains visible
An old contract pins an earlier pack	A strict version mismatch, with an explicit upgrade proposal available on a separate copy

The experimental package includes a 27-rule general baseline, selected geometry/material/motion/texture checks, a bounded CPU simulation worker, coverage reports and two exportable adopter skills. It works as a CLI or a Python library. An HTTP wrapper remains a proposal.

The packaged reproduction passed 668 software tests with no failures, errors or skips, then matched the retained replay decisions and numerical physics observations. It used a fresh source extraction and a noneditable wheel with recorded, retained dependency distributions. This was not a fresh network dependency installation. No model calls or GPU renders were part of that replay; software and replay counts overlap.

The public source preserves the earlier Astra mechanical-assembly example. Local environment prefixes were removed from retained diagnostic text; scenes, implementation and observed outcomes were preserved. Later owner-local study packets and private session records are excluded.

The MDL route records caller-attested availability, not observed shader compilation or rendered appearance. Motion sampling does not prove every intermediate pose. The simulation worker supports its admitted block-and-ramp model and does not establish physical calibration. These controls complement deeper validators and simulators.

See README.md for installation and the application protocol, test inventory, integration notes and runtime dependency policy for the caller contract. Attached verification.json, MANIFEST.json and SHA256SUMS.txt identify the tested source and wheel.

v0.3.0: Initial public prototype

The first public release of Scene Acceptance, an experimental harness for evaluating saved OpenUSD content against application-owned requirements.

Includes five built-in pack families, an independently installable mesh-budget example, architecture diagrams, setup instructions and retained evidence. The evaluator produces findings; the calling application owns producer retries and release.

The runtime is the same tested 0.3.0 implementation as the earlier article companion. This release adds standalone project documentation, licensing and automated replay. The archive includes source, fixtures, historical results and a file-hash manifest.

Validation from a fresh archive extraction and packaged installation on Python 3.12 / macOS arm64:

• 195 software tests passed.
• 22 pack cases and 32 mesh cases retained their expected decisions and check statuses.
• Four content probes, three physics-configuration cases and four CPU simulations reproduced the retained observations.
• Both README quickstart commands produced the expected acceptance/rejection.
• All 796 manifest-listed files verified before and after replay. Counts overlap; no new model calls were made.

The first pack run's three false acceptances remain available beside the corrected results. Historical model outputs are separate from constructed faults. Texture decoding and the restricted MuJoCo behavior runner are separate experiments, not installed runtime packs. No physical validation, robot policy performance or comparative productivity is established.

Download scene-acceptance-0.3.0.zip for the complete project. scene-acceptance-SHA256SUMS.txt verifies that archive. verification.json records the fresh replay; README and docs/EVIDENCE.md explain how to repeat it. The GitHub-generated source archives contain the same tagged source but have different archive hashes.