AI Agent Engineer

AI Agent Engineer

Project: Video Generation Job Package Validation

4 min read

Project: Video Generation Job Package Validation

Project objective

Use the Python standard library to validate a structured job package for an 8-second short video: output specifications, shot coverage, continuity anchors, caption timing, audio rights, human review, a provenance plan, version/release governance, acceptance dimensions, and failure recovery. It runs fully offline: it does not call a generation API or create video, audio, caption, or image files.

Files

  • video_job_package.json: a three-shot teaching fixture that declares only an offline adapter and does not impersonate a real provider.
  • validate_video_job.py: strictly reads UTF-8 JSON, checks the closed-field contract first, then audits timeline, assets, governance, and budget relationships.
  • test_validate_video_job.py: retained compatibility smoke tests.
  • test_contract_and_cli.py: strict tests for JSON, field types, asset versions, object-level ACL, revocation/deletion propagation, evidence boundaries, cross-field rules, CLI behavior, and read-only operation.

How to run it

In PowerShell 7, run the following from the project root that contains docs-CN/, docs-EN/, and .website/:

Push-Location -LiteralPath 'docs-EN\video-generation\03-project-and-self-assessment\examples'
python -B .\validate_video_job.py .\video_job_package.json
python -B -m unittest discover -s . -p 'test_*.py'
python -B -O -m unittest discover -s . -p 'test_*.py'
python -B -W error -m unittest discover -s . -p 'test_*.py'
Pop-Location

The suite currently contains 133 tests. Normal mode, optimized mode, and warnings-as-errors mode should all pass. A valid job returns 0; a cross-field or governance audit failure returns 1; file, UTF-8, strict-JSON, or field-contract errors return 2. The script is read-only, and -B avoids .pyc caches.

Referenced asset objects contain an asset_id, source_revision, role, source reference, SHA-256, rights record, and acl_reference. Top-level lineage distinguishes a frozen requirement version, shot/post-production transform_id values, and candidate release_id values. governance requires object-level authorization/ACL before scoring, a revocation/deletion propagation plan, and limits factual release to evidence_supported. This is a teaching business contract, not a mirror of any vendor API.

Experiments

  1. Move the start of the second shot before the end of the first to observe an overlap error.
  2. Change the end time of one caption cue to 9.0 to observe an out-of-range error.
  3. Delete one shot’s continuity anchor to understand why “keep it consistent” needs checkable fields.
  4. Set human_review_required to false and confirm that the pre-release human gate cannot be omitted.
  5. Copy the fixture to design a 12-second version; shot times must continuously cover the entire film rather than merely changing total duration.
  6. Add a reference-asset object to one shot, filling asset_id, source_revision, role, source reference, rights record, acl_reference, and a 64-character lowercase SHA-256; then deliberately duplicate the asset_id to observe a contract error.
  7. Set object_acl_required to false and explain why the system must reject the job before frame extraction or model scoring.
  8. Change evidence_policy to another value and explain why Content Credentials or a correct timeline cannot independently support factual release.

Project acceptance

  • Shot IDs are unique, sorted by time, free of overlaps and gaps, and cover the total duration completely.
  • Every shot has shot scale, camera movement, subject, action, setting, prompt, and continuity anchors.
  • Referenced assets have a valid role, stable source, source_revision, content SHA-256, rights record, acl_reference, and globally unique ID.
  • Caption cues are within the valid duration, do not overlap, and declare WebVTT format.
  • Audio rights, reference-asset rights, consent, human review, and provenance planning are explicit.
  • source_revision, transform_id, and release_id distinguish requirements, transformations, and release candidates; object-level ACL and revocation/deletion propagation both have a plan.
  • A factual release is marked evidence_supported only when independently supported by evidence.
  • Acceptance covers prompt adherence, image quality, motion, identity, continuity, audio/captions, safety, and rights.
  • Retries have per-shot limits and a human-escalation condition.
  • There are no API keys, real-person assets, media files, concrete prices, or model weights.

Integrated self-check

  1. Why does a passing job package prove only that the plan is internally consistent, not that a model will certainly generate it?
  2. If a crossfade is required, how should the current “no overlap” rule be extended rather than simply removed?
  3. When a vendor endpoint shuts down, which fields can remain and which must be replaced by the adapter?
  4. How can validation results be connected to an Agent state machine to prevent endless retries after failure?

References

Backlinks