Inspiration
AI video creators rarely lose only a prompt. They lose the relationship between a creative decision, the exact model run, the useful output, and the reason a retry worked. Provider links are temporary, results become detached from their shot plan, and teams cannot reliably inspect or hand off how an asset was produced.
What it does
Jingci turns one creative idea into a DirectorKit: a feasibility diagnosis, three rewrite directions, structured shot cards, platform guidance, stability checks, and recovery strategies. Provenance Vault extends the selected-shot workflow so a creator can inspect provider and model evidence, verify asset and manifest hashes, and retry without losing the parent run.
How we built it
The product uses Next.js, React, TypeScript, and Tailwind. A Python adapter implements Genblaze's official SyncProvider, Pipeline, ObjectStorageSink, and StorageBackend extension points. The credential-free path executes a deterministic provider, stores exact media bytes and a canonical manifest through an in-memory backend, and returns a strict versioned provenance record to the browser.
During an approved private verification, Runway gen4.5 generated one five-second 1280x720 H.264 shot. After that provider task succeeded, a separate recovery phase passed the verified MP4 through Genblaze, uploaded the content-addressed asset and provenance manifest to private Backblaze B2, read both back, verified their hashes and lineage, and deleted the two scoped objects. These were two evidence-preserving phases, not one uninterrupted atomic transaction. This proves scoped integration, not public B2 serving, durable retention, version-level erasure, or general provider reliability.
For judges, we ship an anonymous zero-network Fixture demo that makes its deterministic-data boundary visible. It exercises the DirectorKit, selected-shot provenance UI, failure state, retry lineage, and sanitized project receipt without exposing credentials or invoking external APIs.
Challenges
The hardest problem was preventing mocked, partial, or recovered evidence from being presented as a successful live transaction. We designed fail-closed contracts for malformed responses, interrupted provider requests, ambiguous uploads, cleanup failures, duplicate approvals, and retry lineage. We also kept credentials, signed URLs, private object keys, and raw prompts out of logs, browser state, public evidence, and repository artifacts.
Accomplishments
- DirectorKit-to-shot provenance workflow
- Genblaze pipeline and object-storage integration
- Content-addressed media and canonical manifest contracts
- Parent-child retry lineage that retains failed attempts
- Public zero-network judge mode with desktop/mobile E2E
- One evidence-bounded Runway generation and separate Genblaze-to-B2 recovery verification
- Public bilingual 147-second demo
What we learned
Reliable generative media workflows need more than prompt history. They need inspectable lineage across creative intent, provider execution, storage integrity, retries, and cleanup. Provenance works best when it is part of the workflow contract, not metadata added after the asset has become detached.
What's next
We will turn the selected-shot evidence boundary into a project-wide creative ledger, add platform feedback calibration, and expand the verified recovery workflow without weakening the distinction between Fixture, private live evidence, and public runtime claims.
Existing work disclosure
Jingci existed before the hackathon began on June 22, 2026. Beginning July 13, 2026, the campaign branch added Provenance Vault: the Genblaze adapter, content-addressed storage boundary, fail-closed B2 configuration, provenance contracts, selected-shot UI, retry lineage, HTTP service, evidence gates, and desktop/mobile tests. Campaign work begins at commit 3e42c78 on spike/backblaze-provenance.
Built With
- b2
- backblaze
- cloudflare
- css
- d1
- deepseek
- genblaze
- next.js
- openai
- playwright
- python
- react
- runway
- tailwind
- typescript
- workers
