Visible workflow state
Pending, correction, ready-for-approval, and approved are different product states—not one success counter.
Source-free engineering case study · private alpha.24
Research Studio turns AI output into reviewed evidence, not automatic truth.
A Windows desktop workflow for bilingual short-drama catalog enrichment. It protects the source SQLite database, keeps ChatGPT inside one application window, monitors live response state, stages validated drafts, and requires explicit approval before any result becomes exportable.
Original conceptual illustration · no private catalog or account data
31,521series verified
2.24Mepisodes inspected
20focused verification programs
0moderate-or-higher npm vulnerabilities
Product tour · deterministic synthetic catalog
The UI makes trust boundaries visible.
Every title, description, identifier, result, and count below is fabricated. The captures exclude the browser/account pane and machine paths.
Versioned prompt provenance
Active or used prompts are immutable. Revisions preserve the exact contract attached to historical jobs.
Human approval as storage authority
A valid capture is only a staged job draft. Approval revalidates it and creates the immutable result version.
Approved-only exports
Every format shares one eligibility rule, so staged or correction records cannot escape through another route.
Authority map
Each layer has a job—and a deliberate limit.
Checked snapshot + query-only SQLite
No schema or enrichment write
Correlated complete JSON + domain gates
No file, database, approval, or export authority
Explicit approval revalidates a staged draft
Only this transition creates a result version
Target-runtime SQLite query + transactional activation
Prior build remains recoverable until verification passes
Evidence-first prompt v3
Ask for identity, evidence, confidence, uncertainty, and canonical tags.
The prompt treats catalog text as untrusted, requests title-plus-premise matching, falls back conservatively when identity is uncertain, and forbids unsupported cast, platform, episode, ending, supernatural, pregnancy, and identity-twist claims. ChatGPT returns canonical tag IDs; Research Studio derives the Chinese and English labels locally.
Review the prompt contractWHOLE-LIBRARY SAFETY1-title pilot → 2-title batches
- Workload review before launch
- Checkpoint after every batch
- Safe stop and resume
- Validation and zero-save circuit breaker
- Correction records excluded
- No automation path can approve
Packaging is executable evidence
A build is not accepted because the UI opens.
- 01Verify source
Types, lint, formatting, tests, prompt evals, build, and dependency audit.
- 02Execute native SQLite
Rebuild for Electron ABI 146, query SQLite, then restore Node ABI 137.
- 03Render production
Bundled React over an application-private named pipe—not an external dev server.
- 04Activate transactionally
Preserve durable data only; restore the prior build after injected failure.
Ownership and limits
I owned the decisions; AI-assisted output still had to earn trust.
I owned product direction, extraction architecture, database and browser safety, prompt/evaluation strategy, human-approval design, QA criteria, licensing boundary, technical review, and release approval. AI agents assisted with research, implementation, and iteration; their output was reviewed and verified.
The application remains private. The case study does not claim public distribution rights, code signing, universal model accuracy, provider-policy approval, or production readiness. Research Studio was extracted from a module in SilkReel Windows 5.8.214; I do not claim authorship of the entire upstream product.
Follow the bounded evidence