Continuity Check

Inspiration

Solo video creators often have to be their own writer, editor, and fact-checker. A wrong date, misidentified person, or shaky technical explanation can slip through when research competes with getting a video finished. We built Continuity Check to help catch those errors before publishing, without needing a researcher on staff.

## What it does

Accepts a script or scene excerpt for a factual review.
Extracts checkable claims with Gemini, including statements about real people, places, dates, organizations, and technical or historical facts.
Searches for evidence by sending each claim to Parallel's Search API to retrieve live web sources.
Returns cited verdicts from Gemini: CONFIRMED, CONTRADICTED, or UNVERIFIABLE, so creators can see which claims have support, conflict with evidence, or remain unresolved.

## How we built it

Gemini via Vertex AI extracts factual claims and evaluates the evidence retrieved through the Parallel Search API. Google ADK (Agent Development Kit) wires the extract -> search -> verdict pipeline into an agent (ContinuityCheckAgent), run for real on every request through google.adk.runners.InMemoryRunner with a live session -- the demo page calls this path (POST /check-agent), not a bypass. FastAPI provides the application's API layer, and the service is deployed on Cloud Run.

## Challenges we ran into

Getting a Google Cloud project authenticated and Vertex AI enabled was the actual bottleneck, not the agent logic -- see SETUP_TODO.md. Keeping the pipeline provider-agnostic in its unit tests (fakes for Gemini and Parallel, zero network calls) meant the control-flow logic could be proven correct before real credentials existed, so the live-credential step was purely mechanical once it happened.

## Accomplishments we're proud of

A fixed, multi-step agent whose decision rules are specified and proven in Lean 4 (no evidence ⇒ UNVERIFIABLE, verdicts passed through never invented, bounded fan-out, bounded rate) with Python conformance tests — and where every verdict is traceable to a specific cited source -- no hallucinated citations, and claims the search can't confirm are explicitly labeled UNVERIFIABLE rather than guessed at.

## Already in use by two other systems (real, not planned)

Carbon Footprint of Capital (public calculator, https://carbon-footprint-calc-wine.vercel.app) runs its own footnotes and methodology notes through Continuity Check and shows a verified badge that links back here. The first audit (2026-09-08) graded 25 claims: 23 confirmed, 1 unverifiable, and 1 contradicted — a mislabeled Cambridge index in a footnote that was corrected the same day. Audit tool + report live in that repo.
Portfolio Carbon Steward (a Strands agent for volunteer finance committees) has a --verify flag that sends the sourced facts in each brief to Continuity Check and appends a live fact-check section.

Both integrations are one-way HTTP calls into this service; nothing here depends on them.

## What we learned

Grounding an LLM's verdict strictly on retrieved evidence (not its own training knowledge) is a small prompt change with an outsized trust payoff for a fact-checking tool specifically.

## Artist statement

We built Continuity Check for the solo video creator who has no researcher on staff. For Google Cloud's Agentic Cinema hackathon, in the Parallel track, we wanted to make checking a script's factual claims a practical model invent a citation. When a search or model call fails, the system fails closed: it returns UNVERIFIABLE with the real reason, rather than guessing.

We also chose a test that could expose failure. Our sample script has a character misremember the Berlin Wall's fall date by twenty years. We want the tool to catch a real error, not merely confirm easy facts. For us, useful assistance means making uncertainty visible.

## What's next for Continuity Check

Batching multiple claims into fewer Parallel Search calls to cut latency further; a browser extension that runs continuity checks on a Google Doc script draft inline.

## Built with

Google Cloud (Vertex AI / Gemini), Google ADK, Parallel Search API, Python, FastAPI, Cloud Run, Docker, Lean 4 (specification + proofs of the decision rules; not part of the runtime).

## Links