The Rumor Room

A noir web-investigation game where live evidence—not trivia—decides which entertainment claim cannot survive scrutiny.

Built for the Agentic Cinema: The Blockbuster Hackathon, Parallel partner track.

Play the game

Launch The Rumor Room → · Watch the 3-minute demo →

No installation is required. This is the live Google Cloud deployment using Gemini and Parallel Search.

Proof this is live

Every research turn on the hosted site runs Gemini through Google ADK, which calls Parallel Search as a tool. The board can only show URLs that came back from that Parallel call.

What	Evidence
Hosted app	rumor-room-dpq2d26l7q-uc.a.run.app on Google Cloud Run, revision rumor-room-00006-2tj
Status badge in-game	Top right reads Parallel live search (fixture mode cannot boot in production)
Runtime Parallel call	Cloud Run log entry parallel_search_completed with search ID search_85ade4fdd5244f13f2e9fa33bed17c66, six results, consumed by gemini-3.7-flash; the entry lists Gemini's three queries (agentQueries) and the two case-authored coverage queries (coverageQueries) separately
Three more live search IDs, one per case	search_714f7e26…, search_72d1bc94…, search_0344fc17… in docs/internal/LIVE_VALIDATION.md
Google AI layer	server/providers/live-provider.ts: LlmAgent, Runner, FunctionTool from @google/adk; no other model provider in the dependency tree
Citation trust boundary	requireParallelCitations() drops any URL Gemini returns that Parallel did not
Demo video	youtu.be/ovbrx_9RvN8 — 179-second walkthrough with three live production Parallel searches, captioned

See the case briefing and evidence receipt

The pitch

Four plausible claims arrive at the newsroom. Exactly one is unsupported, outdated, or materially misleading within the case's stated time window.

The player gets four research tokens—meaning four total research turns. On each turn, select any claim and apply one of four move types:

• Trace It — find the earliest discoverable origin.
• Second Source — look for genuinely independent corroboration.
• Studio Line — prioritize the organizations closest to the facts.
• Fresh Cut — search after the claim for corrections, denials, and changed plans.

You may switch claims between turns, focus multiple turns on one claim, or reuse a move type. You may also accuse at any time—including before using every turn.

Every research turn calls a Gemini investigator built with Google ADK. Gemini chooses a research objective, calls Parallel Search as a tool, evaluates provenance and freshness, and returns dated source receipts that physically change the case board. The player then accuses one claim and receives a scored evidence receipt.

Why the integrations matter

Parallel is not an ornamental search box. Fresh web evidence is the game state.

• A stale date can look perfectly sourced until Fresh Cut finds the later announcement.
• Ten headlines can collapse into one origin when Second Source traces circular reporting.
• A famous producer can be promoted into the director's chair until Studio Line retrieves the exact credit.

Gemini is the investigator: it translates each strategic move into a focused research objective, operates the Parallel tool, and turns the tool output into a small, provenance-aware evidence bundle. The server rejects any citation that was not present in the actual Parallel response.

Current build

• Three hand-designed cases. One teaches outdated information, one teaches how many articles can repeat the same source, and one teaches how a headline can overstate what its source actually says.
• Four claims, four research turns, one accusation, a per-case score, and an accumulating three-case campaign score.
• Responsive physical evidence-board interface.
• Adaptive Tone.js electro-noir soundtrack and semantic cues.
• Mute control, audio-independent play, visible state communication, keyboard focus, and reduced-motion support.
• Saved-evidence test mode for repeatable local development and automated tests. The alternative is live mode, where Gemini calls Parallel Search on every research turn.
• A deployment safety check that refuses to start production unless live Gemini + Parallel mode is enabled. This prevents accidentally submitting a version that uses saved test evidence.
• Google ADK 2.0 + Gemini 3.7 Flash + Parallel Search live provider.
• Unit, integrity, trust-boundary, desktop browser, and mobile browser tests.
• Cloud Run container and Cloud Build deployment configuration.

The complete Gemini → Google ADK → Parallel path has been exercised across all three hand-designed cases and through a full browser verdict on the public Cloud Run deployment.

Game rules at a glance

• There are three cases in the current game.
• Each case presents four claims. Exactly one is unsupported, outdated, or materially misleading for that case—not necessarily a completely invented statement.
• You have four research turns per case. The round tokens are simply a visual counter for those four turns.
• Select any claim, then apply a research move to that selected claim.
• You can switch claims between turns, reuse a move type, or focus several turns on one claim.
• You can accuse at any time. A strong first search can solve a case; using all four turns is not mandatory.
• The receipt shows a case score. The campaign score accumulates across all three cases and resets when you choose Start over or finish the campaign.

Runtime integration proof

Gemini and Google ADK

• server/providers/live-provider.ts imports LlmAgent, Runner, FunctionTool, and InMemorySessionService from @google/adk, creates the Gemini investigator, and executes it for every live research turn.
• cloudbuild.yaml configures Gemini Enterprise mode, the Google Cloud project, Gemini 3.7 Flash, and the dedicated rumor-room-runtime service account.
• Google does not use a committed API key. Local development uses Application Default Credentials. Production uses the Cloud Run service account identity.

Parallel Search

• The parallel_search FunctionTool in server/providers/live-provider.ts calls parallel.search(...) from the official parallel-web SDK.
• The Parallel key is stored in Google Secret Manager as parallel-api-key and injected into Cloud Run by cloudbuild.yaml. Its value is never sent to the browser or committed.
• docs/internal/LIVE_VALIDATION.md records real search IDs, effective queries, cited URLs, and all three live case results.
• docs/internal/DEPLOYMENT_REPORT.md records the Cloud Run revision and the structured production log proving the Parallel tool executed.

Architecture

flowchart LR
    P[Player] --> UI[React investigation board]
    UI --> API[Express API]
    API --> ADK[Google ADK LlmAgent]
    ADK -->|FunctionTool| PAR[Parallel Search]
    PAR -->|dated URLs + excerpts| ADK
    ADK -->|structured evidence| GUARD[Citation allowlist + schema validation]
    GUARD --> UI
    UI --> SCORE[Session-scoped verdict scoring]

Loading

The client never receives the answer key. Evidence and scores are scoped to a generated player session. Production requires INVESTIGATION_MODE=live; saved test evidence is intentionally limited to local development and automated testing.

See Architecture for the full request path and trust boundaries.

Run locally

Requires Node.js 22 or newer.

npm install
cp .env.example .env
npm run dev

Open http://127.0.0.1:5173. By default, local development uses saved evidence so every test run receives the same sources and scores. Set live mode to use the real Gemini + Parallel path instead.

Run the live investigator locally

Authenticate with Google Cloud Application Default Credentials, add a Parallel API key, and change the following values in .env:

APP_ENV=development
INVESTIGATION_MODE=live
GOOGLE_CLOUD_PROJECT=your-project-id
GOOGLE_CLOUD_LOCATION=global
GOOGLE_GENAI_USE_ENTERPRISE=true
GOOGLE_MODEL=gemini-3.7-flash
PARALLEL_API_KEY=your-key

Then run npm run dev. The top-right status changes from Saved test evidence to Parallel live search.

Quality gates

npm run check       # lint + unit tests + production build
npm run test:e2e    # Chromium desktop + mobile flows
npm audit           # dependency advisory check

Validated locally on August 28, 2026:

• 16 unit/integrity/trust-boundary tests passing.
• 14 browser scenarios passing across desktop and mobile with zero skips.
• Full three-case campaign and incorrect-verdict path covered in Chromium.
• Automated WCAG A/AA scans passing on the briefing and board at desktop and mobile sizes.
• Clean initial-load console and successful investigation/verdict network flow.
• All 22 unique authored source URLs returning HTTP 200.
• Zero npm audit vulnerabilities.
• Production fixture-mode guard verified.
• Production bundle budgets enforced: 206.8 KiB initial JavaScript, 23.8 KiB CSS, and lazy-loaded audio.
• Credential-free ADK simulation verifies that the real Google runner invokes the Parallel tool contract and returns grounded evidence.
• Credentialed live validation verifies all three intended contradictions, three runtime Parallel search IDs, an error-free live browser verdict, and successful live citation URLs.
• Production verification confirms a live Parallel search record in Cloud Run logs, an error-free public verdict, and zero WCAG A/AA violations on desktop and mobile.

Deployment

The repository includes a multi-stage Dockerfile and cloudbuild.yaml for Google Cloud Run. Follow the deployment runbook; it keeps the Parallel key in Secret Manager and runs the service with a dedicated identity that can call Vertex AI.

Documentation

Start here:

• Architecture and trust boundaries — request path, query composition, citation allowlist, session state.
• Devpost submission
• Three-minute demo script
• Case authoring guide
• Original design document

Verification and operations records, in docs/internal/:

• Live Gemini + Parallel validation — search IDs and results for all three cases.
• Production deployment report — Cloud Run revision, logs, performance, accessibility.
• Testing and QA evidence
• Case research and source record
• Google Cloud deployment runbook
• Product decisions
• Changelog · Security policy · Contributing

License

MIT. See LICENSE.