Newsroom Fleet
Build storyAugust 2026

The model never
publishes. Your editor does.

Newsroom Fleet gives a five-person local newsroom the institutional checks of a national masthead — without pretending software should have the final word.

Local newsrooms carry the same public-interest responsibility and legal exposure as national publications, but many no longer have a data desk, standards editor, dedicated fact-checker, or security team. Newsroom Fleet rebuilds that missing institutional structure as a set of independent specialist agents.

The decisive moment

A reporter presses publish.
The server says no.

The fleet has found a statistic that contradicts the authority’s release, a quote the transcript does not support, risky guilty-before-conviction wording, and a leaked memo carrying a hidden prompt injection.

The refusal is not a model’s recommendation. It is deterministic policy evaluated over persisted verdicts. A reporter cannot clear it. An editor must resolve every blocking verdict and leave a recorded decision.

Why a fleet

Independence is enforced in code.

Each desk receives only the evidence its job requires. Different boundaries make disagreement meaningful — not theatrical.

01

Claim extractor

Turns a draft into atomic, checkable claims. It sees the draft — and nothing else.

02

Source verifier

Tests whether cited or researched evidence actually supports the words on the page.

03

Data checker

Recomputes selected figures against an approved authoritative source.

04

Standards reviewer

Flags attribution and legal-status risks against the newsroom’s own rules.

Under the hood

Agentic reasoning inside deterministic guardrails.

Newsroom Fleet system architecture showing its security boundary, agent plane, policy gate, persistence and Google Cloud services
Every cloud capability sits behind an interface and degrades safely to a local implementation. Fixture mode needs no API keys or network.
0

Unsafe false verifications

Safety is measured, not implied.

The curated evaluation checks extraction, evidence correctness, abstention, quarantine containment, recovery, and publish-gate integrity. The most important result is zero high-risk claims incorrectly cleared.

01

Hostile evidence is quarantined

Screening runs before orchestration. Reviewers receive the quarantine record, never the injected instruction.

02

Failures stay visible

A crashed worker becomes an explicit error that needs a human. It can never quietly turn into a pass.

03

Published work stays alive

A scheduled watcher rechecks selected figures and drafts a correction candidate. An editor still decides.

Built for the All Things Agentic Hackathon

Institutional checks.
Human judgment.

Built with Google ADK, Gemini, Cloud Run, Firestore, Pub/Sub, Cloud Scheduler, and Cloud Trace — with a reproducible local mode that runs without keys or cloud access.