Inspiration

Volunteer event teams receive partial, contradictory reports. A polished summary can hide the most important fact: nobody has resolved the disagreement. SignalGlass keeps that uncertainty visible.

What it does

Reports form an evidence constellation. Replay reveals when a disagreement appeared. A coordinator chooses a source and records a reason, valid only for the exact evidence set reviewed. New evidence reopens the decision. The deterministic handover preserves original report IDs, roles and times.

The optional AI draft cannot change evidence or human decisions. Topic coverage and citation IDs are validated; unresolved labels come from application state. A conservative certainty-word guard rejects some overconfident phrasing. These checks do not prove factual accuracy.

How we built it

A new HTML/CSS/JavaScript app with SVG visualization and in-memory state. Firebase AI Logic SDK 13 runs ONLY_ON_DEVICE by default, or PREFER_ON_DEVICE after explicit consent to Google cloud fallback. Production App Check uses a domain-restricted reCAPTCHA Enterprise site key with enforcement retained. The exact SDK inferenceSource is displayed.

Built during the submission period, beginning 17 November 2026 Brisbane. OpenAI Codex assisted with research, design, implementation and testing. Firebase SDKs are dependencies. No ScopeKit, Substrate or Heliora code was reused. Development began in the portfolio repository; the same new project is now packaged in its own Apache-2.0 repository.

Challenges and learning

A model initially called reported agreement verified. We tightened source attribution and added a conservative rejection guard; this does not eliminate hallucination. Topic matching normalizes capitalization without relaxing report-ID validation.

The recording browser lacks the local LanguageModel API. Device-only testing returned an explicit unsupported result while the evidence board remained usable. Consented tests generated real cloud drafts with inferenceSource in_cloud. Successful on-device execution and offline reload have not been verified.

What we verified

Six reports/two conflicts; replay of earlier evidence; human source-and-reason decision; a seventh report reopening that decision; actual AI generation and citation checks; removal of stale drafts; JSON export. Import roundtrip was blocked by browser file-read permissions and remains unverified.

The 1:17 captioned video renders real app state while performing that workflow. It is a guided visualization, not a desktop screen recording. It uses synthetic data and shows actual AI output excerpts with citations and the local-runtime limitation.

Next steps

Test on compatible Chrome hardware with its local model, evaluate semantic grounding, and work with volunteer coordinators after appropriate consent. This prototype is for event coordination, not emergency response.

Built With

Share this project:

Updates

Submission history