Winner Signal
Winner Signal connects AI-assisted work to evidence that its outputs function, get used, and create value—developing better signals for AI evaluation and training beyond benchmark pass rates.
Core trick
Argon reviews demo videos against commit-pinned code and agent session traces, and an evidence layer checks the inputs' integrity.
Description
Winner Signal connects AI-assisted work to evidence that its outputs function, get used, and create value—developing better signals for AI evaluation and training beyond benchmark pass rates.
Builders share a demo and public GitHub repository, optionally adding reviewed prompts and coding-agent session histories. Gemini 4 Argon reviews video claims alongside commit-pinned code and supplied traces, producing timestamped findings, source references, actor/action/outcome summaries, and follow-up tests. An evidence layer preserves the review inputs and checks their integrity.
Our own project demonstrates a curated backtrace from video claims to actual build decisions, failed tests, fixes, and commits. A live review of our uploaded demo, pinned code, and a curated session excerpt returned five video claims and six trace findings with a verified evidence seal. The backtrace displayed in the workspace is curated, not an automatically reconstructed full execution history.
A public demo gallery gives builders initial viewers. Optional owner-authorized YouTube data provides a separate video-response signal; it does not prove product value or agent capability. The research objective is to discover which real-world signals are useful for evaluating and improving the agents that built the artifacts. Training gains have not yet been demonstrated.
Live app: https://winner-signal.vercel.app
Built during this hackathon on a reused, pinned Apache-2.0 AERead evidence-sealing component.
README excerpt
# Winner Signal A measurement prototype connecting AI-assisted builds to their outputs and audience response. Our goal is to develop validated real-world feedback for AI evaluation and training beyond benchmark pass rates. Builders get a demo gallery, feedback and a private analytics dashboard; training usefulness remains a research question. ## Run Requirements: Node.js 20.9+, npm, Python 3.12, `uv`, and FFmpeg (`ffprobe`). ```sh cp .env.example .env # only when you do not already have .env ./start.sh ``` Open **http://localhost:3000**. Create a local builder account (username and a password of at least 10 characters). The API and worker run on the same machine; the API binds to loopback. Browser requests use the Next.js same-origin proxy. The gallery, file upload, publishing, feedback and playback analytics work without a model API key. In the workspace, **Try a controlled example** creates a private example video, an intentionally mocked booking function, and an explicitly authored fixture review. Its review is not attributed to Gemini, and fixture traffic is excluded from audience totals. Create a real submission with a public GitHub repository. The app resolves an immutable commit and snapshots at most 16 small source files (160 KB total). This is bounded source coverage, not a full audit. Add a hosted demo URL, then explicitly publish. No video upload is required or offered in the current UI. Video-to-code review supports existing uploaded assets and public YouTu