Consensus Gap: Measuring the Distance Between What Experts Know and What People Believe
We built Consensus Gap for the Agent Harness Hackathon by WeMakeDevs, TrueFoundry, and Qodo — an agent that researches any topic and shows what experts actually say versus what the public actually believes, side by side, with sourced evidence.
The problem
There's no tool that systematically shows the gap between expert consensus and public belief. Nutrition myths, economic misconceptions, health practices — this space is largely empty. We wanted to fix that, at least for a hackathon-sized slice of it.
What we built
Give the agent a topic — say, "Does sugar cause hyperactivity in children?" — and it does the following:
- Two subagents research in parallel. One searches academic, institutional, and scientific sources for the expert consensus. The other searches forums, social sentiment, and survey data for what the public believes.
- A sandboxed script scores the gap. Once both subagents return sourced claims, the agent hands off to a Daytona sandbox running a Python script that scores the agreement gap and generates a comparison chart.
- Nothing publishes without approval. Before anything is saved or committed, the agent stops and asks explicitly: do you approve this report? No file gets written, no commit happens, without a yes.
How TrueForge fit in
This was built entirely on TrueForge. Dynamic sub-agents were the single most useful feature for us — the whole concept depends on researching two very different kinds of sources at once, and being able to spawn parallel subagents for that was a direct match for the idea. The sandbox (via Daytona) handled the scoring script in isolation, and the approval-gate capability gave us the human-in-the-loop safety step that we think matters a lot for anything that claims to summarize "consensus."
What broke along the way
A few things slowed us down. Running TrueForge locally on Windows hit an ESM/path resolution bug on startup — a known Node/Windows issue. We also burned time on API key setup: one provider rejected a key with a cryptic "workspace ID required" error, and we hit tight free-tier rate limits on more than one provider mid-session. Separately, a reasoning model's response format wasn't accepted back into conversation history by a provider, which broke a session with a generic 400 error.
None of these were fatal, but they ate into our build time — a reminder that a good chunk of hackathon time goes into infrastructure plumbing rather than the actual idea.
Working with Qodo
We connected Qodo to our GitHub repo early and opened our sandbox script as a pull request rather than pushing straight to main. Qodo reviewed it automatically — an AI summary, a diagram of the change, and a high-level assessment — before we merged. It's a small thing, but it kept us from merging unreviewed code into main even under time pressure, which is exactly the discipline this track is meant to encourage.
What's next
Consensus Gap works for any topic where popular belief and expert evidence might not line up. We'd like to expand it with a persistent report archive, source-credibility weighting, and a lighter-weight public dashboard so results are browsable without running the agent yourself.
GitHub repo: https://github.com/NaitikMorajkar/consensus-gap
Qodo-reviewed PR: https://github.com/NaitikMorajkar/consensus-gap/pull/1
Built by Team Astella — Vedant, Naitik, and Swaroop.
Top comments (0)