Research Radar

Inspiration

Experimental researchers rarely lack notes. They lack a reliable way to turn failed runs, task boards, deadlines, and reflections into the next decision. After several failed experiments, it is easy to repeat a protocol with minor changes, lose the evidence trail, or discover a blocked milestone too late.

What it does

Research Radar is a Codex plugin that reviews experiments and research tasks from Notion. It detects repeated failure sequences, missing next steps, blocked or deadline-exposed work, and mismatches between experimental evidence and the task board. It then produces a Rescue Plan containing:

• A transparent risk level.
• Source-linked factual evidence.
• Clearly labeled inferences.
• At most three prioritized actions.
• An exact preview of proposed Notion changes.

Nothing is written until the researcher confirms the displayed plan ID. Judges can run the complete workflow with anonymous fixture data and no account setup.

How we built it

Research Radar is an installable Codex plugin with a reusable Skill and a TypeScript stdio MCP server. The MCP server reads Notion through an internal integration token, normalizes database records, and emits a stable structured evidence contract. Deterministic checks identify supported risk signals. GPT‑5.6 in Codex uses the Skill to explain those signals, distinguish fact from inference, and collaborate with the user on the final plan.

The same workflow supports a zero-credential demo mode. Approved demo plans become local Markdown files; approved live plans create a Notion page and annotate selected task pages. Plan IDs and duplicate checks make application idempotent.

How we used Codex and GPT‑5.6

Codex helped transform the initial concept into a scoped product, scaffold the plugin and Skill, implement the MCP server and Notion adapter, generate representative anonymous data, and build unit plus protocol-level tests. It also helped iterate on safety boundaries and judge setup.

We made the key product decisions: target experimental graduate researchers, lead with the failed-experiment rescue workflow, require provenance for every factual claim, cap daily actions, and require preview-before-write.

GPT‑5.6 is not used as a generic summarizer. It operates over a typed evidence package, explains the implications, and helps select a discriminating next experiment without inventing causes or sources.

Challenges

• Designing a demo that is useful without exposing a real research workspace.
• Preserving provenance across heterogeneous Notion property types.
• Making agentic writes feel safe and auditable.
• Separating deterministic risk signals from model interpretation.

Accomplishments

• A complete anonymous hero workflow with no credentials.
• Four focused MCP tools with structured output.
• Explicit confirmation and idempotency for every write.
• Partial-access warnings and actionable Notion setup diagnostics.
• Automated stdio MCP client/server and safety tests.

What we learned

The most valuable research assistant is not the one that creates the longest summary. It is the one that shows exactly why an action is recommended, links to the evidence, and gives the researcher control at the point of change.

What's next

Future versions could add public Notion OAuth, configurable lab templates, scheduled weekly review, experiment-instrument imports, and team-level portfolio risk views. These are intentionally outside the hackathon MVP.

Built With

• notion
• research