Research Mode (Anti-Hallucination)

A toggleable anti-hallucination mode that enforces citation requirements, source grounding, and explicit 'I don't know' behavior based on Anthropic's guardrail guidance. Turn it on for factual work and off for creative work.

Research Mode activates a set of anti-hallucination constraints derived from Anthropic's documentation on reducing hallucinations. While active, the agent is required to cite sources, ground claims, distinguish what it knows from what it infers, and say 'I don't know' rather than guess. It stays on until you explicitly exit, making it useful for high-stakes factual or compliance work.

When to use

Use when accuracy matters more than fluency — fact-checking, literature review, compliance answers, or any task where you want sources cited and uncertainty surfaced instead of confident guesses.

Examples

Toggle on for facts

Enforce grounding for a factual task

Enter research mode, then tell me what the GDPR retention requirements are with citations

Surface uncertainty

Make the agent admit gaps

Stay in research mode and answer only what you can support from sources — flag anything you're unsure about