Feature: Experiment Novelty Score to Reduce Repeated Search Patterns

Is your feature request related to a problem? Please describe.

As the number of experiments grows, the agent may revisit similar regions of the search space despite maintaining a history of previous runs.

While repeated validation is sometimes useful, it can also reduce exploration efficiency by spending compute on ideas that are only minor variations of previously rejected experiments.

Describe the solution you'd like

Introduce an optional experiment novelty score.

Before launching a new experiment, compare the proposed modification against recent accepted and rejected experiments.

The score could consider:

• Hyperparameter similarity
• Architectural similarity
• Optimization strategy similarity

Low-novelty experiments could be flagged in the experiment log, allowing users to better understand whether the agent is genuinely exploring new directions or repeatedly searching nearby solutions.

Describe alternatives you've considered

Users can manually inspect experiment histories, but this becomes increasingly difficult as runs scale to dozens or hundreds of experiments.

Additional context

This feature would not change the keep-or-revert workflow. Instead, it would provide additional visibility into search diversity and exploration behavior during long autoresearch runs.