novelty-audit

name	novelty-audit
description	LLM guidance for mechanism-aware novelty auditing.

Novelty Audit Guidance

Use this guidance before judging a hypothesis as novel or collided.

LLM Responsibilities

• Extract the hypothesis mechanism as a causal claim, not just a method description.
• Search by method terms, task terms, and mechanism terms.
• Compare candidates along mechanism, task, data regime, experimental setup, claimed contribution, and evidence strength.
• Decide the final novelty label yourself after reading the candidate evidence.

Tool Boundary

• extract_mechanism_tuple persists your tuple and may add normalization hints.
• compare_mechanism_tuples returns a mechanical similarity hint only.
• A possible_true_collision hint is not a final collision. Confirm whether the candidate actually covers the same contribution.

Reporting Rules

• If overlap is real, name the exact paper and explain the shared mechanism.
• If overlap is partial, turn it into a required baseline or crucial experiment.
• If evidence is abstract-only, lower confidence unless a later metadata/full-text check supports it.