Research: findings phrasing — moderate explanation maximizes agreement (arXiv:2607.24601, 2607.09524) #525

Context

Two empirical findings relevant to how Cora presents findings:

1. XAI & Trust study (arXiv:2607.24601, n=34): full explanations yield highest perceived trust (3.99/5) but NOT highest agreement; moderate explanations achieve highest agreement (89.22%). More explanation makes developers question AI recommendations more often. No explanation = lowest trust and agreement.

2. CuREV curation pipeline (arXiv:2607.09524): separating high/low-quality review comments via an evaluation framework, then using high-quality ones as in-context exemplars to reformulate low-quality ones, produces more actionable and more stylistically diverse comments.

Proposal

• Keep findings format at 'moderate explanation': severity + short reason + code evidence; avoid long reasoning dumps → shipped in feat(review): moderate-explanation style in findings prompt #527 (b018eaf)
• Long-term: use human accept/dismiss history on findings as quality labels; store accepted findings as few-shot exemplars for findings phrasing (exemplar-guided reformulation)

References

• http://arxiv.org/abs/2607.24601
• http://arxiv.org/abs/2607.09524
• Related positioning evidence (agentic review = speed, not quality): http://arxiv.org/abs/2607.13196