research-ideation-screening

Research Ideation Screening

This skill prevents vague novelty claims. An idea is only useful if it is tied to evidence, implementable under constraints, and falsifiable by an experiment.

---

1. Idea Generation Lenses

Generate candidates using at least five lenses:

failure case repair
metric-target mismatch
data efficiency
compute efficiency
robustness and distribution shift
architecture simplification
training objective change
evaluation protocol improvement
tooling or reproducibility gap
negative result worth documenting

Avoid ideas that only rename existing methods.

---

2. Idea Card

Write each candidate in:

research_workspace/ideas/idea_<N>.md

Template:

# Idea <N>: <title>

- Gap:
- Evidence:
- Hypothesis:
- Method sketch:
- Minimal implementation:
- Baseline:
- Dataset/benchmark:
- Primary metric:
- Ablation:
- Expected failure mode:
- Resource estimate:
- Novelty risk:
- License/data risk:
- Stop condition:

---

3. Screening Matrix

Maintain:

research_workspace/ideas/screening_matrix.md

Score 1-5:

evidence strength
novelty
implementation simplicity
evaluation clarity
resource fit
expected impact
risk containment
paper-worthiness

Select the idea with the best combination of evidence, feasibility, and falsifiability, not the most impressive wording.

---

4. Red-Team Questions

Before committing to an idea, answer:

Has this already been done?
Can a small experiment disprove it?
Is the baseline fair?
Would a null result still teach something?
Can the implementation be isolated?
Can results be plotted in one clear table or figure?
What would make us abandon the idea?

If these cannot be answered, return to literature mapping or scope refinement.