research-depth-control

Research Depth Control

This skill raises the research-depth floor for auto research tasks. Use it together with literature-evidence-mapping before ideation, before experiment planning, and again before final paper writing.

---

1. Depth Targets

Default targets unless Boss narrows the scope:

candidate papers: 30-60
included papers: 15-30
deep-read papers: 8-12
baseline/source papers: 3-6
recent papers: at least 5 from the last 24 months when the field is active
negative/limitation papers: at least 3
code/data/benchmark sources: at least 5 combined

For small or niche fields, record why the target cannot be met and list the exact searches that were exhausted.

---

2. Four-Pass Search

Run and log four distinct passes:

breadth pass: broad field terms, surveys, benchmark papers, classic baselines
depth pass: citation chasing from top papers, authors, labs, official repos
gap pass: limitations, failure modes, robustness, negative results, open problems
recency pass: latest 6-24 month papers from arXiv/OpenReview/venue pages/Papers With Code

Each pass must record:

query
source/index
date searched
number of hits screened
inclusion/exclusion rationale
new papers added

Do not stop after a single search engine or a single query phrasing.

---

3. Source Stratification

The final paper set should cover these buckets where applicable:

foundational work
strong baseline methods
most recent frontier methods
benchmark/dataset/evaluation papers
reproducibility or implementation sources
failure/limitation/negative evidence
adjacent methods that solve a similar bottleneck

If one bucket is empty, explain whether it is irrelevant or still an evidence gap.

---

4. Evidence Density Gate

Before idea selection, require:

at least 15 included papers or a documented niche-field exception
at least 8 deep-read evidence cards
at least 3 papers tied to the chosen benchmark/dataset
at least 3 explicit limitations or open problems cited from papers
at least 2 reusable baselines or implementation references