research

name	research
description	Conduct keyword-based research for a new alpha topic and produce a durable note in research/notes/. Invoke before implementing a new strategy when no existing note covers the topic.

Research skill

Goal: ground every alpha attempt in external sources before code is written. Output: a single research/notes/<topic>.md file plus an index update.

When to invoke

• Before implementing a new alpha whose intended cell vector or idea family is not already covered by an existing research/notes/*.md.
• Skip and reuse the existing note if the planned idea_family already has one.

Inputs

• topic slug (snake_case): e.g. vpin_crypto, kyle_lambda_intraday
• short hypothesis sentence

Steps

1. Read existing notes — research/index.csv plus any related notes. If a prior note covers the topic, return its path and stop.
2. Search with WebSearch (3–5 queries):
   • one academic-leaning query (e.g. "VPIN" crypto futures intraday alpha)
   • one practitioner query (e.g. crypto orderflow strategy backtest 2025)
   • one community query (e.g. reddit algotrading CVD divergence reversal)
3. Fetch with WebFetch:
   • 1–2 academic sources (arxiv / ssrn / jstor)
   • 1–2 practitioner or exchange-research pages
   • 0–1 community thread (treat as hypothesis source only)
4. Theory check with /market-microstructure if the topic touches Kyle, Glosten-Milgrom, VPIN, bid-ask spread, or related microstructure concepts.
5. Synthesize into research/notes/<topic>.md using research/notes/_template.md schema. Required fields:
   • hypothesis: one sentence
   • data_required: which panel / manifest / bar fields are needed
   • applicability: explicit yes/no/conditional given our 7-symbol crypto 1m bars and 0.20% taker fee
   • ## Sources: minimum 2 sources, each with category tag, URL, access date
   • ## Mechanism: 3–5 sentences explaining the signal's information content
   • ## Verdict: actionable recommendation
6. Append one row to research/index.csv: <topic>,open,<sources_count>,,<today>
7. Return the absolute path of the new note.

Source whitelist

See research/README.md. Forbidden: anonymous Telegram signals, paid YouTube ads, vendor sales pages.

Saturation guard

Before doing the full search, check whether the intended cell vector (bar, transform, horizon, universe, exit, idea_family) is saturated by running:

uv run python scripts/governance/check.py --only coverage --json

If saturated, do not produce a new note for that cell — pick a different cell first.

Anti-pattern

Do not write a note that simply paraphrases an existing one with a different parameter setting. Each note represents a distinct mechanism, not a parameter variant.