math-research

Answer a methodology question about a mathematical or algorithmic choice with a cited, tradeoff-explicit findings brief. Research-only - this skill never edits code, and the subagent is instructed not to. Know the limit of that guarantee: it is enforced by instruction, not mechanically, whenever the subagent has write tools. Step 2 covers what to do when that is not good enough.

The question: $ARGUMENTS

1. Locate the specialization - do not assume it

The branch of applied math is a finding, not an input. Before dispatching, spend a few cheap reads establishing:

• The numeric stack. Grep the manifest and imports for what the project actually computes with - numpy/scipy/scikit-learn/statsmodels, nalgebra/faer/ndarray, Eigen/BLAS/LAPACK, torch/jax, R packages. The stack constrains which recommendations are implementable.
• The call site. The specific function, module, or pipeline stage the question is about. If the user named one, start there; otherwise find it.
• The problem class. Name it explicitly - optimization, dimensionality reduction, linear-system solving, eigenproblems, graph algorithms, probabilistic inference, time-series estimation, quadrature, signal processing. State the specialization in your dispatch so the subagent researches as a specialist in that, not as a generalist.
• The problem scale. Rough data size (n, dimensionality, sparsity) and any latency or throughput constraint, taken from config, tests, fixtures, or sample data. Step 4 asks whether an alternative is practically better at the sizes this project actually sees - without a number here, that judgment is a guess. If no scale is discoverable, say so explicitly rather than assuming one.

If the question is purely theoretical with no implementation to inspect, say so and skip to step 2 - but do not invent a call site that is not there.

2. Dispatch one research subagent

Use the Task tool with general-purpose, and spawn one - do not fan out. It is the type built for open-ended research across code and the web, which is this skill's whole job.

Do not substitute a read-only agent type to get a write guarantee. Plan is scoped to designing implementation plans and closes with a critical-files list; Explore is scoped to locating code, not analyzing it. Both would be working against their own instructions here, and the findings contract in step 4 is what loses. Where a mechanical guarantee is genuinely required, define an agent whose tools: frontmatter grants read and web access only - a real restriction from a purpose-built persona, rather than a borrowed one.

State in the dispatch:

• The question, the specialization and problem scale from step 1, and the concrete call site.
• The numeric stack, so implementation notes are written in the host language rather than an unusable one.
• The read-only mandate. It reads code and searches the literature. It never edits, creates, or reformats a file. A researcher that can write is a researcher that starts implementing.
• Read the implementation before recommending anything. Never reason from file or function names alone - what the code does and what it is called routinely diverge, and a recommendation aimed at the wrong current behavior is worse than none.

3. Hold the subagent to source discipline

This is what separates methodology review from confident recall. Require that every recommendation trace to one of:

1. Peer-reviewed literature - a journal or established conference, or an arXiv preprint carrying real citation weight (roughly 50+ citations, or independently corroborated by a source in one of the tiers below). "It sounds authoritative" is not a tier.
2. Established numerical libraries - LAPACK, scipy, scikit-learn, MATLAB algorithm documentation, and the equivalents for the host stack.
3. Official library documentation for the stack in use.
4. Textbooks, cited by author and title.

Blogs, tutorials, and forum answers may be read to discover leads and must never appear as the justification for a change - trace back to a primary source first, or drop the claim. An uncited recommendation is a finding that did not survive.

4. Fix the return contract

Require findings in this shape, one block per topic:

• Current approach - what the code does now, cited as path/file.ext:line.
• Proposed alternative - the method recommended, or an explicit "keep the current approach" with the reason.
• Citation - authors, title, venue, year, or a documentation URL.
• Tradeoffs - benefits, drawbacks, scaling behavior as the problem size grows (with concrete time and space bounds), and numerical stability or conditioning concerns.
• Implementation notes - which library or from-scratch approach in the host stack, the integration points, and a rough complexity estimate. Verify any recommended library actually exists and is still maintained.
• Compatibility - what this changes for downstream consumers of the output.

Require one explicit judgment per finding: is the alternative practically better here, or only theoretically superior? Say which. Most methods that win on paper lose on problem sizes this project actually sees.

5. Relay

Return the brief with citations intact - they are the deliverable, not decoration. Keep your own commentary minimal.

State plainly when the literature does not settle the question, when the current approach is already the right one, or when a recommendation rests on an assumption about problem size or data distribution that nobody has verified. A brief that recommends a change in every case is not a review.