scaling-group/eve/matlas-search

Matlas Search Manual

Use this skill when a proof or evaluation needs a second retrieval source for:

• a named theorem, lemma, proposition, corollary, definition, or standard fact;
• a textbook-standard result that is hard to identify through TheoremSearch;
• canonical citation metadata for a result from books or journal papers;
• a check that a cited theorem's hypotheses match the current proof obligation.

Matlas is supporting evidence only. It does not prove that a dependency is valid for the current proof. Always compare the returned statement against the exact proof obligation.

Public references:

• Web app: https://matlas.ai/
• API docs: https://matlas.ai/docs
• Search endpoint: POST https://matlas.ai/api/search
• Paper: https://arxiv.org/abs/2604.17484

Matlas searches published mathematical literature, including journal papers and textbooks. It complements TheoremSearch: TheoremSearch is often better for theorem-level statements from arXiv and structured web corpora; Matlas is often better for published or textbook sources with DOI-style provenance.

Basic Search

Use natural language, LaTeX, or a mix. The API returns a JSON array of hits. Request at least 10 results, then inspect the relevant top hits.

curl -fsSL https://matlas.ai/api/search \
  -H 'Content-Type: application/json' \
  -d '{
    "query": "Banach fixed point theorem contraction complete metric space",
    "num_results": 10
  }'

Typical hit fields include:

{
  "type": "book",
  "entity_name": "Theorem 2.12",
  "doi": "10....",
  "title": "An Introduction to Nonlinear Analysis",
  "authors": "Martin Schechter",
  "journal": "",
  "year": "1991",
  "statement": "Let ...",
  "candidate_id": "..."
}

Trust the returned statement and source metadata more than the rank alone. A high-ranking hit can still have the wrong hypotheses, and a relevant hit can be useful even if metadata such as DOI or year is missing.

Query Strategy

1. State the needed result as a theorem: "If [hypotheses], then [conclusion]."
2. Start with a broad query using standard terminology.
3. Search again with distinctive words from a promising hit's statement,

title, or named theorem.

1. Use Matlas alongside TheoremSearch when the dependency matters: agreement

across independent retrieval routes is stronger evidence than either alone.

1. If the proof needs a citation, prefer hits with enough metadata to audit:

title, authors, journal or book, year, DOI, and exact statement.

Good query patterns:

• "spectral theorem compact self-adjoint operator Hilbert space"
• "Nakayama lemma finitely generated module local ring"
• "radical splitting theorem finite dimensional algebra"
• "Frobenius reciprocity finite group representations"

Poor query patterns:

• "prove this"
• proof paragraphs with irrelevant exposition;
• local notation without mathematical context;
• citation fishing when the proof has not stated the needed result.

Validation Checklist

Before using a Matlas hit in a proof or score, check:

• objects match: groups, rings, schemes, spaces, modules, categories, fields;
• hypotheses match: finiteness, completeness, compactness, smoothness,

noetherian assumptions, characteristic, topology, boundedness;

• conclusion matches: existence, uniqueness, equivalence, vanishing, density,

openness, decomposition, exactness;

• direction and variance match: source/target, left/right, covariant/

contravariant, injection/surjection;

• notation imported from the source is compatible with the current proof;
• no necessary condition is only present in surrounding text and absent from

the returned statement.

If a result is only close, treat it as a lead. Do not silently strengthen, weaken, or restate a theorem to fit the current proof.

Reporting Provenance

When Matlas affects a proof step or evaluation, record enough for audit:

Matlas: RESOLVED
Need: <proof obligation>
Statement: <exact returned statement>
Source: <title>, <authors>, <journal/book>, <year>, <doi if present>
Candidate id: <candidate_id if present>
Use: <why the hypotheses and conclusion match>

If it is only partial:

Matlas: PARTIAL
Need: <proof obligation>
Closest statement: <exact returned statement>
Source: <metadata>
Mismatch: <specific missing/extra hypothesis or wrong conclusion>
Next query: <sharper query>

Failure Policy

If search fails:

• Network, TLS, or API error: report Matlas unavailable with the endpoint and

error text.

• Empty results: retry once with simpler language and fewer local symbols.
• Low relevance: report unresolved, summarize the closest hit, and explain

the mismatch.

• Missing DOI or metadata: use the available title/authors/statement, but say

what is missing.

Never fabricate:

• theorem names;
• statement bodies;
• authors;
• titles;
• DOI values;
• years;
• candidate IDs;
• search scores.