Sci Literature Manager

SCI Literature Manager

Make literature discoverable and reusable without mixing it with experiment records. Follow project-local rules and the ownership boundary in sci-research-manager.

Classify the task

• literature_discovery: find candidate papers from primary/official sources.
• literature_indexing: organize global PDFs, source packets, notes, and indexes.
• deep_reading: hand the located source to sci-paper-reader.
• route_synthesis: return source-grounded options for a project question and competing causes.
• citation_verification: verify title, author, venue, year, DOI/arXiv, code, and exact claims.
• experiment_support: return verified source material for a literature-to-experiment brief.
• maintenance: deduplicate, repair links, or archive superseded notes.

Retrieve before reading broadly

Read the relevant global query route/index, then matching records, then decisive primary-source sections. Reuse unchanged source IDs, notes, manifests, and tool capabilities already read in this task; inspect deltas on resumption. Do not batch-read the entire library unless the user asks for a full audit. Context reuse never replaces exact claim verification.

For a known paper, resolve its supplied DOI/arXiv/Source ID or exact title, check identity/version/duplicates, verify missing metadata, and register only the requested delta. Do not restart domain discovery, create project claims, or deep-read every section as an intake prerequisite. Reuse an existing record; preserve preprint/publication and alias relations rather than deleting them as duplicates.

For discovery, state the question and start with the mechanism, synonyms, and closest citation neighbors. Expand only to resolve a named uncertainty. Stop when another expansion no longer changes the nearest overlap, consequential protocol difference, or next evidence-needed action; report coverage and unresolved sources, not an exhaustive novelty claim. Access failures are not evidence of absence. Reuse valid results, retry only for a diagnosed recoverable cause or useful authorized alternate source, and stop retrying unchanged failures.

Request identifiers, selected metadata, exact sections, page/table anchors, or short relevant passages from tools. Keep full text and large catalogs on disk when available. Mark incomplete/truncated output; do not repeatedly fetch the same full response or treat an abstract as proof of mechanism or exact results. Use deterministic lookup, extraction, deduplication, and verification scripts before considering bounded delegation; no particular provider, model, or notification tool is required.

Keep a portable paper packet

Only when a reusable reading packet is requested, prefer:

<category>/<paper-slug>/
  paper.pdf
  paper_understanding_YYYYMMDD.md
  paper_visual_YYYYMMDD.html
  assets/
  README.md

Use library-relative paths and existing conventions. Generate only requested artifacts; ordinary intake needs a verified source record and available source location, not Markdown/HTML/PPT/Word derivatives. For a reading packet, keep Markdown canonical. Return candidate project relevance to sci-research-manager only when project adoption or state is in scope; do not embed it in global identity or reusable understanding artifacts.

Maintain verification states

Use clear labels such as:

• verified;
• cached_pdf;
• needs_download;
• needs_bibliographic_verification;
• needs_metric_verification;
• paywalled_or_access_limited.

Do not invent bibliographic fields or exact metrics. Record page/table/figure anchors for numerical claims. Distinguish preprints from accepted publications.

Build problem-driven reading routes

Use:

problem -> competing causes -> paper groups -> variables/controls -> evidence limits -> validation implication -> next question

Do not turn a mature framework or paper module into the project's contribution by default.

Return a typed source handback

For intake or a narrow lookup, return only the source ID/location, verified fields, changed records, and unresolved gaps. For a project decision or experiment-support request, return a compact typed source handback with the relevant fields below; do not create a second plan or registry:

• stable source identities, canonical locations, versions, and bibliographic verification status;
• bounded source claims with exact locations and protocol limits;
• theoretical motivation separated from any proposed project hypothesis;
• variables/cues and required controls;
• success/failure criteria;
• likely confounds and non-transferable assumptions;
• do-not-overclaim/do-not-do-next;
• unresolved identity, access, or citation gaps.

Include the nearest substantive overlap, remaining testable difference, decisive source locations actually read, and search coverage. A close precedent may make a method a reproduction or baseline; it does not automatically invalidate the whole direction. Do not invent novelty from a new name, benchmark, or untested condition. The research owner integrates the decision before proposal-specific asset preparation or costly execution.

Return to the current requester. Use sci-research-manager only when project adoption, tags, LiteratureClaim verification, experiment allocation, or downstream state actually enters scope; a direct lookup or intake can finish without a coordinator call. Literature-only items never enter the experiment index or claim map as project evidence.