PubMed-centred research workflow

deep-research turns scientific questions and papers into verified, reusable evidence artifacts.

Use it for full literature reviews, manuscript-oriented evidence projects, and lightweight one-paper or selected-paper summaries — all backed by immutable source snapshots, character-offset extraction spans, and appraisal records that a script re-checks rather than a reader trusting.

Full review

Question to report

Configure a PICO/PECO-style question, search PubMed, screen records, retrieve full text, extract, appraise, synthesize, verify, and render the report.

View pipeline
Proof layer

Evidence kernel

Write-once snapshots, character-offset spans, an append-only retrieval log, and an assembler that re-slices every excerpt instead of trusting one. Agent-written text is never itself evidence.

View evidence kernel
Manuscript repo

Standalone storage

Create a repo with canonical paper metadata, source snapshots, reusable extractions, project-scoped appraisals, manuscript files, and exports.

View architecture
Lightweight profile

Paper summaries

Summarize one paper or a user-specified set by reusing the same registry, extraction, appraisal, and verifier machinery without running a full review.

View summaries
Reference manager

Search, ask, tag, export

Ranked full-text and facet search over the registry, questions answered from papers you already hold with span-verified citations, saved PubMed searches that report what's new, personal tags/star-ratings/notes, and a bridge into an OKF knowledge-graph bundle.

View reference manager
ReadCube export

Publish an import bundle

ReadCube is the library UI and PDF viewer, not deep-research. Prepare a RIS + PDF import bundle from the registry, with checksum-verified staged publishing and incremental re-export.

View ReadCube export

What It Is For

NeedUseOutput
Literature review or evidence synthesisFull Stage 0-8 pipelineVerified report, evidence table, PRISMA flow, optional HTML/PDF/docx
One article summary with appraisal/deep-research:summarizeSingle-paper summary plus appraisal and verification
Several specified papers/deep-research:summarize-setOne summary per paper plus optional bounded overview
Just the structured data, no write-up yet/deep-research:extractVerified extraction fetched and promoted into the registry — no appraisal, no summary
A question answered from papers you already hold/deep-research:askAn answer built only from spans that still verify — not a literature search
A reusable extracted-paper pool, no reportA run with pipeline.stop_after_stage: 5Extractions every later run reuses instead of redoing
Manuscript project without wiki-manager/deep-research:init and /deep-research:project createStandalone project, reusable paper pool, exports
Boundary: selected-paper summaries are not systematic reviews. They may orient the reader across a bounded set, but they do not claim comprehensive retrieval, pooled effects, or body-of-evidence certainty — and the verifier rejects that wording outside a run that actually produced it.

How The Pieces Fit

        your question                     papers you already have
              │                                     │
              ▼                                     ▼
      ┌───────────────┐                   ┌───────────────────┐
      │  Stage 0-8    │ ─── extractions ─▶ │  canonical store  │ ◀── summarize / summarize-set
      │  pipeline     │ ◀── reuse ──────── │  registry.jsonl   │
      └───────┬───────┘                   │  extractions/     │ ──▶ search · ask · annotate
              │                           │  appraisals/<p>/   │ ──▶ bib · RIS+PDF · OKF bundle
              │                           └─────────┬─────────┘
              ▼                                     │
      report · digest                               ▼
      verification.json                 data/sources/  snapshots + events.jsonl
                                        (every span re-sliced from here, twice)

Work done once is available everywhere. A paper extracted by a review is reused by a later one-paper summary; a paper registered for a summary is already in the pool the next review seeds from. Appraisals are the deliberate exception — they are scoped to the project whose question they answer.

The Core Rule

Scientific claims flow through structured records: source snapshot → extraction span → appraisal domain → summary record → verification. The tool does not free-form summarize from a PDF when a structured extraction should exist, and it does not let an agent supply its own quotation. Every excerpt in a finished artifact was cut from an immutable snapshot by a script — once at assembly, once at publish.

Invariants That Never Bend

#Rule
1No PubMed-backed claim without a PMID actually retrieved this run.
2No claim from any source without a retrieved evidence id: PMID, DOI, PMCID, preprint DOI, guideline URL, report URL, or bundle-relative concept path.
3Abstract-only vs. full-text is tagged on every extraction. Abstract-only is never appraised as if full text.
4Null and negative findings are actively searched for. Screeners must not exclude on outcome positivity.
5Conflicts are surfaced with an explanation, never averaged away.
6"New insights / hypotheses" is hard-walled from "what the evidence shows".
7Retracted and Expression-of-Concern papers are flagged at screening and at every point of use.
8PubMed attribution is honored: cite PubMed and DOIs.
9No paywall circumvention of any kind. No credentials, sessions, proxies, captcha solving, or mirrors; no retrying a 401/402/403 by another route. Fail closed.
10-11Bundle concepts preserve full PubMed metadata; per-claim attribution uses footnotes keyed to source ids. A body-only citation list is not acceptable.
12-13The wiki's own wiki/ tree is never written to; standard markdown links are the graph layer, wikilinks additive only.
14Unknown values are null. Never invent a bibliographic field, a number, or a citation.
15Whenever a PDF is handed to you, its DOI is stated alongside the PMID — or the absence of a DOI is stated.

Start Here

If you are…Read
Setting up for the first timeQuickstart — prerequisites, a real repo, a first summary, and a troubleshooting table
About to ask a research questionScientific questions — pick the framework before searching, not after
Running or debugging a reviewFull pipeline — every stage, every gate, every failure mode
Wondering whether to trust a quoteEvidence kernel — how a claim is made checkable
Deciding where papers and judgements liveRepo architecture — source of truth vs. derived, and what a backup must contain
Looking for a flagCommand reference, then python3 scripts/<name>.py --help