The full upstream README, mirrored here for reference. Install config, tool schemas, adoption signals, and an original overview live on the DWIN Research Adoption listing page.
Turn AI research into measured setup improvements without letting papers silently rewrite your system.
This repository contains one local MCP entrypoint plus three reusable skills. The workflow keeps five states separate:
The MCP combines seven bounded capsules: Research Radar, Document Intelligence, Evidence Graph, Setup Intelligence, Research Adoption, Memory Graph, and Factory Health. It does not include personal session history, financial data, employer artifacts, or private source corpora.
Published package: dwin-research-adoption on npm. Official Registry identity: io.github.vidvinianinov/dwin-research-adoption.
Generic MCP client configuration:
No API key is required. The default data directory is private application data; override it with DWIN_FACTORY_DATA when isolation is needed. Network access is deny-by-default. Only official arXiv endpoints and an optional loopback local-model endpoint are allowlisted by the bundled policy.
The default basic profile exposes 38 tools across discovery, PDF evidence, setup inventory, hybrid retrieval, reconciliation, evaluation, and governed Memory. Set DWIN_TOOL_PROFILE=full to expose additional category scans, watches, retrieval experiments, and the pinned demonstration corpus.
Registration is explicit: the package never scans your home directory. From a stable checkout or global installation:
Call setup_sync to classify and index instructions, skills, MCP manifests, capsules, contracts, policies, hooks, steerings, SOPs, scripts, evals, fixtures, evidence, and configuration. setup_search uses FTS5/BM25; setup_reopen verifies exact source text; setup_graph exposes bounded typed relationships; and setup_reconcile returns observed or gap candidates without editing the setup.
For optional semantic recall, call setup_bridge_evidence, explicitly provision the pinned model, and then call evidence_embed:
The embedding cache binds exact text, role, model revision, preprocessing/configuration, runtime, and architecture. BM25 stays the default; vectors and graph edges remain candidate-only.
The lightweight package does not bundle model weights. Explicitly provision pinned Docling and register one PDF:
Then call document_sync, document_search, and document_reopen. The parser runs locally and returns page/block provenance, hashes, and a private lossless Docling export. Provisioning is intentionally separate because the tested runtime uses about 1.34 GB of model artifacts.
The public package includes a bounded daily runner plus a macOS launchd installer. A stable checkout or global installation is required; ephemeral npx paths are rejected.
The scheduled run executes only jobs declared daily in capsule manifests and writes hashed receipts under private application data. It does not publish posts, install research changes, or promote Memory.
Current stable release: v0.2.0. Skill catalog: skills.sh/vidvinianinov/dwin-research-adoption.
Included skills:
triage-ai-research: route a question to an audited arXiv template and produce candidate evidence.adopt-research-into-setup: turn a selected paper into a baseline/intervention experiment.review-evidence-and-memory: verify provenance and promote only accepted, approved lessons.npm run demo uses a sanitized offline arXiv fixture and an isolated index of the package itself. It discovers the unified tools, finds papers, builds a typed setup index, reopens exact evidence, reports three explicit gap candidates, runs the baseline/intervention reconciliation case, and verifies that Memory remains empty. npm run audit:release inspects the exact npm file list for secrets, personal paths, employer terms, symlinks, and metadata drift.
The first real research-to-setup result is the memory portability launch case. It includes a negative adoption decision and makes no productivity or token-saving claim.
The default retrieval path is lexical and source-cited. Optional embeddings are local, pinned, hash-cached, and experimental; they do not confer truth or authority. Parsed PDFs and external abstracts are untrusted inputs. Graph edges are candidate relationships. Only evaluator receipts plus a current human approval can authorize durable memory.
The local stdio package is Apache-2.0. A future hosted service may add managed scheduling, organization policy, collaborative approvals, and hosted evaluation while preserving the public contracts.
See SECURITY.md, Privacy, and Terms. Report vulnerabilities privately rather than opening a public issue.
The dependency-ordered product milestones and exit criteria are in the goal chain.
MCP Registry name: io.github.vidvinianinov/dwin-research-adoption