Management Research Notes
binqi20/management-research-notes
Zettelkasten acadêmico file-based para pesquisa em gestão e sustentabilidade
Ver no GitHubBase de conhecimento estilo Zettelkasten para pesquisa acadêmica em gestão e sustentabilidade empresarial. Organiza artigos peer-reviewed em notas Markdown com metadados estruturados, abstratos íntegros e índices SQLite/CSV/BibTeX reconstruíveis, pronto para agentes de IA.
Instalação
npx convez-ai binqi20/management-research-notes1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Construir literatura de pesquisa rastreável para trabalhos acadêmicos
- Alimentar agentes de IA com conhecimento estruturado e verificável
- Organizar artigos peer-reviewed com metadados confiáveis e índices derivados
Veredito

Projeto robusto para quem precisa de literatura organizada com rigor acadêmico e proveniência verificável. Diferencia-se pela auditoria de fidelidade e arquitetura file-first que permite reconstrução completa dos índices — essencial para pesquisadores que confiam seus dados críticos a sistemas.
README
A tradução para o português deste README ainda está na fila.
Management Research Notes
A file-based academic knowledge base for management and business sustainability research.
One Markdown note per peer-reviewed paper. Trusted bibliographic metadata from manifests, analytic distillation from a verbatim-anchored extraction prompt, derived SQLite / CSV / BibTeX indexes that you can rebuild from scratch in seconds.
Author
Management Research Notes is authored and maintained by
Binqi Tang
,
a researcher in management and business sustainability. Collaborations are
welcome — whether that means
using this knowledge base in your own research workflow, adapting it for a
new source or subfield, or discussing joint projects. Please
open an issue
or reach out on GitHub. Citations welcome — see How to cite
below.
What it is
This project is a Zettelkasten-style knowledge base for academic literature. Each
peer-reviewed article is stored as a single Markdown file in notes/ with two
parts:
- YAML frontmatter — structured metadata: trusted bibliographic fields
(title, authors, year, journal, DOI), a controlled-vocabulary topic list,
the methods, the sample, and three custom analytic fields
(
unit_of_analysis,level_of_theory,dependent_variable_family) that make cross-paper queries actually useful. - Markdown body — a human-readable distillation: verbatim abstract, research question, mechanism / process, theoretical contribution, practical implication, limitations, future research, and a clean APA 7th citation.
A small Python pipeline derives a SQLite index, a flat CSV, and a BibTeX file
from the notes — none of which are the source of truth, all of which can be
rebuilt with python tools/build_index.py && python tools/export_csv.py && python tools/export_bibtex.py.
Why it exists
Large language models have made it tempting to build a literature review by dropping a pile of PDFs into a chat window and trusting the response. For scholars in management, economics, and adjacent citation-heavy fields — disciplines where every substantive claim is expected to be traceable to a specific prior study — that naive workflow breaks down for three reasons:
- Context windows are finite. Even a 200K-token model cannot hold a few hundred full papers at once. You need distilled notes small enough to load in bulk but faithful enough to support real argument.
- Provenance is non-negotiable. Every analytic claim in a note is traceable
to a cited paper. The verbatim abstract is checked as a contiguous substring
of the extracted PDF text, every factual claim in a v2 note carries a ≤25-word
evidence anchor verified against the PDF by
tools/validate_note.py, and a two-layer faithfulness audit (see below) has a fresh-eyes LLM subagent cross-check the prose fields. Hallucinated or paraphrased content does not reach the published library — which is what makes these notes safe to cite in a defensible review. - Cross-platform and agent-ready. The same note is grep-able from the
command line, uploadable as project knowledge to an LLM chat, and citable
from a Word / LaTeX / Typst draft via the BibTeX export. Via the SQLite + FTS5
index and the
AGENTS.mdentry point, any AI agent — Claude Code, Cursor, Windsurf, a custom SDK app — can consume the library to run retrieval-augmented analysis with known faithfulness guarantees. That's the piece that matters for researchers: it lets one scholar (or an agent acting on their behalf) reason across a literature much larger than any single head can hold, without sacrificing the citation trail a peer reviewer, committee, or journal editor will expect.
The corpus shipped here is management and business sustainability, but the pattern — one Markdown file per paper, trusted bibliographic manifests, a verbatim-anchored extraction prompt, a two-layer audit — transfers straightforwardly to any scholarly corpus where paraphrastic drift is costly: economics working papers, organization theory, policy analysis, STS, labor studies. Fork the repo, point it at your own manifest, and the validator and audit layer come along for free.
The current design targets libraries growing from dozens to tens of thousands of papers. Whether the shape holds at that scale is an empirical question rather than a settled assumption: some parts — one file per paper, SQLite-derived indexes, the controlled-vocabulary topic list, even the extraction prompt's field set — may need to evolve as the library grows, new sources appear, or research workflows change. Treat the current architecture as a working hypothesis refined by each release, not a frozen spec.
Faithfulness audit
Every analytic field in every note is checked by a two-layer faithfulness audit before it is accepted into the library:
- Layer 1 — Evidence anchors (mechanical, deterministic). Each factual claim in the frontmatter (sample size, country, industry, time period, theories, methods, keywords) carries a ≤25-word verbatim quote from the PDF. The validator checks each quote is a substring of the extracted PDF text under hyphen-tolerant normalization. A fabricated quote fails deterministically — there is no way to pass Layer 1 with invented evidence.
- Layer 2 — Semantic audit (fresh cold-context subagent). For the
prose fields (research question, mechanism, theoretical contribution,
practical implication, limitations, future research) a fresh Claude
subagent reads the PDF, reads the note, and emits a per-field verdict
against the rubric at
docs/audit-rubric.md:SUPPORTED/PARTIAL/UNSUPPORTED/CONTRADICTED. A note is rejected if any verdict isUNSUPPORTEDorCONTRADICTED.
The full library has been swept three times across releases:
- v0.2.0 sweep (2026-04-17, 90 notes): 88 / 90 initial PASS; two
fails (Mahringer 2025, Li 2026) repaired via re-extraction, and the
validator was tightened in one place (extended the
Not reported in paperescape valve tofuture_research, mirroring thelimitationsexemption) based on what the audit found. - v0.3.0 sweep (2026-04-17, full 105 notes): 105 / 105 PASS, 0 UNSUPPORTED, 0 CONTRADICTED. One PARTIAL verdict (Castelló 2025 citing the EU Digital Markets Act instead of the EU Digital Services Act) was surfaced and fixed before release.
- v0.4.0 sweep (2026-04-29, 113 notes): 113 / 113 PASS, 0 UNSUPPORTED, 0 CONTRADICTED. Eight new notes added from AMJ vol. 68 no. 4 (Grégoire editorial + 7 substantive papers).
- v0.5.0 sweep (2026-04-30, 121 notes): 121 / 121 PASS, 0 UNSUPPORTED, 0 CONTRADICTED. Eight new notes added from AMJ vol. 68 no. 3 (Bermiss editorial + 7 substantive papers).
- v0.6.0 sweep (2026-04-30, 129 notes): 129 / 129 PASS, 0 UNSUPPORTED, 0 CONTRADICTED. Eight new notes added from AMJ vol. 68 no. 2 (Rouse editorial + 7 substantive papers). Cleanest batch yet — only 1 PARTIAL across the 8 new notes.
- v0.7.0 sweep (2026-04-30, 139 notes): 139 / 139 PASS, 0 UNSUPPORTED, 0 CONTRADICTED. Ten new notes added from AMJ vol. 68 no. 1 (Gruber + Cronin editorials + 8 substantive papers — first 10-paper issue we've processed).
- v0.8.0 sweep (2026-05-05, 147 notes): 147 / 147 PASS, 0 UNSUPPORTED, 0 CONTRADICTED. Eight new notes added from AMJ vol. 67 no. 6 (Reinecke editorial + 7 substantive papers — first issue from volume 67 we've processed).
- v0.9.0 sweep (2026-05-05, 156 notes): 156 / 156 PASS, 0 UNSUPPORTED, 0 CONTRADICTED. Nine new notes added from AMJ vol. 67 no. 5 (Dorobantu editorial + 8 substantive papers). Zero PARTIAL across all 54 prose-field audits — first perfect-SUPPORTED batch.
- v0.10.0 sweep (2026-05-06, 165 notes): 165 / 165 PASS, 0 UNSUPPORTED, 0 CONTRADICTED. Nine new notes added from AMJ vol. 67 no. 4 (Bliese editorial + 8 substantive papers). 1 PARTIAL.
- v0.11.0 sweep (2026-05-07, 175 notes): 175 / 175 PASS, 0 UNSUPPORTED, 0 CONTRADICTED. Ten new notes added from AMJ vol. 67 no. 3. Three-batch ≤1-PARTIAL streak.
- v0.11.1 patch (2026-05-07, 175 notes):
Bibliographic-integrity correction. A user-flagged check against
Google Scholar surfaced that the Bednar 2024 note (and many
others) were citing the online-first publication year rather
than the issue year per APA 7. Built
tools/verify_years.pyto cross-check every DOI against CrossRef'spublished-printfield. Found 48 papers (27% of library) carrying the online-first year. All 48 corrected: manifest rows, frontmatteryear:fields, and APA citations updated to the issue year. 175/175 now align with CrossRef. The audit prose fields are unchanged (year doesn't affect any audited content claim), so no re-audit was needed. - v0.11.2 patch (2026-05-09, 175 notes):
Bibliographic cross-check extended to all seven CrossRef fields
(year, title, journal, volume, issue, pages, authors) via the
new
tools/verify_metadata.py. The first full sweep on the v0.11.1 library surfaced 21 papers needing manifest backfills (mostly online-first papers in the NBS digest and AMJ vol-69-1 whose issues had since published) plus one wrong-page error (Bamberger's AMR address) and one missing-Oxford-comma title typo in the same paper. After fixes: 175/175 MATCH for year, journal, and authors, and 0 MISMATCH across volume / issue / pages on every paper where CrossRef carries those fields. One remaining flag (Reinecke book review) is documented as a CrossRef-side data corruption, not a note error. Like v0.11.1, no re-audit needed — the audit prose fields are independent of bibliographic metadata. - v0.12.0 sweep (2026-05-10, 186 notes):
Adds AMJ vol. 67 no. 2 (11 substantive papers: Anderson, Arslan,
Grégoire, Gruber, Li, Liao, Muethel, Pache, Piazza, Preston,
Rostain). 186 / 186 PASS, 0 UNSUPPORTED, 0 CONTRADICTED, 0
PARTIAL — the cleanest batch outcome to date.
Also introduces Tier 3 of the bibliographic-integrity prevention
strategy: the new
tools/populate_manifest.pyauto-populates each manifest'svolume/issue/pagescolumns from CrossRef BEFORE extraction, and (with--fix-year) auto-corrects any wrong years the manifest may have inherited from the publisher's online-first metadata. This is now Step 0 of/synapse-ingest. The v0.12.0 batch proved the mechanism: Tier 3 caught and fixed 7 wrong-year manifest rows in vol-67-no-2 before extraction started — exactly the v0.11.1 bug class, prevented at the source instead of patched retroactively. - v0.13.0 sweep (2026-05-12, 197 notes):
Adds AMJ vol. 67 no. 1 (10 substantive papers: Barkema, Bettinazzi,
Cao, Dorobantu, Han, Jia, two Park papers, Soublière, To, plus
Gruber editorial on the new Research Methods Articles portfolio).
197 / 197 PASS, 0 UNSUPPORTED, 0 CONTRADICTED, 1 PARTIAL (Cao
practical_implication — minor extrapolation from a contributions
paragraph). Second consecutive batch ingested under the Tier 3
prevention infrastructure (introduced v0.12.0) and the post-extraction
column-merge heuristic (introduced commit
59a9ace). Tier 3 found zero year mismatches in this batch (the manifest had already been captured after print publication), but still backfilled 11 vol/issue/pages columns. Every extraction agent independently usedgrepon the extracted text to verify anchor contiguity before writing — Layer 1 passed on first try for all 11 papers, the first batch since the column-merge heuristic went into the extraction prompt to do so. - v0.13.1 patch (2026-05-12, 197 notes):
slugify()updated to fold diacritics via NFKD and strip apostrophes, resolving fragmented paper_ids likegr-goire-2024→gregoire-2024. 5 paper_ids renamed as a deliberate, one-time exception to hard rule 4 (stable IDs) while the library was small and had no known external citations by paper_id. - v0.13.2 patch (2026-05-12, 197 notes):
Adds
tools/lint_manifests.py— a per-row structural + CrossRef audit of every manifest. First run surfaced 4 legitimate compound surnames (von Krogh, Lee Cunningham, Ter Wal, van den Oever — added toKNOWN_COMPOUND_SURNAMESallowlist) and 6 latent D'Amico-class bugs in the NBS-2026-02 manifest (full given+family name captured instead of just the family name). 6 more paper_ids renamed; combined with v0.13.1 this closes the v0.13.x cleanup cohort (11 papers total). Hard rule 4's footnote in CLAUDE.md revised to use cohort framing. - v0.14.0 sweep (2026-05-16, 210 notes):
Adds AMJ vol. 66 no. 6 (12 substantive papers + Grimes editorial on
generative AI's impact on management scholarship). 210 / 210 PASS,
0 UNSUPPORTED, 0 CONTRADICTED, 3 PARTIAL (Couture and Sajjadiani
mechanism_process — minor attribution drift; Zhang limitations —
added "cross-sectional" label not in paper). First batch ingested
under the complete three-gate pipeline: Step 0
populate_manifest.py(Tier 3), Step 0.5lint_manifests.py(structural), Step 4.5verify_metadata.py(Tier 2 content). All three gates exit 0. Also backfills v0.11.2-pattern drift on two ASQ papers (Lee Democratic Deviations, Reinecke book review) whose vol-71 issue 2 published after their extraction; their note frontmatter + APA citations now reflect the issued version. Xu 2023 title added toKNOWN_CROSSREF_DATA_ERRORS(CrossRef has malformedForest<i>and</i>the Treeswith no spaces around the italic; note matches the published version correctly). - v0.15.0 sweep (2026-05-20, 222 notes): Adds AMJ vol. 66 no. 5 (11 substantive papers + Dencker editorial on positioning research on novel phenomena). 222 / 222 PASS, 0 UNSUPPORTED, 0 CONTRADICTED, 0 PARTIAL — a perfectly clean batch: 12/12 Layer 1 first-try PASS and 12/12 Layer 2 PASS with no PARTIAL verdicts at all. Ingested under the complete three-gate pipeline (populate_manifest → lint_manifests → verify_metadata), all exit 0 with no drift findings or new false positives. Notable papers include the Lander/Voronov pair on Scottish and Canadian whisky authenticity, and the Carnabuci/Rua-Gomez network-formation studies.
- v0.16.0 sweep (2026-05-20, 233 notes): Adds AMJ vol. 66 no. 4 (10 substantive papers + Wellman editorial on publishing multimethod research). 233 / 233 PASS, 0 UNSUPPORTED, 0 CONTRADICTED, 1 PARTIAL (Kundro future_research — mild overgeneralization of the paper's actual future directions). Ingested under the complete three-gate pipeline; all gates exit 0 with no drift findings or new false positives. 11/11 Layer 1 first-try PASS (fifth consecutive batch at 100%). The Banerjee audit is a nice illustration of the Layer 2 value: H4 in that paper predicted one direction but the data showed the opposite (significant) — the note correctly reports the empirical result rather than the failed hypothesis, which the auditor verified as faithful rather than CONTRADICTED.
- v0.17.0 sweep (2026-05-20, 244 notes): Adds AMJ vol. 66 no. 3 (10 substantive papers + Langley editorial on opening up AMJ's research methods repertoire). 244 / 244 PASS, 0 UNSUPPORTED, 0 CONTRADICTED, 1 PARTIAL (Toivonen practical_implication — practitioner prescriptions synthesized from a paper with no dedicated implications section). Ingested under the complete three-gate pipeline; all gates exit 0 with no drift findings or new false positives. 11/11 Layer 1 first-try PASS (sixth consecutive batch at 100%). This sweep completes the AMJ vol. 66 + vol. 67 coverage from no. 3 onward — the pilot now spans 13 consecutive AMJ issues plus vol. 68 and vol. 69.
- v0.18.0 sweep (2026-05-23, 257 notes):
Adds AMJ vol. 66 no. 2 (12 substantive papers + Krogh editorial on AI
research opportunities). 257 / 257 PASS, 0 UNSUPPORTED, 0
CONTRADICTED, 1 PARTIAL (Zhang mechanism_process — minor mediator-vs-
DV imprecision in one study). 13/13 Layer 1 first-try PASS (seventh
consecutive batch at 100%). Three of this issue's papers (Krogh→von
Krogh, Cunningham→Lee Cunningham, Wal→Ter Wal) are the compound
surnames already in the
KNOWN_COMPOUND_SURNAMESallowlist; Step 0.5 suppressed all three cleanly, and verify_metadata's note-level author check passed without new false positives (notes carry the full byline forms that match CrossRef; the short form lives only in the paper_id slug). This release also promotes theextraction_model-verbatim instruction intodocs/extraction-prompt.md— a validated systemic fix (13/13 agents kept the field uniform this batch vs. up to 7 outliers needing manual cleanup in prior batches), eliminating a recurring per-batch provenance-field normalization. - v0.19.0 sweep (2026-05-23, 270 notes):
Adds AMJ vol. 66 no. 1 (12 substantive papers + Gruber editorial on
the 23rd editorial term's strategic priorities). 270 / 270 PASS, 0
UNSUPPORTED, 0 CONTRADICTED, 0 PARTIAL — the third perfectly-clean
batch (after v0.12.0 and v0.15.0). This release also served as a
standalone validation of two infrastructure fixes from prior
releases:
extraction_modelverbatim (codified in v0.18.0, commitcc16c71): extraction agents were deliberately dispatched WITHOUT the inline reminder; 13/13 still kept the field uniform purely from readingdocs/extraction-prompt.md. The codified location is independently load-bearing.- Editorial-abstract escape (added in commit
5b4ae81, pre-v0.19.0): Gruber editorial — the first editorial extracted after the fix — wrote"Not reported in paper"for Abstract instead of the sentence-fragment workaround used by all five earlier editorials. 13/13 Layer 1 first-try PASS (eighth consecutive batch at 100%). All three gates exit 0; zero post-extraction remediation needed.
- v0.19.1 patch (2026-05-27, 270 notes):
Provenance correction —
extraction_modelfield. No content changes; no new notes. The v0.18.0 fix had standardized this field toclaude-opus-4-6(the bundle's stale default), but the actual extraction model for batches from 2026-04-18 onward was Claude Opus 4.7 (the parent session's model, with 57 notes carrying direct agent self-reports as evidence). Updated 108 notes that predated my v0.18.0 fix-to-verbatim with the accurateclaude-opus-4-7value; left 105 pre-cutover notes atclaude-opus-4-6(no direct evidence of model identity for those earlier batches). Constant intools/prepare_paper.pyupdated toclaude-opus-4-7so future bundles record the accurate model. Library is now provenance-correct for everything where we have direct evidence; pre-cutover notes retain their original metadata as historical record. All three gates still exit 0 after the sweep (provenance field is not validated or audited).
Run the audit on a single note with:
python tools/audit_note.py notes/<paper_id>.md
Or, from inside a Claude Code session: /audit-note <paper_id>.
What's in this release
This release contains 270 curated notes across two source streams:
- NBS 2026-02 — 77 notes distilled from the Network for Business Sustainability (NBS) February 2026 monthly research digest. (62 notes shipped in v0.2.0; 15 previously-missing papers were recovered and added in v0.3.0.)
- AMJ pilot — 193 notes across nineteen consecutive recent issues of the Academy of Management Journal (vol. 66 no. 1 through vol. 66 no. 6, vol. 67 no. 1 through vol. 67 no. 6, vol. 68 no. 1 through vol. 68 no. 6, and vol. 69 no. 1). v0.19.0 added vol. 66 no. 1 (13 notes); v0.18.0 added vol. 66 no. 2 (13 notes); v0.17.0 added vol. 66 no. 3 (11 notes); v0.16.0 added vol. 66 no. 4 (11 notes); v0.15.0 added vol. 66 no. 5 (12 notes); v0.14.0 added vol. 66 no. 6 (13 notes); v0.13.0 added vol. 67 no. 1 (11 notes); v0.12.0 added vol. 67 no. 2 (11 notes); v0.11.0 added vol. 67 no. 3 (10 notes); v0.10.0 added vol. 67 no. 4 (9 notes); v0.9.0 added vol. 67 no. 5 (9 notes); v0.8.0 added vol. 67 no. 6 (8 notes); v0.7.0 added vol. 68 no. 1 (10 notes); v0.6.0 added vol. 68 no. 2 (8 notes); v0.5.0 added vol. 68 no. 3 (8 notes); v0.4.0 added vol. 68 no. 4 (8 notes); the original pilot covered the remaining three issues (28 notes).
| Paper type | Count |
|---|---|
| empirical-quantitative | 113 |
| empirical-qualitative | 75 |
| empirical-mixed | 33 |
| editorial | 23 |
| conceptual | 19 |
| review | 4 |
| book-review | 3 |
| Total | 270 |
Every note carries a v2 evidence: anchor block (Layer 1) and has passed
the Layer 2 semantic audit. See Faithfulness audit
above.
Repository layout
management-research-notes/
├── README.md ← you are here
├── AGENTS.md ← tool-agnostic entry point for AI agents
├── LICENSE ← MIT
├── CITATION.cff ← cite-this-repo metadata
├── CLAUDE.md ← rules for any Claude Code session in this folder
├── docs/
│ ├── extraction-prompt.md ← the canonical extraction prompt (v2)
│ └── audit-rubric.md ← rubric the Layer 2 auditor uses
├── notes/ ← 175 curated paper notes (the source of truth)
│ └── nbs-2026-02-spoor-2026.md
├── index/ ← derived views, all rebuildable
│ ├── synapse.db ← SQLite + FTS5 (~1.7 MB)
│ ├── papers.csv ← flat tabular export
│ ├── library.bib ← BibTeX, one @article per note
│ └── topics.json ← 14-domain controlled vocabulary
├── tools/ ← the Python pipeline
│ ├── pdf_to_text.py ← PDF → plain text (pdftotext + pdfplumber)
│ ├── prepare_paper.py ← bundle a paper for extraction
│ ├── ingest_batch.py ← walk a folder of PDFs
│ ├── validate_note.py ← verbatim-abstract + bib + taxonomy + anchors
│ ├── audit_note.py ← two-layer faithfulness audit (Layer 1 + Layer 2)
│ ├── build_index.py ← rebuild SQLite from notes/
│ ├── export_csv.py ← rebuild papers.csv
│ └── export_bibtex.py ← rebuild library.bib
├── library/
│ ├── NBS/2026-02/
│ │ ├── manifest.tsv ← trusted bibliographic source for the batch
│ │ └── missing.tsv ← papers NBS listed but PDFs unavailable
│ └── AMJ/{vol-68-5,vol-68-6,vol-69-1}/
│ └── manifest.tsv ← per-issue manifests for the AMJ pilot
│ (pdfs/ and text/ are intentionally NOT published — see Copyright below)
└── .synapse/
├── config.yaml ← validator policy + custom-field schema
└── schema.sql ← SQLite schema
Quick start
git clone https://github.com/binqi20/management-research-notes.git
cd management-research-notes
# Optional: rebuild the SQLite index from the notes
python tools/build_index.py
# Query the library — answers come back in milliseconds
sqlite3 index/synapse.db "SELECT COUNT(*) FROM papers;"
sqlite3 index/synapse.db \
"SELECT paper_type, COUNT(*) FROM papers GROUP BY 1 ORDER BY 2 DESC;"
sqlite3 index/synapse.db \
"SELECT p.id, p.title FROM papers p JOIN topics t ON t.paper_id = p.id
WHERE t.topic = 'circular-economy' ORDER BY p.year DESC;"
# Full-text search across abstracts and analytic fields
sqlite3 index/synapse.db \
"SELECT id FROM papers_fts WHERE papers_fts MATCH 'stakeholder AND legitimacy';"
# Re-export BibTeX or CSV
python tools/export_bibtex.py
python tools/export_csv.py
Optional dependencies (for ingesting new PDFs locally)
brew install poppler # gives you pdftotext
pip install pyyaml pdfplumber # pyyaml: validator; pdfplumber: extraction fallback
The validator (tools/validate_note.py) reads the corresponding extracted
text file under library/<source>/<issue>/text/ to perform its verbatim-abstract
substring check. That folder is intentionally absent from the public repo
because the text is derived from copyrighted PDFs — see below. For a public
clone, the SQLite, CSV, and BibTeX queries above all work without it.
For AI agents
If you are an AI agent (or writing one) and want to consume or contribute
to this knowledge base, start with AGENTS.md — it is the
tool-agnostic entry point covering:
- the data formats you can consume without running any tooling (Markdown notes, SQLite + FTS5, CSV, BibTeX),
- the 7 hard rules any agent must follow when creating or editing notes,
- the faithfulness guarantees this library carries (every abstract verbatim-anchored, every prose field audited, zero contradictions),
- and how to cite the knowledge base if your agent surfaces a note.
For Claude Code-specific operational conventions (slash commands,
subagent dispatch, tool names), see CLAUDE.md as well.
Copyright and licensing
The code, the curated notes, the topic taxonomy, and the derived indexes in this repository are released under the MIT license and are © 2026 Binqi Tang.
The original journal articles that the notes summarize are NOT included.
They remain the copyright of their respective publishers and authors. To
reproduce a note end-to-end, obtain the source PDF through legitimate channels
(your institution, the publisher, or the author's preprint), drop it into
incoming/, and run the pipeline locally.
Each note contains a short verbatim abstract (treated as fair-use quotation/commentary) and an original analytic distillation written by the author of this repository. Notes cite the original work via DOI in every file's frontmatter and APA citation block.
How to cite
This repository ships with a CITATION.cff, which GitHub uses
to render a "Cite this repository" button in the right sidebar. It will give
you both APA and BibTeX automatically. Or, manually:
@software{tang_mgmt_research_notes_2026,
author = {Tang, Binqi},
title = {Management Research Notes: A File-Based Academic Knowledge
Base for Management and Business Sustainability Research},
year = {2026},
version = {0.19.1},
doi = {10.5281/zenodo.19564336},
url = {https://doi.org/10.5281/zenodo.19564336},
license = {MIT}
}
If you use a specific note, please also cite the original paper using the DOI in that note's APA citation block — the analytic distillation is a commentary, not a substitute.
Roadmap
This project is intended to be a long-running research-infrastructure project, not a one-shot data drop. The near-term roadmap:
- More monthly batches — extend NBS coverage with each new digest and continue the AMJ sweep beyond the current three-issue pilot, keeping paper IDs stable across updates.
- Additional journal sources — add Web of Science exports and journal
RSS feeds as parallel
library/{source}/{issue}/trees, reusing the same pipeline. - Audit layer hardening — cross-model auditing (run Layer 2 with a
second model vendor), automated re-audit on prompt changes, and a
lightweight public audit-summary CSV so consumers can see which prose
fields carry
PARTIALverdicts without needing to regenerate. - Vector search — only after the library passes ~5,000 notes; the SQLite FTS5 index is plenty for now.
- Community contributions — issues and pull requests welcome from collaborators who want to share extraction prompts, topic taxonomies, or curated subsets.
Acknowledgements
- The Network for Business Sustainability (NBS) for curating the monthly research digest that seeds this library.
- Anthropic Claude Code for the extraction-and-validation workflow that drives the ingestion pipeline.
- The authors of every paper in
notes/— without their original scholarship, there is nothing to distill.
Maintained by Binqi Tang. MIT licensed. Issues and PRs welcome.