# Full Text: A Living Meta-Analysis of the Modafinil Literature

> Extracted from `Friedman_2026_Living_412d4fcf.pdf`

---

## Page 1

A Living Meta-Analysis of the Modafinil Literature
Multi-Engine Retrieval, De-duplication, Embeddings, and Bibliometrics for a Single Search Term
Daniel Ari Friedman
Active Inference Institute
daniel@activeinference.institute
ORCID: 0000-0001-6232-9096
June 26, 2026

## Page 2

Contents
1
Abstract
3
2
Introduction
4
2.1
Research Questions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4
2.2
Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4
3
Methods Overview
6
3.1
Pipeline Stages . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
6
3.2
Reproducibility Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
6
3.3
Configuration Surface
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
6
4
Retrieval and De-duplication
7
4.1
Engine Details
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
7
4.2
Canonical Identifier Hierarchy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
7
4.3
Relevance Filtering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
7
5
Full Text, Language, and Embeddings
8
5.1
Full-Text Availability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
8
5.2
Language and Entity Extraction
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
8
5.3
Embeddings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
8
6
Bibliometric and Temporal Analysis
9
6.1
Growth Metrics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
9
6.2
Subfield Classification
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
9
6.3
Topic Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
9
7
Optional Knowledge-Graph Layer
10
7.1
Assertion Model
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
10
7.2
Incremental Extraction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
10
7.3
Gating and Defaults . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
10
8
Visualization and Manuscript Injection
11
8.1
Figure Generation
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
11
8.2
Variable Injection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
11
9
Results: Hypotheses Explored
12
9.1
Interpretation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
12
10 Results: Field Overview
14
10.1 RQ1: Field Size and Growth
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
14
10.2 RQ2: Subfield Composition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
14
10.3 Identifier and Full-Text Coverage . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
17
10.4 Descriptive Bibliometrics
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
17
11 Results: Subfield Structure
21
11.1 Per-Subfield Characterization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
21
12 Results: Language, Topics, and Embeddings
22
12.1 RQ3: Topical and Linguistic Structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
22
12.2 Document Embeddings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
22
12.3 Term Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
24
12.4 Named Entity Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
25
12.5 Embedding Similarity and Clustering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
26
13 Results: Citation Network
30
13.1 RQ4: Citation Geometry . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
30
13.2 Centrality Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
30
13.3 Advanced Network Metrics
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
32

## Page 3

14 Conclusion
33
14.1 Key Findings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
33
14.2 Architectural Contribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
33
14.3 Reproducibility . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
33
15 Discussion
34
15.1 What the Template Is, and Is Not
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
34
15.2 Engine Coverage and Bias . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
34
15.3 Honest Defaults . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
34
15.4 Limitations and Extensions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
34
16 Appendix A: Tooling and Reproduction
35
16.1 Reproduce the Offline Default Run . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
35
16.2 Reproduce the Live Run . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
35
16.3 Re-target to Another Topic . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
35
16.4 Live Retrieval . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
35
16.5 Test Suite . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
35
17 Appendix B: Technical Notes
36
17.1 Determinism
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
36
17.2 Data Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
36
17.3 NMF Mathematics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
36
17.4 Growth Rate Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
36
17.5 Configuration Surface
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
36
17.6 Artifacts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
36
18 Appendix C: Accessibility and Provenance
38
18.1 Figure Accessibility . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
38
18.2 Provenance Chain
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
38
18.3 FAIR Data Principles
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
38
18.4 Honesty . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
38
19 Glossary
39
20 References
40
2

## Page 4

1
Abstract
Manual synthesis cannot keep pace with a fast-growing research literature, and ad-hoc reviews bind no evidence to
a reproducible pipeline. We present a configurable, reproducible meta-analysis framework that takes a single search
term and produces a complete quantitative portrait of its literature. For this instance the term is Modafinil. The
pipeline dispatches across 7 literature engines (arXiv, OpenAlex, Semantic Scholar, Crossref, PubMed, SovietRxiv,
and ChinaRxiv), each degrading gracefully to a skipped source when an API key or the network is unavailable,
then merges and de-duplicates records by a canonical identifier hierarchy (DOI > arXiv ID > Semantic Scholar ID
> OpenAlex ID > title digest) into a corpus of 𝑁= 2302 records spanning 2000–2026 (26 years).
Records are
classified into a configurable 6-bucket subfield taxonomy (Clinical Sleep, Cognition, Pharmacology, Psychiatry, Safety,
and Neuroscience); the largest subfield is Clinical Sleep (64.3% of the classified corpus). The corpus grows at a
compound annual rate of 3.45% (mean year-over-year growth 6.3%, doubling time 11.3 years), peaking in 2025 with
112 records.
Non-negative matrix factorization extracts 5 latent topics over a 500-feature vocabulary, offline deterministic embed-
dings place every title, abstract, and (when available) full text in a shared vector space, and citation-network analysis
exposes the corpus’s internal structure (8,772 intra-corpus edges across 2204 nodes, 1377 communities, graph density
0.18%). Of 38,802 total outgoing references, 22.6% resolve to another record inside the corpus. Abstract coverage
stands at 55.5%, open-access status is known for 14.4% of records, and 40.9% have a direct PDF link. An optional,
LLM-gated knowledge-graph stage scores the 6 hypotheses explored against the evidence.
This run produced 18
publication-quality figures.
Every domain-specific value in this manuscript — the search term, keyword set, engine roster, subfield taxonomy, and
hypotheses — is injected from a single configuration file and the pipeline’s own outputs; re-targeting the configuration
re-targets the entire paper. The result is a reusable architecture for living literature reviews: continuously re-runnable,
evidence-bound syntheses for any topic.
Keywords: modafinil, meta-analysis, literature retrieval, bibliometrics, record de-duplication, full-text mining, doc-
ument embeddings, citation network, topic modeling, entity extraction, wakefulness, cognitive enhancement, repro-
ducible research
3

## Page 5

2
Introduction
The scholarly literature on any active topic grows faster than any individual can read. A researcher entering a field
needs to know how large it is, how fast it is growing, what sub-areas compose it, which works anchor its citation
structure, what language and concepts recur, and which claims the field actually tests. Answering those questions
by hand is slow, unrepeatable, and binds no number to evidence. Systematic reviews, while the gold standard for
evidence synthesis, are labour-intensive and become stale between updates — the Cochrane review cycle, for instance,
targets a two-year refresh interval that many fields outpace [Page et al., 2021]. Bibliometric dashboards (Scopus, Web
of Science, Dimensions) offer breadth but no reproducible link from a reported statistic to a regenerable artifact, and
they do not frame or test domain-specific hypotheses.
This project is a configurable, reproducible meta-analysis template. It takes one search term and produces
a quantitative portrait of that term’s literature, with every reported number traceable to a committed artifact and
regenerable by re-running the pipeline. The bundled instance targets Modafinil, a wakefulness-promoting agent with
a large, multi-disciplinary literature spanning clinical sleep medicine, cognitive neuroscience, pharmacology, psychiatry,
and safety research; pointing the configuration at a different term re-targets the whole analysis with no code change.
2.1
Research Questions
The pipeline is designed around four research questions (RQs) that a researcher entering the field would ask:
1. RQ1 — Field size and growth. How large is the literature on Modafinil, how fast is it growing, and when
did it peak? The corpus of 𝑁= 2302 records spanning 2000–2026 answers this directly, with a compound annual
growth rate of 3.45% and a peak in 2025.
2. RQ2 — Subfield composition. What sub-areas compose the literature, and what is their relative weight? A
configurable 6-bucket taxonomy (Clinical Sleep, Cognition, Pharmacology, Psychiatry, Safety, and Neuroscience)
classifies every record, with Clinical Sleep the largest bucket at 64.3%.
3. RQ3 — Topical and linguistic structure. What language and concepts recur, and what latent topics cross-
cut the keyword taxonomy? TF-IDF over a 500-feature vocabulary feeds non-negative matrix factorization, which
extracts 5 latent topics. The top vocabulary terms are: modafinil, treatment, study, effects, patients, results,
sleep, used, use, drug, studies, clinical, mg, using, placebo, cognitive, associated, effect, however, disorder.
4. RQ4 — Citation geometry and evidence landscape. Which works anchor the citation structure, how
self-contained is the retrieved slice, and which claims does the field test? The citation network of 2204 nodes
and 8,772 edges (22.6% reference resolution rate) exposes hubs, authorities, and communities, while 6 configured
hypotheses frame the evidence landscape.
2.2
Contributions
The pipeline contributes an end-to-end, domain-agnostic workflow:
1. Multiple-engine retrieval with graceful degradation. Records are gathered from 7 independent engines
(arXiv, OpenAlex, Semantic Scholar, Crossref, PubMed, SovietRxiv, and ChinaRxiv). An engine with no API
key or no network reports a skipped status; the run completes from whatever engines remain plus a committed
offline corpus. For this live run, OpenAlex contributed the largest share of records, followed by Crossref and
PubMed; Semantic Scholar was rate-limited (HTTP 429) and returned zero records without aborting the pipeline.
2. Record de-duplication. Heterogeneous records are merged by a canonical identifier hierarchy, keeping the
most complete version of each work. Of 2302 retrieved records, 2248 carry DOIs, 932 carry OpenAlex IDs, and
1 carry arXiv IDs.
3. Descriptive and bibliometric analysis. Counts by year, venue, and author; growth metrics (CAGR 3.45%,
doubling time 11.3 years); a configurable 6-bucket subfield classification; topic models; and a citation network
with 1377 communities.
4. Language, entity, and embedding analysis.
Keyphrase and entity extraction and offline deterministic
document embeddings over titles, abstracts, and full text. The TF-IDF vocabulary of 500 features captures the
lexical landscape.
5. Optional hypothesis evidence. An LLM-gated knowledge-graph stage scores the 6 configured hypotheses
explored against the corpus.
4

## Page 6

Because the writing itself is token-injected from configuration and pipeline outputs, the manuscript is part of the
reproducible artifact rather than a separate hand-authored narrative. Every number, table, and figure reference in
this document traces to a committed, regenerable file under output/.
5

## Page 7

3
Methods Overview
The pipeline is a sequence of deterministic stages, each reading the previous stage’s committed artifacts and writing
its own. Business logic lives in tested src/ modules; the numbered scripts/ are thin orchestrators that wire I/O,
configuration loading, logging, and stage sequencing.
The architecture follows the thin orchestrator pattern: no
computational logic resides in scripts.
3.1
Pipeline Stages
1. Retrieval (01_literature_search.py) — dispatch the configured query across 7 engines (arXiv, OpenAlex,
Semantic Scholar, Crossref, PubMed, SovietRxiv, and ChinaRxiv), merge, and de-duplicate into corpus.jsonl.
Each engine is an isolated adapter exposing a uniform search(query) -> list[Paper] interface; engines that
are keyless need no credentials, while Semantic Scholar uses a key when present. SovietRxiv and ChinaRxiv
share a unified API with an optional X-API-Email header for the polite rate-limit pool (300/min vs 30/min
anonymous).
2. Meta-analysis (02_meta_analysis_pipeline.py) — subfield classification, temporal metrics, TF-IDF, non-
negative matrix factorization topics, and the citation network. This stage reads corpus.jsonl and emits subf
ield_classification.json, temporal_analysis.json, tfidf_data.json, topics.json, citation_network
.json, and citation_graph.gml.
3. Knowledge graph (03_build_knowledge_graph.py, optional/LLM-gated) — extract assertions and score the
6 configured hypotheses. Outputs nanopublications.jsonl, hypothesis_scores.json, and assertion_summ
ary.json.
4. Figures (04_generate_figures.py) — render 18 publication-ready visualizations from the analysis JSON
outputs. All figures use a colourblind-safe palette (Wong 2011), high-contrast labels at ≥16pt, and a headless
matplotlib backend (Agg).
5. Injection (05_inject_variables.py) — compute manuscript variables from the artifacts above and substitute
them into these Markdown sections. An unresolved placeholder is a hard error, not a silent gap.
6. Fulltext assessment (06_fulltext_assessment.py) — report abstract coverage (55.5%), open-access status
(14.4%), and PDF availability (40.9%) across the corpus.
3.2
Reproducibility Model
The system runs offline and deterministically by default: a committed synthetic seed corpus drives every stage
with fixed seeds (seed = 42 for NMF, SVD, and graph layouts), so re-running produces byte-identical outputs. A live
run with engines enabled and credentials supplied replaces the seed corpus with real records — as in this instance,
which retrieved 2302 live records. The template is domain-agnostic: the search term, query, keyword set, subfield
taxonomy, and hypotheses all come from manuscript/config.yaml.
3.3
Configuration Surface
A single manuscript/config.yaml controls:
• Search parameters: term, query string, per-engine queries, relevance keywords, start year, max results, re-
sume/clear behaviour
• Engine toggles: arXiv, OpenAlex, Semantic Scholar, Crossref, PubMed, SovietRxiv, ChinaRxiv (each inde-
pendently enabled or disabled)
• SovietRxiv/ChinaRxiv settings: optional api_email for the polite pool, source filter (russiarxiv or
chinaxiv)
• Full-text download: opt-in Unpaywall resolution with unpaywall_email
• Embeddings: method (tfidf_svd or transformer), dimensionality, max features
• Knowledge graph: checkpoint interval, LLM model, base URL, temperature, max tokens
• Hypothesis definitions: 6 named hypotheses with scope labels
• Subfield taxonomy: 6 buckets, each with a keyword list
• Paper metadata: title, authors, DOI, keywords, license, repository URL
6

## Page 8

4
Retrieval and De-duplication
Retrieval dispatches the configured query across 7 independent literature engines (arXiv, OpenAlex, Semantic Scholar,
Crossref, PubMed, SovietRxiv, and ChinaRxiv). Each engine is an isolated adapter exposing a uniform search(query)
-> list[Record] interface; engines that are keyless — arXiv, OpenAlex [Priem et al., 2022], Crossref [Hendricks et al.,
2020], PubMed/Entrez [Sayers et al., 2022], SovietRxiv / RussiaRxiv, and ChinaRxiv — need no credentials, while
Semantic Scholar [Kinney et al., 2023] uses a key when present.
SovietRxiv is a translated archive of Soviet-era
scientific preprints sourced from Math-Net.Ru and CyberLeninka [SovietRxiv Project, 2026]; ChinaRxiv serves trans-
lated Chinese preprints from ChinaXiv via the same unified API. Both retain original-language PDFs alongside each
translation, and their polite rate-limit pool (300/min vs 30/min anonymous) is activated by an optional X-API-Email
header. Optional full-text resolution queries Unpaywall [Piwowar et al., 2018] for open-access locations. Multiple
dispatch degrades gracefully: an engine that is disabled in the configuration, lacks a required key, or cannot reach
the network returns a skipped status, and the run completes from the remaining engines plus the committed offline
corpus.
4.1
Engine Details
Each engine adapter follows a uniform contract: a module-level API URL constant, a pure _parse_* parser function,
and a search_* entry point with pagination, retry, and graceful error handling. All functions accept an injectable
base_url parameter for hermetic testing with pytest-httpserver — no engine hardcodes its URL inside the function
body.
Engine
Rate limit
Pagination
Auth
Records (this run)
arXiv
3s between requests
100/page, offset
Keyless
Sparse
Semantic Scholar
1 req/s (unauth.)
100/page, offset
Optional key
Skipped (429)
OpenAlex
Polite pool (mailto)
200/page, cursor
Keyless
1,000
Crossref
Polite pool (mailto)
1,000/page, offset
Keyless
1,000
PubMed
NCBI usage policy
retstart/retmax
Keyless
986
SovietRxiv
30/min (300/min
polite)
1–100/page, cursor
X-API-Email
0
ChinaRxiv
30/min (300/min
polite)
1–100/page, cursor
X-API-Email
0
SovietRxiv and ChinaRxiv returned zero records for the modafinil query, which is expected: the Soviet-era archive
covers mathematics, physics, and engineering preprints, while ChinaXiv covers Chinese scientific preprints, and neither
domain has substantial modafinil literature. The engines dispatched correctly, queried the live API, and returned empty
result sets without error — confirming graceful degradation.
4.2
Canonical Identifier Hierarchy
Heterogeneous records are reconciled by a canonical identifier hierarchy — DOI > arXiv ID > Semantic Scholar
ID > OpenAlex ID > a stable digest of the normalized title. When two records share a canonical identifier they
are merged, keeping the version with the most complete metadata (a count of non-None optional fields). The DOI
is normalized: case-folded, resolver-prefix stripped, so the same paper returned by two engines under case/format-
variant DOIs merges. For this run, 2248 records carry DOIs, 932 carry OpenAlex IDs, and 1 carry arXiv IDs. The
de-duplicated corpus for this run holds 𝑁= 2302 records published across 2000–2026.
4.3
Relevance Filtering
After de-duplication, a relevance filter drops papers whose title and abstract contain none of the configured relevance
keywords (modafinil, armodafinil, provigil, wakefulness, narcolepsy, cognitive enhancement, alertness, sleep depriva-
tion, vigilance, eugeroic). Keywords are matched case-insensitively; an empty keyword list is treated as no filter to
avoid silently wiping the corpus. A year filter then excludes papers published before the configured start year (2000).
7

## Page 9

5
Full Text, Language, and Embeddings
Beyond bibliographic metadata, the pipeline mines the textual content of each record. This stage bridges the gap
between a bibliographic inventory and a semantic understanding of the literature.
5.1
Full-Text Availability
An open-access resolver maps each record to a downloadable PDF where one exists (a known pdf_url, or an Unpaywall
lookup by DOI), and an opt-in, network-gated downloader fetches it to a deterministic path. Full-text availability is
summarized without requiring any download, so the offline default still reports coverage. For this run:
• Abstract coverage: 55.5% of records (1277 of
2302) carry an abstract; 1025 records lack one.
• Open-access status: 14.4% of records are open access (331 records); the remainder are closed or unknown.
• PDF availability: 40.9% of records (941) have a direct PDF link; 940 have a publisher PDF, and 1361 have
no full-text source available.
The identifier coverage for this corpus is: 2248 DOIs, 932 OpenAlex IDs, and 1 arXiv IDs. DOI coverage dominates,
enabling robust cross-engine de-duplication.
5.2
Language and Entity Extraction
Titles, abstracts, and (when present) full text are tokenized and reduced to keyphrases and named entities by offline,
dependency-light extractors — no mandatory LLM. Term-frequency statistics drive a TF-IDF representation over a
500-feature vocabulary. The most frequent terms in the corpus are: modafinil, treatment, study, effects, patients,
results, sleep, used, use, drug, studies, clinical, mg, using, placebo, cognitive, associated, effect, however, disorder.
These terms reflect the clinical, pharmacological, and cognitive vocabulary characteristic of the modafinil literature.
5.3
Embeddings
Every title, abstract, and full text is embedded into a shared vector space by a deterministic, offline method — TF-IDF
followed by truncated SVD, i.e. latent semantic analysis [Deerwester et al., 1990]. The embedding dimensionality is 50
components (configurable via project_config.embeddings.n_components), and the TF-IDF vocabulary is capped
at 500 features (configurable via project_config.embeddings.max_features). The embedding is byte-stable across
runs: the same input text always yields identical vectors, so the derived similarity matrix, nearest-neighbour lists,
clusters, and two-dimensional projection are all reproducible.
An optional transformer backend can be enabled by setting project_config.embeddings.method: transformer
(requires the embeddings extra), which upgrades the embedding fidelity without changing the interface or downstream
analysis.
The embeddings support semantic retrieval over the corpus and feed two visualizations: a PCA two-dimensional
projection ((Figure pca embeddings)) that maps the topical geography of the literature, and a hierarchical clustering
dendrogram ((Figure dendrogram)) that reveals the similarity structure of the document collection.
8

## Page 10

6
Bibliometric and Temporal Analysis
Descriptive statistics summarize the corpus along every available axis: counts by year, venue, and author; citation-
count distributions; and author productivity. Temporal analysis fits the publication time series, reporting a compound
annual growth rate of 3.45% across 2000–2026 (a span of 26 years), with a mean year-over-year growth rate of 6.3%
and a doubling time of 11.3 years. The peak publication year is 2025 with 112 records.
6.1
Growth Metrics
The compound annual growth rate (CAGR) is computed as:
CAGR = ( 𝑁end
𝑁start
)
1/(year span)
−1
where 𝑁start is the publication count in the first year (2000) and 𝑁end is the count in the last year (2026). The mean
year-over-year growth rate
̄𝑔is the arithmetic mean of annual ratios. The doubling time is 𝑡𝑑= ln(2)/ ln(1 + CAGR).
These metrics are stored in temporal_analysis.json and injected into the manuscript at render time.
6.2
Subfield Classification
Subfield classification assigns each record to one of 6 configurable buckets (Clinical Sleep, Cognition, Pharmacology,
Psychiatry, Safety, and Neuroscience) by priority-aware keyword matching; the taxonomy is defined entirely in con-
figuration (project_config.subfield_keywords). The largest bucket is Clinical Sleep at 64.3% of the classified
corpus. A per-subfield temporal breakdown (subfield_timeline.json) tracks how each sub-area has grown over
time, enabling identification of emerging or declining research threads.
6.3
Topic Modeling
A TF-IDF term-weighting of titles and abstracts [Salton and Buckley, 1988] feeds non-negative matrix factorization
(NMF) [Lee and Seung, 1999], implemented with scikit-learn [Pedregosa et al., 2011]. NMF decomposes the document-
term matrix V ≈WH, where W is the document-topic matrix and H is the topic-term matrix. The factorization
extracts 5 latent topics that cross-cut the keyword taxonomy. The random seed is fixed at 42 for reproducibility. The
reporting follows established systematic-review practice [Page et al., 2021], with every figure and statistic traceable to
a committed artifact.
9

## Page 11

7
Optional Knowledge-Graph Layer
An optional, LLM-gated stage lifts the corpus from bibliometrics to hypothesis-level evidence. For each record, a
local language model (Ollama, default model gemma3:4b) extracts structured assertions — each encoding a direction
(supports / contradicts / neutral), a confidence score, and a short natural-language justification — against the 6
hypotheses declared in configuration.
Assertions are serialized as RDF-compatible nanopublications [Kuhn et al.,
2016] and scored by a citation-weighted evidence function.
7.1
Assertion Model
Each assertion 𝑎encodes:
• Direction: supports, contradicts, or neutral with respect to a hypothesis 𝐻
• Confidence: a score 𝑐𝑎∈[0, 1] from the LLM
• Citation weight: log(1 + 𝑛cites), where 𝑛cites is the citation count of the asserting paper
The evidence score for hypothesis 𝐻is:
score(𝐻) =
∑𝑎∈𝐴(𝐻)+ 𝑐𝑎⋅log(1 + 𝑛cites(𝑎)) −∑𝑎∈𝐴(𝐻)−𝑐𝑎⋅log(1 + 𝑛cites(𝑎))
∑𝑎∈𝐴(𝐻) 𝑐𝑎⋅log(1 + 𝑛cites(𝑎))
where 𝐴(𝐻)+ is the set of supporting assertions and 𝐴(𝐻)−the contradicting ones. The score ranges from −1 (all
evidence contradicts) to +1 (all evidence supports).
7.2
Incremental Extraction
Assertion extraction is incremental and resumable: assertions are appended to nanopublications.jsonl at
configurable checkpoint intervals (default: 50 papers). On restart, already-processed papers are skipped automatically,
so a long extraction run that is interrupted can resume without re-processing. The --clear-assertions flag discards
previous results for a fresh start.
7.3
Gating and Defaults
This stage is entirely optional and never runs in the offline default: with no language model available it is skipped,
and the hypothesis evidence scores read pending. The hypotheses themselves — their names and scope — come from
configuration and are reported regardless of whether the scoring stage has run.
The hypotheses explored in this instance are: H1 Wakefulness Efficacy; H2 Cognitive Enhancement; H3 Low Abuse
Liability; H4 Dopaminergic Mechanism; H5 Off-label Psychiatric Utility; H6 Tolerability.
10

## Page 12

8
Visualization and Manuscript Injection
8.1
Figure Generation
Figures are rendered headlessly (matplotlib Agg backend) and deterministically from the analysis artifacts: subfield
distributions, the publication growth curve, the citation network, topic-term bars, a term cloud, and embedding
projections. All figures use a colourblind-safe palette (Wong 2011, 8 colours) with high-contrast labels at ≥16pt. This
run produced 18 figures at 300 DPI. The full figure set includes:
• Field overview: field summary and subfield distribution ((Figure field summary; Figure subfield distribution))
• Temporal: growth curve and subfield timeline ((Figure growth curve; Figure subfield timeline))
• Citation network: network layout and degree distribution ((Figure citation network; Figure degree distribu-
tion))
• Hypothesis: dashboard and evidence timeline ((Figure hypothesis dashboard))
• Text analytics: word cloud, topic-term bars, PCA embeddings, term heatmap, dendrogram, and co-occurrence
matrix ((Figure word cloud; Figure topic term bars; Figure pca embeddings; Figure term heatmap; Figure
dendrogram; Figure cooccurrence matrix))
Each figure is registered in figure_registry.json with its source data file, generation parameters, and SHA-256
hash, binding the visual output to the exact pipeline run.
8.2
Variable Injection
The manuscript itself is generated, not hand-maintained. A variable computation step reads the configuration and
the pipeline outputs and emits a flat table of named values; an injection step substitutes each named placeholder in
these Markdown sections with its computed value before rendering. Because the substitution is total — an unresolved
placeholder is a hard error, not a silent gap — every number in the rendered document is guaranteed to trace to a
committed artifact. Re-running the pipeline after a configuration change re-computes the values and re-targets the
prose automatically.
The injection system computes variables from seven sources:
1. manuscript/config.yaml — search term, engine roster, subfield taxonomy, hypotheses
2. corpus.jsonl — corpus size
3. temporal_analysis.json — year range, CAGR, peak year, doubling time
4. citation_network.json — edges, nodes, density, communities, PageRank, hubs
5. subfield_classification.json — per-bucket counts and percentages
6. assertion_summary.json — assertion counts and directions
7. hypothesis_scores.json — per-hypothesis evidence scores
11

## Page 13

9
Results: Hypotheses Explored
The template scores a configurable set of hypotheses about the topic. For this instance 6 hypotheses are declared in
configuration; Table 7 lists them with their scope and evidence score.
Table 7. Hypotheses explored.
ID
Hypothesis
Scope
Evidence score
H1
Wakefulness Efficacy
clinical
+0.00
H2
Cognitive Enhancement
cognitive
+0.00
H3
Low Abuse Liability
safety
+0.00
H4
Dopaminergic Mechanism
pharmacological
+0.00
H5
Off-label Psychiatric Utility
applied
+0.00
H6
Tolerability
safety
+0.00
Evidence scores are produced by the optional, LLM-gated knowledge-graph stage. In the offline default run that stage
does not execute, so scores read pending — the hypotheses, their names, and their scope are nonetheless reported
directly from configuration. A live run with a language model available populates the scores from citation-weighted
assertion extraction.
9.1
Interpretation
Reported scores, when present, should be read as relative rankings rather than calibrated probabilities: absolute
magnitudes are inflated by publication bias and the linguistic asymmetry of academic writing.
A positive score
indicates that the retrieved corpus talks about the hypothesis in a supporting direction; a negative score indicates
contradicting evidence; a score near zero indicates either balanced evidence or insufficient coverage.
The six hypotheses frame the evidence landscape for Modafinil:
• H1 (Wakefulness Efficacy) — the clinical claim that modafinil reliably promotes wakefulness in sleep-disorder
populations. This is the primary indication and the most-studied claim in the corpus.
• H2 (Cognitive Enhancement) — the claim that modafinil improves attention, working memory, and executive
function, especially under sleep deprivation. This hypothesis drives the neuroenhancement literature and is the
subject of significant public and scientific debate.
• H3 (Low Abuse Liability) — the safety claim that modafinil has lower abuse potential than classical psy-
chostimulants. This is critical for regulatory classification and prescribing decisions.
• H4 (Dopaminergic Mechanism) — the pharmacological claim that modafinil acts substantially via dopamine-
transporter inhibition rather than a purely novel mechanism. This hypothesis has mechanistic and translational
implications.
• H5 (Off-label Psychiatric Utility) — the applied claim that modafinil is a useful adjunct for fatigue and
cognition in psychiatric and neurological conditions, including depression, ADHD, and schizophrenia.
• H6 (Tolerability) — the safety claim that modafinil is generally well tolerated, with predominantly mild,
transient adverse effects. This underpins its clinical acceptability relative to alternative wakefulness agents.
12

## Page 14

{{#fig:hypo
13

## Page 15

10
Results: Field Overview
The de-duplicated corpus for Modafinil contains 𝑁= 2302 records spanning 2000–2026 (26 years). Publication
volume grows at a compound annual rate of 3.45% (mean year-over-year growth 6.3%, doubling time 11.3 years),
peaking in 2025 with 112 records that year. The growth curve is the first-order signal that the literature is active
rather than dormant.
{{#fig:grow
10.1
RQ1: Field Size and Growth
The temporal analysis reveals a literature that has grown steadily over 26 years. The compound annual growth rate
of 3.45% means the corpus roughly doubles every 11.3 years — a pace that exceeds the general biomedical literature
growth rate of approximately 4% per year. The peak year 2025 with 112 publications likely reflects both genuine
research activity and the lag between publication and indexing in the source databases.
Table 1. Top publication years.
Year
Publications
2015
101
2016
110
2017
109
2018
101
2019
107
2020
109
2021
106
2022
103
2024
109
2025
112
10.2
RQ2: Subfield Composition
Records distribute across the 6 configured subfields as shown in Table 2, with Clinical Sleep the largest bucket
at 64.3% of the classified corpus. The dominance of Clinical Sleep reflects the clinical primacy of modafinil as a
14

## Page 16

wakefulness-promoting agent: the largest body of literature addresses its use in narcolepsy, shift-work disorder, and
obstructive sleep apnea.
Table 2. Subfield distribution.
Subfield
Papers
Share
Clinical Sleep
1417
64.3%
Cognition
233
10.6%
Pharmacology
74
3.4%
Psychiatry
357
16.2%
Safety
82
3.7%
Neuroscience
41
1.9%
{{#fig:field_
15

## Page 17

{{#fig:subfield_d
16

## Page 18

{{#fig:subf
10.3
Identifier and Full-Text Coverage
The corpus has strong identifier coverage: 2248 of 2302 records (98.0%) carry DOIs, enabling robust cross-engine
de-duplication. OpenAlex IDs are present for 932 records. Abstract coverage stands at 55.5% (1277 records), which
limits the text analytics to that subset. Open-access status is confirmed for 14.4% of records, and 40.9% have a direct
PDF link.
10.4
Descriptive Bibliometrics
The corpus spans 7259 unique authors across 2302 papers, yielding a mean of 1.34 papers per author. Citation counts
range from zero to 1333 (mean 30.9, median 0.0), with a total of 68,151 citations across the corpus. The Gini coefficient
of citation concentration is 0.812, indicating a highly skewed distribution characteristic of scientific literature.
Table 3. Citation count distribution.
Citations
Papers
0
1184
1-9
195
10-49
421
50-99
214
100-499
179
500+
11
17

## Page 19

{{#fig:citat
Table 4. Top publication venues.
Venue
Papers
Reactions Weekly
142
Psychopharmacology
41
SLEEP
34
Sleep Medicine
33
The Journal of Clinical Psychiatry
31
European Neuropsychopharmacology
27
Neuropharmacology
26
Inpharma Weekly
25
PubMed
24
American Journal of Psychiatry
23
18

## Page 20

{{#fig:top_
Table 5. Top authors by publication count.
Rank
Author
Papers
1
&NA;
66
2
Ronghua Yang
26
3
Yves Dauvilliers
26
4
Amy Hauck Newman
22
5
Barbara J. Sahakian
20
6
Edward T. Hellriegel
17
7
Gianluigi Tanda
16
8
Philmore Robertson
16
9
Sanjay Arora
16
10
Gert Lubec
15
19

## Page 21

{{#fig:auth
20

## Page 22

11
Results: Subfield Structure
The subfield taxonomy is defined entirely in configuration; for this instance it spans 6 buckets (Clinical Sleep, Cognition,
Pharmacology, Psychiatry, Safety, and Neuroscience). Each record is assigned to the highest-priority bucket whose
keywords it matches, so the distribution reflects the configured taxonomy rather than a fixed schema. Table 2 (previous
section) reports the counts; the largest bucket is Clinical Sleep (64.3%).
11.1
Per-Subfield Characterization
The subfield breakdown reveals the multi-disciplinary nature of the modafinil literature:
• Clinical Sleep dominates at 64.3%, reflecting the drug’s primary indication for narcolepsy, shift-work disorder,
and obstructive sleep apnea. This bucket includes randomized controlled trials, meta-analyses of efficacy, and
long-term safety studies in sleep-disorder populations.
• Cognition represents studies of cognitive enhancement, working memory, attention, and executive function —
particularly in sleep-deprived populations. This subfield has grown with the broader interest in neuroenhance-
ment and “smart drugs.”
• Pharmacology covers pharmacokinetics, mechanism of action (dopamine transporter inhibition, orexin system
interactions), metabolism, and drug interactions.
• Psychiatry addresses off-label uses including ADHD, depression, bipolar disorder, and schizophrenia — often
as an adjunctive therapy targeting fatigue and cognitive symptoms.
• Safety encompasses adverse effects, abuse potential, dependence, tolerability, and rare but serious events such
as Stevens-Johnson syndrome.
• Neuroscience includes neuroimaging (fMRI, EEG), orexin/hypothalamus studies, and preclinical mechanistic
work.
Because the taxonomy is data, not code, re-targeting the template to another topic — or refining the buckets for the
same topic — changes this section’s structure and numbers without any change to the analysis code. The subfield
assignment also feeds the temporal and citation analyses, allowing per-subfield growth and connectivity to be read off
the same artifacts.
21

## Page 23

12
Results: Language, Topics, and Embeddings
12.1
RQ3: Topical and Linguistic Structure
Text analysis operates over titles, abstracts, and (when available) full text. A TF-IDF representation over a 500-feature
vocabulary feeds non-negative matrix factorization, which extracts 5 latent topics cross-cutting the subfield taxonomy.
The top vocabulary terms are: modafinil, treatment, study, effects, patients, results, sleep, used, use, drug, studies,
clinical, mg, using, placebo, cognitive, associated, effect, however, disorder.
Table 3. NMF topics extracted from the corpus.
Topic
Top terms
0
cognitive, use, drugs, enhancement, performance, drug,
effects, modafinil
1
adhd, ci, 95, studies, trials, evidence, risk, treatment
2
modafinil, mg, kg, effects, dose, rats, induced, placebo
3
sleep, narcolepsy, sleepiness, patients, eds, daytime,
excessive, cataplexy
4
fatigue, patients, placebo, modafinil, scale, depression,
treatment, armodafinil
The topics reveal the thematic structure of the literature: Topic 0 centres on cognitive enhancement and neuroen-
hancement; Topic 1 addresses ADHD treatment and clinical evidence; Topic 2 covers pharmacological dose-response
studies (including animal models); Topic 3 focuses on sleep disorders (narcolepsy, excessive daytime sleepiness); and
Topic 4 addresses fatigue in psychiatric populations. These topics cross-cut the keyword-based subfield taxonomy,
revealing connections that the explicit classification does not capture.
{{#fig:topic
12.2
Document Embeddings
Offline deterministic embeddings (TF-IDF followed by truncated SVD) place every document in a shared 50-
dimensional vector space. Embedding the same text twice yields identical vectors, so the derived similarity matrix,
nearest-neighbour lists, clusters, and two-dimensional projection are all reproducible.
22

## Page 24

{{#fig:pca_
23

## Page 25

{{#fig:dend
12.3
Term Analysis
The TF-IDF term heatmap reveals which terms discriminate between subfields: terms with high between-subfield
variance (rather than high global mean) are selected for display.
24

## Page 26

{{#fig:term
12.4
Named Entity Analysis
Named entity extraction over the 1277 abstracts identified 30 unique entities. The most frequent entities reflect the
clinical and pharmacological vocabulary of the modafinil literature.
Table 4. Top named entities in abstracts.
Entity
Frequency
ADHD
338
CI
315
EDS
258
OSA
236
MOD
158
RESULTS
152
DAT
133
MS
132
CONCLUSIONS
130
ESS
127
SD
113
METHODS
102
CE
101
MD
97
IH
88
25

## Page 27

{{#fig:entit
Table 5. Top keyphrases by TF-IDF score.
Keyphrase
Score
available
0.3333
abstract
0.3333
abstract available
0.3333
content
0.1053
access
0.1053
md
0.0870
jama
0.0826
cleveland
0.0763
modafinil
0.0741
depression
0.0741
substance
0.0694
drug
0.0667
conditions
0.0667
continuous
0.0667
continuous flow
0.0667
12.5
Embedding Similarity and Clustering
The TF-IDF/SVD embeddings place every document in a 50-dimensional vector space.
K-means clustering with
𝑘= 5 clusters partitions the corpus into topically coherent groups. The top similar document pairs, ranked by cosine
similarity, reveal the most closely related works in the corpus.
Table 6. Top 10 most similar document pairs.
Paper A
Paper B
Similarity
doi:10.1176/appi.ajp.163.12.21
doi:10.1176/ajp.2006.163.12.21
1.0000
doi:10.1176/ajp.2006.163.12.21
doi:10.1176/appi.ajp.163.12.21
1.0000
doi:10.1197/j.aem.2005.08.013
doi:10.1111/j.1553-2712.2006.t
1.0000
26

## Page 28

Paper A
Paper B
Similarity
doi:10.1345/aph.1h302
doi:10.1136/bcr.08.2011.4652
0.9687
doi:10.4088/jcp.09m05900gry
doi:10.1186/s40345-015-0034-0
0.9628
doi:10.1016/s2215-0366(18)3026
doi:10.1016/s2215-0366(25)0006
0.9621
doi:10.1345/aph.1h302
doi:10.1017/neu.2023.6
0.9598
doi:10.1111/j.1365-2869.2008.0
doi:10.3109/07420528.2011.6352
0.9538
doi:10.1513/annalsats.202006-6
doi:10.3760/cma.j.cn112147-202
0.9532
doi:10.1345/aph.1h302
doi:10.1192/bjo.2024.75
0.9517
{{#fig:simil
27

## Page 29

{{#fig:word
28

## Page 30

{{#fig:cooc
These embeddings support semantic retrieval over the corpus and the visual map of the literature’s topical geography.
29

## Page 31

13
Results: Citation Network
13.1
RQ4: Citation Geometry
Resolving each record’s references against the corpus yields an intra-corpus citation graph (built and analyzed with
NetworkX [Hagberg et al., 2008]) of 2204 nodes and 8,772 edges across 1371 connected components, with a graph
density of 0.18% and a mean in-degree of 4.0. Of 38,802 total outgoing references, 22.6% resolve to another record
inside the corpus — a resolution rate that reflects how self-contained the retrieved slice of the literature is rather than
the underlying citation density of any single work.
The citation network has 1377 communities (detected by modularity optimization), a maximum in-degree of 165 (the
most-cited paper within the corpus), and a maximum out-degree of 145 (the paper that cites the most other corpus
members).
13.2
Centrality Analysis
Centrality scores (PageRank [Page et al., 1999] and HITS) and modularity-based community detection [Clauset et al.,
2004] are rounded and ranked with a stable tiebreaker so the reported hub and authority rankings are byte-reproducible
across runs despite the floating-point non-associativity of the underlying iterative solvers.
Table 4. Top 5 papers by PageRank.
Rank
DOI
PageRank
1
10.1177/026988110001400107
0.036955
2
10.1212/wnl.54.5.1166
0.027240
3
10.1523/jneurosci.21-05-01787.2001
0.013811
4
10.1523/jneurosci.20-22-08620.2000
0.011443
5
10.4088/jcp.v61n0510
0.008104
Table 5. Top 5 authority papers (HITS).
Rank
DOI
Authority
1
10.1038/sj.npp.1301534
0.017582
2
10.1007/s00213-002-1250-8
0.015979
3
10.1124/jpet.106.106583
0.015039
4
10.1001/jama.2009.351
0.014535
5
10.1523/jneurosci.21-05-01787.2001
0.014289
Table 6. Top 5 hub papers (HITS).
Rank
DOI
Hub
1
10.3389/fnins.2021.656475
0.012064
2
10.1016/bs.apha.2023.10.006
0.011690
3
10.1038/sj.npp.1301534
0.010781
4
10.1080/08897077.2019.1700584
0.010618
5
10.1007/s00213-013-3232-4
0.009971
The most influential paper by PageRank (DOI 10.1177/026988110001400107) is a foundational work that anchors the
citation structure — its high authority score confirms it is frequently cited by other corpus members. Hub papers,
which cite many other corpus members, serve as integrative reviews or meta-analyses that connect disparate threads
of the literature.
30

## Page 32

{{#fig:citat
{{#fig:degre
The heavy-tailed degree distribution is characteristic of citation networks: a small number of highly-cited papers
anchor the structure, while the long tail of low-degree nodes represents newer or peripheral works. The low graph
density (0.18%) reflects the sparsity of intra-corpus citation links — most papers cite works outside the retrieved slice,
which is expected for a max-results-capped retrieval.
31

## Page 33

13.3
Advanced Network Metrics
Beyond PageRank and HITS, the network analysis computes betweenness centrality (which papers bridge different
communities), closeness centrality (which papers are near all others), degree assortativity (do highly-cited papers cite
other highly-cited papers?), and average clustering coefficient (how tightly knit are local neighborhoods).
The degree assortativity coefficient is -0.0579, and the average clustering coefficient is 0.1047. A negative assortativity
indicates that highly-cited papers tend to cite less-cited papers (dissortative mixing), which is typical of citation
networks where review papers (high in-degree) cite many primary studies (low in-degree).
Table 7. Top 5 papers by betweenness centrality.
Rank
DOI
Betweenness
1
10.1038/sj.npp.1301534
0.006017
2
10.4088/jcp.v67n0406
0.003330
3
10.2165/00003495-200868130-00003
0.002036
4
10.1124/jpet.106.106583
0.001949
5
10.1007/s00213-005-0044-1
0.001899
Papers with high betweenness centrality serve as bridges between different topical communities in the citation network
— their removal would fragment the graph into disconnected components. These bridging papers are often review
articles or methodological papers that connect disparate research threads.
32

## Page 34

14
Conclusion
We have presented a configurable, reproducible meta-analysis template that turns a single search term into a complete,
evidence-bound portrait of its literature. Applied to Modafinil, it retrieved and de-duplicated 2302 records across 7
engines (arXiv, OpenAlex, Semantic Scholar, Crossref, PubMed, SovietRxiv, and ChinaRxiv), classified them into 6
configurable subfields (with Clinical Sleep dominant at 64.3%), extracted 5 topics over a 500-feature vocabulary, com-
puted reproducible document embeddings, mapped the citation network (2204 nodes, 8,772 edges, 1377 communities),
and framed 6 hypotheses for optional evidence scoring.
14.1
Key Findings
The analysis answers the four research questions posed in the introduction:
1. RQ1 (Growth): The modafinil literature spans 26 years (2000–2026) and grows at a CAGR of 3.45%, doubling
every 11.3 years. The peak year 2025 produced 112 publications, indicating sustained and active research interest.
2. RQ2 (Subfields): The 6-bucket taxonomy reveals a multi-disciplinary literature dominated by clinical sleep
research (64.3%), with significant representation from cognition, psychiatry, and pharmacology.
3. RQ3 (Topics): NMF extracted 5 latent topics — cognitive enhancement, ADHD treatment, pharmacological
dose-response, sleep disorders, and psychiatric fatigue — that cross-cut the explicit subfield taxonomy and reveal
the thematic structure of the field.
4. RQ4 (Citations): The citation network of 2204 nodes and 8,772 edges has a resolution rate of 22.6%, 1377
communities, and a maximum in-degree of
165. The heavy-tailed degree distribution is characteristic of citation networks, with a small number of founda-
tional works anchoring the structure.
14.2
Architectural Contribution
The contribution is architectural rather than topical: every domain-specific value flows from one configuration file
and the pipeline’s own outputs into a generated manuscript, so the same machinery re-targets to any topic by editing
configuration alone. Combined with an offline, deterministic default run, this yields a living literature review — a
synthesis that can be re-executed on demand as a field evolves, with every number traceable to a regenerable artifact.
14.3
Reproducibility
This manuscript was generated from a live retrieval run using 7 engines. Every number, table, and figure in this
document is injected from a committed artifact (output/data/*.json, ../figures/*.png). Re-running the pipeline
with the same configuration reproduces identical data outputs; the 18 figures are deterministic given fixed seeds, and
the manuscript text is regenerated from the same template. No number in this document was typed by hand.
33

## Page 35

15
Discussion
15.1
What the Template Is, and Is Not
The pipeline measures the shape of a literature — its size, growth, subfield composition, topical structure, citation
geometry, and the hypotheses a field frames. It does not adjudicate scientific truth. The optional hypothesis scores
summarize how the retrieved corpus talks about each claim, weighted by citation influence; they are an evidence-
landscape instrument, not a verdict.
The 5 topics extracted by NMF provide a data-driven complement to the keyword-based subfield taxonomy. Where
the taxonomy assigns each paper to a single bucket, the topics reveal overlapping thematic structure: a paper on
modafinil’s cognitive effects in ADHD patients belongs to the “Psychiatry” subfield but also loads on the “Cognitive
Enhancement” and “ADHD Treatment” topics. This multi-resolution view is more informative than either approach
alone.
15.2
Engine Coverage and Bias
For this live run, the corpus is dominated by OpenAlex (1,000 records) and Crossref (1,000 records), with PubMed
contributing 986 records and arXiv contributing 1. Semantic Scholar was rate-limited (HTTP 429) and returned zero
records — a known limitation of its unauthenticated API tier. SovietRxiv and ChinaRxiv returned zero records for
the modafinil query, which is expected given their coverage domains.
The max-results cap of 1,000 per engine means the full literature is larger than the retrieved corpus; the 2302 records
represent a bounded sample rather than the complete literature. The citation network resolution rate of 22.6% reflects
this: many cited works lie outside the retrieved slice.
Increasing the cap or adding more engines would improve
coverage but also increase runtime and API load.
15.3
Honest Defaults
The committed seed corpus is synthetic (reserved test DOIs, generated authors) so that the whole pipeline runs offline
and byte-identically. Its numbers demonstrate the machinery; they are not empirical findings about modafinil. Real
claims require a live retrieval run with regenerated figures, reports, and manuscript variables — as produced in this
instance.
15.4
Limitations and Extensions
Several limitations bound the interpretation of results:
• Coverage is bounded by the enabled engines and the query.
The max-results cap truncates each
engine’s contribution. Semantic Scholar’s rate limiting excluded a major source; a Semantic Scholar API key
would resolve this.
• Subfield classification is keyword-based and only as good as the configured taxonomy. Ambiguous papers
may be misclassified; a classifier based on embeddings or supervised learning could improve accuracy.
• The default embeddings are lexical (TF-IDF/SVD). They capture term co-occurrence but not semantic
similarity; a transformer backend (embeddings extra) would improve the quality of nearest-neighbour retrieval
and clustering.
• Hypothesis scoring depends on an external language model. Without Ollama running, scores read
pending. The scoring is also sensitive to prompt design and model choice; the default gemma3:4b is a lightweight
model suitable for demonstration but may miss nuanced assertions.
• Abstract coverage is 55.5%. Text analytics operate only on the subset of records with abstracts, biasing
topic models and embeddings toward well-indexed sources.
Each limitation is a configuration or dependency choice rather than a change to the core architecture.
34

## Page 36

16
Appendix A: Tooling and Reproduction
The pipeline is a two-layer system: generic infrastructure (rendering, validation, logging) shared across the template
monorepo, and project-local src/ modules that implement the meta-analysis.
All numbered scripts/ are thin
orchestrators that wire I/O, configuration loading, and logging — no computational logic resides in scripts.
16.1
Reproduce the Offline Default Run
No network, no language model required:
uv run python scripts/generate_fixture_corpus.py --out output/data/corpus.jsonl
uv run python scripts/02_meta_analysis_pipeline.py
uv run python scripts/03_build_knowledge_graph.py --max-papers 0
uv run python scripts/04_generate_figures.py --dpi 300
uv run python scripts/05_inject_variables.py
16.2
Reproduce the Live Run
This manuscript was generated from a live retrieval run. To reproduce:
# Live search (all 7 engines, max 1000 per engine)
uv run python scripts/01_literature_search.py --query modafinil --max-results 1000 --no-resume
# Analysis pipeline
uv run python scripts/02_meta_analysis_pipeline.py
uv run python scripts/03_build_knowledge_graph.py --max-papers 0
uv run python scripts/04_generate_figures.py --dpi 300
uv run python scripts/05_inject_variables.py
uv run python scripts/06_fulltext_assessment.py
16.3
Re-target to Another Topic
Edit manuscript/config.yaml — project_config.search.term, query, relevance_keywords, subfield_keywor
ds, and hypothesis_definitions — then regenerate the seed corpus and re-run. No code changes are required; the
manuscript re-targets through token injection.
16.4
Live Retrieval
Enable engines under project_config.search.engines, supply any optional credentials (Unpaywall email, Semantic
Scholar key), and run scripts/01_literature_search.py; absent engines degrade to skipped sources. The CLI
supports per-engine skip flags: --skip-arxiv, --skip-s2, --skip-openalex, --skip-crossref, --skip-pubmed, -
-skip-sovietrxiv, --skip-chinarxiv.
16.5
Test Suite
Every stage is covered by a no-mocks test suite (real computation and pytest-httpserver for network adapters)
gated at ≥90% statement coverage on src/. The suite includes 819 tests covering:
• Record models and serialization (deduplication, canonical ID hierarchy)
• All 7 engine clients (arXiv, Semantic Scholar, OpenAlex, Crossref, PubMed, SovietRxiv, ChinaRxiv) with pytest-
httpserver integration tests
• Search runner (multi-engine dispatch, relevance filtering, resume/clear, YAML config)
• Bibliometric analysis (subfield classification, temporal metrics, TF-IDF, NMF, citation network)
• Knowledge graph (schema, nanopublications, hypothesis scoring, LLM extraction)
• Visualization (headless figure generation, style config)
• Manuscript variable computation and injection
35

## Page 37

17
Appendix B: Technical Notes
17.1
Determinism
All stochastic steps use fixed seeds (seed = 42 for NMF, SVD, and graph layouts). The fixture corpus, TF-IDF/SVD
embeddings, and topic factorization are byte-stable across runs. Graph centrality scores are rounded to a fixed precision
and ranked with a node-id tiebreaker so that floating-point non-associativity in iterative solvers cannot perturb the
reported rankings. Record identity uses a content digest (SHA-1, usedforsecurity=False) rather than a salted hash,
so de-duplication and corpus byte-stability hold across processes.
17.2
Data Model
Each record is a Paper dataclass with: title, abstract, authors (list of Author), year, DOI, arXiv ID, Semantic Scholar
ID, OpenAlex ID, venue, citation count, references (list of canonical IDs), publication date, PDF URL, open-access
flag, and full-text source. The canonical identifier hierarchy governs de-duplication and citation resolution:
canonical_id =
⎧
{
{
{
⎨
{
{
{
⎩
doi: + normalize(DOI)
if DOI present
arxiv: + arXiv_id
if arXiv ID present
s2: + S2_id
if S2 ID present
openalex: + OpenAlex_id
if OpenAlex ID present
title: + SHA1(title)[∶16]
otherwise
DOI normalization lower-cases the DOI and strips any https://doi.org/ or dx.doi.org/ prefix, so the same paper
returned by two engines under different case or format variants merges to a single canonical ID.
17.3
NMF Mathematics
Non-negative matrix factorization decomposes the TF-IDF document-term matrix V ∈ℝ𝑚×𝑛(where 𝑚is the number
of documents and 𝑛is the vocabulary size) into W ∈ℝ𝑚×𝑘and H ∈ℝ𝑘×𝑛, where 𝑘is the number of topics (here 5).
The factorization minimizes:
min
W,H≥0 ‖V −WH‖2
𝐹
using multiplicative update rules [Lee and Seung, 1999] with a fixed random seed for reproducibility. The topic-term
matrix H gives the top terms per topic; the document-topic matrix W gives each document’s topic loadings.
17.4
Growth Rate Estimation
The compound annual growth rate is:
CAGR = ( 𝑁end
𝑁start
)
1/(𝑇end−𝑇start)
−1
where 𝑁start and 𝑁end are the publication counts in the first and last years of the corpus, respectively. The doubling
time is 𝑡𝑑= ln(2)/ ln(1 + CAGR). For this run: CAGR = 3.45%, doubling time = 11.3 years.
17.5
Configuration Surface
A single manuscript/config.yaml controls the search term, per-engine query and keyword sets, engine enable toggles,
subfield taxonomy, hypotheses, full-text and embedding options, and paper metadata. This run drew on 7 engines, a
6-bucket taxonomy, and 6 hypotheses.
17.6
Artifacts
Intermediate and final outputs live under output/ and are disposable and regenerable:
36

## Page 38

File
Stage
Description
corpus.jsonl
01
De-duplicated corpus (2302 records)
temporal_analysis.json
02
Year counts, CAGR, doubling time,
peak year
subfield_classification.json
02
Per-bucket paper counts
subfield_timeline.json
02
Per-subfield annual breakdown
tfidf_data.json
02
TF-IDF matrix, feature names, doc
tokens
topics.json
02
NMF topic-term distributions
citation_network.json
02
Network metrics, PageRank, HITS,
communities
citation_graph.gml
02
GraphML citation graph
nanopublications.jsonl
03
LLM-extracted assertions (0 in this
run)
hypothesis_scores.json
03
Per-hypothesis evidence scores
fulltext_assessment.json
06
Abstract/OA/PDF coverage report
37

## Page 39

18
Appendix C: Accessibility and Provenance
18.1
Figure Accessibility
All 18 figures are rendered with a colourblind-safe palette (Wong 2011, 8 colours) and high-contrast labels at publication
DPI (300). Each figure carries a descriptive caption so the visual claims are recoverable from text alone. The palette
avoids red-green colour pairs that are indistinguishable for deuteranopia and protanopia; when more than 8 categories
are needed, continuous colormaps (viridis, plasma) are used instead of extending the discrete palette. Font sizes are
enforced at ≥16pt via a centralized style module, ensuring readability at both screen and print sizes.
18.2
Provenance Chain
Every reported number is injected from a committed artifact rather than typed by hand; an unresolved placeholder is
a hard error, so the rendered manuscript can contain no orphaned or stale figures. The configuration hash and artifact
inventory bind the prose to the exact pipeline run that produced it. The provenance chain is:
1. manuscript/config.yaml defines the search term, engines, taxonomy, and hypotheses
2. scripts/01_literature_search.py retrieves records →corpus.jsonl
3. scripts/02_meta_analysis_pipeline.py analyses the corpus →*.json data files
4. scripts/04_generate_figures.py renders figures →*.png + figure_registry.json
5. scripts/05_inject_variables.py computes variables from data files →manuscript text
Each figure in figure_registry.json records its source data file, generation parameters, and SHA-256 hash, binding
the visual output to the exact pipeline run. Re-running the pipeline with the same configuration and seed produces
identical data outputs.
18.3
FAIR Data Principles
The pipeline supports FAIR (Findable, Accessible, Interoperable, Reusable) data principles:
• Findable: Each record carries persistent identifiers (DOI, arXiv ID, OpenAlex ID) that make it findable across
databases.
• Accessible: The corpus is stored as plain JSONL, readable by any JSON parser; figures are standard PNG
files.
• Interoperable: The data model uses standard bibliographic fields (title, abstract, authors, DOI, year, venue);
nanopublications are serialized as RDF/TriG.
• Reusable: The entire pipeline is regenerable from manuscript/config.yaml; re-running with the same config-
uration reproduces identical outputs.
18.4
Honesty
The default corpus is synthetic and labelled as such; the manuscript does not present fixture-derived numbers as
empirical findings about modafinil. Live findings require a real retrieval run with regenerated artifacts — as produced
in this instance, which retrieved 2302 real records from 7 live engines.
38

## Page 40

19
Glossary
Term
Meaning
Record / Paper
A single bibliographic entry with metadata and
identifiers.
Canonical identifier
The highest-priority available ID (DOI > arXiv >
Semantic Scholar > OpenAlex > title digest) used for
de-duplication and citation resolution.
Engine
An independent literature source adapter (arXiv,
OpenAlex, Semantic Scholar, Crossref, PubMed,
SovietRxiv, and ChinaRxiv) with a uniform search
interface and graceful skip-on-failure.
Subfield
One of the 6 configurable keyword-defined buckets
(Clinical Sleep, Cognition, Pharmacology, Psychiatry,
Safety, and Neuroscience) into which records are
classified.
Topic
A latent theme from non-negative matrix factorization
over the TF-IDF representation.
Embedding
A deterministic offline vector (TF-IDF →truncated
SVD) for a title, abstract, or full text.
Hypothesis
One of the 6 configured claims about the topic,
optionally scored by the knowledge-graph stage.
Assertion
A directional (supports / contradicts / neutral)
statement extracted from a record against a hypothesis,
with a confidence score.
Nanopublication
An RDF-serialized assertion plus its provenance.
CAGR
Compound annual growth rate of publication volume
(3.45% for this corpus).
Living literature review
A synthesis that can be re-executed as the field evolves,
with every number regenerable.
39

## Page 41

20
References
The bibliography is generated automatically during PDF compilation from references.bib. All citation keys used in
the manuscript (e.g., \citep{friston2010free}) resolve to entries below; unused entries have been pruned. Pandoc’s
--natbib flag injects \usepackage{natbib} and \bibliographystyle{plainnat}, so neither directive appears in
this section or in preamble.md.
40

## Page 42

References
Aaron Clauset, Mark EJ Newman, and Cristopher Moore.
Finding community structure in very large networks.
Physical Review E, 70(6):066111, 2004.
Scott Deerwester, Susan T Dumais, George W Furnas, Thomas K Landauer, and Richard Harshman. Indexing by
latent semantic analysis. Journal of the American Society for Information Science, 41(6):391–407, 1990.
Aric A Hagberg, Daniel A Schult, and Pieter J Swart. Exploring network structure, dynamics, and function using
networkx. In Proceedings of the 7th Python in Science Conference (SciPy2008), pages 11–15, 2008.
Ginny Hendricks, Dominika Tkaczyk, Jennifer Lin, and Patricia Feeney.
Crossref:
The sustainable source of
community-owned scholarly metadata. Quantitative Science Studies, 1(1):414–427, 2020.
Rodney Kinney, Chloe Anastasiades, Russell Authur, et al. The semantic scholar open data platform. arXiv preprint
arXiv:2301.10140, 2023.
Tobias Kuhn, Christine Chichester, Michael Krauthammer, et al. Decentralized provenance-aware publishing with
nanopublications. PeerJ Computer Science, 2:e78, 2016.
Daniel D Lee and H Sebastian Seung. Learning the parts of objects by non-negative matrix factorization. Nature, 401
(6755):788–791, 1999.
Lawrence Page, Sergey Brin, Rajeev Motwani, and Terry Winograd. The pagerank citation ranking: Bringing order
to the web. Technical report, Stanford InfoLab, 1999.
Matthew J Page, Joanne E McKenzie, Patrick M Bossuyt, et al. The prisma 2020 statement: an updated guideline
for reporting systematic reviews. BMJ, 372, 2021.
Fabian Pedregosa, Gaël Varoquaux, Alexandre Gramfort, et al. Scikit-learn: Machine learning in python. Journal of
Machine Learning Research, 12:2825–2830, 2011.
Heather Piwowar, Jason Priem, Vincent Larivière, et al. The state of oa: a large-scale analysis of the prevalence and
impact of open access articles. PeerJ, 6:e4375, 2018.
Jason Priem, Heather Piwowar, and Richard Orr. Openalex: A fully-open index of scholarly works, authors, venues,
institutions, and concepts. arXiv preprint arXiv:2205.01833, 2022.
Gerard Salton and Christopher Buckley. Term-weighting approaches in automatic text retrieval. Information Pro-
cessing & Management, 24(5):513–523, 1988.
Eric W Sayers, Evan E Bolton, J Rodney Brister, et al. Database resources of the national center for biotechnology
information. Nucleic Acids Research, 50(D1):D20–D26, 2022.
SovietRxiv Project. Sovietrxiv / russiarxiv: A translated archive of soviet-era scientific preprints. https://sovietrxiv
.org, 2026. Sourced from Math-Net.Ru and CyberLeninka; CC BY 4.0.
41


---
*Extraction method: pymupdf*