Per-domain lexicon files. Each domain has one file containing a comma-separated term list plus a generated prose section that reframes the domain in thermodynamic terms.

Author: Christopher Betances — catqualia.com
License: CC BY 4.0 (see LICENSE)
Language: English
Format: UTF-8 text. 223 of the 353 domain files are Markdown; 130 carry the same structure without Markdown heading markers.

Read the "Content note" section below before loading this corpus. A subset of domains covers extremism, weapons, CBRN and bioweapons, trafficking, and illicit finance at a level of detail above a neutral glossary.

What is in this directory

	Files
Corpus mirror	733
This card (`README.md`) + `LICENSE`	2
Total on disk	735

The corpus mirror is 26,095,368 bytes. The byte total for the whole directory is not quoted here because it depends on the length of this card; read it with the second command below, which excludes nothing.

# corpus only (excludes this card and the license)
find . -type f ! -name README.md ! -name LICENSE | wc -l
# -> 733
find . -type f ! -name README.md ! -name LICENSE -printf '%s\n' \
  | python3 -c "import sys;print(sum(int(l) for l in sys.stdin))"
# -> 26095368

# whole directory, including this card and the license
find . -type f | wc -l
# -> 735
find . -type f -printf '%s\n' \
  | python3 -c "import sys;print(sum(int(l) for l in sys.stdin))"

Corpus layout

04_thermo_vocab/
├── README.md                     # this card
├── LICENSE                       # CC BY 4.0
├── <355 corpus files at top level>
├── tier_1_excellent/     41 files
├── tier_2_good/          95 files
├── tier_3_needs_upgrade/ 126 files
├── tier_4_poor/          91 files
└── _BLOOM/               25 files   # 6 at top level + 19 in _BLOOM/reports/

The tier directories are duplicates, not additional data

Every file inside tier_1_excellent/, tier_2_good/, tier_3_needs_upgrade/ and tier_4_poor/ is a byte-for-byte copy of a file that also exists at the top level of the directory. Verified by MD5 over the corpus only — the commands below exclude this card and the license, which are unique files that would otherwise inflate the distinct-content count by 2:

find . -type f ! -name README.md ! -name LICENSE -exec md5sum {} + \
  | awk '{print $1}' | sort -u | wc -l                                 # -> 380
find . -type f ! -name README.md ! -name LICENSE | wc -l               # -> 733

380 distinct file contents behind 733 paths. Every one of the 353 domain files appears exactly twice: once at the top level, once inside the single tier directory that classifies it (41 + 95 + 126 + 91 = 353).

Counts that follow from this:

Quantity	Value
Files on disk (corpus)	733
Distinct file contents	380
Distinct domain lexicons	353
Tier classification copies	353
`_BLOOM/` working files	25
Auxiliary data files	3

If you want 353 unique files, take the top level and drop the four tier directories; the tier directories are a curation overlay recording how each domain was graded. If you want the tier labels, read each domain from its tier directory and ignore the top-level copy.

Auxiliary files (not domain lexicons)

MANIFEST.jsonl (281 rows) — 71,030 B — manifest of the corpus.
thermodynamic_training_corpus.jsonl (1,991 rows) — 5,715,238 B — derived training corpus.
thermodynamic_training_corpus.jsonl copy — 2,752,504 B — an earlier, smaller revision under a copy-style filename. Not identical to the file above.

`_BLOOM/` working files

Pipeline scaffolding, not domain lexicons: SPEC.md, LEDGER.md, NEW_DOMAINS.md, QUEUE.txt, checksums_now.txt, pools_batch3.json, and 19 short per-worker reports under _BLOOM/reports/.

Record format

Two layouts are present. Both carry the same fields.

Layout A — Markdown (223 of 353 domains)

# <Domain> Vocabulary

*Enhanced with Thermodynamic Architecture*

---

## Section I: <Descriptive Section Title>

### Terms

term one, term two, term three, ...

### Thermodynamic Enhancement

<generated prose>

Headings quoted verbatim from the corpus:

# Acoustics Vocabulary
# Nuclear Terrorism Vocabulary
## Section I: Wave Propagation, Impedance, and Structural Acoustics
## Section I: Atmospheric Structure, Stability, and Convective Initiation
## Section I: ADME Fundamentals and Drug Transport
## Section II: Fissile Material, Proliferation Pathways & Nuclear Smuggling
## Coda — The Thermodynamics of Cardiology
### Thermodynamic Enhancement

Section numbering is Roman in most files (## Section I, ## Section II, …) and Arabic in others (## Section 1, ## Section 2, …). Some files use a descriptive interval instead, e.g. ## Section I — Avian Taxonomy, Anatomy, and Plumage Biophysics, and some introduce a section with ### Section I: … (three hashes) rather than two. 706 Section headings occur across the corpus.

Layout B — Plain text (130 of 353 domains)

The same content without Markdown markers:

<Domain> Vocabulary
Enhanced with Thermodynamic Architecture
1. <Descriptive Section Title>

Terms:
term one, term two, ...

Thermodynamic Enhancement:
<generated prose>

Field semantics

Title line — <Domain> Vocabulary, always the first line.
Section heading — groups terms into thematic blocks.
Term list — a single comma-separated block. Entries are domain vocabulary, acronyms, named entities and, in some domains, short attributive phrases (for example, botulinum neurotoxin sits alongside a quoted LD50 figure). Ordering is thematic, not alphabetical. 215 of the 353 domains contain at least one term repeated inside a list.
Thermodynamic Enhancement — generated prose, one block per section, 823 headings across the corpus, that restates the domain's concepts in thermodynamics vocabulary (entropy, free energy, dissipation, phase transitions). This is generated commentary, not sourced reference material. It applies thermodynamic metaphors to domains where no quantitative thermodynamic model exists, and it states specific numeric values and historical attributions without inline sourcing. Verify anything load-bearing; treat it as a vocabulary aid, not a citation source.

Corpus content and completeness

Measured over the 353 distinct domain lexicons:

Content state	Domains	Notes
Real, domain-specific term list	242	the usable part of the corpus
Coda-only (zero terms)	75	title plus one boilerplate paragraph
Placeholder (generic list)	36	never enriched; see below
Total	353

144,711 terms across the 242 real domains. Median 517 terms per such domain; range 2 to 3,620.
244 distinct term-list payloads.

Placeholder domains (36)

These files were never enriched. Each contains an identical generic thermodynamic term list beginning placeholder, thermodynamic, entropy, energy, free energy, …, with the domain name substituted into the title and the closing paragraph only. They contain no domain-specific content:

grep -rl "^placeholder, thermodynamic, entropy" . | wc -l   # -> 72 files

72 files = 36 domains × 2 copies (top level + tier). All 36 are classified tier_3_needs_upgrade. Affected domains include Explosives & Detonation Physics Vocabulary, Ballistics & Terminal Effects Vocabulary, Conspiracy Theory Psychology & Radicalization Vocabulary and Interrogation & Resistance to Interrogation Vocabulary.

Note the distinction: the generic placeholder list begins with the literal token placeholder. These 36 domains are ~3.6 KB each and are not the same thing as the 75 coda-only domains, which contain no term list at all.

Completeness tracks the tier label

The tier directories are consistent with content completeness, which is the most useful thing this corpus documents about itself:

Tier	Domains	Placeholder	Coda-only	Real
`tier_1_excellent`	41	0	3	38
`tier_2_good`	95	0	2	93
`tier_3_needs_upgrade`	126	36	5	85
`tier_4_poor`	91	0	65	26

Every placeholder sits in tier_3_needs_upgrade; tier_4_poor is 71% coda-only.

Domain overlap

Domains are not disjoint in subject matter. Near-duplicate pairs exist with distinct content, e.g. Genetics & Genomics Vocabulary and Genetics Genomics Vocabulary, and Meteorology & Atmospheric Dynamics Vocabulary and Meteorology Atmospheric Dynamics Vocabulary (the latter pair share an identical term list). Cryptography Deep Vocabulary and Cryptography & Zero-Knowledge Vocabulary also overlap. Selecting a domain by name may not match the file you expect.

File size distribution (all 733 corpus files, bytes)

Statistic	Bytes	Bucket	Files
Total	26,095,368	< 600 B	140
Min	173	600 B – 5 KB	75
p10	381	5 KB – 20 KB	64
p25	3,643	20 KB – 40 KB	423
Median	24,981	40 KB – 100 KB	24
p75	29,313	> 100 KB	7
p90	33,807
p99	91,269
Max	5,715,238
Mean	35,601

The 140 files under 600 B are 124 domain files (the coda-only stubs, in both naming forms) plus 16 small _BLOOM/reports/ worker notes. The largest file is thermodynamic_training_corpus.jsonl.

Two naming conventions coexist: extensionless (Cardiology Vocabulary) and .md with underscores (Aerospace_Engineering_Vocabulary.md). No canonical domain name appears in both forms, so a naive filename-based dedup will not reconcile the two sets.

Content note

The corpus covers a broad range of domains on equal terms, including domains not suitable for unrestricted release. The following domain files cover subject matter where the term lists go past a neutral glossary into operational vocabulary, application-level framing, and section structure organised around how the activity is carried out. They are listed by filename so a reader can decide before loading. Their content is not reproduced here.

These files are present unmodified. Nothing was removed, filtered or rewritten for this release.

Extremism, terrorism and political violence (15)

Accelerationism Vocabulary
Anarchism & Far-Left Extremism Vocabulary
Anti-Government Extremism Vocabulary
Conspiracy Theory Psychology & Radicalization Vocabulary   [placeholder — no domain content]
Eco-Terrorism Vocabulary
Guerrilla Warfare & Insurgency Vocabulary
Incel Violence Vocabulary
Jihadist Terrorism Vocabulary
Lone Wolf Attacks Vocabulary
QAnon & Conspiracy Theory Vocabulary
Sovereign Citizen Movement Vocabulary
Stochastic Terrorism Vocabulary
White Supremacy & Neo-Nazism Vocabulary
Cyberterrorism Vocabulary
Cyber Weapons Vocabulary

Weapons and military systems (12)

Autonomous Weapons Vocabulary
Directed Energy Vocabulary
Wound Ballistics Vocabulary
Space Warfare Vocabulary
Tunnel Warfare & Smuggling Vocabulary
Maritime Piracy Vocabulary
Infrastructure Warfare Vocabulary
Economic Warfare Vocabulary
Nuclear & Radiological Vocabulary
Nuclear Engineering Vocabulary
Ballistics & Terminal Effects Vocabulary                 [placeholder — no domain content]
Explosives & Detonation Physics Vocabulary               [placeholder — no domain content]

Chemical, biological, radiological and nuclear (6)

Bioweapons & Weaponized Biology Vocabulary
Bioterrorism Vocabulary
Chemical Terrorism Vocabulary
Nuclear Terrorism Vocabulary
Radiological Terrorism Vocabulary
Poison & Toxicology Vocabulary

Trafficking and illicit trade (10)

Arms Trafficking Vocabulary
Human Trafficking Vocabulary
Drug Trafficking Vocabulary
Organ Trafficking Vocabulary
Wildlife Trafficking Vocabulary
Waste Trafficking Vocabulary
Antiquities Looting Vocabulary
Illegal Fishing Vocabulary
Illegal Logging Vocabulary
Illegal Mining Vocabulary

Illicit finance and cybercrime-as-a-service (18)

Money Laundering Vocabulary
Terrorist Financing Vocabulary
Carding & Payment Fraud Vocabulary
Insurance Fraud Vocabulary
Investment Fraud & Ponzi Vocabulary
Pig Butchering & Romance Scams Vocabulary
Blockchain & DeFi Exploitation Vocabulary
Ransomware & Digital Extortion Vocabulary
Phishing-as-a-Service Vocabulary
Bulletproof Hosting Vocabulary
Botnet Architecture Vocabulary
DDoS Infrastructure Vocabulary
Initial Access Brokerage Vocabulary
Supply Chain & Trust Exploitation Vocabulary
Darknet Market Operations Vocabulary
Sextortion Vocabulary
Doxing & Harassment Vocabulary
Swatting Vocabulary

61 files named above; 58 carry a domain-specific term list. Three — marked [placeholder] — contain only the generic placeholder list and no domain-specific content.

Sections observed in the named files, quoted as headings only, to document the basis for this note: Section 2: Weaponization, Dispersal Systems, and Delivery; Section II: Production, Dissemination & Delivery Systems; Section II: Fissile Material, Proliferation Pathways & Nuclear Smuggling; Section III: Radiological Dispersal, Exposure Devices & Dirty Bombs; Section 1: Recruitment, Transport, and Debt Bondage; Section 1: Card Data Acquisition & Processing; Section 4: Victim Profiling & Target Valuation; Section 3: Cryptocurrency Mechanics & Money Laundering.

Operational framing outside the categories above

For completeness: further files outside those five categories also organise their sections around application rather than definition, principally in offensive-security and intelligence-analysis domains. Named here so the list is not misleading — for example Exploit Development Vocabulary, Exploitation & Intrusion Vocabulary, Escape & Evasion Vocabulary, HUMINT Vocabulary (Section 1: Agent Recruitment and Tradecraft), IMINT Vocabulary (Section 5: Targeting Intelligence, Battle Damage Assessment, and Legal Frameworks), TECHINT Vocabulary, DNINT Vocabulary, Surveillance & Counter-Surveillance Vocabulary (Section 5: Covert Communications & Tradecraft), and Darknet Market Operations Vocabulary (listed above). No editorial judgement is implied by the categorisation.

Method used to build the list

The title line and every Markdown heading of all 353 domain files were read. A file was selected when it (a) falls inside one of the five named categories and (b) contains at least one section whose heading addresses operational application — delivery, dispersal, weaponization, production, recruitment, targeting, laundering, smuggling, tradecraft or evasion. Files identified as placeholders by the grep -rl "^placeholder, thermodynamic, entropy" check were kept in the list and marked, rather than silently dropped, so the counts reconcile with the corpus. The boundary is a judgement; the criterion and its evidence are stated so a reader can redraw it.

This note is a factual disclosure of scope. It is not a statement about the accuracy, legality or safety of the material, and no domain is endorsed by its inclusion.

Not determined from the content

Why each domain was generated; how the sensitive-topic domains were selected; the intended downstream use; whether the numeric values in the generated prose were checked against any source; what produced the Thermodynamic Enhancement text. No provenance metadata, generation script or model identifier accompanies the files.

Verification commands

Every number in this card was re-measured against this directory. Run these from inside it.

# corpus size, excluding this card and the license
find . -type f ! -name README.md ! -name LICENSE | wc -l                  # -> 733
find . -type f ! -name README.md ! -name LICENSE -printf '%s\n' \
  | python3 -c "import sys;print(sum(int(l) for l in sys.stdin))"         # -> 26095368

# duplicate tier copies: distinct contents vs paths (corpus only)
find . -type f ! -name README.md ! -name LICENSE -exec md5sum {} + \
  | awk '{print $1}' | sort -u | wc -l                                    # -> 380

# files per directory
for d in . _BLOOM tier_1_excellent tier_2_good tier_3_needs_upgrade tier_4_poor; do
  printf "%-24s %s\n" "$d" "$(find "$d" -maxdepth 1 -type f | wc -l)"
done

# placeholder domains
grep -rl "^placeholder, thermodynamic, entropy" . | wc -l                # -> 72

# content-state split over the 353 distinct domains
python3 - <<'PY'
import os, re
tiers = ['tier_1_excellent','tier_2_good','tier_3_needs_upgrade','tier_4_poor']
ph = coda = real = 0
for d in tiers:
    for f in os.listdir(d):
        t = open(os.path.join(d, f), encoding='utf-8', errors='replace').read()
        n = sum(len(x.split(',')) for x in
                re.findall(r'(?:### Terms|^Terms:)\s*\n(.+?)\n', t, re.S | re.M))
        if re.search(r'^placeholder, thermodynamic, entropy', t, re.M): ph += 1
        elif n == 0: coda += 1
        else: real += 1
print('placeholder', ph, '| coda-only', coda, '| real', real)   # -> 36 | 75 | 242
PY

# size distribution (corpus only; matches the table above)
find . -type f ! -name README.md ! -name LICENSE -printf '%s\n' | sort -n \
  | python3 -c "
import sys, statistics
v=[int(l) for l in sys.stdin]
print('n',len(v),'total',sum(v),'median',statistics.median(v),'max',max(v))
"
# -> n 733 total 26095368 median 24981 max 5715238

Attribution

Betances, Christopher. Thermodynamic Vocabulary Corpus. catqualia.com, 2026.
Licensed under CC BY 4.0.

Downloads last month: 4

Total file size:

26.1 MB

You need to agree to share your contact information to access this dataset

Thermodynamic Vocabulary Corpus