docs/ADVERSARIAL_DATABASE_AUDIT_2026-03-12.md

Adversarial Database Audit: ancient-scripts-datasets

Date: 2026-03-12 Scope: Full database critique — IPA accuracy, missing languages, source quality, data integrity, scholarly accuracy, public presentation Method: 6 parallel research agents, each attacking a different dimension Database: HuggingFace / GitHub

---

Table of Contents

1. Executive Summary
2. CRITICAL Issues (Must Fix)
3. IPA & Phonetic Transcription Errors
4. Missing Ancient Languages
5. Source Quality & Provenance
6. Data Integrity Issues
7. Scholarly Accuracy of Lexical Entries
8. HuggingFace / GitHub Presentation
9. Recommendations Summary

---

1. Executive Summary

The ancient-scripts-datasets database is an ambitious project covering 1,136 languages with 3.3M entries. Its pipeline architecture (Iron Law, adversarial auditing, transliteration maps) is well-designed. However, this audit identified serious issues across all dimensions:

Category	Critical	High	Medium	Low
IPA/Phonetic Errors	3	5	5	—
Missing Languages	9 (Tier 1)	9 (Tier 2)	15+ (Tier 3-4)	20+
Source Quality	2	3	2	—
Data Integrity	2	3	4	2
Scholarly Accuracy	2	2	2	—
Presentation	3	3	3	—

The three most damaging findings:

1. Avestan claims 3,455 entries but only has 739 — the avesta_org source (2,716 entries) is entirely missing
2. The HuggingFace dataset is private/gated with 0 downloads — no researcher can access it
3. 55+ ancient languages with freely available digital datasets are absent, including Sumerian (15K+ entries), Akkadian (28K+), Egyptian (60K+), Sanskrit (2.5M+), and Ancient Greek (116K+)

---

2. CRITICAL Issues (Must Fix)

C1. Avestan Entry Count Mismatch

• DATABASE_REFERENCE.md claims 3,455 entries including 2,716 from avesta_org
• Actual ave.tsv has only 739 entries from wiktionary and wiktionary_cat
• The avesta_org source is entirely absent — either never ingested, or data was lost
• Impact: 78.6% of claimed Avestan data does not exist

C2. Bogus Data Artifact in Avestan

• Entry inprogress with IPA inproɡress is a data processing placeholder left in production data
• Entry phoneticvalue is another processing artifact
• Impact: Corrupts any downstream analysis

C3. HuggingFace Dataset is Private

• Set to Private: True with Gated: auto
• 0 downloads, 0 likes
• No researcher can discover or access the data without owner approval
• Impact: The dataset effectively does not exist as a public resource

C4. No License File

• GitHub repo has license: null — no LICENSE file at repo root
• HuggingFace YAML claims CC-BY-SA-4.0 but no actual license file exists
• Impact: Researchers cannot legally use the dataset with confidence

C5. SCA Tokenizer Bug — Labiovelar Modifier

• The SCA diacritic regex does NOT include ʷ (U+02B7, MODIFIER LETTER SMALL W)
• Result: every labiovelar in PIE, Lycian, Carian, Etruscan produces a spurious 0 character
• Example: kʷ → SCA K0 instead of just K
• Impact: Systematic data corruption in SCA column for all languages with labiovelars

---

3. IPA & Phonetic Transcription Errors

3.1 Phonological Accuracy Issues

Issue	Language	Severity	Details
š → ʃ disputed	Hittite	HIGH	Most Hittitologists (Kloekhorst 2008, Melchert) consider Hittite š = [s], not [ʃ]. Creates a spurious s/ʃ contrast.
ç → θ disputed	Old Persian	HIGH	Kent (1953) says /θ/, Kloekhorst (2008) says /ts/. SCA consequence: θ→T vs ts→S — significant for cognate detection.
h₃ → ɣʷ speculative	PIE	HIGH	Highly speculative phonetic value from Leiden school. Maps to SCA G+0 (bug, see C5) instead of H like other laryngeals. Creates artificial asymmetry.
Etruscan θ inconsistent	Etruscan	CRITICAL	θ mapped as fricative [θ] while φ→[pʰ] and χ→[kʰ] are aspirated stops. Same phoneme series, inconsistent treatment.
Lydian ś/š collapse	Lydian	CRITICAL	Two distinct sibilant graphemes both map to [ʃ], losing a phonemic distinction Gusmani (1964) maintains. Compare: Carian correctly distinguishes ś→[ɕ] from š→[ʃ].
Carian ỳ/ý self-mapping	Carian	MEDIUM	Map to themselves (non-IPA output), producing SCA class "0".
PIE voiced aspirates as breathy	PIE	MEDIUM	bʰ→bʱ uses breathy-voice diacritic ʱ (U+02B1) not in SCA regex. May produce spurious "0" segments.
Urartian ejectives assumed	Urartian	MEDIUM	ṣ→tsʼ, ṭ→tʼ assumes ejective realization. Could be pharyngealized or glottalized instead.
Avestan TTE collapsed	Avestan	MEDIUM	U+10B1D (TTE) collapsed with plain T, losing a potential phonemic distinction.

3.2 Missing Transliteration Mappings

Gap	Language	Impact
Cuneiform determinatives not handled	Hit, Xlw, Xhu, Xur, Elx	Sumerograms pass through untransliterated
Missing Old Persian signs (U+103AE, U+103B8, U+103BB)	Peo	Source text with these signs passes through unconverted
Missing Phrygian Greek letters (ξ, ψ, φ, χ)	Xpg	New Phrygian Greek-alphabet inscriptions partially unhandled
Missing Proto-Kartvelian aspirated affricates (cʰ, čʰ)	Ccs-pro	Incomplete three-way contrast
Missing Tocharian aspirated stops (kh, ph, th)	Txb/Xto	Brahmi-derived aspirates not mapped
Missing Dravidian alveolar nasal (ṉ)	Dra-pro	Four-way coronal nasal distinction incomplete

3.3 SCA Sound Class Systematic Losses

These are by-design limitations of the 18C+5V system, but they affect downstream cognate detection:

Collapse	Languages Affected	Significance
Ejectives = plain stops	Proto-Kartvelian, Urartian	Three-way Kartvelian contrast lost
Pharyngealized = plain	Ugaritic, Proto-Semitic, Phoenician	Semitic emphatic series lost
Uvulars = velars (q=K)	All Semitic languages	Fundamental q/k contrast lost
Retroflex = dental/alveolar	Proto-Dravidian, Tocharian	Diagnostic etymological feature lost
Sibilant voicing collapsed (s=z=S)	Avestan, Proto-Kartvelian	But labial/dental voicing preserved — inconsistent
Precomposed nasalized vowels dropped	Lycian, Lydian, Avestan	ã, ẽ, ũ may fail tokenizer regex — possible bug

---

4. Missing Ancient Languages

Tier 1: Critical Omissions (9 languages)

Language	ISO	Family	Available Entries	Best Source	Already Acknowledged?
Sumerian	sux	Isolate	15,944+ lemmas	ePSD2 (Penn)	Yes
Akkadian	akk	Afroasiatic	28,000+ words	CAD (Chicago) — free PDFs	Yes
Ancient Egyptian	egy	Afroasiatic	60,647 lemmas	TLA — downloadable HF datasets	Yes
Sanskrit	san	Indo-European	2,500,000+ items	Digital Corpus of Sanskrit	Yes
Mycenaean Greek	gmy	Indo-European	630+ words, 6K tablets	DAMOS (Oslo)	Yes
Ancient Greek	grc	Indo-European	116,502 entries	LSJ, Perseus	No
Gothic	got	Indo-European	3,600 lemmas	Project Wulfila — full TEI corpus	No
Old Church Slavonic	chu	Indo-European	130,000+ items	GORAZD digital dictionary	No
Old Norse	non	Indo-European	50,000+ words	ONP, Cleasby-Vigfusson	No

Key observation: The database includes 8 languages with <500 known entries (Lycian, Lydian, Carian, Rhaetic, Messapic, Lemnian, Phrygian, Phoenician) but omits Ancient Greek (116K entries) and Gothic (3.6K lemmas with complete annotated corpus). This is a severe coverage bias.

Tier 2: Important Omissions (9 languages)

Language	ISO	Family	Available Entries	Best Source
Coptic	cop	Afroasiatic	11,263 entries	Coptic Dictionary Online
Hattic	xht	Isolate	~300 words	Palaeolexicon, Wiktionary
Pali	pli	Indo-European	Tens of thousands	PTS Dictionary, DPD
Classical Armenian	xcl	Indo-European	Thousands	Calfa.fr, Bedrosian
Old English	ang	Indo-European	40,000+ entries	Bosworth-Toller
Ge'ez	gez	Afroasiatic	Thousands	Leslau dictionary
Syriac	syc	Afroasiatic	Tens of thousands	SEDRA
Aramaic	arc	Afroasiatic	3M parsed words	CAL (HUC)
Biblical Hebrew	hbo	Afroasiatic	8,000+ entries	BDB on Sefaria

Tier 3: Notable Omissions (15+ languages)

Middle Persian, Sogdian, Parthian, Khwarezmian, Gandhari Prakrit, Old Japanese, Classical Tibetan, Gaulish, Venetic, Faliscan, Eblaite, Old Irish, Classical Nahuatl, Classic Mayan, Sabaic, Oscan, Umbrian.

Tier 4: Missing Reconstructed Proto-Languages (4+)

Proto-Austronesian (ACD: 5K stems), Proto-Uralic (580+ lemmas), Proto-Bantu (BLR3: 10K reconstructions), Proto-Sino-Tibetan (STEDT: 1M records).

Geographic Coverage Gap

Region	In DB	Missing
Ancient Near East	4 languages	8+ (Sumerian, Akkadian, Egyptian, Eblaite, Aramaic, Hebrew, etc.)
Ancient Mediterranean	8 languages	7+ (Mycenaean, Ancient Greek, Oscan, Umbrian, Coptic, etc.)
South/East Asia	0 languages	6+ (Sanskrit, Pali, Old Japanese, Tibetan, Classical Chinese, etc.)
Africa	0 languages	4+ (Egyptian, Coptic, Ge'ez, Meroitic)
Americas	0 languages	2+ (Classical Nahuatl, Classic Mayan)

---

5. Source Quality & Provenance

5.1 Source Reliability Ratings

Source	IPA Reliability	Academic Rigor	Best Available?
WikiPron	Moderate (varies by language)	Peer-reviewed tool (LREC 2020)	Yes, for modern languages
Wiktionary (direct)	Low-Moderate	None (wiki, crowd-sourced)	No
eDiAna (LMU Munich)	N/A (transliteration only)	Excellent	Yes, for Anatolian
Palaeolexicon	Low	None (volunteer project, no institutional affiliation)	No
Oracc eCUT	N/A (transliteration only)	Excellent	Yes, for Urartian
TIR (Vienna)	N/A (epigraphy)	Excellent	Yes, for Rhaetic
avesta.org	Low	None (personal website, non-specialist author)	No
NorthEuraLex	Moderate (auto-generated)	Peer-reviewed	Yes, for coverage
ABVD	Low (~50% fake-IPA)	Peer-reviewed DB, but IPA issues	Yes, for Austronesian scope
WOLD	Good (expert)	Expert-curated	Yes, for loanwords

5.2 Key Source Concerns

Palaeolexicon (503 Etruscan + 482 Lycian + 304 Carian + 259 Hurrian + 225 Luwian + 187 Lydian = 1,960 entries):

• Independent volunteer project with no institutional affiliation
• No formal editorial board or peer review
• Site itself warns: "there is no way to be sure about how words were pronounced"
• Data should be cross-referenced against primary academic sources before use

avesta.org (2,716 claimed entries, currently missing):

• Personal website of a Chemical Engineering graduate, not an Iranist
• Based on Kanga (1900) — over 125 years old
• Modern standard: Bartholomae's Altiranisches Wörterbuch or Kellens/Skjaervo

Wiktionary for ancient languages:

• No editorial gatekeeping — anyone can edit
• Low attestation bar for extinct languages
• Contributor expertise is unverified
• Hittite page explicitly states it is "for people who know Hittite to fill up"
• Period mixing (OAv/YAv, Classical/Ecclesiastical Latin) without distinction

ABVD ~50% fake-IPA (confirmed):

• Known problem in computational historical linguistics community
• ABVD stores orthographic forms, not IPA, for many languages
• Lexibank 2 standardized version should be preferred over raw ABVD data

5.3 Superior Sources NOT Used

Source	What It Covers	Why Better
TITUS (Frankfurt)	Ancient IE languages	Most comprehensive digital corpus for ancient IE
ePSD2 (Penn)	Sumerian	15,944 lemmas, authoritative
CAD (Chicago)	Akkadian	28,000 words, 26 volumes free online
TLA (Berlin)	Egyptian	60,647 lemmas, downloadable
PHOIBLE (UW/MPI)	Phonological inventories	3,020 inventories for validation
Lexibank 2 (MPI)	Standardized CLDF	CLTS-mapped versions of ABVD/NorthEuraLex/etc.
STEDT (Berkeley)	Sino-Tibetan	1M records from 500+ sources
Project Wulfila	Gothic	Complete corpus in TEI, freely downloadable

---

6. Data Integrity Issues

6.1 Per-File Issues

File	Entries	Issue	Severity
ave.tsv	739 (claimed 3,455)	Missing 2,716 entries from avesta_org	CRITICAL
ave.tsv	—	Bogus inprogress and phoneticvalue entries	CRITICAL
ave.tsv	—	74.4% empty Concept_IDs	HIGH
ave.tsv	—	Old/Young Avestan not distinguished	MEDIUM
xlw.tsv	2,230	581 SCA zeros from Sumerograms (26.1%)	HIGH
xur.tsv	748	171 SCA zeros from Sumerograms (22.9%)	HIGH
xur.tsv	—	54.4% identity rate — transliteration map likely incomplete	HIGH
ett.tsv	753	250+ proper nouns mixed with vocabulary (33%)	HIGH
ett.tsv	—	3 duplicate Word entries	MODERATE
ett.tsv	—	162 SCA zeros from capitalized proper names	MODERATE
ine-pro.tsv	1,704	50.6% empty Concept_IDs	MODERATE
uga.tsv	371	6 SCA zeros from 'V' placeholder characters	MINOR
hit.tsv	266	Trailing periods on some Concept_IDs	MINOR

6.2 Metadata Issues

• 7 of 8 audited ancient languages are absent from languages.tsv — only Etruscan appears, and only for its WikiPron subset (207 of 753 entries)
• Lexicon count mismatch on HuggingFace: README claims 1,136 but only 1,135 TSV files exist

6.3 Sumerogram Contamination

Luwian (581 entries) and Urartian (171 entries) contain cuneiform Sumerograms — uppercase logograms like LUGAL, URU, DINGIR, MUNUS.LUGAL that represent Sumerian words used as shorthand in cuneiform texts. These are NOT phonemic data in the target language and:

• Produce SCA class "0" (unknown)
• Would corrupt any phonological comparison pipeline (e.g., PhaiPhon)
• Should be filtered or flagged before analysis

---

7. Scholarly Accuracy of Lexical Entries

7.1 Per-Language Verdicts

Language	Verdict	Key Issues
Hittite	CONCERNS	Sumerograms/Akkadograms mixed in as Hittite words (IM, LU, URU, KU6, ITUD, ZAG, TUR, MUNUS.LUGAL, DUMU.MUNUS — at least 10-12 entries). One Avestan word (xshap = "night") in wrong file entirely.
PIE	MOSTLY ACCURATE	Reconstructions consistent with LIV/Pokorny/NIL. Laryngeal IPA values represent one theory but applied consistently.
Etruscan	CONCERNS	~33% proper nouns (personal names, theonyms, Hellenized names like Hercle/Aplu/Achile). Some over-confident meanings for words scholars consider uncertain (antar="eagle", arimos="monkey").
Ugaritic	MOSTLY ACCURATE	Vocabulary aligns with DUL. Vowel patterns are reconstructed (inherent to Ugaritic studies). A few place/divine names mixed in.
Luwian	MOSTLY ACCURATE	Honest "(unknown)" labels for uncertain glosses (~35%). Good source (eDiAna). Cuneiform/hieroglyphic not improperly mixed.
Avestan	CONCERNS	Data artifacts in production data. OAv/YAv not distinguished. Single-character alphabet entries treated as words. 74.4% empty Concept_IDs.

7.2 Cross-Language Contamination

Entry	In File	Actually Belongs To
xshap ("night")	hit.tsv (Hittite)	Avestan (xshap- is Old Iranian)
GE ("ina: in, by, from")	hit.tsv (Hittite)	Akkadian (preposition ina)
Sumerograms (LUGAL, URU, etc.)	hit.tsv, xlw.tsv, xur.tsv	Sumerian logograms

---

8. HuggingFace / GitHub Presentation

8.1 HuggingFace Problems

Issue	Severity
Dataset is private/gated — 0 downloads, 0 access	CRITICAL
No LICENSE file on either platform	CRITICAL
README is a 46-line stub — empty Quick Start, no citations, no limitations, no examples	CRITICAL
sources/ directory leaked (2,928 files, 0.44 GB of upstream CLDF repos)	HIGH
.pytest_cache/ directories present (dev artifacts)	HIGH
Copyrighted PDFs included (Trask dictionary, Rodriguez-Ramos 2014)	HIGH
PyTorch .pth and .pkl files present (security risk)	MEDIUM
No datasets.load_dataset() support (no loading script/Parquet)	MEDIUM
Lexicon count: 1,135 files vs 1,136 claimed	MEDIUM

8.2 GitHub Problems

Issue	Severity
No LICENSE file at repo root	CRITICAL
GitHub README describes original Luo et al. datasets; HuggingFace README describes 1,136-language expansion — narratives disconnected	HIGH
No GitHub topics/tags set	LOW
0 stars, 0 forks, 0 contributors beyond author	— (not a quality issue)

8.3 Strengths

• GitHub README is thorough and well-organized (158 lines)
• 18 adversarial audit documents show genuine quality control effort
• Scripts for full reproduction are present and documented
• Active development (commits every 1-2 days)
• DATABASE_REFERENCE.md (this document's subject) is exceptionally detailed

---

9. Recommendations Summary

Immediate Fixes (Critical)

1. Investigate and restore the missing 2,716 Avestan avesta_org entries, or correct DATABASE_REFERENCE.md
2. Remove bogus entries (inprogress, phoneticvalue) from ave.tsv
3. Make HuggingFace dataset public (or document the gating rationale)
4. Add a LICENSE file to both repos
5. Fix SCA tokenizer: add \u02B7 (ʷ) to diacritic regex to prevent spurious "0" on labiovelars

High Priority

6. Fix Etruscan θ/φ/χ inconsistency: either all fricatives or all aspirated stops
7. Fix Lydian ś/š collapse: distinguish as ś→[ɕ] and š→[ʃ] (matching Carian treatment)
8. Filter/flag Sumerograms in Hittite, Luwian, and Urartian before phonetic analysis
9. Remove cross-language contamination: xshap from Hittite, Akkadian ina from Hittite
10. Filter/flag proper nouns in Etruscan (~250 entries)
11. Remove leaked sources/ directory and .pytest_cache/ from HuggingFace
12. Remove copyrighted PDFs from the repository
13. Expand HuggingFace README to match GitHub quality (Quick Start, citations, limitations)

Medium Priority

14. Add Sumerian, Akkadian, and Egyptian using ePSD2, CAD, and TLA (highest-impact expansions)
15. Add Ancient Greek and Gothic using LSJ/Perseus and Project Wulfila
16. Replace avesta.org with Bartholomae or modern Iranian references
17. Cross-reference Palaeolexicon data against eDiAna and other primary sources
18. Use Lexibank 2 standardized versions of ABVD/NorthEuraLex instead of raw data
19. Distinguish Old vs Young Avestan entries
20. Add precomposed nasalized vowels (ã, ẽ, ũ) to SCA tokenizer
21. Document controversial phonological choices (Hittite š, OP ç, PIE laryngeals) in transliteration maps
22. Add ancient languages to languages.tsv metadata
23. Add breathy-voice diacritic ʱ (U+02B1) to SCA regex

Design Considerations (Long-term)

24. Consider adding ejective/emphatic/uvular classes to SCA to preserve Kartvelian/Semitic contrasts
25. Consider PHOIBLE integration for phonological inventory validation
26. Consider adding geographic coverage for South/East Asian, African, and American ancient languages
27. Consider adding confidence indicators for IPA quality per entry (attested vs. reconstructed)

---

This audit was conducted by 6 parallel AI research agents. All findings are based on file reading, web research, and scholarly knowledge. No code was written or executed.