docs/prd/PRD_DATABASE_RECTIFICATION.md

PRD: Database Rectification & Expansion Plan

Date: 2026-03-12 Triggered by: Adversarial Database Audit 2026-03-12 Reference: DATABASE_REFERENCE.md (protocols, schema, Iron Law) Status: DRAFT — awaiting approval

---

IRON LAW (UNCHANGED — SUPERSEDES ALL GOALS)

DATA MAY ONLY ENTER THE DATASET THROUGH CODE THAT DOWNLOADS IT
FROM AN EXTERNAL SOURCE.

NO EXCEPTIONS. NO "JUST THIS ONCE." NO "IT'S FASTER."

Every phase below produces Python scripts that fetch data via HTTP. No hardcoded word lists. No direct TSV edits. No LLM-generated linguistic content. Every script must contain urllib.request.urlopen(), requests.get(), or equivalent HTTP fetch. Transliteration maps are CODE (transformation rules from cited grammars), not DATA.

---

ADVERSARIAL PIPELINE v2 (ENHANCED)

Every phase uses the Dual-Agent Adversarial Pipeline. This PRD upgrades the adversarial auditor from v1 (surface-level checks) to v2 (deep cross-reference validation).

Team A: Extraction Agent

• Writes and runs Python scripts following the script template
• Produces TSV data via HTTP fetch → parse → transliterate → write
• NEVER writes data values directly

Team B: Critical Adversarial Auditor (v2 — ENHANCED)

Runs after EACH step with VETO POWER. The v2 auditor performs deep validation, not surface-level checks.

What Team B MUST Do (Deep Checks)

Check	Method	Pass Criteria
50-Word Cross-Reference	Select 50 random entries from the newly scraped data. For each, fetch the LIVE source URL and verify the word appears there with the same form and meaning.	>= 48/50 match (96%). Any mismatch = STOP.
IPA Spot-Check	For 20 random entries, manually apply the transliteration map character-by-character and verify the output matches the IPA column.	20/20 match. Any mismatch = flag map bug.
SCA Consistency	For 20 random entries, verify ipa_to_sound_class(IPA) == SCA column.	20/20 match.
Source Provenance	For 10 random entries, construct the exact URL where each entry can be found in the original source. Verify it loads.	10/10 accessible.
Concept ID Accuracy	For 20 random entries with non-empty Concept_IDs, verify the gloss matches the source's definition.	>= 18/20 match.
Dedup Verification	Count unique Word values in the output. Compare to total rows.	0 duplicates.
Entry Count Plausibility	Verify count is non-round and matches expected range from source research.	Not exactly round (100, 200, 500).

What Team B Must NOT Do (Banned Checks — Wastes Time)

• "Does the file have a header?" (Always yes by construction)
• "Are there HTML tags in the data?" (Parsing handles this)
• "Is the file UTF-8?" (Always yes by construction)
• "Does the script import urllib?" (Obvious from code review)
• Any check that doesn't touch real data

Auditor Report Format

# Adversarial Audit v2: {Phase} — {Language} ({iso})

## 50-Word Cross-Reference
- Sampled: [list 50 words]
- Source URL pattern: {url}
- Matches: N/50
- Mismatches: [list any failures with details]

## IPA Spot-Check (20 entries)
| Word | Expected IPA | Actual IPA | Match? |
|------|-------------|------------|--------|
| ... | ... | ... | ... |

## SCA Consistency (20 entries)
- All match: YES/NO

## Source Provenance (10 entries)
| Word | Source URL | Accessible? |
|------|-----------|-------------|
| ... | ... | ... |

## Concept ID Accuracy (20 entries)
- Matches: N/20

## Dedup Check
- Unique words: N
- Total rows: N
- Duplicates: 0

## Verdict: PASS / FAIL
## Blocking Issues: [list if any]

---

PROPER NOUNS POLICY

Proper nouns (theonyms, toponyms, anthroponyms) are VALUED DATA, not contamination.

All ancient language lexicons SHOULD include:

• Theonyms (divine names): gods, goddesses, mythological figures
• Toponyms (place names): cities, rivers, mountains, temples, regions
• Anthroponyms (personal names): rulers, historical figures, common name elements
• Ethnonyms (people/tribe names): tribal and ethnic designations

Concept_ID should tag these as theonym:{name}, toponym:{name}, anthroponym:{name}, ethnonym:{name}.

Where specialist proper noun databases exist (see Phase 5), they MUST be scraped alongside regular vocabulary.

---

PHASE 0: Critical Bug Fixes

Priority: IMMEDIATE — blocks all other phases Estimated effort: 1 session No adversarial audit needed (code changes only, no data ingestion)

0.1 Fix SCA Tokenizer — Labiovelar Bug

File: cognate_pipeline/src/cognate_pipeline/normalise/sound_class.py Bug: ʷ (U+02B7) missing from diacritic regex → produces spurious "0" for every labiovelar Fix: Add \u02B7 to the diacritic character class on the tokenizer regex (line ~95) Also add: \u02B1 (breathy voice ʱ) for PIE voiced aspirates Test: Run ipa_to_sound_class("kʷ") → should produce "K" not "K0"

0.2 Fix SCA Tokenizer — Precomposed Nasalized Vowels

Bug: Precomposed ã (U+00E3), ẽ (U+1EBD), ũ (U+0169) may fail tokenizer regex Fix: Either (a) NFC-decompose input before tokenizing, or (b) add precomposed nasalized vowels to the character class Test: Run ipa_to_sound_class("ã") → should produce "A" not ""

0.3 Write Cleaning Script — Remove Bogus Entries

Iron Law compliance: Write a Python script scripts/clean_artifacts.py that:

1. Reads each ancient language TSV
2. Identifies known artifact patterns: inprogress, phoneticvalue, entries where Word matches ^[a-z]+progress$ or similar processing placeholders
3. Writes cleaned TSV (preserving all legitimate entries)
4. Logs removed entries to audit trail
5. Reports counts

NOT a direct edit — this is a deterministic cleaning script.

0.4 Fix Metadata — Add Ancient Languages to languages.tsv

Script: scripts/update_metadata.py — reads all TSVs in lexicons/, counts entries, updates languages.tsv with ISO, name, family, entry count, source breakdown. Run after every data change.

0.5 Presentation Fixes

Task	Action
Add LICENSE file	Create LICENSE at repo root with CC-BY-SA-4.0 full text
Make HuggingFace public	Change dataset visibility to public (manual step)
Fix HuggingFace README	Expand to include Quick Start, citations, limitations, loading examples
Remove leaked files	Add sources/, .pytest_cache/, *.pth, *.pkl to HF .gitignore; remove copyrighted PDFs
Fix lexicon count	Identify which of 1,136 claimed files is missing; create or correct count

---

PHASE 1: IPA & Transliteration Map Corrections

Priority: HIGH — affects all downstream phonetic analysis Estimated effort: 1 session Adversarial audit: YES (Team B verifies 20 entries per map change via IPA spot-check)

1.1 Transliteration Map Fixes

Each fix below modifies scripts/transliteration_maps.py (CODE, not DATA). After all fixes, run scripts/reprocess_ipa.py to propagate changes.

Fix	Language	Change	Academic Reference
Etruscan θ/φ/χ consistency	ett	Change θ→tʰ (aligning with φ→pʰ, χ→kʰ as aspirated stop series) OR change all three to fricatives (θ, f, x). Pick ONE. Recommended: all aspirated stops per Bonfante early-period analysis: θ→tʰ	Bonfante & Bonfante (2002), Rix (1963)
Lydian ś/š distinction	xld	Change ś→ɕ (alveolopalatal, matching Carian treatment), keep š→ʃ	Gusmani (1964), Melchert
Carian ỳ/ý placeholders	xcr	Map to best-guess IPA or explicit unknown marker. Recommended: ỳ→ə, ý→e (tentative vocalic values) with comment noting uncertainty	Adiego (2007)
Hittite š controversy	hit	ADD COMMENT documenting the debate. Keep š→ʃ as the current choice but note: "Kloekhorst (2008) argues for [s]. Hoffner & Melchert (2008) use the conventional symbol." Do NOT change the value without user decision.	Hoffner & Melchert (2008), Kloekhorst (2008)
Old Persian ç controversy	peo	ADD COMMENT documenting the debate. Keep ç→θ per Kent but note Kloekhorst's /ts/ argument.	Kent (1953), Kloekhorst (2008)
PIE h₃ value	ine	ADD COMMENT noting the speculative nature of h₃→ɣʷ. Note: "Leiden school reconstruction. Many scholars leave h₃ phonetically unspecified."	Beekes (2011), Fortson (2010)
Missing Phrygian Greek letters	xpg	Add: ξ→ks, ψ→ps, φ→pʰ, χ→kʰ	Brixhe & Lejeune (1984)
Missing PK aspirated affricates	ccs	Add: cʰ→tsʰ, čʰ→tʃʰ	Klimov (1998)
Missing Old Persian signs	peo	Add: U+103AE (di), U+103B8 (mu), U+103BB (vi)	Kent (1953)

1.2 Post-Fix Reprocessing

# Dry run first (ALWAYS)
python scripts/reprocess_ipa.py --dry-run

# Verify: identity rates should decrease or stay the same, NEVER increase
# Verify: no regressions (Never-Regress Rule)

# Run live
python scripts/reprocess_ipa.py

1.3 Adversarial Audit for Phase 1

Team B verifies:

• For each modified map: take 20 entries from that language's TSV, manually apply the updated map, verify IPA matches
• Verify no regressions: compare before/after identity rates
• Verify SCA correctness for 20 entries per language

---

PHASE 2: Data Restoration & Cleanup

Priority: HIGH — fixes audit-identified data problems Estimated effort: 1–2 sessions Adversarial audit: YES (full v2 pipeline)

2.1 Avestan — Re-scrape avesta.org (Restore Missing 2,716 Entries)

Problem: DATABASE_REFERENCE.md claims 3,455 entries including 2,716 from avesta_org, but ave.tsv only has 739 entries. The avesta_org data was either never ingested or was lost.

Script: scripts/scrape_avesta_org.py (already exists — re-run or debug)

Steps:

1. Team A: Verify scrape_avesta_org.py still works against live site
2. Team A: Run --dry-run to confirm expected entry count
3. Team A: Run live scrape, deduplicating against existing 739 entries
4. Team B: 50-word cross-reference against live avesta.org/avdict/avdict.htm
5. Team B: IPA spot-check 20 entries against AVESTAN_MAP
6. Update DATABASE_REFERENCE.md with actual count

Acceptance: ave.tsv has 2,500+ entries (the 3,455 was an aspiration, actual may differ)

2.2 Sumerogram Handling Script

Problem: Hittite (10+ entries), Luwian (581), and Urartian (171) contain Sumerograms — uppercase cuneiform logograms (LUGAL, URU, DINGIR, etc.) that are NOT phonemic data in the target language.

Script: scripts/tag_sumerograms.py

Approach: Do NOT delete Sumerograms — they are legitimate scholarly data. Instead:

1. Write a script that identifies likely Sumerograms (all-uppercase ASCII, known Sumerogram patterns)
2. Add a tag to the Concept_ID field: prefix with sumerogram: (e.g., sumerogram:king for LUGAL)
3. This allows downstream pipelines to filter them if needed while preserving the data
4. Log all tagged entries to audit trail

Sumerogram detection heuristic:

def is_sumerogram(word: str) -> bool:
    """Detect cuneiform Sumerograms (uppercase sign names)."""
    if word.isupper() and word.isascii() and len(word) >= 2:
        return True
    if re.match(r'^[A-Z]+(\.[A-Z]+)+$', word):  # MUNUS.LUGAL pattern
        return True
    if re.match(r'^[A-Z]+\d+$', word):  # KU6, AN2 pattern
        return True
    return False

Team B checks: Verify 20 tagged entries are actually Sumerograms (not coincidentally uppercase native words).

2.3 Cross-Language Contamination Fix

Problem: hit.tsv contains at least one Avestan word (xshap = "night") and Akkadian entries (GE = "ina").

Script: scripts/clean_cross_contamination.py

1. For each ancient language TSV, check every entry against a known-contamination list (populated from audit findings)
2. Remove entries confirmed to be from wrong language
3. Log removals to audit trail

Known contamination (from audit):

• hit.tsv: xshap (Avestan), GE/ina (Akkadian)

Team B checks: Verify each removed entry is genuinely from the wrong language by checking Wiktionary source pages.

---

PHASE 3: New Language Ingestion — Tier 1

Priority: HIGH — the 9 most critical missing languages Estimated effort: 3–5 sessions (can parallelize across languages) Adversarial audit: YES (full v2 pipeline per language)

General Protocol (applies to all Tier 1 languages)

For each new language:

1. Create transliteration map in transliteration_maps.py (if needed) with cited academic reference
2. Write extraction script following the standard template:
   • Must use urllib.request.urlopen() or requests.get()
   • Must deduplicate against existing entries
   • Must apply transliterate() and ipa_to_sound_class()
   • Must save raw JSON/HTML to data/training/raw/
   • Must save audit trail to data/training/audit_trails/
3. Run --dry-run first
4. Deploy Team B adversarial auditor (full v2: 50-word cross-ref, IPA spot-check, etc.)
5. Run live
6. Add to language_configs.py
7. Run reprocess_ipa.py --language {iso}
8. Update metadata (languages.tsv)
9. Commit & push to both GitHub and HuggingFace

---

3.1 Sumerian (sux)

Field	Value
ISO	sux
Family	Isolate
Primary Source	ePSD2 — oracc.museum.upenn.edu/epsd2/sux (JSON API)
Secondary Source	DCCLT lexical texts via Oracc
Expected entries	10,000–15,944 lemmas
Script name	scripts/scrape_epsd2_sumerian.py
Transliteration map	New: SUMERIAN_MAP — cuneiform transliteration → IPA (Jagersma 2010, Edzard 2003)
IPA type	Partial (phonology reconstructed via Akkadian scribal conventions)
Special handling	Strip determinatives (superscript d, GIS, etc.). Tag Sumerograms vs. phonemic entries. Separate emesal (women's dialect) from emegir (main dialect).
Proper nouns to include	Divine names (Enlil, Inanna, Enki, Utu, Nanna, etc.), city names (Ur, Uruk, Lagash, Nippur, Eridu, etc.), royal names (Gilgamesh, Ur-Nammu, Shulgi, etc.)

Scraping approach:

• ePSD2 exposes a JSON API at oracc.museum.upenn.edu/epsd2/json/
• Fetch the full glossary index, then individual lemma pages
• Parse: headword, citation form, base, morphology, English gloss
• The ePSD2 provides transliterations in standard Assyriological conventions

3.2 Akkadian (akk)

Field	Value
ISO	akk
Family	Afroasiatic > Semitic (East)
Primary Source	AssyrianLanguages.org — assyrianlanguages.org/akkadian/ (searchable dictionary)
Secondary Source	Oracc glossaries, Wiktionary Category:Akkadian_lemmas
Expected entries	5,000–10,000 (from online searchable sources; full CAD is 28K but PDF-only)
Script name	scripts/scrape_akkadian.py
Transliteration map	New: AKKADIAN_MAP — standard Assyriological transliteration → IPA (Huehnergard 2011, von Soden 1995)
IPA type	Broad phonemic (well-understood via comparative Semitic + cuneiform orthography)
Special handling	Distinguish Old Babylonian, Middle Babylonian, Neo-Assyrian, etc. via source metadata if available. Handle determinatives.
Proper nouns to include	Divine names (Marduk, Ishtar, Shamash, Ea, Sin, Nabu, etc.), city names (Babylon, Nineveh, Assur, Sippar, etc.), royal names (Hammurabi, Sargon, Nebuchadnezzar, etc.)

3.3 Ancient Egyptian (egy)

Field	Value
ISO	egy
Family	Afroasiatic > Egyptian
Primary Source	TLA — thesaurus-linguae-aegyptiae.de (API or web scrape)
Secondary Source	TLA HuggingFace datasets (huggingface.co/datasets/thesaurus-linguae-aegyptiae/)
Expected entries	10,000–49,037 lemmas
Script name	scripts/scrape_tla_egyptian.py
Transliteration map	New: EGYPTIAN_MAP — Egyptological transliteration (Manuel de Codage) → IPA (Allen 2014, Loprieno 1995)
IPA type	Partial (consonantal skeleton well-known; vowels reconstructed from Coptic, cuneiform transcriptions, and comparative Afroasiatic)
Special handling	Egyptian had no written vowels. Provide consonantal IPA skeleton. Consider separate entries for different periods (Old/Middle/Late/Demotic). Hieroglyphic Unicode signs (U+13000–U+1342F) should be mapped if present.
Proper nouns to include	Pharaoh names (Khufu, Ramesses, Thutmose, etc.), deity names (Ra, Osiris, Isis, Horus, Thoth, Anubis, etc.), place names (Thebes, Memphis, Heliopolis, etc.)

3.4 Sanskrit (san)

Field	Value
ISO	san
Family	Indo-European > Indo-Iranian > Indo-Aryan
Primary Source	Wiktionary Category:Sanskrit_lemmas (massive category)
Secondary Source	WikiPron Sanskrit entries, DCS (Digital Corpus of Sanskrit) if API accessible
Expected entries	5,000–20,000 from Wiktionary alone
Script name	scripts/scrape_sanskrit.py
Transliteration map	New: SANSKRIT_MAP — IAST/Devanagari → IPA (Whitney 1896, Mayrhofer 1986)
IPA type	Full phonemic (Sanskrit phonology is comprehensively documented)
Special handling	Handle both Devanagari (U+0900–U+097F) and IAST romanization. Vedic Sanskrit vs Classical Sanskrit distinction desirable.
Proper nouns to include	Divine names (Indra, Agni, Varuna, Vishnu, Shiva, etc.), place names (Hastinapura, Ayodhya, Lanka, etc.), epic names (Arjuna, Rama, Krishna, etc.)

3.5 Ancient Greek (grc)

Field	Value
ISO	grc
Family	Indo-European > Hellenic
Primary Source	Wiktionary Category:Ancient_Greek_lemmas
Secondary Source	WikiPron Ancient Greek entries, Perseus Digital Library
Expected entries	10,000+ from Wiktionary
Script name	scripts/scrape_ancient_greek.py
Transliteration map	New: ANCIENT_GREEK_MAP — Greek alphabet → reconstructed Classical Attic IPA (Allen 1987, Smyth 1920)
IPA type	Full phonemic (Classical Attic pronunciation well-reconstructed)
Special handling	Use Classical Attic pronunciation (not Koine or Modern). Handle polytonic orthography (breathing marks, accents). Distinguish from Modern Greek WikiPron entries.
Proper nouns to include	Theonyms (Zeus, Athena, Apollo, Hermes, etc.), place names (Athens, Sparta, Thebes, Troy, etc.), hero names (Achilles, Odysseus, Herakles, etc.)

3.6 Gothic (got)

Field	Value
ISO	got
Family	Indo-European > Germanic (East)
Primary Source	Project Wulfila — wulfila.be (TEI corpus + glossary)
Secondary Source	Wiktionary Category:Gothic_lemmas
Expected entries	3,000–3,600 lemmas
Script name	scripts/scrape_wulfila_gothic.py
Transliteration map	New: GOTHIC_MAP — Gothic alphabet (U+10330–U+1034F) + transliteration → IPA (Wright 1910, Braune/Heidermanns 2004)
IPA type	Full phonemic (Gothic phonology well-understood from comparative Germanic)
Special handling	Handle Gothic script Unicode block. Project Wulfila provides downloadable TEI XML — use cached-fetch pattern if needed.
Proper nouns to include	Biblical proper nouns in Gothic form (Iesus, Xristus, Pawlus, Iairusalem, etc.), tribal names (Gutans, etc.)

3.7 Mycenaean Greek (gmy)

Field	Value
ISO	gmy
Family	Indo-European > Hellenic
Primary Source	DAMOS — damos.hf.uio.no (complete annotated Mycenaean corpus)
Secondary Source	Palaeolexicon Linear B section
Expected entries	500–800
Script name	scripts/scrape_damos_mycenaean.py
Transliteration map	New: MYCENAEAN_MAP — Linear B syllabary → reconstructed IPA (Ventris & Chadwick 1973, Bartonek 2003)
IPA type	Partial (Linear B is a syllabary that obscures many consonant clusters and final consonants)
Special handling	Linear B is a syllabary — each sign represents a CV syllable. The underlying Greek word must be reconstructed from the syllabic spelling. Many readings are uncertain.
Proper nouns to include	Place names from tablets (pa-ki-ja-ne/Sphagianai, ko-no-so/Knossos, etc.), divine names (di-wo/Zeus, a-ta-na-po-ti-ni-ja/Athena Potnia, etc.)

3.8 Old Church Slavonic (chu)

Field	Value
ISO	chu
Family	Indo-European > Slavic (South)
Primary Source	Wiktionary Category:Old_Church_Slavonic_lemmas
Secondary Source	GORAZD digital dictionary (gorazd.org) if API accessible
Expected entries	2,000–5,000 from Wiktionary
Script name	scripts/scrape_ocs.py
Transliteration map	New: OCS_MAP — Cyrillic/Glagolitic → IPA (Lunt 2001)
IPA type	Full phonemic (OCS phonology well-established)
Special handling	Handle both Cyrillic and Glagolitic scripts. OCS Cyrillic uses characters not in modern Cyrillic (ѣ, ъ, ь, ѫ, ѧ, etc.).
Proper nouns to include	Place names from OCS texts, biblical proper nouns in OCS form

3.9 Old Norse (non)

Field	Value
ISO	non
Family	Indo-European > Germanic (North)
Primary Source	Wiktionary Category:Old_Norse_lemmas
Secondary Source	Cleasby-Vigfusson online if scrapable
Expected entries	5,000–10,000
Script name	scripts/scrape_old_norse.py
Transliteration map	New: OLD_NORSE_MAP — Old Norse orthography → IPA (Gordon 1957, Noreen 1923)
IPA type	Full phonemic (Old Norse phonology well-documented)
Special handling	Handle Old Norse special characters (ð, þ, æ, ø, ǫ). Distinguish Old West Norse (Old Icelandic) from Old East Norse if possible.
Proper nouns to include	Divine names from Eddas (Oðinn, Þórr, Freyr, Freyja, Loki, Baldr, etc.), place names (Ásgarðr, Miðgarðr, Jǫtunheimr, etc.), hero names (Sigurðr, Ragnarr, etc.)

---

PHASE 4: Proper Noun Expansion

Priority: MEDIUM-HIGH — enhances all existing and new languages Estimated effort: 2–3 sessions (parallelizable) Adversarial audit: YES (full v2 pipeline)

4.1 Strategy

For each language already in the database (and each new Tier 1 language), identify and scrape specialist proper noun sources. Proper nouns are tagged in Concept_ID as:

• theonym:{name} — divine/mythological names
• toponym:{name} — place names
• anthroponym:{name} — personal names (rulers, historical figures)
• ethnonym:{name} — tribal/ethnic names

4.2 Proper Noun Sources by Language (Detailed — from specialist research)

Tier 1 Sources: Structured Data with API/Download (Best Targets)

#	Language	Source	URL	API Type	Est. Proper Nouns	Notes
1	Sumerian	ORACC ePSD2 QPN glossaries	oracc.museum.upenn.edu/epsd2/names/	JSON API (build-oracc.museum.upenn.edu/json/)	1,000+ (qpn-x-divine, qpn-x-placeN, qpn-x-people, qpn-x-temple, qpn-x-ethnic, qpn-x-celestial)	Sub-glossaries by type code. Best structured source in entire survey.
2	Sumerian	ETCSL proper nouns (Oxford)	etcsl.orinst.ox.ac.uk/cgi-bin/etcslpropnoun.cgi	Scrapable HTML tables	917 unique (12,537 occurrences): ~400 DN, ~200 RN, ~150 SN, ~120 TN, ~80 PN	Categorized by type (DN/RN/SN/TN/PN/GN/WN).
3	Akkadian	ORACC QPN glossaries (all sub-projects)	oracc.museum.upenn.edu (rinap, saao, cams, etc.)	JSON API	Thousands across dozens of sub-projects	Same JSON structure as Sumerian QPN. Covers Neo-Assyrian, Neo-Babylonian, Old Babylonian.
4	Egyptian	TLA proper noun lemmas	thesaurus-linguae-aegyptiae.de	JSON/TEI XML API + HuggingFace JSONL	Thousands (subset of 49,037 + 11,610 lemmas)	Categories for kings, deities, persons, places, titles. Raw JSON + TEI XML in lasting repository.
5	Egyptian	Pharaoh.se king list	pharaoh.se	Scrapable HTML	300–350 pharaoh names (with variants)	Turin Canon (223), Abydos (76), Karnak (61), Saqqara (58), Manetho. Per-pharaoh URLs.
6	Ancient Greek	LGPN (Lexicon of Greek Personal Names, Oxford)	search.lgpn.ox.ac.uk	REST API (clas-lgpn5.classics.ox.ac.uk:8080/exist/apps/lgpn-api/)	35,982 unique personal names (~400,000 individuals across 8 volumes)	Single richest source for ancient Greek anthroponyms. Data also in ORA (Oxford Research Archive).
7	Ancient Greek	Pleiades Gazetteer	pleiades.stoa.org	JSON + CSV bulk download (daily dumps at atlantides.org/downloads/pleiades/json/)	36,000+ places, 26,000+ ancient names	GitHub releases. CC-BY licensed. Coordinates, time periods, citations.
8	Ancient Greek	Theoi.com mythology	theoi.com	Scrapable HTML (consistent structure)	1,000–1,500 mythological figures	Gods, daimones, creatures, heroes. Alphabetical pages.
9	Gothic	Project Wulfila	wulfila.be/gothic/download/	TEI XML download with POS tags	200–300 biblical proper nouns	Nouns tagged "Noun, proper." Most machine-friendly source in survey.
10	Etruscan	CIE/TLE Digital Concordance (Zenodo)	Zenodo (search "Etruscan Faliscan concordance")	CSV download	1,000+ unique names (from 12,000+ inscriptions)	~67% of inscriptions contain personal names. Far exceeds current ~250.

Tier 2 Sources: Structured HTML, Easily Scrapable

#	Language	Source	URL	Est. Proper Nouns	Notes
11	Sumerian	AMGG (Ancient Mesopotamian Gods & Goddesses)	oracc.museum.upenn.edu/amgg/listofdeities/	~100 major deity profiles	Scholarly profiles with epithets, iconography.
12	Hittite	HDN (Hittite Divine Names)	cuneiform.neocities.org/HDN/outline	~1,000+ divine name entries	Updates van Gessel's 3-volume Onomasticon. HTML tables + PDF.
13	Hittite	HPN + LAMAN (Hittite Name Finder)	cuneiform.neocities.org/HPN/outline / cuneiform.neocities.org/laman/start	Hundreds of personal names	Unified divine + geographical + personal name retrieval.
14	Ugaritic	Wikipedia List of Ugaritic Deities	en.wikipedia.org/wiki/List_of_Ugaritic_deities	200–234 divine names	MediaWiki API. Cuneiform/alphabetic writings + functions.
15	Ugaritic	Sapiru Project deity lists	sapiru.wordpress.com	~60–80 per list (multiple lists)	Actual Ras Shamra sacrificial deity lists (~1250 BCE).
16	Avestan	Avesta.org Zoroastrian Names	avesta.org/znames.htm	400+ personal names + divine names	Single long page. Based on Bartholomae.
17	Avestan	Encyclopaedia Iranica	iranicaonline.org	400+ names (article "Personal Names, Iranian ii")	Per-deity articles (Anahita, Mithra, Verethragna, Amesha Spentas).
18	Etruscan	ETP (Etruscan Texts Project, UMass)	scholarworks.umass.edu/ces_texts/	200+ (from 300+ post-1990 inscriptions)	Searchable by keyword, location, date.
19	Etruscan	Godchecker Etruscan Mythology	godchecker.com/etruscan-mythology/list-of-names/	89 deity names	Static HTML list.
20	Old Norse	Nordic Names	nordicnames.de/wiki/Category:Old_Norse_Names	Substantial subset of 50,000+ total	MediaWiki API. Name, meaning, etymology, gender.
21	Old Norse	Eddic proper nouns (Voluspa.org / Sacred-Texts)	voluspa.org/poeticedda.htm	500–800 unique (deities, giants, dwarves, places, weapons)	Dvergatal alone lists ~70 dwarf names. Requires NLP extraction.

Tier 3 Sources: Existing + Wiktionary Expansion

Language	Source	URL	Est. Names
Hurrian	Palaeolexicon	palaeolexicon.com	50+
Urartian	Oracc eCUT	oracc.museum.upenn.edu/ecut/	100+
Lycian/Lydian/Carian	eDiAna	ediana.gwi.uni-muenchen.de	50+ each
Phoenician	Wiktionary	en.wiktionary.org	50+
PIE	Wiktionary reconstructed theonyms	en.wiktionary.org	30+
Mycenaean	DAMOS	damos.hf.uio.no	100+
Sanskrit	Wiktionary proper nouns	en.wiktionary.org	500+
OCS	Wiktionary proper nouns	en.wiktionary.org	100+

4.3 Per-Language Script

Create scripts/scrape_proper_nouns.py — a unified script with per-language configs:

PROPER_NOUN_CONFIGS = {
    "grc": {
        "sources": [
            {"type": "wiktionary_cat", "category": "Category:Ancient_Greek_proper_nouns"},
            {"type": "theoi", "url": "https://www.theoi.com/greek-mythology/..."},
        ],
        "iso_for_translit": "grc",
        "tsv_filename": "grc.tsv",
    },
    ...
}

4.4 Adversarial Audit for Proper Nouns

Team B checks (in addition to standard v2):

• Verify 20 proper nouns are attested in the source language (not modern inventions)
• Verify Concept_ID tags are correct (theonym vs toponym vs anthroponym)
• Verify no modern-language proper nouns leaked in (e.g., English "John" in a Gothic file)

---

PHASE 5: Source Quality Upgrades

Priority: MEDIUM — replaces weak sources with stronger ones Estimated effort: 2 sessions Adversarial audit: YES

5.1 Replace avesta.org with Bartholomae

Problem: avesta.org is a personal website by a non-specialist, based on a 125-year-old dictionary. Solution: After Phase 2 restores the avesta_org data, write a SECOND script that cross-references against Bartholomae's Altiranisches Wörterbuch entries available via:

• TITUS Frankfurt digitized texts
• Wiktionary entries that cite Bartholomae

Script: scripts/crossref_avestan_bartholomae.py

• For each avesta_org entry, search Wiktionary for a matching Avestan entry with Bartholomae citation
• Flag entries that appear in avesta_org but NOT in any academic source
• Add bartholomae_verified: true/false to audit trail

5.2 Cross-Reference Palaeolexicon Against eDiAna

Problem: Palaeolexicon (1,960 entries across 6 languages) is a volunteer project with no peer review. Solution: For Anatolian languages where eDiAna overlaps (Lycian, Lydian, Carian, Luwian), verify Palaeolexicon entries against eDiAna.

Script: scripts/crossref_palaeolexicon_ediana.py

• Load both Palaeolexicon and eDiAna entries for each Anatolian language
• Flag Palaeolexicon entries with no eDiAna match
• Log verification status to audit trail

5.3 Upgrade ABVD Data via Lexibank 2

Problem: ABVD entries are ~50% orthographic (fake-IPA). Solution: Where Lexibank 2 provides CLTS-standardized versions of ABVD languages, prefer those.

Script: scripts/upgrade_abvd_lexibank.py

• Download Lexibank 2 standardized forms for ABVD languages
• For each ABVD entry where Lexibank provides a CLTS-standardized IPA, update the IPA column
• Apply Never-Regress Rule: only update if Lexibank IPA differs from Word (i.e., is not identity)

---

PHASE 6: New Language Ingestion — Tier 2

Priority: MEDIUM — important but less critical than Tier 1 Estimated effort: 3–4 sessions (parallelizable) Adversarial audit: YES (full v2 pipeline per language)

Languages

Language	ISO	Family	Primary Source	Est. Entries
Coptic	cop	Afroasiatic	Coptic Dictionary Online (coptic-dictionary.org)	5,000–11,263
Hattic	xht	Isolate	Palaeolexicon + Wiktionary	100–300
Pali	pli	Indo-European	PTS Dictionary (dsal.uchicago.edu), Digital Pali Dict	5,000+
Classical Armenian	xcl	Indo-European	Wiktionary Category:Old_Armenian_lemmas, Calfa.fr	2,000+
Old English	ang	Indo-European	Wiktionary Category:Old_English_lemmas	5,000+
Ge'ez	gez	Afroasiatic	Wiktionary + Leslau dictionary if accessible	1,000+
Syriac	syc	Afroasiatic	SEDRA (sedra.bethmardutho.org) + Wiktionary	3,000+
Aramaic (Imperial/Biblical)	arc	Afroasiatic	CAL (cal.huc.edu) + Wiktionary	3,000+
Biblical Hebrew	hbo	Afroasiatic	Wiktionary Category:Biblical_Hebrew_lemmas	3,000+

Per-Language Protocol

Same as Phase 3: create transliteration map → write extraction script → dry-run → adversarial audit → run live → update metadata.

Each language needs:

1. Transliteration map in transliteration_maps.py with cited reference
2. Extraction script in scripts/
3. Entry in language_configs.py
4. Proper noun scraping (gods, places, rulers) from the same sources

---

PHASE 7: New Language Ingestion — Tier 3 & Proto-Languages

Priority: LOW-MEDIUM — expansion after core is solid Estimated effort: 4+ sessions Adversarial audit: YES

Tier 3 Ancient Languages

Language	ISO	Source	Est. Entries
Middle Persian	pal	MPCD (mpcorpus.org)	3,000+
Sogdian	sog	Gharib Dictionary (Internet Archive)	1,000+
Old Japanese	ojp	ONCOJ (oncoj.ninjal.ac.jp)	2,000+
Gaulish	xtg	Lexicon Leponticum	500+
Oscan	osc	CEIPoM (Zenodo)	500+
Umbrian	xum	CEIPoM	300+
Venetic	xve	CEIPoM	300+
Classical Nahuatl	nci	Wiktionary + colonial dictionaries	2,000+
Eblaite	xeb	Oracc/DCCLT	1,000+
Old Irish	sga	eDIL (dil.ie)	5,000+
Palaic	plq	eDiAna	50+

Reconstructed Proto-Languages

Language	ISO	Source	Est. Entries
Proto-Austronesian	map	ACD (acd.clld.org)	3,000–5,000
Proto-Uralic	urj-pro	Wiktionary + Starostin	500+
Proto-Bantu	bnt-pro	BLR3 (africamuseum.be)	5,000+
Proto-Sino-Tibetan	sit-pro	STEDT (stedt.berkeley.edu)	1,000+
Proto-Celtic	cel-pro	Matasovic dictionary (Internet Archive)	1,000+
Proto-Germanic	gem-pro	Wiktionary Category:Proto-Germanic_lemmas	2,000+

---

PHASE 8: Ongoing Quality Assurance

8.1 Automated Validation Suite

Write scripts/validate_all.py — a comprehensive validation script that runs after ANY data change:

def validate_all():
    for tsv in LEXICON_DIR.glob("*.tsv"):
        # 1. Header check
        # 2. No empty IPA
        # 3. No duplicate Words
        # 4. SCA matches ipa_to_sound_class(IPA) for all entries
        # 5. No '0' in SCA (flag but don't fail — may be legitimate unknowns)
        # 6. Source field is non-empty
        # 7. Entry count matches languages.tsv
        # 8. No known artifact patterns (inprogress, phoneticvalue, etc.)

8.2 Pre-Push Validation Gate

Add to the HuggingFace push workflow:

1. Run validate_all.py — must pass with 0 errors
2. Run reprocess_ipa.py --dry-run — verify no regressions
3. Verify all TSV files have correct header
4. Verify languages.tsv entry counts match actual

8.3 DATABASE_REFERENCE.md Auto-Update

After every phase completion, update DATABASE_REFERENCE.md with:

• New language entries in the Ancient Languages table
• Updated entry counts
• New source entries in the Source Registry
• New transliteration maps in the Map Registry

---

Execution Order & Dependencies

PHASE 0 (Critical Bugs)
    ├── 0.1 SCA tokenizer fix
    ├── 0.2 Nasalized vowel fix
    ├── 0.3 Clean artifacts script
    ├── 0.4 Metadata update
    └── 0.5 Presentation fixes
         ↓
PHASE 1 (IPA Map Fixes)  ──→  reprocess_ipa.py  ──→  validate_all.py
         ↓
PHASE 2 (Data Restoration)
    ├── 2.1 Avestan re-scrape
    ├── 2.2 Sumerogram tagging
    └── 2.3 Contamination fix
         ↓
PHASE 3 (Tier 1 Languages)  ←──  can run 9 languages in PARALLEL
    ├── 3.1 Sumerian
    ├── 3.2 Akkadian
    ├── 3.3 Egyptian
    ├── 3.4 Sanskrit
    ├── 3.5 Ancient Greek
    ├── 3.6 Gothic
    ├── 3.7 Mycenaean Greek
    ├── 3.8 OCS
    └── 3.9 Old Norse
         ↓
PHASE 4 (Proper Nouns)  ←──  runs AFTER Phase 3 (needs Tier 1 TSVs to exist)
         ↓
PHASE 5 (Source Upgrades)  ←──  independent, can run in parallel with Phase 4
         ↓
PHASE 6 (Tier 2 Languages)
         ↓
PHASE 7 (Tier 3 + Proto-Languages)
         ↓
PHASE 8 (Ongoing QA)  ←──  continuous after all phases

---

Success Criteria

Metric	Current	Target
Ancient/reconstructed languages	23	42+ (Tier 1+2)
Total ancient language entries	17,567	100,000+
Languages with >80% non-identity IPA	10	30+
Languages with 0% empty Concept_IDs	~5	25+
SCA "0" rate across all ancient langs	~5%	<1%
Proper noun coverage per language	Variable	All languages have theonym + toponym entries
Adversarial audit pass rate	—	100% (all phases pass v2 audit)
HuggingFace accessibility	Private	Public
License	None	CC-BY-SA-4.0 (file present)

---

Appendix A: Script Naming Convention

scripts/scrape_{source}_{language}.py     # Single-source, single-language
scripts/scrape_{source}.py                # Single-source, multi-language
scripts/scrape_proper_nouns.py            # Unified proper noun scraper
scripts/clean_{issue}.py                  # Cleaning/fixing scripts
scripts/crossref_{source1}_{source2}.py   # Cross-reference validation
scripts/upgrade_{source}.py              # Source quality upgrades
scripts/validate_all.py                   # Comprehensive validation
scripts/tag_sumerograms.py               # Sumerogram identification

Appendix B: Transliteration Map Naming Convention

# In transliteration_maps.py:
SUMERIAN_MAP: Dict[str, str] = { ... }      # Jagersma (2010)
AKKADIAN_MAP: Dict[str, str] = { ... }      # Huehnergard (2011)
EGYPTIAN_MAP: Dict[str, str] = { ... }      # Allen (2014)
SANSKRIT_MAP: Dict[str, str] = { ... }      # Whitney (1896)
ANCIENT_GREEK_MAP: Dict[str, str] = { ... } # Allen (1987)
GOTHIC_MAP: Dict[str, str] = { ... }        # Wright (1910)
MYCENAEAN_MAP: Dict[str, str] = { ... }     # Ventris & Chadwick (1973)
OCS_MAP: Dict[str, str] = { ... }           # Lunt (2001)
OLD_NORSE_MAP: Dict[str, str] = { ... }     # Gordon (1957)

Appendix C: Adversarial Auditor Dispatch Template

When deploying the adversarial pipeline for any phase, spawn two parallel agents:

Agent A (Extraction):

You are Team A (Extraction Agent). Your job is to write and run a Python script
that scrapes {SOURCE} for {LANGUAGE} data. Follow the Iron Law: all data must
come from HTTP requests. Use the standard script template from DATABASE_REFERENCE.md.
[... phase-specific instructions ...]

Agent B (Adversarial Auditor v2):

You are Team B (Critical Adversarial Auditor v2). You have VETO POWER.
After Agent A completes, perform the following DEEP checks:

1. 50-WORD CROSS-REFERENCE: Select 50 random entries from the output TSV.
   For each, construct the source URL and verify the word appears there.
   Use WebFetch to check each URL. Report matches and mismatches.

2. IPA SPOT-CHECK: For 20 random entries, manually apply the transliteration
   map character-by-character. Show your work. Report any mismatches.

3. SCA CONSISTENCY: For 20 random entries, verify ipa_to_sound_class(IPA) == SCA.

4. SOURCE PROVENANCE: For 10 random entries, provide the exact URL where
   each entry can be verified. Fetch each URL and confirm.

5. CONCEPT ID ACCURACY: For 20 entries with glosses, verify the gloss matches
   the source definition.

6. DEDUP: Count unique words. Report any duplicates.

7. ENTRY COUNT: Is the count non-round and plausible?

DO NOT perform surface-level checks (header format, encoding, file existence).
Only perform checks that touch REAL DATA and REAL SOURCES.

Produce a full v2 audit report. Verdict: PASS or FAIL with blocking issues.

---

End of PRD