Phase 8: Add 24 new ancient/proto-language lexicons (12,911 entries) + scripts

data/training/lexicons/aav-pro.tsv

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e79666cda701eb7cacdeed5dabd6a42efee4b3f166370098ae9dadbdd64c4531
+size 7221

data/training/lexicons/afa-pro.tsv

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:61133fa6798c15a365e718b5a9661a93a111704ac44835096aa1c6a110002805
+size 1754

data/training/lexicons/alg-pro.tsv

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:688232af1cf00874d4a03d40ee70fdf891f49c0753a550d8ead36f1849e1e25f
+size 12516

data/training/lexicons/ang.tsv

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:97148a592a22d23447219bafbaf162a2a6dfdfa9be5e62fd8de2a79501436e87
+size 1380125

data/training/lexicons/bnt-pro.tsv

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:60e43f675233ad626ce95faf531f2fe62de60ffd1f89861503277313fe15a2d0
+size 21342

data/training/lexicons/cel-pro.tsv

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:edb8dccc90cd4fca4240c2d9b633a157e0d2e286dfda77c38bcbddf9d445caaf
+size 69341

data/training/lexicons/cop.tsv

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4e21fd347ad46ae447d2fa035be6b16bb5ee3909686db37ad7f4d31da8b6546d
+size 612872

data/training/lexicons/gem-pro.tsv

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f4e072d2c7c801551aea6ecbc7ff3990d890a123929d39c7da0810de65b5722a
+size 247613

data/training/lexicons/gez.tsv

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:74a847a4436cda72eca00c8fef513e62b89d6f292e5330e97f89fcac4c333d3b
+size 24220

data/training/lexicons/hbo.tsv

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8b61db4b0beca87cf5295dfd6b2886335cf01fd020f2ff22c80f7ac82e6ce2ff
+size 557976

data/training/lexicons/ira-pro.tsv

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f13c6f351a44148c6756b475dba2ad57bbe51cd22abe5d827aaf473d3f8c175c
+size 16964

data/training/lexicons/itc-pro.tsv

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d261a9c43d834be998cb15ebaba0b5d5dcfb2bad7399c8fc7e77bf1bffb16573
+size 31068

data/training/lexicons/jpx-pro.tsv

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e631d8018dfb1a5f6e49fa0701eecbf4b546fa266ca8df2fc0c340e0de60ae31
+size 15732

data/training/lexicons/map.tsv

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0aa3f80b9d64b9e91525b830a391a1922f56065af0ce8563eaf21de512f33d60
+size 354233

data/training/lexicons/myn-pro.tsv

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fcaea06aab389a39f5bba7c957ca5fe7c378cf33dc0207ab49294db97a29a08e
+size 2744

data/training/lexicons/nci.tsv

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6d9653460ed3fe2dc15a8b845828ab15f3bb024473cba810f2a5a46a14965b98
+size 207573

data/training/lexicons/obm.tsv

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:456818e7df5c705bceb3ffcf07ec8d622ce305880042911f0c667d4d9f8a4413
+size 1423

data/training/lexicons/ojp.tsv

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fc09525f1db3c46b6b7a06e2d7e5f32a8716032e0425127ea63ec62eb7cd27e1
+size 186624

data/training/lexicons/osc.tsv

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1e2dabe2edd32d53ae87518f265c4d3b200596dd3e04c2cbb87201f1e7200e95
+size 75402

data/training/lexicons/pal.tsv

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cb7237468dbb09443f993b90f8428226150ab8747aac22d338f5ed7062ecad50
+size 11959

data/training/lexicons/pli.tsv

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:309bed32d4a1a30ab025762ead709019230ee559caa1e55588ca665a2ff50578
+size 130234

data/training/lexicons/poz-oce-pro.tsv

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cc107d7ac5cb3a1b86e4708bcaae85e68cb8b351fa0105fc997906db496f93ab
+size 4462

data/training/lexicons/poz-pol-pro.tsv

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:02312e2e280f98f3c1310d729cd36ddb772daa209a56cbed7e620778c0af081e
+size 5655

data/training/lexicons/sga.tsv

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:acf7b4bb75992066e0f8853300ea83df8b940f8b37541b7f79b2ac3b552ee357
+size 1501984

data/training/lexicons/sit-pro.tsv

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:12bd35e6fb1bdea808cb741f50c53e9cbf82b6440f8fb2f110877af0ab84b012
+size 13739

data/training/lexicons/sla-pro.tsv

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d1908678551d9fb10053004ce7a91a04d76a7c7f4f2dec03ff5b02a52fd534f7
+size 220064

data/training/lexicons/sog.tsv

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6500b1de893c8c643c5a81ce22ec249da693c67c3b4c9ca710c64a349f6f2c82
+size 6883

data/training/lexicons/sqj-pro.tsv

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b66f53ceec08d01ee57913b7c6a1259f301650d19cc2754cab53f781c2e6680b
+size 8540

data/training/lexicons/tai-pro.tsv

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a11ac8b7d9287870b00a2455c2620c1fdf3b4d2f34d4841af2bcce2a08ec7df2
+size 5849

data/training/lexicons/trk-pro.tsv

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:49fe64e6da2e399c85b29255c124836215099d18863357eddb12cd2be82b176c
+size 39499

data/training/lexicons/urj-pro.tsv

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e89b1c38e6c5c370f949a88f7801ea6e4e1f881d5fc73cda95195767dde31d86
+size 22682

data/training/lexicons/xce.tsv

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6c341342ebb3dfa3293e109a91221252b3087b510b34b859b75b5b2d13ef589a
+size 497

data/training/lexicons/xcl.tsv

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:701be1b023101c427700053fca0c26feb572980e0c56a7b71873cc24f70a22c5
+size 311750

data/training/lexicons/xeb.tsv

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7028c905272098ae4d9e4e963c305d372577d24d90b5dbcb8236f3539062ff5e
+size 30017

data/training/lexicons/xfa.tsv

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:17ad979a03bce750cb6fdc11931aa4eb8b594601e9e6afa485bd6d81a49e0d4b
+size 17401

data/training/lexicons/xgn-pro.tsv

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c41c4de9488420092713d0d0fc55fa3718b155f6f74dd5558520ed4fde6ddd8e
+size 5029

data/training/lexicons/xht.tsv

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4809b7c19c84179c2620c39f5583b2352b86a1f0f7a75569c6b489c4b7d7d67a
+size 9337

data/training/lexicons/xib.tsv

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d510524e0a000a5e0a2f46a3642f0b7d7f47cbb616cf9e4f789fd9efa5334e9d
+size 1416

data/training/lexicons/xlp.tsv

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f5df739a8ea99c7f596df81c611470f7c5d17b3146b9d683753853ee950a1a0c
+size 14942

data/training/lexicons/xmr.tsv

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f4c7f024775825b84326561eb9807d5860914977efcdaea91b85a6410c1293f1
+size 98939

data/training/lexicons/xsa.tsv

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a98e9845399f78284581435401155e20020233e3fcdf05e72e1f22780c2c89bd
+size 4935

data/training/lexicons/xtg.tsv

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ff1215f061de476ef0d0a7fc67047179a88a6257d2cc5a21604677d9dac286d0
+size 9388

data/training/lexicons/xto-pro.tsv

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b4fc7376fb22ee052059363f36bdf934e670d5794b5fa9c6018e259e5f17be03
+size 5721

data/training/lexicons/xum.tsv

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3e8fd6d438f641afbdefd4ec727d1e1e97701f99d2a3bec3729250b19c6d7bb5
+size 60372

data/training/lexicons/xve.tsv

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e02116a96d18c5571c2d407c55e234f114c6f6e2a7499c7bd1d3af7a6a36dc97
+size 24924

data/training/metadata/languages.tsv

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9a3f345f6180a9acffa6feef04d3181c706d9e8b1539b6f318783d1e1e0fa8a7
+size 39342

docs/DATABASE_REFERENCE.md

@@ -0,0 +1,1003 @@
+# Ancient Scripts Datasets — Master Database Reference
+
+> **Last updated:** 2026-03-13 | **Commit:** `3e3fdf1` | **Total entries:** 3,466,000+ across 1,178 languages
+
+This document is the single source of truth for understanding, modifying, and extending this database. It is designed for both human researchers and AI agents.
+
+---
+
+## Table of Contents
+
+1. [Database Overview](#1-database-overview)
+2. [TSV Schema & Format](#2-tsv-schema--format)
+3. [Ancient Languages — Complete Registry](#3-ancient-languages--complete-registry)
+4. [Non-Ancient Languages — Summary](#4-non-ancient-languages--summary)
+5. [Source Registry](#5-source-registry)
+6. [IPA & Phonetic Processing Pipeline](#6-ipa--phonetic-processing-pipeline)
+7. [Transliteration Maps System](#7-transliteration-maps-system)
+8. [Sound Class (SCA) System](#8-sound-class-sca-system)
+9. [Scripts & Data Flow](#9-scripts--data-flow)
+10. [PRD: Adding New Data](#10-prd-adding-new-data)
+11. [PRD: Adding New Languages](#11-prd-adding-new-languages)
+12. [Data Acquisition Rules (Iron Law)](#12-data-acquisition-rules-iron-law)
+13. [Adversarial Review Protocol](#13-adversarial-review-protocol)
+14. [Re-processing & Cleaning Runbook](#14-re-processing--cleaning-runbook)
+15. [Known Limitations & Future Work](#15-known-limitations--future-work)
+
+---
+
+## 1. Database Overview
+
+### Locations
+
+| Location | Path / URL | What |
+|----------|-----------|------|
+| **HuggingFace dataset** | `https://huggingface.co/datasets/PhaistosLabs/ancient-scripts-datasets` | **PRIMARY cloud copy.** All lexicons, cognate pairs, metadata, sources, scripts, docs. Push here after any data change. |
+| **HuggingFace local clone** | `C:\Users\alvin\hf-ancient-scripts\` | Local clone of HuggingFace repo. Use `huggingface_hub` API or `git push` to sync. |
+| **GitHub repo** | `https://github.com/Nacryos/ancient-scripts-datasets.git` | Scripts, docs, pipeline code. Lexicon data is gitignored but committed via force-add for some ancient langs. |
+| **Local working copy** | `C:\Users\alvin\ancient-scripts-datasets\` | Full repo + generated data + CLDF sources |
+| **CLDF sources** | `sources/` (593 MB) | **Gitignored.** Cloned separately: `northeuralex`, `ids`, `abvd`, `wold`, `sinotibetan`, `wikipron` |
+| **Total local footprint** | 2.2 GB | Includes all generated data + CLDF source repos |
+
+### What IS Tracked in Git (GitHub)
+
+- `scripts/` — All extraction and processing scripts
+- `cognate_pipeline/` — Python package for phonetic processing
+- `docs/` — PRDs, audit reports, this reference doc
+- `data/training/metadata/` — `languages.tsv`, `source_stats.tsv` (small summary files)
+- `data/training/validation/` — Validation sets (via Git LFS)
+- `data/training/lexicons/*.tsv` — Ancient language TSVs (force-added despite gitignore)
+
+### What is NOT Tracked in Git (gitignored)
+
+- `data/training/lexicons/` — Modern language TSVs (1,113 files, regenerated from scripts)
+- `data/training/cognate_pairs/` — Cognate pair datasets (regenerated)
+- `sources/` — CLDF source repositories (cloned separately, ~593 MB)
+
+### What IS on HuggingFace (everything)
+
+**HuggingFace is the single source of truth for ALL data files.** It contains:
+- All 1,136 lexicon TSVs (ancient + modern)
+- All cognate pair datasets
+- All metadata files
+- All scripts, docs, and pipeline code
+- All CLDF source repos (2,928 files in `sources/`)
+- Raw audit trails and intermediate extraction files
+
+### HuggingFace Push Rules
+
+1. **After any data change** (new entries, IPA reprocessing, map fixes): push updated TSVs to HF
+2. **After any script change** that affects output: push scripts to HF
+3. **Use `huggingface_hub` API** for individual file uploads:
+   ```python
+   from huggingface_hub import HfApi
+   api = HfApi()
+   api.upload_file(
+       path_or_fileobj="data/training/lexicons/ave.tsv",
+       path_in_repo="data/training/lexicons/ave.tsv",
+       repo_id="PhaistosLabs/ancient-scripts-datasets",
+       repo_type="dataset",
+       commit_message="fix: reprocess Avestan IPA with expanded transliteration map"
+   )
+   ```
+4. **For bulk uploads** (many files): use `upload_large_folder()` from the HF local clone at `C:\Users\alvin\hf-ancient-scripts\`
+5. **Always push to BOTH** GitHub (scripts/docs) and HuggingFace (data + scripts/docs)
+6. **Never let HF fall behind** — if data exists locally but not on HF, it's not deployed
+
+**To reconstruct all data from scratch:**
+```bash
+# 1. Clone CLDF sources
+git clone https://github.com/lexibank/northeuralex sources/northeuralex
+git clone https://github.com/lexibank/ids sources/ids
+git clone https://github.com/lexibank/abvd sources/abvd
+git clone https://github.com/lexibank/wold sources/wold
+git clone https://github.com/lexibank/sinotibetan sources/sinotibetan
+# WikiPron: download from https://github.com/CUNY-CL/wikipron
+
+# 2. Run extraction pipeline
+python scripts/expand_cldf_full.py        # Modern languages from CLDF
+python scripts/ingest_wikipron.py          # WikiPron IPA data
+python scripts/run_lexicon_expansion.py    # Ancient language extraction (requires internet)
+python scripts/reprocess_ipa.py            # Apply transliteration maps
+python scripts/assemble_lexicons.py        # Generate metadata
+```
+
+### Directory Structure
+
+```
+ancient-scripts-datasets/
+  data/training/
+    lexicons/           # 1,136 TSV files (one per language) [GITIGNORED]
+    metadata/           # languages.tsv, source_stats.tsv, etc. [TRACKED]
+    cognate_pairs/      # inherited, similarity, borrowing pairs [GITIGNORED]
+    validation/         # stratified ML training/test sets [GIT LFS]
+    language_profiles/  # per-language markdown profiles
+    raw/                # raw JSON audit trails
+    audit_trails/       # JSONL provenance logs
+  scripts/              # 23 extraction scripts + 7 parsers [TRACKED]
+  cognate_pipeline/     # Python package for phonetic processing [TRACKED]
+  docs/                 # PRDs, audit reports, this file [TRACKED]
+  sources/              # CLDF repos [GITIGNORED, clone separately]
+```
+
+**Scale:**
+- 1,178 languages (68 ancient/reconstructed + 1,113 modern — 3 overlap)
+- 3,466,000+ total lexical entries
+- 170,756 ancient language entries (68 languages)
+- 3,296,156 modern language entries (1,113 languages)
+
+---
+
+## 2. TSV Schema & Format
+
+Every lexicon file follows this 6-column tab-separated schema:
+
+```
+Word	IPA	SCA	Source	Concept_ID	Cognate_Set_ID
+```
+
+| Column | Description | Example |
+|--------|-------------|---------|
+| **Word** | Orthographic/transliterated form | `pahhur`, `*wódr̥`, `𐬀𐬵𐬎𐬭𐬀` |
+| **IPA** | Broad phonemic IPA transcription | `paxːur`, `wodr̩`, `ahura` |
+| **SCA** | Sound Class Alphabet encoding (18C + 5V) | `PAKUR`, `WOTR`, `AHURA` |
+| **Source** | Data provenance identifier | `wiktionary`, `ediana`, `wikipron` |
+| **Concept_ID** | Semantic concept (first gloss word, snake_case) | `fire`, `water`, `-` |
+| **Cognate_Set_ID** | Cognate grouping identifier | `PIE_fire_001`, `-` |
+
+**Rules:**
+- Header row MUST be present as line 1
+- UTF-8 encoding, Unix line endings preferred
+- No empty IPA fields — use Word as fallback if no conversion possible
+- Source field must accurately reflect actual data origin
+- `-` for unknown/unavailable fields
+
+---
+
+## 3. Ancient Languages — Complete Registry
+
+### Entry Counts & IPA Quality (as of 2026-03-12)
+
+| # | Language | ISO | Family | Entries | Identity% | Top Sources | IPA Type |
+|---|----------|-----|--------|---------|-----------|-------------|----------|
+| 1 | Avestan | ave | Indo-Iranian | 3,455 | 14.4% | avesta_org (2,716), wiktionary_cat (384), wiktionary (355) | Broad phonemic (Hoffmann & Forssman) |
+| 2 | Tocharian B | txb | Indo-European | 2,386 | 25.2% | wiktionary_cat (2,386) | Broad phonemic (Tocharian map) |
+| 3 | Luwian | xlw | Anatolian | 2,230 | 26.2% | ediana (1,985), palaeolexicon (225) | Broad phonemic (Luwian map) |
+| 4 | Proto-Indo-European | ine-pro | Indo-European | 1,704 | 0.2% | wiktionary_cat (863), wiktionary (841) | Broad phonemic (reconstructed) |
+| 5 | Lycian | xlc | Anatolian | 1,098 | 36.7% | ediana (517), palaeolexicon (482) | Broad phonemic (Melchert 2004) |
+| 6 | Etruscan | ett | Tyrsenian | 753 | 25.5% | palaeolexicon (503), wikipron (207) | Broad phonemic (Bonfante) |
+| 7 | Urartian | xur | Hurro-Urartian | 748 | 54.4% | oracc_ecut (704), wiktionary (44) | Partial (cuneiform sign names) |
+| 8 | Lydian | xld | Anatolian | 693 | 53.0% | ediana (447), palaeolexicon (187) | Broad phonemic (Gusmani 1964) |
+| 9 | Carian | xcr | Anatolian | 532 | 39.7% | palaeolexicon (304), ediana (174) | Broad phonemic (Adiego 2007) |
+| 10 | Proto-Kartvelian | ccs-pro | Kartvelian | 504 | 22.2% | wiktionary (254), wiktionary_cat (250) | Broad phonemic (Klimov 1998) |
+| 11 | Old Persian | peo | Indo-Iranian | 486 | 10.5% | wiktionary (244), wiktionary_cat (242) | Broad phonemic (Kent 1953) |
+| 12 | Tocharian A | xto | Indo-European | 467 | 23.1% | wiktionary_cat (467) | Broad phonemic (Tocharian map) |
+| 13 | Proto-Dravidian | dra-pro | Dravidian | 406 | 7.1% | wiktionary_cat (235), wiktionary (171) | Broad phonemic (Krishnamurti) |
+| 14 | Proto-Semitic | sem-pro | Afroasiatic | 386 | 26.9% | wiktionary_cat (247), wiktionary (139) | Broad phonemic (Huehnergard) |
+| 15 | Ugaritic | uga | Afroasiatic | 371 | 15.6% | wiktionary (344), wiktionary_cat (27) | Broad phonemic (Tropper 2000) |
+| 16 | Hittite | hit | Anatolian | 266 | 20.3% | wiktionary (266) | Broad phonemic (Hoffner & Melchert) |
+| 17 | Hurrian | xhu | Hurro-Urartian | 260 | 50.4% | palaeolexicon (259) | Broad phonemic (Wegner 2007) |
+| 18 | Elamite | elx | Isolate | 301 | 71.1% | wiktionary (301) | Minimal (transparent orthography) |
+| 19 | Rhaetic | xrr | Tyrsenian | 187 | 55.1% | tir_raetica (142), wiktionary (45) | Partial (North Italic alphabet) |
+| 20 | Phoenician | phn | Afroasiatic | 180 | 18.3% | wiktionary (180) | Broad phonemic (abjad reconstruction) |
+| 21 | Phrygian | xpg | Indo-European | 79 | 36.7% | wiktionary (79) | Partial (small corpus, Greek-script support) |
+| 22 | Messapic | cms | Indo-European | 45 | 88.9% | wiktionary (45) | Minimal (Greek-alphabet, mostly identity) |
+| 23 | Lemnian | xle | Tyrsenian | 30 | 53.3% | wiktionary (30) | Minimal (very small corpus) |
+| | | | | | | | |
+| **--- Tier 2 (Phase 6) ---** | | | | | | | |
+| 24 | Old English | ang | Germanic | 31,319 | 10.5% | wiktionary_cat (31,319) | Broad phonemic (Hogg 1992) |
+| 25 | Biblical Hebrew | hbo | Semitic | 12,182 | 0.1% | wiktionary_cat (12,182) | Broad phonemic (Blau 2010) |
+| 26 | Coptic | cop | Egyptian | 11,180 | 0.1% | wiktionary_cat (7,987), kellia (3,193) | Broad phonemic (Layton 2000) |
+| 27 | Old Armenian | xcl | Indo-European | 6,277 | 0.0% | wiktionary_cat (6,277) | Broad phonemic (Meillet 1913) |
+| 28 | Pali | pli | Indo-Aryan | 2,792 | 19.1% | wiktionary_cat (2,792) | Broad phonemic (Geiger 1943) |
+| 29 | Ge'ez | gez | Semitic | 496 | 0.0% | wiktionary_cat (496) | Broad phonemic (Dillmann 1857) |
+| 30 | Hattic | xht | Isolate | 269 | 37.9% | wiktionary_cat (269) | Partial (cuneiformist conventions) |
+| | | | | | | | |
+| **--- Tier 3 (Phase 7) ---** | | | | | | | |
+| 31 | Old Irish | sga | Celtic | 41,300 | 39.4% | edil (40,309), wiktionary_cat (991) | Broad phonemic (Thurneysen) |
+| 32 | Old Japanese | ojp | Japonic | 5,393 | 59.7% | oncoj (4,974), wiktionary_cat (419) | Broad phonemic (Frellesvig 2010) |
+| 33 | Classical Nahuatl | nci | Uto-Aztecan | 3,873 | 5.7% | wiktionary_cat (3,873) | Broad phonemic |
+| 34 | Oscan | osc | Italic | 2,122 | 15.1% | ceipom (2,122) | Broad phonemic (CEIPoM Standard_aligned) |
+| 35 | Umbrian | xum | Italic | 1,631 | 3.7% | ceipom (1,631) | Broad phonemic (CEIPoM Standard_aligned) |
+| 36 | Venetic | xve | Italic | 721 | 86.5% | ceipom (721) | Minimal (Latin transliteration) |
+| 37 | Gaulish | xtg | Celtic | 271 | 92.3% | diacl (183), wiktionary_cat (88) | Minimal (Latin transliteration) |
+| 38 | Middle Persian | pal | Indo-Iranian | 242 | 62.8% | wiktionary_cat (242) | Broad phonemic (MacKenzie 1971) |
+| 39 | Sogdian | sog | Indo-Iranian | 194 | 37.1% | iecor (161), wiktionary_cat (33) | Broad phonemic (Gharib 1995) |
+| | | | | | | | |
+| **--- Proto-Languages (Phase 7) ---** | | | | | | | |
+| 40 | Proto-Austronesian | map | Austronesian | 11,624 | 41.1% | acd (11,624) | Broad phonemic (Blust notation) |
+| 41 | Proto-Germanic | gem-pro | Germanic | 5,399 | 32.9% | wiktionary_cat (5,399) | Broad phonemic (reconstructed) |
+| 42 | Proto-Celtic | cel-pro | Celtic | 1,584 | 68.3% | wiktionary_cat (1,584) | Partial (mixed Latin/IPA) |
+| 43 | Proto-Uralic | urj-pro | Uralic | 585 | 50.3% | wiktionary_cat (585) | Broad phonemic (Sammallahti 1988) |
+| 44 | Proto-Bantu | bnt-pro | Niger-Congo | 467 | 54.0% | wiktionary_cat (467) | Broad phonemic (BLR notation) |
+| 45 | Proto-Sino-Tibetan | sit-pro | Sino-Tibetan | 358 | 100.0% | wiktionary_cat (358) | Already IPA (Wiktionary provides IPA) |
+| | | | | | | | |
+| **--- Phase 8 Batch 1 (Proto-Languages + Italic/Celtic) ---** | | | | | | | |
+| 46 | Proto-Slavic | sla-pro | Balto-Slavic | 5,068 | 18.4% | wiktionary_cat (5,068) | Broad phonemic (reconstructed) |
+| 47 | Proto-Turkic | trk-pro | Turkic | 1,027 | 27.8% | wiktionary_cat (1,027) | Broad phonemic (reconstructed) |
+| 48 | Proto-Italic | itc-pro | Italic | 739 | 46.7% | wiktionary_cat (739) | Broad phonemic (reconstructed) |
+| 49 | Faliscan | xfa | Italic | 566 | 67.1% | ceipom (566) | Partial (CEIPoM Standard_aligned) |
+| 50 | Proto-Japonic | jpx-pro | Japonic | 426 | 70.2% | wiktionary_cat (426) | Partial (mixed notation) |
+| 51 | Lepontic | xlp | Celtic | 421 | 27.6% | lexlep (421) | Broad phonemic (Lexicon Leponticum) |
+| 52 | Proto-Iranian | ira-pro | Indo-Iranian | 366 | 4.6% | wiktionary_cat (366) | Broad phonemic (reconstructed) |
+| 53 | Ancient South Arabian | xsa | Semitic | 127 | 25.2% | wiktionary (127) | Broad phonemic (Musnad abjad) |
+| 54 | Celtiberian | xce | Celtic | 11 | 100.0% | wiktionary_cat (11) | Minimal (very small corpus) |
+| | | | | | | | |
+| **--- Phase 8 Batch 2 (Proto-Languages + Ancient) ---** | | | | | | | |
+| 55 | Meroitic | xmr | Nilo-Saharan | 1,978 | 39.8% | meroitic-corpus (1,978) | Broad phonemic (Rilly 2007) |
+| 56 | Proto-Algonquian | alg-pro | Algic | 258 | 28.7% | wiktionary_cat (258) | Broad phonemic (reconstructed) |
+| 57 | Proto-Albanian | sqj-pro | Albanian | 210 | 43.8% | wiktionary_cat (210) | Broad phonemic (reconstructed) |
+| 58 | Proto-Austroasiatic | aav-pro | Austroasiatic | 180 | 100.0% | wiktionary_cat (180) | Already IPA (Wiktionary provides IPA) |
+| 59 | Proto-Polynesian | poz-pol-pro | Austronesian | 157 | 100.0% | wiktionary_cat (157) | Already IPA (Wiktionary provides IPA) |
+| 60 | Proto-Tai | tai-pro | Kra-Dai | 148 | 0.7% | wiktionary_cat (148) | Broad phonemic (Li 1977) |
+| 61 | Proto-Tocharian | xto-pro | Tocharian | 138 | 22.5% | wiktionary_cat (138) | Broad phonemic (reconstructed) |
+| 62 | Proto-Mongolic | xgn-pro | Mongolic | 126 | 41.3% | wiktionary_cat (126) | Broad phonemic (reconstructed) |
+| 63 | Proto-Oceanic | poz-oce-pro | Austronesian | 114 | 92.1% | wiktionary_cat (114) | Minimal (transparent orthography) |
+| 64 | Moabite | obm | Semitic | 31 | 0.0% | wiktionary_cat (31) | Broad phonemic (Canaanite abjad) |
+| | | | | | | | |
+| **--- Phase 8 Batch 3 (Proto-Languages + Iberian) ---** | | | | | | | |
+| 65 | Proto-Mayan | myn-pro | Mayan | 65 | 20.0% | wiktionary_cat (65) | Broad phonemic (Kaufman 2003) |
+| 66 | Proto-Afroasiatic | afa-pro | Afroasiatic | 48 | 54.2% | wiktionary_cat (48) | Broad phonemic (Ehret 1995) |
+| 67 | Iberian | xib | Isolate | 39 | 74.4% | wiktionary_cat (39) | Partial (undeciphered script) |
+| | | | | | | | |
+| **--- Phase 8 Eblaite ---** | | | | | | | |
+| 68 | Eblaite | xeb | Semitic | 667 | 0.3% | dcclt-ebla (667) | Broad phonemic (Krebernik 1982) |
+
+**Total ancient + classical: 170,756 entries across 68 languages | Overall identity rate: ~30%**
+
+### Understanding Identity Rate
+
+**Identity rate = % of entries where Word == IPA** (no phonetic conversion applied).
+
+| Rate | Meaning | Example Languages |
+|------|---------|-------------------|
+| <10% | Excellent IPA conversion | ine-pro (0.2%), dra-pro (7.1%) |
+| 10-30% | Good conversion | peo (10.5%), ave (14.4%), hit (20.3%), ccs-pro (22.2%), txb (25.2%) |
+| 30-50% | Moderate — some chars unmapped | xlc (36.7%), xcr (39.7%), xhu (50.4%) |
+| 50-70% | Partial — significant gaps | xld (53.0%), xur (54.4%), elx (71.1%) |
+| >70% | Minimal — mostly passthrough | cms (88.9%) |
+
+**Causes of high identity:**
+- **Cuneiform sign notation** (xur): Uppercase Sumerograms like `LUGAL`, `URU` aren't phonemic — 156 entries in xur
+- **Already-IPA characters** (cms): Some scripts use characters that ARE IPA (θ, ə, ŋ)
+- **Transparent orthography** (elx): Latin letters already map 1:1 to IPA
+- **eDiAna pre-transliterated forms** (xlc, xld): Source provides Latin transliterations that are already close to IPA
+- **Plain ASCII stems** (txb, xto): Short roots like `ak`, `aik` are valid in both orthography and IPA
+
+### IPA Quality Categories
+
+| Category | Definition | Ancient Languages |
+|----------|-----------|-------------------|
+| **FULL** | >80% WikiPron-sourced IPA | (none — ancient langs don't have WikiPron) |
+| **BROAD PHONEMIC** | Scholarly transliteration → IPA via cited map | hit, uga, phn, ave, peo, ine-pro, sem-pro, ccs-pro, dra-pro, xlw, xhu, ett, txb, xto, xld, xcr, xpg |
+| **PARTIAL** | Some chars converted, gaps remain | xlc, xrr |
+| **MINIMAL** | Mostly identity / transparent orthography | elx, xle, cms |
+| **CUNEIFORM MIXED** | Mix of converted transliterations + unconverted sign names | xur |
+
+**Important:** For dead languages, **broad phonemic is the ceiling**. Narrow allophonic IPA is not possible because allophonic variation is unrecoverable from written records. The IPA column represents the best scholarly reconstruction of phonemic values, not actual pronunciation.
+
+---
+
+## 4. Non-Ancient Languages — Summary
+
+- **1,113 languages** with 3,296,156 entries
+- **Dominant source:** WikiPron (85.3% of entries = 2,822,808)
+- **Other sources:** ABVD (6.7%), NorthEuraLex (5.7%), WOLD (1.8%), sinotibetan (0.1%)
+
+**WikiPron entries** have true broad phonemic IPA (scraped from Wiktionary pronunciation sections by trained linguists). These are the gold standard.
+
+**ABVD entries** are often orthographic (Word == IPA). The `fix_abvd_ipa.py` script applies rule-based G2P conversion for Austronesian languages.
+
+---
+
+## 5. Source Registry
+
+| Source ID | Full Name | Type | URL | Languages Covered |
+|-----------|-----------|------|-----|-------------------|
+| `wikipron` | WikiPron Pronunciation Dictionary | Scraped IPA | `sources/wikipron/` (local) | 800+ modern languages |
+| `abvd` | Austronesian Basic Vocabulary Database | CLDF | `sources/abvd/` (local) | 500+ Austronesian |
+| `northeuralex` | NorthEuraLex | CLDF | `sources/northeuralex/` (local) | 100+ Eurasian |
+| `wold` | World Loanword Database | CLDF | `sources/wold/` (local) | 40+ worldwide |
+| `sinotibetan` | Sino-Tibetan Etymological Database | CLDF | `sources/sinotibetan/` (local) | 50+ Sino-Tibetan |
+| `wiktionary` | Wiktionary (appendix/lemma pages) | Web scrape | `en.wiktionary.org` | All ancient langs |
+| `wiktionary_cat` | Wiktionary (category pagination) | MediaWiki API | `en.wiktionary.org/w/api.php` | ine-pro, uga, peo, ave, dra-pro, sem-pro, ccs-pro, txb, xto |
+| `ediana` | eDiAna (LMU Munich) | POST API | `ediana.gwi.uni-muenchen.de` | xlc, xld, xcr, xlw |
+| `palaeolexicon` | Palaeolexicon | REST API | `palaeolexicon.com/api/Search/` | xlc, xld, xcr, xlw, xhu, ett |
+| `oracc_ecut` | Oracc eCUT (Urartian texts) | JSON API | `oracc.museum.upenn.edu/ecut/` | xur |
+| `tir_raetica` | TIR (Thesaurus Inscriptionum Raeticarum) | Web scrape | `tir.univie.ac.at` | xrr |
+| `wikipedia` | Wikipedia vocabulary tables | Web scrape | `en.wikipedia.org` | xur (supplement) |
+| `avesta_org` | Avesta.org Avestan Dictionary | Web scrape | `avesta.org/avdict/avdict.htm` | ave |
+| `kaikki` | Kaikki Wiktionary Dump | JSON dump | `kaikki.org` | Various |
+| `kellia` | Kellia Coptic Lexicon | XML | `data.copticscriptorium.org` | cop |
+| `ceipom` | CEIPoM (Italian Epigraphy) | CSV | `zenodo.org` (CC BY-SA 4.0) | osc, xum, xve |
+| `edil` | eDIL (Electronic Dict of Irish Lang) | XML | `github.com/e-dil/dil` | sga |
+| `acd` | ACD (Austronesian Comparative Dict) | CLDF | `github.com/lexibank/acd` (CC BY 4.0) | map |
+| `oncoj` | ONCOJ (Oxford-NINJAL OJ Corpus) | XML | `github.com/ONCOJ/data` (CC BY 4.0) | ojp |
+| `diacl` | DiACL (Diachronic Atlas of Comp Ling) | CLDF | `github.com/lexibank/diacl` (CC BY 4.0) | xtg |
+| `iecor` | IE-CoR (IE Cognate Relationships) | CLDF | `github.com/lexibank/iecor` (CC BY 4.0) | sog |
+| `lexlep` | Lexicon Leponticum (Zurich) | Web/CSV | `lexlep.univie.ac.at` | xlp |
+| `meroitic-corpus` | Meroitic Language Corpus (GitHub) | JSON/CSV | `github.com/MeroiticLanguage/Meroitic-Corpus` | xmr |
+| `dcclt-ebla` | DCCLT/Ebla (ORACC) | JSON ZIP | `oracc.museum.upenn.edu/dcclt-ebla/` (CC0) | xeb |
+
+---
+
+## 6. IPA & Phonetic Processing Pipeline
+
+### Pipeline Architecture
+
+```
+Source Data (Word column)
+    ↓
+transliterate(word, iso)          ← scripts/transliteration_maps.py
+    ↓                                (greedy longest-match, NFC-normalized)
+IPA string (broad phonemic)
+    ↓
+ipa_to_sound_class(ipa)           ← cognate_pipeline/.../sound_class.py
+    ↓                                (tokenize → segment_to_class → join)
+SCA string (e.g., "PATA")
+```
+
+### IPA Generation Methods (by source type)
+
+| Source | IPA Method | Quality |
+|--------|-----------|---------|
+| WikiPron | Pre-extracted from Wiktionary pronunciation | True broad IPA |
+| Wiktionary (ancient) | `transliterate(word, iso)` via language-specific map | Broad phonemic |
+| ABVD | Orthographic passthrough → `fix_abvd_ipa.py` G2P | Variable |
+| eDiAna | `transliterate(word, iso)` | Broad phonemic |
+| Palaeolexicon | Source IPA if available, else `transliterate()` | Broad phonemic |
+| Oracc | `transliterate(word, iso)` | Partial (cuneiform) |
+| NorthEuraLex/WOLD | CLDF Segments column → joined IPA | Good |
+
+### Never-Regress Re-processing Rule
+
+When re-applying transliteration maps to existing data (`scripts/reprocess_ipa.py`):
+
+```python
+candidate_ipa = transliterate(word, iso)
+
+if candidate_ipa != word:
+    final_ipa = candidate_ipa       # New map converts — use it
+elif old_ipa != word:
+    final_ipa = old_ipa             # New map can't, but old was good — keep
+else:
+    final_ipa = word                # Both identity — nothing to do
+```
+
+**This ensures:** IPA quality can only improve or stay the same. It never regresses.
+
+---
+
+## 7. Transliteration Maps System
+
+**File:** `scripts/transliteration_maps.py` (~800 lines)
+
+### How It Works
+
+Each ancient language has a `Dict[str, str]` mapping scholarly transliteration conventions to broad IPA. The `transliterate()` function applies these via **greedy longest-match**: keys sorted by descending length, first match consumed at each position.
+
+### Map Registry (updated 2026-03-13 — 180+ new rules across 13 original maps + 15 new maps in Phases 6-7 + 24 new maps in Phase 8)
+
+| ISO | Language | Keys | Academic Reference |
+|-----|----------|------|--------------------|
+| `hit` | Hittite | 49 | Hoffner & Melchert (2008) — added š, ḫ, macron vowels |
+| `uga` | Ugaritic | 68 | Tropper (2000) — added ʾ, macron/circumflex vowels, ḫ, ṣ, Ugaritic script (U+10380-1039F) |
+| `phn` | Phoenician | 23 | Standard 22-letter abjad |
+| `xur` | Urartian | 27 | Wegner (2007) — added ṣ, ṭ, y, w, ə, ʾ |
+| `elx` | Elamite | 19 | Grillot-Susini (1987), Stolper (2004) |
+| `xlc` | Lycian | 33 | Melchert (2004) — added x, j, o, long vowels |
+| `xld` | Lydian | 38 | Gusmani (1964), Melchert — added ã, ẽ, ũ (nasalized vowels), c, h, z, x |
+| `xcr` | Carian | 35 | Adiego (2007) — added β, z, v, j, f, ŋ, ĺ, ỳ, ý |
+| `ave` | Avestan | 97 | Hoffmann & Forssman (1996) + Unicode 5.2 (U+10B00-10B3F) |
+| `peo` | Old Persian | 68 | Kent (1953) — added z, č, Old Persian cuneiform syllabary (U+103A0-103C3, 31 signs) |
+| `ine` | Proto-Indo-European | 61 | Fortson (2010), Beekes (2011) — added ḗ, ṓ, morpheme boundaries, accented syllabic sonorants |
+| `sem` | Proto-Semitic | 44 | Huehnergard (2019) |
+| `ccs` | Proto-Kartvelian | 66 | Klimov (1998) — added s₁/z₁/c₁/ʒ₁ subscript series, morpheme boundaries |
+| `dra` | Proto-Dravidian | 49 | Krishnamurti (2003) |
+| `xpg` | Phrygian | 55 | Brixhe & Lejeune (1984), Obrador-Cursach (2020) — added Greek alphabet support (22 letters) |
+| `xle` | Lemnian | 24 | Greek-alphabet reconstruction |
+| `xrr` | Rhaetic | 26 | North Italic alphabet reconstruction |
+| `cms` | Messapic | 25 | Greek-alphabet reconstruction |
+| `xlw` | Luwian | 39 | Melchert (2003), Yakubovich (2010) |
+| `xhu` | Hurrian | 31 | Wegner (2007), Wilhelm (2008) |
+| `ett` | Etruscan | 61 | Bonfante & Bonfante (2002), Rix (1963) + Old Italic Unicode — added z, o, d, g, b, q, σ→s |
+| `txb`/`xto` | Tocharian A/B | 35 | Krause & Thomas (1960), Adams (2013), Peyrot (2008) — added retroflex series (ṭ, ḍ, ṇ, ḷ) |
+| | | | |
+| **--- Phase 6: Tier 2 Maps ---** | | | |
+| `cop` | Coptic | 40+ | Layton (2000), Loprieno (1995) — Sahidic dialect |
+| `pli` | Pali (IAST) | 30+ | Geiger (1943), Oberlies (2001) |
+| `xcl` | Old Armenian | 40+ | Meillet (1913), Schmitt (1981) |
+| `ang` | Old English | 30+ | Hogg (1992), Campbell (1959) |
+| `gez` | Ge'ez (Ethiopic) | 50+ | Dillmann (1857), Tropper (2002) |
+| `hbo` | Biblical Hebrew | 40+ | Blau (2010), Khan (2020) |
+| | | | |
+| **--- Phase 7: Tier 3 + Proto Maps ---** | | | |
+| `osc` | Oscan | 12 | CEIPoM Standard_aligned conventions |
+| `xum` | Umbrian | 12 | CEIPoM Standard_aligned conventions |
+| `xve` | Venetic | 6 | CEIPoM Token_clean conventions |
+| `sga` | Old Irish | 25 | Thurneysen (1946), Stifter (2006) — lenition + macron vowels |
+| `xeb` | Eblaite | 20 | Standard Semitist notation |
+| `nci` | Classical Nahuatl | 15 | Andrews (2003), Launey (2011) |
+| `ojp` | Old Japanese | 20 | Frellesvig (2010), ONCOJ conventions |
+| `pal` | Middle Persian | 25 | MacKenzie (1971), Skjærvø (2009) |
+| `sog` | Sogdian | 25 | Gharib (1995), Sims-Williams (2000) |
+| `xtg` | Gaulish | 15 | Delamarre (2003) |
+| `gem-pro` | Proto-Germanic | 20 | Ringe (2006), Kroonen (2013) |
+| `cel-pro` | Proto-Celtic | 15 | Matasović (2009) |
+| `urj-pro` | Proto-Uralic | 12 | Sammallahti (1988), Janhunen (1981) |
+| `bnt-pro` | Proto-Bantu | 20 | Bastin et al. (2002), Meeussen (1967) |
+| `sit-pro` | Proto-Sino-Tibetan | 18 | Matisoff (2003), Sagart (2004) |
+| | | | |
+| **--- Phase 8 Maps ---** | | | |
+| `sla-pro` | Proto-Slavic | 25+ | Shevelov (1964), Holzer (2007) |
+| `trk-pro` | Proto-Turkic | 20+ | Clauson (1972), Róna-Tas (1991) |
+| `itc-pro` | Proto-Italic | 15+ | Meiser (1998), Bakkum (2009) |
+| `jpx-pro` | Proto-Japonic | 15+ | Vovin (2005), Frellesvig (2010) |
+| `ira-pro` | Proto-Iranian | 20+ | Cheung (2007), Lubotsky (2001) |
+| `xfa` | Faliscan | 12 | CEIPoM Standard_aligned conventions |
+| `xlp` | Lepontic | 25 | Lexicon Leponticum (Stifter et al.) |
+| `xce` | Celtiberian | 15+ | De Bernardo Stempel (1999) |
+| `xsa` | Ancient South Arabian | 30+ | Stein (2003), Beeston (1984) |
+| `alg-pro` | Proto-Algonquian | 15+ | Bloomfield (1946), Goddard (1994) |
+| `sqj-pro` | Proto-Albanian | 15+ | Orel (1998), Demiraj (1997) |
+| `aav-pro` | Proto-Austroasiatic | 10+ | Shorto (2006), Sidwell (2015) |
+| `poz-pol-pro` | Proto-Polynesian | 10+ | Biggs (1978), Pawley (1966) |
+| `tai-pro` | Proto-Tai | 20+ | Li (1977), Pittayaporn (2009) |
+| `xto-pro` | Proto-Tocharian | 15+ | Adams (2013), Peyrot (2008) |
+| `poz-oce-pro` | Proto-Oceanic | 10+ | Ross et al. (1998, 2003, 2008) |
+| `xgn-pro` | Proto-Mongolic | 15+ | Poppe (1955), Nugteren (2011) |
+| `xmr` | Meroitic | 30+ | Rilly (2007), Griffith (1911) |
+| `obm` | Moabite | 22 | Canaanite abjad (shares Phoenician map base) |
+| `myn-pro` | Proto-Mayan | 20+ | Kaufman (2003), Campbell & Kaufman (1985) |
+| `afa-pro` | Proto-Afroasiatic | 15+ | Ehret (1995), Orel & Stolbova (1995) |
+| `xib` | Iberian | 25+ | De Hoz (2010), Untermann (1990) |
+| `xeb` | Eblaite | 20+ | Krebernik (1982), Fronzaroli (2003) |
+
+### NFC Normalization
+
+All map keys and input text are NFC-normalized before comparison. This ensures `š` (U+0161, composed) matches `s` + combining caron (U+0073 + U+030C, decomposed). Cache is per-ISO to prevent cross-language leakage.
+
+### ISO Code Mapping for Proto-Languages
+
+TSV filenames use hyphenated codes but `ALL_MAPS` uses short codes:
+
+| TSV filename ISO | Map ISO |
+|-----------------|---------|
+| `ine-pro` | `ine` |
+| `sem-pro` | `sem` |
+| `ccs-pro` | `ccs` |
+| `dra-pro` | `dra` |
+| `gem-pro` | `gem-pro` |
+| `cel-pro` | `cel-pro` |
+| `urj-pro` | `urj-pro` |
+| `bnt-pro` | `bnt-pro` |
+| `sit-pro` | `sit-pro` |
+
+### Adding a New Map
+
+1. Add the `Dict[str, str]` constant (e.g., `NEW_LANG_MAP`) with cited reference
+2. Register in `ALL_MAPS`: `"iso_code": NEW_LANG_MAP`
+3. Clear `_nfc_cache` implicitly (happens on next call with new ISO)
+4. Run `reprocess_ipa.py --language iso_code` to apply
+5. Deploy adversarial auditor to verify
+
+---
+
+## 8. Sound Class (SCA) System
+
+**File:** `cognate_pipeline/src/cognate_pipeline/normalise/sound_class.py`
+
+### Class Inventory
+
+| Class | IPA Segments | Description |
+|-------|-------------|-------------|
+| A | a, ɑ, æ, ɐ | Open vowels |
+| E | e, ɛ, ə, ɘ, ø, œ | Mid vowels |
+| I | i, ɪ, ɨ | Close front vowels |
+| O | o, ɔ, ɵ | Mid back vowels |
+| U | u, ʊ, ʉ, ɯ, y | Close back vowels |
+| P/B | p, b, ɸ, β | Labial stops |
+| T/D | t, d, ʈ, ɖ | Coronal stops |
+| K/G | k, g, ɡ, q, ɢ | Dorsal stops |
+| S | s, z, ʃ, ʒ, ɕ, ʑ, f, v, θ, ð, x, ɣ, χ, ts, dz, tʃ, dʒ | Fricatives + affricates |
+| M/N | m, n, ɲ, ŋ, ɳ, ɴ | Nasals |
+| L/R | l, ɫ, ɭ, ɬ, r, ɾ, ɽ, ʀ, ɹ, ʁ | Liquids |
+| W/Y | w, ʋ, ɰ, j | Glides |
+| H | ʔ, h, ɦ, ʕ, ħ | Glottals/pharyngeals |
+| 0 | (anything unmapped) | Unknown |
+
+### Processing Chain
+
+```python
+ipa_to_sound_class("paxːur")
+  → tokenize_ipa("paxːur")  →  ["p", "a", "xː", "u", "r"]
+  → [segment_to_class(s) for s in segments]  →  ["P", "A", "K", "U", "R"]
+  → "PAKUR"
+```
+
+---
+
+## 9. Scripts & Data Flow
+
+### Data Flow Diagram
+
+```
+EXTERNAL SOURCES
+  ├── Wiktionary API ──────────→ extract_ave_peo_xpg.py
+  │                              extract_phn_elx.py
+  │                              extract_pie_urartian.py
+  │                              extract_wiktionary_lexicons.py
+  │                              expand_wiktionary_categories.py
+  │                              expand_xpg.py
+  ├── eDiAna API ──────────────→ scrape_ediana.py
+  ├── Palaeolexicon API ───────→ scrape_palaeolexicon.py
+  ├── Oracc JSON API ──────────→ scrape_oracc_urartian.py
+  ├── avesta.org ──────────────→ scrape_avesta_org.py
+  ├── TIR (Vienna) ────────────→ scrape_tir_rhaetic.py
+  ├── WikiPron TSVs ───────────→ ingest_wikipron.py
+  └── CLDF Sources ────────────→ expand_cldf_full.py
+                                  convert_cldf_to_tsv.py
+                     ↓
+         data/training/lexicons/{iso}.tsv
+                     ↓
+         normalize_lexicons.py (NFC, dedup, strip stress)
+         reprocess_ipa.py (re-apply updated transliteration maps)
+         fix_abvd_ipa.py (Austronesian G2P fix)
+                     ↓
+         assemble_lexicons.py → metadata/languages.tsv
+         assign_cognate_links.py → cognate_pairs/*.tsv
+         build_validation_sets.py → validation/*.tsv
+```
+
+### Script Quick Reference
+
+| Script | Purpose | Languages |
+|--------|---------|-----------|
+| `extract_ave_peo_xpg.py` | Wiktionary Swadesh + category | ave, peo, xpg |
+| `extract_phn_elx.py` | Wiktionary + appendix | phn, elx |
+| `extract_pie_urartian.py` | Wiktionary + Wikipedia | ine-pro, xur |
+| `extract_wiktionary_lexicons.py` | Wiktionary appendix | sem-pro, ccs-pro, dra-pro, xle |
+| `extract_anatolian_lexicons.py` | Multi-source | xlc, xld, xcr |
+| `expand_wiktionary_categories.py` | Wiktionary category pagination | ine-pro, uga, peo, ave, dra-pro, sem-pro, ccs-pro |
+| `expand_xpg.py` | Wiktionary category + appendix | xpg |
+| `scrape_ediana.py` | eDiAna POST API | xlc, xld, xcr, xlw |
+| `scrape_palaeolexicon.py` | Palaeolexicon REST API | xlc, xld, xcr, xlw, xhu, ett |
+| `scrape_avesta.py` | avesta.org (old, superseded) | ave |
+| `scrape_avesta_org.py` | avesta.org dictionary (current, adversarial-audited) | ave |
+| `scrape_oracc_urartian.py` | Oracc eCUT JSON API | xur |
+| `scrape_tir_rhaetic.py` | TIR web scrape | xrr |
+| `ingest_wikipron.py` | WikiPron TSV ingestion | 800+ modern |
+| `expand_cldf_full.py` | CLDF full extraction | All CLDF languages |
+| `reprocess_ipa.py` | Re-apply transliteration maps | 23 ancient |
+| `fix_abvd_ipa.py` | G2P for Austronesian | ABVD languages |
+| `normalize_lexicons.py` | NFC + dedup + SCA recompute | All |
+| `assemble_lexicons.py` | Generate metadata | All |
+| `ingest_wiktionary_tier2.py` | Wiktionary category ingestion (Tier 2+) | Phase 6-8 Wiktionary languages |
+| `fetch_wiktionary_raw.py` | Fetch raw Wiktionary category JSON | Phase 6-8 Wiktionary languages |
+| `ingest_dcclt_ebla.py` | ORACC DCCLT/Ebla extraction | xeb |
+| `ingest_meroitic.py` | Meroitic Language Corpus | xmr |
+| `ingest_lexlep.py` | Lexicon Leponticum extraction | xlp |
+| `ingest_ceipom_italic.py` | CEIPoM italic epigraphy | osc, xum, xve, xfa |
+| `update_metadata.py` | Update languages.tsv from disk | All |
+| `validate_all.py` | Comprehensive TSV validation | All |
+| `push_to_hf.py` | Push files to HuggingFace | All Phase 6-8 |
+
+---
+
+## 10. PRD: Adding New Data to Existing Languages
+
+### Prerequisites
+
+- The language already has a TSV file in `data/training/lexicons/`
+- You have identified a new external source with verifiable data
+- A transliteration map exists in `transliteration_maps.py` (if ancient)
+
+### Step-by-Step
+
+#### Step 1: Identify Source
+- Find a publicly accessible online source (API, web page, database)
+- Verify it returns real lexical data (not AI-generated)
+- Document the URL, API format, and expected entry count
+
+#### Step 2: Write Extraction Script
+```python
+# Template: scripts/scrape_{source}_{iso}.py
+#!/usr/bin/env python3
+"""Scrape {Source Name} for {Language} word lists.
+Source: {URL}
+"""
+import urllib.request  # MANDATORY — proves data comes from HTTP
+...
+
+def fetch_data(url):
+    """Fetch from external source."""
+    req = urllib.request.Request(url, headers={"User-Agent": "..."})
+    with urllib.request.urlopen(req) as resp:
+        return json.loads(resp.read())
+
+def process_language(iso, config, dry_run=False):
+    """Process and deduplicate."""
+    existing = load_existing_words(tsv_path)  # MUST deduplicate
+    entries = fetch_data(url)
+    new_entries = [e for e in entries if e["word"] not in existing]
+    ...
+    # Apply transliteration
+    ipa = transliterate(word, iso)
+    sca = ipa_to_sound_class(ipa)
+    f.write(f"{word}\t{ipa}\t{sca}\t{source_id}\t{concept_id}\t-\n")
+```
+
+**Critical:** Script MUST contain `urllib.request.urlopen()`, `requests.get()`, or equivalent HTTP fetch. No hardcoded word lists.
+
+#### Step 3: Run with --dry-run
+```bash
+python scripts/scrape_new_source.py --dry-run --language {iso}
+```
+
+#### Step 4: Run Live
+```bash
+python scripts/scrape_new_source.py --language {iso}
+```
+
+#### Step 5: Re-process IPA (if map was updated)
+```bash
+python scripts/reprocess_ipa.py --language {iso}
+```
+
+#### Step 6: Deploy Adversarial Auditor
+See [Section 13](#13-adversarial-review-protocol).
+
+#### Step 7: Commit & Push to Both Repos
+```bash
+# GitHub
+git add scripts/scrape_new_source.py data/training/lexicons/{iso}.tsv
+git commit -m "Add {N} entries to {Language} from {Source}"
+git push
+
+# HuggingFace (MANDATORY — HF is the primary data host)
+python -c "
+from huggingface_hub import HfApi
+api = HfApi()
+for f in ['data/training/lexicons/{iso}.tsv', 'scripts/scrape_new_source.py']:
+    api.upload_file(path_or_fileobj=f, path_in_repo=f,
+                    repo_id='PhaistosLabs/ancient-scripts-datasets', repo_type='dataset',
+                    commit_message='Add {N} entries to {Language} from {Source}')
+"
+```
+
+---
+
+## 11. PRD: Adding New Languages
+
+### Prerequisites
+
+- ISO 639-3 code identified
+- At least one external source with verifiable word lists
+- Script conventions for the relevant writing system understood
+
+### Step-by-Step
+
+#### Step 1: Create Transliteration Map (if needed)
+
+Add to `scripts/transliteration_maps.py`:
+
+```python
+# ---------------------------------------------------------------------------
+# N. NEW_LANGUAGE  (Author Year, "Title")
+# ---------------------------------------------------------------------------
+NEW_LANGUAGE_MAP: Dict[str, str] = {
+    "a": "a", "b": "b", ...
+    # Every key MUST have a cited academic reference
+}
+```
+
+Register in `ALL_MAPS`:
+```python
+ALL_MAPS = {
+    ...
+    "new_iso": NEW_LANGUAGE_MAP,
+}
+```
+
+#### Step 2: Write Extraction Script
+
+Follow the template in [Section 10](#10-prd-adding-new-data). The script must:
+- Fetch from an external source via HTTP
+- Parse the response (HTML, JSON, XML)
+- Apply `transliterate()` and `ipa_to_sound_class()`
+- Write to `data/training/lexicons/{iso}.tsv`
+- Save raw JSON to `data/training/raw/` for audit trail
+- Deduplicate by Word column
+
+#### Step 3: Add to Language Config (optional)
+
+If the language will be part of the ancient languages pipeline, add to `scripts/language_configs.py`.
+
+#### Step 4: Add to Re-processing List
+
+Add the ISO code to `ANCIENT_LANGUAGES` in `scripts/reprocess_ipa.py` and to `ISO_TO_MAP_ISO` if the TSV filename differs from the map ISO.
+
+#### Step 5: Run Extraction
+```bash
+python scripts/scrape_{source}.py --language {iso} --dry-run
+python scripts/scrape_{source}.py --language {iso}
+```
+
+#### Step 6: Verify
+
+```bash
+# Check entry count and IPA quality
+python scripts/reprocess_ipa.py --dry-run --language {iso}
+```
+
+#### Step 7: Deploy Adversarial Auditor
+
+See [Section 13](#13-adversarial-review-protocol).
+
+#### Step 8: Commit and Push
+
+---
+
+## 12. Data Acquisition Rules (Iron Law)
+
+```
+┌─────────────────────────────────────────────────────────────────────┐
+│  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."             │
+└─────────────────────────────────────────────────────────────────────┘
+```
+
+### What IS Allowed
+
+| Action | Example | Why OK |
+|--------|---------|--------|
+| Write a script with `urllib.request.urlopen()` | `scrape_palaeolexicon.py` | Data comes from HTTP |
+| Parse HTML/JSON from downloaded content | `BeautifulSoup(html)` | Deterministic extraction |
+| Apply transliteration map (CODE, not DATA) | `transliterate(word, "hit")` | Transformation rules are code |
+| Re-compute SCA from IPA | `ipa_to_sound_class(ipa)` | Deterministic function |
+
+### What is FORBIDDEN
+
+| Action | Example | Why Forbidden |
+|--------|---------|---------------|
+| Write data rows directly | `f.write("water\twɔːtər\t...")` | Data authoring |
+| Hardcode word lists from memory | `WORDS = [("fire", "paxːur")]` | LLM knowledge ≠ source |
+| Fill in missing fields with guesses | `ipa = "probably θ"` | Hallucination risk |
+| Generate translations/transcriptions | `ipa = "wɔːtər"  # I know how water sounds` | Not from a source |
+| Pad entries to reach a target count | Adding 13 entries to make it 200 | Fabrication |
+
+### The Cached-Fetch Pattern (Acceptable Gray Area)
+
+If a source requires JavaScript rendering or CAPTCHAs:
+1. Use WebFetch/browser to access the source
+2. Save raw content to `data/training/raw/{source}_{iso}_{date}.html`
+3. Write a parsing script that reads from the saved file
+4. The auditor spot-checks 5 entries against the live source
+
+### Transliteration Maps Are CODE, Not DATA
+
+Transliteration maps (e.g., `"š": "ʃ"`) are **transformation rules** derived from published grammars, not lexical content. Adding or modifying map entries is a code change, not data authoring. However, every map entry MUST cite an academic reference.
+
+---
+
+## 13. Adversarial Review Protocol
+
+### Architecture: Dual-Agent System
+
+```
+Team A (Extraction Agent)     Team B (Adversarial Auditor)
+  ├── Writes code               ├── Reviews code
+  ├── Runs scripts              ├── Spot-checks output
+  ├── Produces TSV data         ├── Verifies provenance
+  └── NEVER writes data         └── Has VETO POWER
+       directly
+```
+
+### When to Deploy
+
+- After ANY new data is added to the database
+- After ANY transliteration map change
+- After ANY re-processing run
+- After ANY script modification that affects output
+
+### Audit Checklist (per modular step)
+
+#### Code Review
+- [ ] Script contains `urllib`/`requests`/`curl` (not hardcoded data)
+- [ ] No literal IPA data in `f.write()` calls
+- [ ] Source attribution matches actual source
+- [ ] Deduplication against existing entries
+
+#### Data Quality
+- [ ] Entry count is non-round and plausible
+- [ ] No duplicate Word values
+- [ ] No empty IPA fields
+- [ ] Identity rate is explainable (not suspiciously low or high)
+- [ ] SCA matches `ipa_to_sound_class(IPA)` for 20 random samples
+
+#### Never-Regress Verification
+- [ ] No entry went from non-identity IPA to identity (regression)
+- [ ] Entry counts did not decrease
+- [ ] Existing Word/Source/Concept_ID/Cognate_Set_ID unchanged
+
+#### Provenance
+- [ ] 20 random entries traced back to source URL
+- [ ] Raw JSON/HTML audit trail saved in `data/training/raw/`
+
+### Red Flags (STOP immediately)
+
+| Red Flag | What It Means |
+|----------|---------------|
+| No `urllib`/`requests` in extraction code | Agent is authoring data |
+| Entry count is exactly round (100, 200, 500) | Likely padded |
+| >90% of entries have empty required fields | Extraction didn't work |
+| Script contains `f.write("word\tipa\t...")` with literal data | Direct data authoring |
+| Transformation output == input for >80% without cited justification | Map not actually applied |
+
+### Report Format
+
+```markdown
+# Adversarial Audit: {Step} — {Language} ({iso})
+## Checks:
+- [ ] No data authoring: PASS/FAIL
+- [ ] Entry count: PASS/FAIL (expected X, got Y)
+- [ ] IPA quality: PASS/FAIL (identity rate: Z%)
+- [ ] SCA consistency: PASS/FAIL (N/N verified)
+- [ ] Provenance: PASS/FAIL (N/20 traced to source)
+## Verdict: PASS / WARN / FAIL
+## Blocking: YES (if FAIL)
+```
+
+---
+
+## 14. Re-processing & Cleaning Runbook
+
+### When to Re-process
+
+- After modifying any transliteration map in `transliteration_maps.py`
+- After fixing a bug in `transliterate()` or `ipa_to_sound_class()`
+- After adding a new language to `ALL_MAPS`
+
+### How to Re-process
+
+```bash
+# Dry run first (ALWAYS)
+python scripts/reprocess_ipa.py --dry-run
+
+# Check: identity rates should decrease or stay the same, NEVER increase
+# Check: "Changed" column shows expected number of modifications
+# Check: "Errors" column is 0
+
+# Run live
+python scripts/reprocess_ipa.py
+
+# Or for a single language
+python scripts/reprocess_ipa.py --language xlw
+```
+
+### Common Cleaning Operations
+
+#### Remove entries with HTML artifacts
+```python
+# Check for HTML entities
+grep -P '&\w+;' data/training/lexicons/{iso}.tsv
+# Remove affected lines via Python script (not manual edit)
+```
+
+#### Remove entries from wrong source (contamination)
+```python
+# Example: Hurrian TSV had Hittite entries from wrong Palaeolexicon ID
+# Write a Python script that identifies and removes contaminated entries
+# Save removed entries to audit trail
+```
+
+#### Deduplicate
+```python
+# reprocess_ipa.py handles dedup by Word column
+# For more complex dedup, use normalize_lexicons.py
+```
+
+#### Fix ABVD fake-IPA
+```bash
+python scripts/fix_abvd_ipa.py
+```
+
+### Post-Cleaning Verification
+
+```bash
+# Verify entry counts
+python -c "
+for iso in ['hit','uga',...]:
+    with open(f'data/training/lexicons/{iso}.tsv') as f:
+        print(f'{iso}: {sum(1 for _ in f) - 1} entries')
+"
+
+# Verify no empty IPA
+python -c "
+for iso in [...]:
+    with open(f'data/training/lexicons/{iso}.tsv') as f:
+        for line in f:
+            parts = line.strip().split('\t')
+            if len(parts) >= 2 and not parts[1]:
+                print(f'EMPTY IPA: {iso} {parts[0]}')
+"
+```
+
+---
+
+## 15. Known Limitations & Future Work
+
+### Linguistic Limitations
+
+| Issue | Languages Affected | Root Cause |
+|-------|-------------------|------------|
+| Broad phonemic only (no allophonic) | All ancient | Dead languages — allophonic variation unrecoverable |
+| Cuneiform sign names as entries | xur, xhu | Source provides sign-level notation, not phonemic. ~156 Sumerograms in xur. |
+| High identity for transparent orthographies | elx, cms, xle | Writing system maps 1:1 to IPA |
+| Old Persian ç → θ debatable | peo | Kent (1953) says /θ/, Kloekhorst (2008) says /ts/ |
+| Old Persian cuneiform inherent vowels | peo | Syllabary signs (𐎣=ka, 𐎫=ta) include inherent vowels that may be redundant in context |
+| eDiAna entries drive high identity | xlc, xld | eDiAna provides already-transliterated forms; identity is expected, not a map gap |
+
+### Technical Debt
+
+| Issue | Priority | Fix |
+|-------|----------|-----|
+| `use_word_for_ipa` dead config in expand_wiktionary_categories.py | Low | Remove the config key |
+| Some extraction scripts have hardcoded word lists from pre-Iron-Law era | Medium | Rewrite with HTTP fetch |
+| ABVD entries still ~50% fake-IPA after G2P fix | Medium | Better G2P or manual review |
+| NorthEuraLex/WOLD join segments with spaces | Low | Handled by normalize_lexicons.py |
+| Combining diacritics in Lycian/Carian (U+0303, U+0302) | Low | Normalize in preprocessing before transliteration |
+| Greek letter leaks in Carian source data | Low | Data cleaning script to normalize σ→s, α→a, etc. |
+| HTML entities in 4 PIE IPA entries | Low | Decode with `html.unescape()` in reprocess_ipa.py |
+| 15 Old Persian proper nouns have wrong-language IPA | Low | Filter or manually correct Akkadian/Greek transcriptions |
+
+### Expansion Opportunities
+
+| Language | Current | Available | Source |
+|----------|---------|-----------|--------|
+| Sumerian | 0 | 5,000+ | EPSD2 (ePSD), Oracc |
+| Akkadian | 0 | 10,000+ | CAD, CDA, ePSD2 |
+| Egyptian | 0 | 3,000+ | TLA (Thesaurus Linguae Aegyptiae) |
+| Sanskrit | (modern only) | 50,000+ | Monier-Williams, DCS |
+| Linear B | 0 | 500+ | DAMOS, Wingspread |
+| Luvian Hieroglyphic | (mixed with xlw) | 500+ | Hawkins (2000) |
+
+---
+
+## Appendix A: Quick Commands
+
+```bash
+# Count entries for a language
+wc -l data/training/lexicons/{iso}.tsv
+
+# Check identity rate
+python -c "
+with open('data/training/lexicons/{iso}.tsv') as f:
+    lines = f.readlines()[1:]
+    total = len(lines)
+    identity = sum(1 for l in lines if l.split('\t')[0] == l.split('\t')[1])
+    print(f'{identity}/{total} = {identity/total*100:.1f}%')
+"
+
+# Test a transliteration map
+python -c "
+import sys; sys.path.insert(0, 'scripts')
+from transliteration_maps import transliterate
+print(transliterate('test_word', 'iso_code'))
+"
+
+# Re-process single language (dry run)
+python scripts/reprocess_ipa.py --dry-run --language {iso}
+
+# Run adversarial audit (deploy via AI agent)
+# See Section 13 for protocol
+```
+
+## Appendix B: File Checksums Reference
+
+Run after any batch operation to create a baseline:
+```bash
+find data/training/lexicons -name "*.tsv" -exec wc -l {} \; | sort -k2 > /tmp/lexicon_counts.txt
+```

docs/prd/PRD_DATABASE_RECTIFICATION.md

@@ -0,0 +1,796 @@
+# PRD: Database Rectification & Expansion Plan
+
+**Date:** 2026-03-12
+**Triggered by:** [Adversarial Database Audit 2026-03-12](../ADVERSARIAL_DATABASE_AUDIT_2026-03-12.md)
+**Reference:** [DATABASE_REFERENCE.md](../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](../DATABASE_REFERENCE.md#10-prd-adding-new-data)
+- 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
+
+```markdown
+# 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
+
+```bash
+# 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:**
+```python
+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](../DATABASE_REFERENCE.md#10-prd-adding-new-data):
+   - 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:
+
+```python
+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:
+
+```python
+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
+
+```python
+# 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*

scripts/fetch_wiktionary_raw.py

@@ -0,0 +1,201 @@
+#!/usr/bin/env python3
+"""Fetch Wiktionary category members and save as raw JSON for later processing.
+
+Uses curl with retry-after header respect. Designed to handle rate limiting
+gracefully by waiting the specified time between retries.
+
+Iron Rule: All data comes from HTTP API responses.
+
+Usage:
+    python scripts/fetch_wiktionary_raw.py [--language ISO]
+"""
+
+from __future__ import annotations
+
+import argparse
+import io
+import json
+import logging
+import subprocess
+import sys
+import time
+from pathlib import Path
+
+sys.stdout = io.TextIOWrapper(sys.stdout.buffer, encoding="utf-8")
+sys.stderr = io.TextIOWrapper(sys.stderr.buffer, encoding="utf-8")
+
+ROOT = Path(__file__).resolve().parent.parent
+RAW_DIR = ROOT / "data" / "training" / "raw"
+
+logger = logging.getLogger(__name__)
+
+API_URL = "https://en.wiktionary.org/w/api.php"
+USER_AGENT = "PhaiPhon/1.0 (ancient-scripts-datasets; academic research)"
+
+# All categories: (category_name, namespace)
+CATEGORIES = {
+    "cop": ("Coptic_lemmas", 0),
+    "pli": ("Pali_lemmas", 0),
+    "xcl": ("Old_Armenian_lemmas", 0),
+    "ang": ("Old_English_lemmas", 0),
+    "gez": ("Ge%27ez_lemmas", 0),
+    "hbo": ("Hebrew_lemmas", 0),
+    "xht": ("Hattic_lemmas", 0),
+    # Tier 3 + Proto-languages
+    "gem-pro": ("Proto-Germanic_lemmas", 118),
+    "cel-pro": ("Proto-Celtic_lemmas", 118),
+    "urj-pro": ("Proto-Uralic_lemmas", 118),
+    "nci": ("Classical_Nahuatl_lemmas", 0),
+    "sga": ("Old_Irish_lemmas", 0),
+    # Phase 7 additions
+    "pal": ("Middle_Persian_lemmas", 0),
+    "bnt-pro": ("Proto-Bantu_lemmas", 118),
+    "sit-pro": ("Proto-Sino-Tibetan_lemmas", 118),
+    "xtg": ("Gaulish_lemmas", 0),
+    "sog": ("Sogdian_lemmas", 0),
+    "ojp": ("Old_Japanese_lemmas", 0),
+    # Phase 8 P0 additions
+    "sla-pro": ("Proto-Slavic_lemmas", 118),
+    "trk-pro": ("Proto-Turkic_lemmas", 118),
+    "itc-pro": ("Proto-Italic_lemmas", 118),
+    "jpx-pro": ("Proto-Japonic_lemmas", 118),
+    "ira-pro": ("Proto-Iranian_lemmas", 118),
+    # Phase 8 P1 proto-languages
+    "alg-pro": ("Proto-Algonquian_lemmas", 118),
+    "sqj-pro": ("Proto-Albanian_lemmas", 118),
+    "aav-pro": ("Proto-Austroasiatic_lemmas", 118),
+    "poz-pol-pro": ("Proto-Polynesian_lemmas", 118),
+    "tai-pro": ("Proto-Tai_lemmas", 118),
+    "xto-pro": ("Proto-Tocharian_lemmas", 118),
+    "poz-oce-pro": ("Proto-Oceanic_lemmas", 118),
+    "xgn-pro": ("Proto-Mongolic_lemmas", 118),
+    # Phase 8 additional ancient languages
+    "obm": ("Moabite_lemmas", 0),
+    # Batch 3: P2 proto-languages + Iberian
+    "myn-pro": ("Proto-Mayan_lemmas", 118),
+    "afa-pro": ("Proto-Afroasiatic_lemmas", 118),
+    "xib": ("Iberian_lemmas", 0),
+}
+
+
+def fetch_one_page(url: str) -> tuple[str, int]:
+    """Fetch one URL via curl. Returns (body, retry_after_secs)."""
+    result = subprocess.run(
+        ["curl", "-s", "-D", "-",
+         "-H", f"User-Agent: {USER_AGENT}",
+         url],
+        capture_output=True, text=True, timeout=60,
+    )
+    output = result.stdout
+    # Split headers from body
+    parts = output.split("\r\n\r\n", 1)
+    if len(parts) < 2:
+        parts = output.split("\n\n", 1)
+
+    headers = parts[0] if parts else ""
+    body = parts[1] if len(parts) > 1 else ""
+
+    # Check for 429
+    retry_after = 0
+    if "429" in headers.split("\n")[0]:
+        for line in headers.split("\n"):
+            if line.lower().startswith("retry-after:"):
+                try:
+                    retry_after = int(line.split(":", 1)[1].strip())
+                except ValueError:
+                    retry_after = 300  # Default 5 min
+        if retry_after == 0:
+            retry_after = 300
+
+    return body, retry_after
+
+
+def fetch_category(iso: str, category: str, namespace: int = 0) -> list[str]:
+    """Fetch all members of a Wiktionary category, respecting rate limits."""
+    members = []
+    base = (
+        f"action=query&list=categorymembers&cmtitle=Category:{category}"
+        f"&cmtype=page&cmnamespace={namespace}&cmlimit=500&format=json"
+    )
+    extra = ""
+    page = 0
+
+    while True:
+        page += 1
+        url = f"{API_URL}?{base}{extra}"
+
+        body, retry_after = fetch_one_page(url)
+
+        if retry_after > 0:
+            logger.warning("%s: Rate limited. Retry-After=%d seconds (%.1f min). Waiting...",
+                           iso, retry_after, retry_after / 60)
+            time.sleep(retry_after + 5)
+            # Retry after waiting
+            body, retry_after = fetch_one_page(url)
+            if retry_after > 0:
+                logger.error("%s: Still rate limited after waiting. Aborting.", iso)
+                return members
+
+        try:
+            data = json.loads(body)
+        except json.JSONDecodeError:
+            logger.error("%s: Invalid JSON on page %d. Body: %s", iso, page, body[:200])
+            return members
+
+        for m in data.get("query", {}).get("categorymembers", []):
+            members.append(m["title"])
+
+        cont = data.get("continue", {})
+        if "cmcontinue" in cont:
+            extra = f"&cmcontinue={cont['cmcontinue']}"
+            if page % 5 == 0:
+                logger.info("  %s page %d: %d members...", iso, page, len(members))
+            time.sleep(1.5)  # Be nice
+        else:
+            break
+
+    return members
+
+
+def main():
+    parser = argparse.ArgumentParser(description="Fetch Wiktionary category raw data")
+    parser.add_argument("--language", "-l", help="Specific ISO code")
+    args = parser.parse_args()
+
+    logging.basicConfig(
+        level=logging.INFO,
+        format="%(asctime)s %(levelname)s: %(message)s",
+        datefmt="%H:%M:%S",
+    )
+
+    RAW_DIR.mkdir(parents=True, exist_ok=True)
+
+    if args.language:
+        cats = {args.language: CATEGORIES[args.language]}
+    else:
+        cats = CATEGORIES
+
+    for iso, cat_info in cats.items():
+        category, namespace = cat_info
+        raw_path = RAW_DIR / f"wiktionary_category_{iso}.json"
+        if raw_path.exists():
+            with open(raw_path, "r", encoding="utf-8") as f:
+                existing = json.load(f)
+            logger.info("%s: Already cached (%d members). Skipping.", iso, len(existing.get("members", [])))
+            continue
+
+        logger.info("%s: Fetching %s (ns=%d)...", iso, category, namespace)
+        members = fetch_category(iso, category, namespace)
+        logger.info("%s: Got %d members", iso, len(members))
+
+        if members:
+            with open(raw_path, "w", encoding="utf-8") as f:
+                json.dump({"category": category, "members": members}, f, ensure_ascii=False)
+            logger.info("%s: Saved to %s", iso, raw_path)
+
+        # Pause between languages to be polite
+        time.sleep(5)
+
+
+if __name__ == "__main__":
+    main()

scripts/ingest_acd.py

@@ -0,0 +1,307 @@
+#!/usr/bin/env python3
+"""Ingest Proto-Austronesian reconstructed forms from the ACD CLDF dataset.
+
+Source: Austronesian Comparative Dictionary (ACD) — CLDF on GitHub
+URL: https://github.com/lexibank/acd
+License: CC BY 4.0
+Citation: Blust, Trussel & Smith (2023), DOI: 10.5281/zenodo.7737547
+
+The CLDF forms.csv contains reconstructed forms for 42 proto-languages.
+Forms use Blust notation (not IPA) — requires transliteration.
+
+Iron Rule: Data comes from downloaded CSV files. No hardcoded word lists.
+
+Usage:
+    python scripts/ingest_acd.py [--dry-run]
+"""
+
+from __future__ import annotations
+
+import argparse
+import csv
+import io
+import json
+import logging
+import re
+import sys
+import unicodedata
+from pathlib import Path
+
+sys.stdout = io.TextIOWrapper(sys.stdout.buffer, encoding="utf-8")
+sys.stderr = io.TextIOWrapper(sys.stderr.buffer, encoding="utf-8")
+
+ROOT = Path(__file__).resolve().parent.parent
+sys.path.insert(0, str(ROOT / "cognate_pipeline" / "src"))
+sys.path.insert(0, str(ROOT / "scripts"))
+
+from cognate_pipeline.normalise.sound_class import ipa_to_sound_class  # noqa: E402
+
+logger = logging.getLogger(__name__)
+
+LEXICON_DIR = ROOT / "data" / "training" / "lexicons"
+AUDIT_TRAIL_DIR = ROOT / "data" / "training" / "audit_trails"
+RAW_DIR = ROOT / "data" / "training" / "raw"
+
+ACD_DIR = RAW_DIR / "acd_cldf"
+ACD_BASE = "https://raw.githubusercontent.com/lexibank/acd/main/cldf/"
+
+# Blust notation → IPA mapping
+# Reference: Blust (2009) The Austronesian Languages, Chapter 2
+BLUST_TO_IPA = {
+    # Capital letters = special proto-phonemes
+    "C": "ts",      # *C — voiceless dental/alveolar affricate
+    "N": "ŋ",       # *N — velar nasal (sometimes ñ)
+    "R": "ʀ",       # *R — uvular trill or retroflex
+    "S": "s",       # *S — voiceless sibilant
+    "Z": "z",       # *Z — voiced sibilant
+    "H": "h",       # *H — laryngeal
+    "L": "ɬ",       # *L — lateral fricative
+    "T": "t",       # *T — voiceless dental stop
+    "D": "d",       # *D — voiced dental stop
+    # Digraphs
+    "ng": "ŋ",
+    "ny": "ɲ",
+    "nj": "ɲ",
+    # Glottal
+    "q": "ʔ",
+    # Vowels with special values
+    "e": "ə",       # Blust *e = schwa in PAN
+    # Subscript digits (used for homonyms) — remove
+    "₁": "", "₂": "", "₃": "", "₄": "", "₅": "",
+    "₆": "", "₇": "", "₈": "", "₉": "", "₀": "",
+}
+
+
+def blust_to_ipa(form: str) -> str:
+    """Convert Blust notation to approximate IPA."""
+    # Remove reconstruction asterisk
+    form = form.lstrip("*")
+    # Remove parenthetical optional segments
+    form = re.sub(r"\([^)]+\)", "", form)
+
+    # Greedy longest-match transliteration
+    keys = sorted(BLUST_TO_IPA.keys(), key=len, reverse=True)
+    result = []
+    i = 0
+    while i < len(form):
+        matched = False
+        for key in keys:
+            if form[i:i + len(key)] == key:
+                result.append(BLUST_TO_IPA[key])
+                i += len(key)
+                matched = True
+                break
+        if not matched:
+            if form[i] not in "- ":  # skip hyphens and spaces
+                result.append(form[i])
+            i += 1
+    return "".join(result)
+
+
+def download_if_needed():
+    """Download ACD CLDF files if not cached."""
+    import urllib.request
+
+    ACD_DIR.mkdir(parents=True, exist_ok=True)
+    for fname in ("forms.csv", "languages.csv", "cognatesets.csv"):
+        local = ACD_DIR / fname
+        if local.exists():
+            logger.info("Cached: %s (%d bytes)", fname, local.stat().st_size)
+            continue
+        url = ACD_BASE + fname
+        logger.info("Downloading %s ...", url)
+        req = urllib.request.Request(url, headers={
+            "User-Agent": "PhaiPhon/1.0 (ancient-scripts-datasets)"
+        })
+        with urllib.request.urlopen(req, timeout=120) as resp:
+            data = resp.read()
+        with open(local, "wb") as f:
+            f.write(data)
+        logger.info("Downloaded %s (%d bytes)", fname, len(data))
+
+
+def load_proto_languages():
+    """Load language metadata to identify proto-languages."""
+    lang_path = ACD_DIR / "languages.csv"
+    protos = {}
+    with open(lang_path, "r", encoding="utf-8") as f:
+        for row in csv.DictReader(f):
+            name = row.get("Name", "")
+            lid = row.get("ID", "")
+            # Proto-languages have names starting with "Proto-"
+            if name.startswith("Proto-"):
+                protos[lid] = name
+    return protos
+
+
+def extract_proto_forms():
+    """Extract reconstructed forms from ACD CLDF."""
+    protos = load_proto_languages()
+    logger.info("Found %d proto-languages in ACD", len(protos))
+
+    forms_path = ACD_DIR / "forms.csv"
+    entries = {}  # (proto_lang, form) -> {gloss, ...}
+
+    with open(forms_path, "r", encoding="utf-8") as f:
+        for row in csv.DictReader(f):
+            lang_id = row.get("Language_ID", "")
+            if lang_id not in protos:
+                continue
+
+            form = row.get("Form", "").strip()
+            value = row.get("Value", "").strip()
+            gloss = row.get("Description", "").strip()
+
+            if not form:
+                continue
+            # Use Form (cleaned) rather than Value (has optional segments)
+            word = form
+
+            # Strip infix angle brackets: C<in>aliS → CinaliS
+            word = re.sub(r"<([^>]+)>", r"\1", word)
+            # Strip parenthetical optional segments: (q)uNah → uNah
+            word = re.sub(r"\([^)]+\)", "", word)
+            # Remove leading asterisk
+            word = word.lstrip("*")
+            # Remove subscript digits (homonym markers)
+            word = re.sub(r"[₀₁₂₃₄₅₆₇₈₉]", "", word)
+            # Remove tilde variants: keep only first form
+            if " ~ " in word:
+                word = word.split(" ~ ")[0]
+            # Remove hyphens (prefix/suffix markers)
+            word = word.strip("-").strip()
+
+            # NFC normalize
+            word = unicodedata.normalize("NFC", word)
+
+            key = (lang_id, word)
+            if key not in entries:
+                entries[key] = {
+                    "word": word,
+                    "gloss": gloss,
+                    "proto_lang": protos[lang_id],
+                    "proto_lang_id": lang_id,
+                }
+
+    return entries
+
+
+def load_existing_words(tsv_path: Path) -> set[str]:
+    """Load existing Word column values."""
+    existing = set()
+    if tsv_path.exists():
+        with open(tsv_path, "r", encoding="utf-8") as f:
+            for line in f:
+                if line.startswith("Word\t"):
+                    continue
+                word = line.split("\t")[0]
+                existing.add(word)
+    return existing
+
+
+def main():
+    parser = argparse.ArgumentParser(description="Ingest ACD Proto-Austronesian")
+    parser.add_argument("--dry-run", action="store_true")
+    args = parser.parse_args()
+
+    logging.basicConfig(
+        level=logging.INFO,
+        format="%(asctime)s %(levelname)s: %(message)s",
+        datefmt="%H:%M:%S",
+    )
+
+    download_if_needed()
+
+    # We ingest all proto-forms into a single map.tsv (Proto-Austronesian family)
+    tsv_path = LEXICON_DIR / "map.tsv"
+    existing = load_existing_words(tsv_path)
+    logger.info("Existing Proto-Austronesian entries: %d", len(existing))
+
+    entries = extract_proto_forms()
+    logger.info("ACD proto-forms: %d", len(entries))
+
+    # Count by proto-language
+    by_lang = {}
+    for (lid, _), info in entries.items():
+        name = info["proto_lang"]
+        by_lang[name] = by_lang.get(name, 0) + 1
+    for name, count in sorted(by_lang.items(), key=lambda x: -x[1])[:10]:
+        logger.info("  %s: %d", name, count)
+
+    # Process
+    new_entries = []
+    audit_trail = []
+    skipped = 0
+
+    for (lid, word), info in sorted(entries.items()):
+        clean_word = word.strip()
+        if not clean_word or len(clean_word) < 2 or len(clean_word) > 50:
+            skipped += 1
+            continue
+
+        if clean_word in existing:
+            skipped += 1
+            continue
+
+        # Convert Blust notation to IPA
+        ipa = blust_to_ipa(word)
+        if not ipa:
+            ipa = clean_word
+
+        try:
+            sca = ipa_to_sound_class(ipa)
+        except Exception:
+            sca = ""
+
+        new_entries.append({
+            "word": clean_word,
+            "ipa": ipa,
+            "sca": sca,
+        })
+        existing.add(clean_word)
+
+        audit_trail.append({
+            "word": clean_word,
+            "raw_form": word,
+            "ipa": ipa,
+            "gloss": info["gloss"],
+            "proto_lang": info["proto_lang"],
+            "source": "acd",
+        })
+
+    logger.info("New: %d, Skipped: %d", len(new_entries), skipped)
+
+    if args.dry_run:
+        print(f"\nDRY RUN: ACD Proto-Austronesian Ingestion:")
+        print(f"  ACD proto-forms:    {len(entries)}")
+        print(f"  Existing:           {len(existing) - len(new_entries)}")
+        print(f"  New:                {len(new_entries)}")
+        print(f"  Total:              {len(existing)}")
+        return
+
+    if new_entries:
+        LEXICON_DIR.mkdir(parents=True, exist_ok=True)
+        if not tsv_path.exists():
+            with open(tsv_path, "w", encoding="utf-8") as f:
+                f.write("Word\tIPA\tSCA\tSource\tConcept_ID\tCognate_Set_ID\n")
+
+        with open(tsv_path, "a", encoding="utf-8") as f:
+            for e in new_entries:
+                f.write(f"{e['word']}\t{e['ipa']}\t{e['sca']}\tacd\t-\t-\n")
+
+    if audit_trail:
+        AUDIT_TRAIL_DIR.mkdir(parents=True, exist_ok=True)
+        audit_path = AUDIT_TRAIL_DIR / "acd_ingest_map.jsonl"
+        with open(audit_path, "w", encoding="utf-8") as f:
+            for r in audit_trail:
+                f.write(json.dumps(r, ensure_ascii=False) + "\n")
+
+    print(f"\nACD Proto-Austronesian Ingestion:")
+    print(f"  ACD proto-forms:    {len(entries)}")
+    print(f"  Existing:           {len(existing) - len(new_entries)}")
+    print(f"  New:                {len(new_entries)}")
+    print(f"  Total:              {len(existing)}")
+
+
+if __name__ == "__main__":
+    main()