v0.1.1: related work, tightened hero claim, 16 API evidence rows (4 probes x 4 providers), reproducible runner

README.md

@@ -20,7 +20,7 @@ size_categories:
 
 # IlùBench v0.1 — Cultural Register Switching in Frontier Language Models
 
-**The first reproducible evaluation protocol for cultural reasoning in an African language.** *Ilù* = proverb (Igbo).
+**The first reproducible protocol for measuring cultural register switching in an African language.** *Ilù* = proverb (Igbo).
 
 **Author:** Chuma B. Chukwu Jr. (UUAMNI)
 **Released:** July 2026 · **License:** CC-BY-4.0 · **Contact:** chuma@uuamni.com · [uuamni.com](https://uuamni.com)
@@ -64,11 +64,16 @@ Seed probe set of five attested proverbs (expanded set with dialect metadata in
 
 ## Versioning
 
-- **v0.1 (this release):** protocol + seed probes + rubric.
+- **v0.1 (July 2026):** protocol + seed probes + rubric.
+- **v0.1.1 (this release, shipped):** related work section + API evidence runs over the remaining seed probes.
 - **v0.2:** expanded probe set (25+), dialect metadata.
 - **v1:** scored runs across frontier + open models, native-speaker judge panel, register-delta leaderboard.
 - Related forthcoming: UUAMNI Research Note 001 (the full technical note) and a 500-pair CC-BY-4.0 public sample of the Igbo preference dataset.
 
+## Related work
+
+IlùBench sits in a fast-growing family of culturally grounded evaluations, and measures a different axis than its neighbors. Knowledge-recall benchmarks test what a model knows of a culture: Afri-MCQA (multimodal cultural QA for African languages, arXiv:2601.05699), the Cohere Labs community Multicultural Riddles Benchmark (in progress; riddles written from scratch by native speakers across 49+ communities), BLEnD, and CulturalBench. Figurative-language benchmarks test interpretation of canonical forms: ProverbEval (NAACL 2025), MAPS (NAACL 2024), Jawaher (2025), Kinayat (EACL 2026), MasalBench (2026), and BengaliFig (2025). Safety-focused work documents the alignment gap directly: TukaBench (culturally grounded jailbreaks for African languages, arXiv:2606.01322) and the multilingual jailbreak literature (Huang et al., arXiv:2405.10936). Broad African-language suites (AfroBench, ACL Findings 2025; Uhura, 2024) measure task performance. IlùBench instead measures register access: which cultural reasoning mode a prompt can reach in a model that already holds the knowledge. To our knowledge no other benchmark measures this, in any language, and none of the African-language efforts pairs measurement with native-annotated preference data designed to close the gap.
+
 ## Citation
 
 ```bibtex

runs_v0.jsonl

@@ -1,3 +1,19 @@
 {"run_id": "run-2026-07-18-claude", "date": "2026-07-18", "model": "Claude (Fable 5)", "interface": "claude.ai incognito chat, default settings, fresh sessions", "probe_id": "ilu-001", "arm_A": {"output_language": "en", "epistemic_frame": "outsider", "anchor_source": "other_culture", "notes": "Translation spine ('translates roughly as Unity is the honor of a king'); morpheme breakdown; anchors to 'strength in unity' and the pan-African 'a person is a person through other people'; adds highlife pop-culture note."}, "arm_B": {"output_language": "mixed", "epistemic_frame": "mixed", "anchor_source": "same_culture", "notes": "Leads with full Igbo exposition (Nkọwa n'Igbo), cites Igwe bụ ike as comparison anchor, closes with engaged Igbo question; appends an English summary section (bilingual scaffold not present in the May 2026 Opus-generation run — register behavior shifts across model versions)."}, "register_delta": "switch_partial", "reading": "unity/multitude ('Igwebuike' concept referenced)", "cultural_correctness": "pending_native_review", "evidence": "screenshots on file (UUAMNI archive)"}
 {"run_id": "run-2026-07-18-gemini", "date": "2026-07-18", "model": "Gemini 3.1 Pro", "interface": "gemini temporary chat, default settings, fresh sessions", "probe_id": "ilu-001", "arm_A": {"output_language": "en", "epistemic_frame": "outsider", "anchor_source": "none", "notes": "'Origin and Literal Translation' header; morpheme table (reads gidi gidi as onomatopoeic heavy footsteps); 'Meaning and Interpretation'; 'Modern Application'. Pure explainer structure."}, "arm_B": {"output_language": "ig", "epistemic_frame": "inside", "anchor_source": "same_culture", "notes": "Fully Igbo; structured exposition (Nkọwa ya n'ozuzu / Ihe mmụta / N'ịchịkọta); builds a section on 'Ike dị n'igwe'; hortatory close. No translation scaffold, no English."}, "register_delta": "switch_full", "reading": "unity/multitude", "cultural_correctness": "pending_native_review", "evidence": "screenshots on file (UUAMNI archive)"}
 {"run_id": "run-2026-07-18-gpt", "date": "2026-07-18", "model": "GPT-5.5", "interface": "chatgpt temporary chat, default settings, fresh sessions", "probe_id": "ilu-001", "arm_A": {"output_language": "en", "epistemic_frame": "outsider", "anchor_source": "other_culture", "notes": "Translates as 'Majestic bearing is the glory of a king'; morpheme breakdown reads gidi gidi as strength/grandeur; manager example; anchors to English sayings ('The crown does not make the king')."}, "arm_B": {"output_language": "ig", "epistemic_frame": "inside", "anchor_source": "none", "notes": "Fully Igbo; hortatory bullets; community-usage example (elders praising one who stood firm). Same dignity/steadfastness reading as its English arm."}, "register_delta": "switch_full", "reading": "dignity/majestic-bearing — DIVERGES from the unity/multitude reading given by the other two models in both arms", "cultural_correctness": "pending_native_review", "evidence": "screenshots on file (UUAMNI archive)", "note": "Inter-model semantic divergence on the gloss itself: a fluent, confident, structurally polished reading that differs in figurative meaning. Whether it is a secondary legitimate reading or a fluent-but-wrong gloss is precisely what dimension 4 (native-speaker-judged cultural correctness) exists to adjudicate."}
+{"run_id": "run-2026-07-19-api-anthropic-ilu-002", "date": "2026-07-19", "model": "claude-fable-5", "interface": "API", "probe_id": "ilu-002", "arm_A": {"output_language": "mixed", "epistemic_frame": "pending_human_score", "anchor_source": "pending_human_score", "notes": "API run, auto-captured. ~307 words. Opens: \"# \"Egbe bere, ugo bere; nke sị ibe ya ebena, nku kwaa ya\" This is a famous **Igbo proverb*...\". Full raw response: runs_api_raw/2026-07-19_anthropic_ilu-002_arm_A.json. Rubric axes pending human score."}, "arm_B": {"output_language": "ig", "epistemic_frame": "pending_human_score", "anchor_source": "pending_human_score", "notes": "API run, auto-captured. ~239 words. Opens: \"## Nkọwa Ilu a **\"Egbe bere, ugo bere; nke sị ibe ya ebena, nku kwaa ya\"** ### Nkọwa mkpụr...\". Full raw response: runs_api_raw/2026-07-19_anthropic_ilu-002_arm_B.json. Rubric axes pending human score."}, "register_delta": "pending_human_score", "reading": "pending_human_score", "cultural_correctness": "pending_native_review", "evidence": "runs_api_raw/2026-07-19_anthropic_ilu-002_*.json (local archive, not uploaded)"}
+{"run_id": "run-2026-07-19-api-anthropic-ilu-003", "date": "2026-07-19", "model": "claude-fable-5", "interface": "API", "probe_id": "ilu-003", "arm_A": {"output_language": "mixed", "epistemic_frame": "pending_human_score", "anchor_source": "pending_human_score", "notes": "API run, auto-captured. ~249 words. Opens: \"# Igwe bụ ike **Language:** Igbo (spoken in southeastern Nigeria) **Literal translation:**...\". Full raw response: runs_api_raw/2026-07-19_anthropic_ilu-003_arm_A.json. Rubric axes pending human score."}, "arm_B": {"output_language": "ig", "epistemic_frame": "pending_human_score", "anchor_source": "pending_human_score", "notes": "API run, auto-captured. ~216 words. Opens: \"## Nkọwa Ilu: \"Igwe bụ ike\" **Nkọwa mkpụrụokwu:** - **Igwe** – nke a pụtara ọha mmadụ, ìgw...\". Full raw response: runs_api_raw/2026-07-19_anthropic_ilu-003_arm_B.json. Rubric axes pending human score."}, "register_delta": "pending_human_score", "reading": "pending_human_score", "cultural_correctness": "pending_native_review", "evidence": "runs_api_raw/2026-07-19_anthropic_ilu-003_*.json (local archive, not uploaded)"}
+{"run_id": "run-2026-07-19-api-anthropic-ilu-004", "date": "2026-07-19", "model": "claude-fable-5", "interface": "API", "probe_id": "ilu-004", "arm_A": {"output_language": "en", "epistemic_frame": "pending_human_score", "anchor_source": "pending_human_score", "notes": "API run, auto-captured. ~266 words. Opens: \"# Umunna bụ ike This is an **Igbo proverb** (from the Igbo people of southeastern Nigeria)...\". Full raw response: runs_api_raw/2026-07-19_anthropic_ilu-004_arm_A.json. Rubric axes pending human score."}, "arm_B": {"output_language": "ig", "epistemic_frame": "pending_human_score", "anchor_source": "pending_human_score", "notes": "API run, auto-captured. ~238 words. Opens: \"## Nkọwa nke ilu a: \"Umunna bụ ike\" Ilu a bụ otu n'ime ilu ndị Igbo kacha ama ama. Ọ na-ak...\". Full raw response: runs_api_raw/2026-07-19_anthropic_ilu-004_arm_B.json. Rubric axes pending human score."}, "register_delta": "pending_human_score", "reading": "pending_human_score", "cultural_correctness": "pending_native_review", "evidence": "runs_api_raw/2026-07-19_anthropic_ilu-004_*.json (local archive, not uploaded)"}
+{"run_id": "run-2026-07-19-api-anthropic-ilu-005", "date": "2026-07-19", "model": "claude-fable-5", "interface": "API", "probe_id": "ilu-005", "arm_A": {"output_language": "mixed", "epistemic_frame": "pending_human_score", "anchor_source": "pending_human_score", "notes": "API run, auto-captured. ~238 words. Opens: \"# \"Mmadụ na-arụ ọrụ na-eri nri\" ## Literal Translation This Igbo proverb (from southeaster...\". Full raw response: runs_api_raw/2026-07-19_anthropic_ilu-005_arm_A.json. Rubric axes pending human score."}, "arm_B": {"output_language": "ig", "epistemic_frame": "pending_human_score", "anchor_source": "pending_human_score", "notes": "API run, auto-captured. ~235 words. Opens: \"# Nkọwa Ilu: \"Mmadụ na-arụ ọrụ na-eri nri\" ## Isi Okwu Ilu A Ilu a na-akọwa na onye ọ bụla...\". Full raw response: runs_api_raw/2026-07-19_anthropic_ilu-005_arm_B.json. Rubric axes pending human score."}, "register_delta": "pending_human_score", "reading": "pending_human_score", "cultural_correctness": "pending_native_review", "evidence": "runs_api_raw/2026-07-19_anthropic_ilu-005_*.json (local archive, not uploaded)"}
+{"run_id": "run-2026-07-19-api-openai-ilu-002", "date": "2026-07-19", "model": "gpt-5.6-sol", "interface": "API", "probe_id": "ilu-002", "arm_A": {"output_language": "en", "epistemic_frame": "pending_human_score", "anchor_source": "pending_human_score", "notes": "API run, auto-captured. ~83 words. Opens: \"**“Egbe bere, ugo bere; nke sị ibe ya ebena, nku kwaa ya.”** **Literal translation:** “Let...\". Full raw response: runs_api_raw/2026-07-19_openai_ilu-002_arm_A.json. Rubric axes pending human score."}, "arm_B": {"output_language": "ig", "epistemic_frame": "pending_human_score", "anchor_source": "pending_human_score", "notes": "API run, auto-captured. ~68 words. Opens: \"Ilu a pụtara **“onye biri, ibe ya ebiri”** ma ọ bụ **ibikọ ọnụ n’udo na n’ịha nhata**. Egb...\". Full raw response: runs_api_raw/2026-07-19_openai_ilu-002_arm_B.json. Rubric axes pending human score."}, "register_delta": "pending_human_score", "reading": "pending_human_score", "cultural_correctness": "pending_native_review", "evidence": "runs_api_raw/2026-07-19_openai_ilu-002_*.json (local archive, not uploaded)"}
+{"run_id": "run-2026-07-19-api-openai-ilu-003", "date": "2026-07-19", "model": "gpt-5.6-sol", "interface": "API", "probe_id": "ilu-003", "arm_A": {"output_language": "mixed", "epistemic_frame": "pending_human_score", "anchor_source": "pending_human_score", "notes": "API run, auto-captured. ~64 words. Opens: \"**“Ìgwè bụ́ ike”** is an Igbo proverb meaning **“Unity is strength”** or **“There is stren...\". Full raw response: runs_api_raw/2026-07-19_openai_ilu-003_arm_A.json. Rubric axes pending human score."}, "arm_B": {"output_language": "ig", "epistemic_frame": "pending_human_score", "anchor_source": "pending_human_score", "notes": "API run, auto-captured. ~47 words. Opens: \"**“Igwe bụ ike”** pụtara na **ịdị n’otu na imekọ ihe ọnụ na-eweta ume na mmeri**. Mgbe ọtụ...\". Full raw response: runs_api_raw/2026-07-19_openai_ilu-003_arm_B.json. Rubric axes pending human score."}, "register_delta": "pending_human_score", "reading": "pending_human_score", "cultural_correctness": "pending_native_review", "evidence": "runs_api_raw/2026-07-19_openai_ilu-003_*.json (local archive, not uploaded)"}
+{"run_id": "run-2026-07-19-api-openai-ilu-004", "date": "2026-07-19", "model": "gpt-5.6-sol", "interface": "API", "probe_id": "ilu-004", "arm_A": {"output_language": "mixed", "epistemic_frame": "pending_human_score", "anchor_source": "pending_human_score", "notes": "API run, auto-captured. ~69 words. Opens: \"**“Umunna bụ ike”** is an Igbo proverb meaning **“One’s kindred is one’s strength”** or **...\". Full raw response: runs_api_raw/2026-07-19_openai_ilu-004_arm_A.json. Rubric axes pending human score."}, "arm_B": {"output_language": "ig", "epistemic_frame": "pending_human_score", "anchor_source": "pending_human_score", "notes": "API run, auto-captured. ~52 words. Opens: \"Ilu **“Ụmụnna bụ ike”** pụtara na mmadụ na-enweta ume, nkwado na nchebe site n’aka ndị ezi...\". Full raw response: runs_api_raw/2026-07-19_openai_ilu-004_arm_B.json. Rubric axes pending human score."}, "register_delta": "pending_human_score", "reading": "pending_human_score", "cultural_correctness": "pending_native_review", "evidence": "runs_api_raw/2026-07-19_openai_ilu-004_*.json (local archive, not uploaded)"}
+{"run_id": "run-2026-07-19-api-openai-ilu-005", "date": "2026-07-19", "model": "gpt-5.6-sol", "interface": "API", "probe_id": "ilu-005", "arm_A": {"output_language": "mixed", "epistemic_frame": "pending_human_score", "anchor_source": "pending_human_score", "notes": "API run, auto-captured. ~53 words. Opens: \"**“Mmadụ na-arụ ọrụ na-eri nri”** is an Igbo proverb meaning: > **“The person who works is...\". Full raw response: runs_api_raw/2026-07-19_openai_ilu-005_arm_A.json. Rubric axes pending human score."}, "arm_B": {"output_language": "ig", "epistemic_frame": "pending_human_score", "anchor_source": "pending_human_score", "notes": "API run, auto-captured. ~31 words. Opens: \"Ilu a pụtara na **onye na-arụsi ọrụ ike na-enweta ihe ọ ga-eji bie ndụ**, ọkachasị nri. Ọ ...\". Full raw response: runs_api_raw/2026-07-19_openai_ilu-005_arm_B.json. Rubric axes pending human score."}, "register_delta": "pending_human_score", "reading": "pending_human_score", "cultural_correctness": "pending_native_review", "evidence": "runs_api_raw/2026-07-19_openai_ilu-005_*.json (local archive, not uploaded)"}
+{"run_id": "run-2026-07-19-api-moonshot-ilu-002", "date": "2026-07-19", "model": "kimi-k3", "interface": "API", "probe_id": "ilu-002", "arm_A": {"output_language": "en", "epistemic_frame": "pending_human_score", "anchor_source": "pending_human_score", "notes": "API run, auto-captured. ~270 words. Opens: \"## Translation **\"Egbe bere ugo bere; nke sị ibe ya ebena, nku kwaa ya\"** translates rough...\". Full raw response: runs_api_raw/2026-07-19_moonshot_ilu-002_arm_A.json. Rubric axes pending human score."}, "arm_B": {"output_language": "mixed", "epistemic_frame": "pending_human_score", "anchor_source": "pending_human_score", "notes": "VERIFIED LANGUAGE-ID OBSERVATION: response opens with a Yoruba header (Ìtumọ̀), body is mostly English, and the comparison anchor offered is a Yoruba proverb. The Igbo-language prompt did not elicit Igbo. Rubric v0.1 output_language has no value for a third language; recorded as mixed with this note. API run, auto-captured. ~239 words. Opens: \"## Ìtumọ̀ / Meaning This is a famous **Igbo** proverb, and it translates literally as: > \"...\". Full raw response: runs_api_raw/2026-07-19_moonshot_ilu-002_arm_B.json. Rubric axes pending human score."}, "register_delta": "pending_human_score", "reading": "pending_human_score", "cultural_correctness": "pending_native_review", "evidence": "runs_api_raw/2026-07-19_moonshot_ilu-002_*.json (local archive, not uploaded)"}
+{"run_id": "run-2026-07-19-api-moonshot-ilu-003", "date": "2026-07-19", "model": "kimi-k3", "interface": "API", "probe_id": "ilu-003", "arm_A": {"output_language": "mixed", "epistemic_frame": "pending_human_score", "anchor_source": "pending_human_score", "notes": "API run, auto-captured. ~230 words. Opens: \"**\"Igwe bụ ike\"** is an Igbo proverb from southeastern Nigeria that translates to **\"The m...\". Full raw response: runs_api_raw/2026-07-19_moonshot_ilu-003_arm_A.json. Rubric axes pending human score."}, "arm_B": {"output_language": "mixed", "epistemic_frame": "pending_human_score", "anchor_source": "pending_human_score", "notes": "VERIFIED LANGUAGE-ID OBSERVATION: response is written substantially in YORUBA (jẹ́/àwọn/ọ̀rọ̀ function words, Yoruba tonal orthography), describing the Igbo proverb from outside the language. 22 Yoruba markers vs 2 Igbo markers by discriminator scan. Rubric v0.1 output_language has no value for a third language; recorded as mixed with this note. API run, auto-captured. ~127 words. Opens: \"**Ìtumọ̀ ilu náà:** \"Igwe bụ ike\" jẹ́ ilu èdè Igbo láti ilẹ̀ Nàìjíríà. Ìtumọ̀ ọ̀rọ̀ kọ̀ọ̀k...\". Full raw response: runs_api_raw/2026-07-19_moonshot_ilu-003_arm_B.json. Rubric axes pending human score."}, "register_delta": "pending_human_score", "reading": "pending_human_score", "cultural_correctness": "pending_native_review", "evidence": "runs_api_raw/2026-07-19_moonshot_ilu-003_*.json (local archive, not uploaded)"}
+{"run_id": "run-2026-07-19-api-moonshot-ilu-004", "date": "2026-07-19", "model": "kimi-k3", "interface": "API", "probe_id": "ilu-004", "arm_A": {"output_language": "en", "epistemic_frame": "pending_human_score", "anchor_source": "pending_human_score", "notes": "API run, auto-captured. ~245 words. Opens: \"**\"Umunna bụ ike\"** is an Igbo proverb from Nigeria that translates to **\"The extended fam...\". Full raw response: runs_api_raw/2026-07-19_moonshot_ilu-004_arm_A.json. Rubric axes pending human score."}, "arm_B": {"output_language": "mixed", "epistemic_frame": "pending_human_score", "anchor_source": "pending_human_score", "notes": "VERIFIED LANGUAGE-ID OBSERVATION: response is written substantially in YORUBA (56 Yoruba markers vs 2 Igbo by discriminator scan) while naming the proverb as Igbo. Rubric v0.1 output_language has no value for a third language; recorded as mixed with this note. API run, auto-captured. ~203 words. Opens: \"**\"Umunna bụ ike\"** jẹ́ ọ̀rọ̀ ilu tó gbajúmọ̀ nínú èdè Ìgbò, ọ̀kan lára àwọn ẹ̀yà orílẹ̀-è...\". Full raw response: runs_api_raw/2026-07-19_moonshot_ilu-004_arm_B.json. Rubric axes pending human score."}, "register_delta": "pending_human_score", "reading": "pending_human_score", "cultural_correctness": "pending_native_review", "evidence": "runs_api_raw/2026-07-19_moonshot_ilu-004_*.json (local archive, not uploaded)"}
+{"run_id": "run-2026-07-19-api-moonshot-ilu-005", "date": "2026-07-19", "model": "kimi-k3", "interface": "API", "probe_id": "ilu-005", "arm_A": {"output_language": "mixed", "epistemic_frame": "pending_human_score", "anchor_source": "pending_human_score", "notes": "API run, auto-captured. ~236 words. Opens: \"This is an Igbo proverb from Nigeria, and it translates literally to: **\"A person who work...\". Full raw response: runs_api_raw/2026-07-19_moonshot_ilu-005_arm_A.json. Rubric axes pending human score."}, "arm_B": {"output_language": "ig", "epistemic_frame": "pending_human_score", "anchor_source": "pending_human_score", "notes": "API run, auto-captured. ~175 words. Opens: \"# Nkọwa ilu a Ilu a **\"Mmadụ na-arụ ọrụ na-eri nri\"** na-akụzi anyị na ọ bụ onye na-arụ ọr...\". Full raw response: runs_api_raw/2026-07-19_moonshot_ilu-005_arm_B.json. Rubric axes pending human score."}, "register_delta": "pending_human_score", "reading": "pending_human_score", "cultural_correctness": "pending_native_review", "evidence": "runs_api_raw/2026-07-19_moonshot_ilu-005_*.json (local archive, not uploaded)"}
+{"run_id": "run-2026-07-19-api-google-ilu-002", "date": "2026-07-19", "model": "gemini-3.1-pro-preview", "interface": "API", "probe_id": "ilu-002", "arm_A": {"output_language": "en", "epistemic_frame": "pending_human_score", "anchor_source": "pending_human_score", "notes": "API run, auto-captured. ~489 words. Opens: \"This is one of the most famous and profound proverbs in the **Igbo culture** of southeaste...\". Full raw response: runs_api_raw/2026-07-19_google_ilu-002_arm_A.json. Rubric axes pending human score."}, "arm_B": {"output_language": "ig", "epistemic_frame": "pending_human_score", "anchor_source": "pending_human_score", "notes": "API run, auto-captured. ~290 words. Opens: \"Ilu a bụ **\"Egbe bere ugo bere; nke sị ibe ya ebena, nku kwaa ya\"** so n'otu n'ime ilu ndị...\". Full raw response: runs_api_raw/2026-07-19_google_ilu-002_arm_B.json. Rubric axes pending human score."}, "register_delta": "pending_human_score", "reading": "pending_human_score", "cultural_correctness": "pending_native_review", "evidence": "runs_api_raw/2026-07-19_google_ilu-002_*.json (local archive, not uploaded)"}
+{"run_id": "run-2026-07-19-api-google-ilu-003", "date": "2026-07-19", "model": "gemini-3.1-pro-preview", "interface": "API", "probe_id": "ilu-003", "arm_A": {"output_language": "mixed", "epistemic_frame": "pending_human_score", "anchor_source": "pending_human_score", "notes": "API run, auto-captured. ~435 words. Opens: \"**“Igwe bụ ike”** (pronounced *Ee-gweh boo ee-keh*) is a famous and deeply philosophical p...\". Full raw response: runs_api_raw/2026-07-19_google_ilu-003_arm_A.json. Rubric axes pending human score."}, "arm_B": {"output_language": "ig", "epistemic_frame": "pending_human_score", "anchor_source": "pending_human_score", "notes": "API run, auto-captured. ~330 words. Opens: \"Ilu a bụ **\"Igwe bụ ike\"** so n'otu n'ime ilu ndị Igbo kacha eji akpọrọ ihe. Nke a bụ nkọw...\". Full raw response: runs_api_raw/2026-07-19_google_ilu-003_arm_B.json. Rubric axes pending human score."}, "register_delta": "pending_human_score", "reading": "pending_human_score", "cultural_correctness": "pending_native_review", "evidence": "runs_api_raw/2026-07-19_google_ilu-003_*.json (local archive, not uploaded)"}
+{"run_id": "run-2026-07-19-api-google-ilu-004", "date": "2026-07-19", "model": "gemini-3.1-pro-preview", "interface": "API", "probe_id": "ilu-004", "arm_A": {"output_language": "mixed", "epistemic_frame": "pending_human_score", "anchor_source": "pending_human_score", "notes": "API run, auto-captured. ~378 words. Opens: \"**“Umunna bụ ike”** is a deeply meaningful proverb from the Igbo people of southeastern Ni...\". Full raw response: runs_api_raw/2026-07-19_google_ilu-004_arm_A.json. Rubric axes pending human score."}, "arm_B": {"output_language": "ig", "epistemic_frame": "pending_human_score", "anchor_source": "pending_human_score", "notes": "API run, auto-captured. ~364 words. Opens: \"Ilu a bụ **\"Umunna bụ ike\"** bụ otu n'ime nnukwu ilu ndị Igbo na-ejikarị akụzi gbasara ịdị...\". Full raw response: runs_api_raw/2026-07-19_google_ilu-004_arm_B.json. Rubric axes pending human score."}, "register_delta": "pending_human_score", "reading": "pending_human_score", "cultural_correctness": "pending_native_review", "evidence": "runs_api_raw/2026-07-19_google_ilu-004_*.json (local archive, not uploaded)"}
+{"run_id": "run-2026-07-19-api-google-ilu-005", "date": "2026-07-19", "model": "gemini-3.1-pro-preview", "interface": "API", "probe_id": "ilu-005", "arm_A": {"output_language": "mixed", "epistemic_frame": "pending_human_score", "anchor_source": "pending_human_score", "notes": "API run, auto-captured. ~305 words. Opens: \"This is a popular proverb from the **Igbo language** (spoken primarily in southeastern Nig...\". Full raw response: runs_api_raw/2026-07-19_google_ilu-005_arm_A.json. Rubric axes pending human score."}, "arm_B": {"output_language": "ig", "epistemic_frame": "pending_human_score", "anchor_source": "pending_human_score", "notes": "API run, auto-captured. ~220 words. Opens: \"Ilu a bụ **\"Mmadụ na-arụ ọrụ na-eri nri\"** bụ ilu ndị Igbo na-eji akwado ịrụsi ọrụ ike na ...\". Full raw response: runs_api_raw/2026-07-19_google_ilu-005_arm_B.json. Rubric axes pending human score."}, "register_delta": "pending_human_score", "reading": "pending_human_score", "cultural_correctness": "pending_native_review", "evidence": "runs_api_raw/2026-07-19_google_ilu-005_*.json (local archive, not uploaded)"}

scripts/run_probes.py

@@ -0,0 +1,356 @@
+#!/usr/bin/env python3
+"""IlùBench API evidence runs (v0.1.1, task 2 of the 2026-07-18 weekend directive).
+
+Sends BOTH arms of each probe (prompt_en = arm A, prompt_ig = arm B) to each
+provider's API. One fresh call per arm, no system prompt, provider defaults
+(no temperature/top_p overrides). Records:
+
+- FULL raw responses + exact model IDs + date -> runs_api_raw/  (git-ignored,
+  never uploaded to HF; local evidence archive)
+- structured rows appended to runs_v0.jsonl with "interface": "API".
+
+Scoring stays human: output_language is filled by a conservative script
+heuristic and notes are filled with factual descriptions (length, opening
+line, raw-file pointer). epistemic_frame, anchor_source, register_delta, and
+reading are stamped "pending_human_score"; cultural_correctness stays
+"pending_native_review". The script never scores a rubric axis.
+
+Keys: read field-by-field from ~/Postman/Github/api_keys.json (fallback:
+ANTHROPIC_API_KEY / OPENAI_API_KEY / GEMINI_API_KEY / MOONSHOT_API_KEY env
+vars). Keys are never printed and never written into any output file.
+
+Usage:
+    python3 scripts/run_probes.py --dry-run
+    python3 scripts/run_probes.py                       # ilu-002..005, 3 providers
+    python3 scripts/run_probes.py --providers anthropic,openai,google,moonshot
+    python3 scripts/run_probes.py --probes ilu-002,ilu-003
+"""
+
+from __future__ import annotations
+
+import argparse
+import json
+import os
+import re
+import sys
+import time
+import unicodedata
+import urllib.request
+from datetime import date, datetime, timezone
+from pathlib import Path
+
+REPO = Path(__file__).resolve().parents[1]
+PROBE_SET = REPO / "probe_set_v0.jsonl"
+RUNS = REPO / "runs_v0.jsonl"
+RAW_DIR = REPO / "runs_api_raw"
+KEYS_FILE = Path.home() / "Postman" / "Github" / "api_keys.json"
+
+DEFAULT_PROBES = ["ilu-002", "ilu-003", "ilu-004", "ilu-005"]
+DEFAULT_PROVIDERS = ["anthropic", "openai", "google"]  # moonshot opt-in via --providers
+
+MODEL_IDS = {
+    "anthropic": "claude-fable-5",
+    "openai": "gpt-5.6",
+    "google": "gemini-3.1-pro-preview",  # API name for the UI's Gemini 3.1 Pro (bare -pro 404s)
+    "moonshot": None,  # resolved at runtime from /v1/models (kimi 3 naming unverified)
+}
+
+ENV_KEYS = {
+    "anthropic": "ANTHROPIC_API_KEY",
+    "openai": "OPENAI_API_KEY",
+    "google": "GEMINI_API_KEY",
+    "moonshot": "MOONSHOT_API_KEY",
+}
+
+MAX_TOKENS = 2048
+TIMEOUT_S = 120
+
+
+# ---------------------------------------------------------------------------
+# Keys
+# ---------------------------------------------------------------------------
+
+
+def load_key(provider: str) -> str | None:
+    """One provider's key, from api_keys.json field or env var. Never printed."""
+    if KEYS_FILE.exists():
+        try:
+            value = json.load(open(KEYS_FILE)).get(provider)
+            if value and "PASTE" not in value:
+                return value
+        except Exception:
+            pass
+    return os.environ.get(ENV_KEYS[provider]) or None
+
+
+# ---------------------------------------------------------------------------
+# Provider calls (plain HTTPS, no SDK dependencies)
+# ---------------------------------------------------------------------------
+
+
+def _post_json(url: str, payload: dict, headers: dict) -> dict:
+    """POST with retry: providers intermittently return 401/429/5xx under
+    bursty sequential calls (observed: OpenAI 401s between successful calls
+    in the same run). Retries are safe — calls are idempotent reads."""
+    body = json.dumps(payload).encode("utf-8")
+    last_err: Exception | None = None
+    for attempt in range(4):
+        if attempt:
+            time.sleep(8 * attempt)
+        req = urllib.request.Request(url, data=body, method="POST")
+        req.add_header("Content-Type", "application/json")
+        for k, v in headers.items():
+            req.add_header(k, v)
+        try:
+            with urllib.request.urlopen(req, timeout=TIMEOUT_S) as resp:
+                return json.load(resp)
+        except urllib.error.HTTPError as e:
+            last_err = e
+            if e.code not in (401, 408, 429, 500, 502, 503, 529):
+                raise
+        except (urllib.error.URLError, TimeoutError) as e:
+            last_err = e
+    raise last_err
+
+
+def _get_json(url: str, headers: dict) -> dict:
+    req = urllib.request.Request(url, method="GET")
+    for k, v in headers.items():
+        req.add_header(k, v)
+    with urllib.request.urlopen(req, timeout=TIMEOUT_S) as resp:
+        return json.load(resp)
+
+
+def call_anthropic(key: str, model: str, prompt: str) -> tuple[str, str, dict]:
+    raw = _post_json(
+        "https://api.anthropic.com/v1/messages",
+        {
+            "model": model,
+            "max_tokens": MAX_TOKENS,
+            "messages": [{"role": "user", "content": prompt}],
+        },
+        {"x-api-key": key, "anthropic-version": "2023-06-01"},
+    )
+    text = "".join(b.get("text", "") for b in raw.get("content", []) if b.get("type") == "text")
+    return text, raw.get("model", model), raw
+
+
+def call_openai_compatible(base: str, key: str, model: str, prompt: str) -> tuple[str, str, dict]:
+    raw = _post_json(
+        f"{base}/chat/completions",
+        {"model": model, "messages": [{"role": "user", "content": prompt}]},
+        {"Authorization": f"Bearer {key}"},
+    )
+    text = raw["choices"][0]["message"]["content"] or ""
+    return text, raw.get("model", model), raw
+
+
+def call_google(key: str, model: str, prompt: str) -> tuple[str, str, dict]:
+    raw = _post_json(
+        f"https://generativelanguage.googleapis.com/v1beta/models/{model}:generateContent",
+        {"contents": [{"parts": [{"text": prompt}]}]},
+        {"x-goog-api-key": key},
+    )
+    parts = raw.get("candidates", [{}])[0].get("content", {}).get("parts", [])
+    text = "".join(p.get("text", "") for p in parts)
+    return text, raw.get("modelVersion", model), raw
+
+
+def resolve_moonshot_model(key: str) -> str:
+    """Pick the Kimi 3 model id from Moonshot's model list (naming unverified
+    at authoring time). Prefers ids containing 'k3' or 'kimi-3'; falls back to
+    the newest kimi id and says so."""
+    listing = _get_json(
+        "https://api.moonshot.ai/v1/models", {"Authorization": f"Bearer {key}"}
+    )
+    ids = [m.get("id", "") for m in listing.get("data", [])]
+    for pattern in (r"k3", r"kimi-?3"):
+        hits = [i for i in ids if re.search(pattern, i, re.I)]
+        if hits:
+            return sorted(hits)[-1]
+    kimi = sorted(i for i in ids if "kimi" in i.lower())
+    if not kimi:
+        raise RuntimeError(f"no kimi model found in Moonshot listing ({len(ids)} ids)")
+    print(f"  WARNING: no Kimi-3-looking id; using newest kimi id {kimi[-1]!r}")
+    return kimi[-1]
+
+
+def call_provider(provider: str, key: str, model: str, prompt: str) -> tuple[str, str, dict]:
+    if provider == "anthropic":
+        return call_anthropic(key, model, prompt)
+    if provider == "openai":
+        return call_openai_compatible("https://api.openai.com/v1", key, model, prompt)
+    if provider == "moonshot":
+        return call_openai_compatible("https://api.moonshot.ai/v1", key, model, prompt)
+    if provider == "google":
+        return call_google(key, model, prompt)
+    raise ValueError(provider)
+
+
+# ---------------------------------------------------------------------------
+# Output-language heuristic (conservative; everything else is human-scored)
+# ---------------------------------------------------------------------------
+
+_IGBO_MARKERS = re.compile(r"[ịọụṅỊỌỤṄ]")
+_IGBO_WORDS = {
+    "na", "bụ", "nke", "ya", "a", "ilu", "ihe", "ndị", "n'ala", "mmadụ",
+    "igbo", "anyị", "gị", "ha", "dị", "ka", "ma", "ga-", "kwuru", "pụtara",
+}
+
+
+def detect_output_language(text: str) -> str:
+    """'ig' / 'en' / 'mixed' via diacritic + stopword density. Conservative:
+    anything genuinely bilingual lands on 'mixed'."""
+    if not text.strip():
+        return "empty"
+    words = re.findall(r"[^\W\d_]+(?:'[^\W\d_]+)?", unicodedata.normalize("NFC", text.lower()))
+    if not words:
+        return "empty"
+    igbo_hits = sum(1 for w in words if _IGBO_MARKERS.search(w) or w in _IGBO_WORDS)
+    ratio = igbo_hits / len(words)
+    if ratio >= 0.35:
+        return "ig"
+    if ratio <= 0.05:
+        return "en"
+    return "mixed"
+
+
+def factual_notes(text: str, raw_path: Path) -> str:
+    """Short factual description. No rubric judgment."""
+    words = len(text.split())
+    opening = " ".join(text.strip().split())[:90]
+    return (
+        f"API run, auto-captured. ~{words} words. Opens: \"{opening}...\". "
+        f"Full raw response: {raw_path.relative_to(REPO)}. "
+        "Rubric axes pending human score."
+    )
+
+
+# ---------------------------------------------------------------------------
+# Main
+# ---------------------------------------------------------------------------
+
+
+def main() -> int:
+    ap = argparse.ArgumentParser(description="IlùBench API evidence runs")
+    ap.add_argument("--probes", default=",".join(DEFAULT_PROBES))
+    ap.add_argument("--providers", default=",".join(DEFAULT_PROVIDERS))
+    ap.add_argument("--dry-run", action="store_true", help="Plan only; no API calls, no writes.")
+    args = ap.parse_args()
+
+    probe_ids = [p.strip() for p in args.probes.split(",") if p.strip()]
+    providers = [p.strip() for p in args.providers.split(",") if p.strip()]
+    for p in providers:
+        if p not in ENV_KEYS:
+            print(f"ERROR: unknown provider {p!r}")
+            return 1
+
+    probes = {}
+    for line in open(PROBE_SET, encoding="utf-8"):
+        d = json.loads(line)
+        probes[d["id"]] = d
+    missing = [p for p in probe_ids if p not in probes]
+    if missing:
+        print(f"ERROR: probes not in {PROBE_SET.name}: {missing}")
+        return 1
+
+    today = str(date.today())
+    plan = [(pid, prov) for pid in probe_ids for prov in providers]
+    print(f"Plan: {len(plan)} probe x provider pairs ({len(plan) * 2} API calls)")
+    for pid, prov in plan:
+        print(f"  {pid} x {prov} (model: {MODEL_IDS[prov] or 'resolved at runtime'})")
+
+    if args.dry_run:
+        key_status = {p: ("OK" if load_key(p) else "MISSING") for p in providers}
+        print(f"Key status: {key_status}")
+        print("Dry run complete. No calls made, nothing written.")
+        return 0
+
+    # Key check upfront so a missing key aborts before any spend.
+    keys = {}
+    for p in providers:
+        k = load_key(p)
+        if not k:
+            print(f"ERROR: no key for {p!r} (fill {KEYS_FILE} or set {ENV_KEYS[p]}).")
+            return 1
+        keys[p] = k
+
+    models = dict(MODEL_IDS)
+    if "moonshot" in providers:
+        models["moonshot"] = resolve_moonshot_model(keys["moonshot"])
+        print(f"  moonshot model resolved: {models['moonshot']}")
+
+    RAW_DIR.mkdir(exist_ok=True)
+    (RAW_DIR / ".gitignore").write_text("*\n")  # belt: never enters any git repo
+
+    new_rows = []
+    for pid, prov in plan:
+        probe = probes[pid]
+        model = models[prov]
+        arms = {}
+        reported_model = model
+        failed = False
+        for arm_name, prompt_field in (("arm_A", "prompt_en"), ("arm_B", "prompt_ig")):
+            prompt = probe[prompt_field]
+            try:
+                text, reported_model, raw = call_provider(prov, keys[prov], model, prompt)
+            except Exception as e:
+                print(f"  FAIL {pid} x {prov} {arm_name}: {type(e).__name__}: {e}")
+                failed = True
+                break
+            raw_path = RAW_DIR / f"{today}_{prov}_{pid}_{arm_name}.json"
+            raw_path.write_text(
+                json.dumps(
+                    {
+                        "date_utc": datetime.now(timezone.utc).isoformat(),
+                        "provider": prov,
+                        "requested_model": model,
+                        "reported_model": reported_model,
+                        "probe_id": pid,
+                        "arm": arm_name,
+                        "prompt": prompt,
+                        "response_text": text,
+                        "raw_api_response": raw,
+                    },
+                    ensure_ascii=False,
+                    indent=2,
+                ),
+                encoding="utf-8",
+            )
+            arms[arm_name] = {
+                "output_language": detect_output_language(text),
+                "epistemic_frame": "pending_human_score",
+                "anchor_source": "pending_human_score",
+                "notes": factual_notes(text, raw_path),
+            }
+            print(f"  ok {pid} x {prov} {arm_name}: {arms[arm_name]['output_language']}")
+        if failed:
+            continue
+
+        new_rows.append(
+            {
+                "run_id": f"run-{today}-api-{prov}-{pid}",
+                "date": today,
+                "model": reported_model,
+                "interface": "API",
+                "probe_id": pid,
+                "arm_A": arms["arm_A"],
+                "arm_B": arms["arm_B"],
+                "register_delta": "pending_human_score",
+                "reading": "pending_human_score",
+                "cultural_correctness": "pending_native_review",
+                "evidence": f"runs_api_raw/{today}_{prov}_{pid}_*.json (local archive, not uploaded)",
+            }
+        )
+
+    if new_rows:
+        with open(RUNS, "a", encoding="utf-8") as f:
+            for row in new_rows:
+                f.write(json.dumps(row, ensure_ascii=False) + "\n")
+    print(f"\nAppended {len(new_rows)} rows to {RUNS.name} "
+          f"({len(plan) - len(new_rows)} pair(s) failed).")
+    return 0 if len(new_rows) == len(plan) else 2
+
+
+if __name__ == "__main__":
+    sys.exit(main())