ratiocine — crack the pattern, then watch the machine try
IOL-AI 2026 · task_type: translation
Apurinã — Verb Agreement
apu (Apurinã)
Context
Study the pattern
| 1 | nhaapitaka | I am going |
| 2 | ãpitaka | you (sg.) are going |
| 3 | apitaka | he/she is going |
Row 10 is the key: kaapitaka = "we (incl.) are going". Compare its kaa- prefix with the person markers in rows 1–3 — that prefix marks we (incl.).
Query
Translate into Apurinã
- 1.we (incl.) are eating
- 2.you (sg.) are speaking
- 3.we (incl.) are speaking
3Scoring
score = √(exact_match × chrF) · Geometric mean of precision and character n-gram overlap. A blank answer scores 0 — always emit your best guess.
Can your AI solve this?
160 problems. 15 endangered languages. 30 minutes. One T4 GPU.
03 / 06 · IOL-AI 2026
The competition
The International Linguistics Olympiad for AI: 160 problems across 15 endangered and low-resource languages, 30 minutes, one T4 GPU. Translate, infer, fill the blanks — under the same √(exact_match × chrF) a human contestant faces, with zero internet at evaluation time.
160
problems
15
languages
30 min
time limit
√EM·chrF
single score
Our best submission
0.1141
chrF 0.2314 · EM 0.0563
Qwen2.5-14B, 4-bit AWQ — the stock model with a parser fix and task-specific prompts, hybrid chain-of-thought for translation and fill-blanks. Two of our eight submissions were selected for the private leaderboard — one high exact-match, one high chrF — to hedge the undisclosed split.
What we learned▶
- Bigger base beat fine-tuned smaller — stock Qwen2.5-14B out-scored our fine-tuned 7B (0.1141 vs 0.075).
- Fix the parser before fine-tuning — a heuristic that dropped answers starting with “The” silently killed ~5% of correct translations.
- Hybrid chain-of-thought is the sweet spot — CoT for translation & fill-blanks, direct for the rest, tiered token budgets to fit the 30-minute wall.
IOL-AI 2026 · result
The Answer
10 bilingual pairs. The model locked onto Apurinã's morphology — agreement prefixes, verb roots, and a progressive suffix.
05 / 06 · Ration — certified reasoning
Every verdict is signed
A reasoning logbook. Each AI verdict is graded by the same √(EM × chrF) algorithm deterministically, in-canister, the problem and assertion are SHA-256-hashed and chain-key-signed on the Internet Computer, and the receipt is appended to an immutable ledger. The model is never in the room when its work is graded.
- Solve14B model on a Modal GPU
- GradeEM + chrF, in-canister
- HashSHA-256
- SignECDSA secp256k1
- Ledgerimmutable append-only
canister cvrwv-mqaaa-aaaai-ax4pa-cai · ratiocine.trustfall.xyz
06 / 06
Ready to play?
Crack today’s language. Then watch the machine try the same puzzle — same grading. A new puzzle every day.
Built with
IOL-AI 2026 competitor · score 0.1141