IOAI / GAITE 2026 Prep

🌐 Browse this repo as a website: https://tapiwamakandigona.github.io/ioai-prep/ — the Mission Hub homepage (countdown, strategy, validated drill cards with copy-paste prompts, intel board) plus the Field Manual (every task decoded with baseline-vs-solution comparisons, beginner explanations, and Gemma 4 playthroughs). Source in site/ (hub in site/hub/), rendered to docs/ — run python3 site/build.py after editing content.

Preparation hub for the GAITE Contest at IOAI 2026 (Astana, Kazakhstan, Aug 2–8, 2026).

Start here

0. plan/01_lock_in.md — 🔒 LOCK IN (Jul 23 → Aug 8): the only 4 things that matter, the E4B-vs-31B model gap settled by live tests on the real contest bot (2026-07-22), and the final day-by-day run-in to Astana. If you read one file, read this. 0.2. plan/00_current_strategy.md — 🎯 the strategy behind it (adopted Jul 11): drill-based, Gemma-first prep. 0.5. plan/playbook.md — 📄 THE ONE-PAGE CHEAT SHEET: 10 prompt templates + 5 recovery moves, live-validated against Gemma 4 across all 3 drills + the real Kaggle run, re-validated Jul 22 against the real contest bot (E4B) with amendments. The only thing to memorize.
1. research/01_competition_intel.md — verified facts about IOAI 2026 & GAITE: format, rules, Gemma 4 access (2000 tokens/query), allowed websites & libraries, hardware, scoring. Read first.
2. research/02_past_tasks_analysis.md — every IOAI 2025 / GAITE task decoded + the patterns they repeat. Tells you exactly what to practice.
3. research/03_gemma4_playbook.md — how to max out the Gemma 4 chatbot under the 2000-token limit with unlimited chats.
4. research/04_2026_home_tasks_deep_dive.md — 🚨 the ACTUAL 2026 At-Home tasks (from IOAI-official/IOAI-2026), decoded cell by cell. Contest Day 1 extends these.
5. research/05_solution_flow.md — the universal task anatomy + 8-step solution flow shared by every IOAI baseline notebook (2024–2026).
6. research/06_discord_intel.md — 🆕 live intel from the IOAI Official Discord (scraped 2026-07-11): official rulings on the Home Tasks, Kaggle links, known issues (Task 1 label noise, Task 3 runtime), and action items.
7. research/07_team_challenge_email.md — official ISC email (Jul 10): Team Challenge practice round on GALBOT simulation, Yandex Contest example task, all essential links, Jul 22 community meeting. 7.5. research/09_galbot_simulation_platform.md — 🆕 first-hand walkthrough of simulation.galbot.com (logged in Jul 22): platform mechanics, time rules, submission flow, and the full ioailab stack (Isaac Sim + Galbot G1, task IDs, collect→mimic→train→evaluate pipeline). ⚠️ practice window ended Jul 22 with ~239 h unused — may reopen. 7.55. research/17_ioailab_repo_deep_dive_2026-08-02.md — 🆕 the ioailab GitHub repo is no longer a stub (cloned + full docs read 2026-08-02): what changed since doc 09 (SortToShelf task family, YOLO+FoundationPose vision baseline, task-flow API refactor, LeRobot export), the 5-line crib, and a concrete minute-by-minute Team Challenge plan for the Aug 3 slot. 7.7. research/11_eastern_practice_round_2026-07-29.md — 🆕 first-hand walkthrough of the actual online judge (contest.app, Eastern Practice Round, solved all 3 on 2026-07-29): platform wiring (chat helper vs judge vs JupyterLab), the Git→Commit→Push→Check submission flow, grade-time rules (hidden test swap, no internet, short time limits), and the winning solution patterns for each archetype (planted-bug vision fix, StandardScaler clustering, precomputed-oracle deduction). Unscored round — read it for the process. 7.8. research/12_gemma_prompt_pack.md — 🆕 the contest-day copy-paste prompt pack (built 2026-07-29 from the real zwe104 chat history with the Gemma assistant): the 6 failure modes that actually happened, the assistant's verified limits, the 7-step game plan per problem, 8 fill-in-the-blank prompts (P1–P8), the "free wins" cram list, and the submit ritual. This is the file to print and cram. Live-tested and hardened 2026-07-29 — see doc 13. 7.85. research/15_day2_wildcards.md + research/16_zwe101_live_test_2026-08-02.md — 🆕 the 6 Day-2 wildcard families from awesome-ioai-tasks not covered by the free-wins table, all live-tested on the real bot (zwe101) 2026-08-02 with every cell executed locally; plus the new P7 false-PASS failure mode. Companion cram sheet: plan/IOAI_2026_Cram_Sheet.pdf — the printable one-night distillation of everything. 7.9. research/13_prompt_pack_live_test_2026-07-29.md — 🆕 live end-to-end test of the pack against the real Gemma assistant: re-solved all 3 practice problems using only the pack's prompts, ran every returned cell locally (A: 77.7%→83.7%; C: ARI 0.05→0.98). Verdict: prompts work, but two dangers found and fixed in doc 12 — P2's #1 suggestion was the augmentation trap itself, and referencing "the same code" without re-pasting made Gemma hallucinate a different problem entirely. 7.6. research/10_community_meeting_2026-07-22.md — 🆕 final community meeting notes + verified official chatbot (chat.ioai2026.kz, Gemma 4 E4B, GAITE limits: 2,000-token replies / 10-msg memory / 60 msg/h) + contest-day procedures (freezes, extra-time forms, stay seated) + on-site resources (1 GPU, 18 GB RAM).
8. plan/30_day_plan.md — the original day-by-day curriculum (superseded by 00_current_strategy.md; kept for the Baseline Improvement Loop and reference).
9. plan/snippets.md — starter personal snippets file; grow it during prep, know it by heart on contest day.

Older material

• ioai_2026_syllabus.md — the official IOAI 2026 syllabus (extracted from ioai-official.org)
• vision-models/ — Vision Models presentation project (deck, study guide, teaching notebooks)

Key external resources

• IOAI-official/IOAI-2025 — all 2025 tasks, baselines, official solutions
• open-cu/awesome-ioai-tasks — task collection across all national olympiads
• batyrq's repos — Kazakhstan team-selection tasks + solutions
• Official 2026 Contest Rules PDF