results: Kimi K2.5 on NEET 2026 — 705/720 (97.9%)

results: Mistral Medium 3.5 on NEET 2026 — 402/720 (55.8%)

config: add Mistral Medium 3.5, Kimi K2.5, Gemma 4 31B

results: Claude Opus 4.7 on NEET 2026 — 605/720 (84.0%)

fix(charts): readable cost axis and label spacing on cost-vs-accuracy

scripts: add Grok 4.3 display name and color to chart generator

results: Grok 4.3 on NEET 2026 — 696/720 (96.7%)

config: add x-ai/grok-4.3 to model list

fix(charts): point watermark to huggingface.co dataset URL

results: GPT-5.4 on NEET 2026 — 497/720 (69.0%)

results: GPT-5.5 on NEET 2026 — 710/720 (98.6%)

scripts: add NEET 2026 chart generator

config: add openai/gpt-5.4 to model list

results: Qwen3-VL 235B Thinking on NEET 2026 — 685/720 (95.1%)

scripts: add patch_one_question.py for re-running single failed questions

results: Claude Sonnet 4.6 on NEET 2026 — 670/720 (93.1%)

results: Claude Haiku 4.5 on NEET 2026 — 516/720 (71.7%)

results: GLM-4.6V on NEET 2026 — 645/720 (89.6%)

config: add anthropic/claude-haiku-4.5 to model list

fix: tolerate provider-specific wrapper tokens and bare answers

feat: extend reasoning_effort to qwen/* and z-ai/*; add Qwen3-VL and GLM-4.6V

results: gemini-3.1-pro-preview on NEET 2026 — 710/720 (178/180)

docs: add NEET 2026 to README dataset coverage (578 → 758 questions)

results: gemini-3-flash-preview on NEET 2026 — 715/720 (179/180)

feat: sort predictions.jsonl and summary.jsonl by question_id at run end

feat: enable thinking effort for Google Gemini models (default high)

data: add NEET 2026 (paper code 13) — 180 questions

feat: enable extended thinking for Claude models (anthropic/*)

config: remove o3, keep gpt-5.5

config: increase max_concurrent_requests to 8

config: raise max_tokens to 32000 and request_timeout to 600s for thinking models

docs: update README for concurrent calls, summary.jsonl fields, partial-credit indicator

fix: write exam_name/exam_year to summary.jsonl; accumulate reprompt cost+tokens

feat: concurrent API calls via ThreadPoolExecutor (default 4 workers)

fix: show partial-credit answers as yellow ~ in progress bar, not red incorrect

fix: leaderboard handles all-exam/all-year runs via summary.jsonl fallback

chore: remove unused io import in test_llm_interface.py

fix: encode question images as PNG (lossless) instead of JPEG

docs: clarify max_tokens comment — high value is intentional for thinking models

fix: failure_unexpected_type scores 0 not -1

Fix remaining old model references in CLAUDE.md, README, llm_interface.py

Update models to May 2026 releases

Fix resume summary bug and add --score-only mode

Log response metadata (tokens, cost, latency) from OpenRouter API

Remove redundant images.tar.gz archive

Add reproducibility controls, leaderboard, and contamination disclosure

Simplify prompts for thinking models

Migrate from pip to uv for dependency management

Move metadata.jsonl into images/ so ImageFolder picks up metadata

Switch to ImageFolder format for HuggingFace dataset viewer