Sync updated single-video evaluation package

Use single-video prompt for Qwen pilot

Use dedicated single-video prompt template

Document single-video prompt and protocol

Match multi-video generation settings in Qwen pilot

Align documented single-video model count and settings

Document gated access for single-video eval

Add 567-question single-video evaluation package and clips

Refresh annotation release archive with handoff README

Expand handoff README and pipeline instructions

Add annotation release archive

Add human annotation web app and continuation pipeline

Add partial VideoMME evaluation artifacts (2 of 10 shards)

Add de-duplicated conductor V4 data: v4_docs

Add de-duplicated conductor V4 data: v4_audit

Add de-duplicated conductor V4 data: v4_archive

Add de-duplicated conductor V4 data: v4_train_mixes

Add de-duplicated conductor V4 data: v4_final_merged

link ANNOTATION_PROCESS.md from top-level overview

add detailed annotation process doc (single-video vs multi-video)

link ANALYSIS.md from top-level overview

model_eval_thinking: update README - all 11 models complete, link analysis

model_eval_thinking: add analysis (think vs nothink, cross-corpus, truncation rates)

model_eval_thinking: Gemma-4-31B-it-think-700v (693/698 complete, 5 gave up)

model_eval_thinking: Gemma-4-31B-it-think v1 (1484/1487 complete, 3 gave up on runaway generation)

model_eval_thinking: Qwen3.5-27B-think v1 (complete)

model_eval_thinking: Qwen3.5-122B-A10B-think v1 (complete)

Add the model-evaluation harness: 24-model real leaderboard on the 1487-question corpus, eval code, and a correction that FLOODGATE_ANNOTATION.md missing-module gap has already been rebuilt

Add analyze/audit_rounds/peek/report/upload_v2_to_hf.py and the manual-review CLI toolkit; flag a real blind-review contamination risk from the gold-bearing data sharing this repo

Add remaining runningbench_handoff scripts (9) covering each layer per-layer repairs, plus 3 general early-era infra scripts

Add blind_test_videobench.py (real blind-guess measurement for videobench layer)

model_eval_thinking: Qwen3.5-9B-think v1 (complete)

model_eval_thinking: Qwen3.5-35B-A3B-think v1 (complete)

Add prep_clips.py and merge_overlimit_clips.py — the clip-budget and >10-file edge case scripts referenced in QA_FIX_REPORT.md but missing from the first upload

Add cut_window_clips.py and make_video_clip.py — the remaining clip-cutting scripts for the 180s-window and earliest-pilot tiers

Add cut_segment_recheck_clips.py — an earlier independent attempt at the same segments_60s visual-recheck gate, confirms the gap this session filled was real

model_eval_thinking: Qwen3-VL-235B-A22B-Thinking-700v (complete)

add dataset overview + repo index

model_eval_thinking: Qwen3.5-9B-think-700v.jsonl

model_eval_thinking: Qwen3.5-35B-A3B-think-700v.jsonl

model_eval_thinking: Qwen3.5-27B-think-700v.jsonl

model_eval_thinking: Qwen3.5-122B-A10B-think-700v.jsonl

model_eval_thinking: Qwen3-VL-30B-A3B-Thinking-700v.jsonl

model_eval_thinking: Kimi-VL-A3B-Thinking-2506-700v.jsonl

model_eval_thinking: GLM-4.6V-think-700v.jsonl

model_eval_thinking: GLM-4.1V-9B-Thinking-700v.jsonl

model_eval_thinking: ERNIE-4.5-VL-28B-A3B-think-700v.jsonl

model_eval_thinking: Qwen3-VL-30B-A3B-Thinking.jsonl

model_eval_thinking: Qwen3-VL-235B-A22B-Thinking.jsonl

model_eval_thinking: Kimi-VL-A3B-Thinking-2506.jsonl