.png&w=256&q=75)
🤓 Latest Submissions
FourTone
FourTone — One clip. Four voices. Track 2 scores every caption twice: accuracy to the video and match to the requested tone. The trap is two-sided — plain descriptions fail the humor styles, and great generic jokes fail accuracy. FourTone's answer: see the video correctly once, then write four genuinely different voices that stay pinned to what's on screen. Pipeline. ffmpeg fast-seeks 6 evenly-spaced frames straight from the video URL — no full download, so 4K 2-minute clips cost seconds. minimax-m3 (multimodal) turns them into one dense factual description under strict anti-hallucination rules: uncertain details get tentative phrasing, and guessing specifics like sign languages or cities is forbidden, since a confident error poisons all four captions. glm-5p2 then writes all four styles in a single JSON call. The craft. Each style is a spec: hard rules plus few-shot examples written for imaginary videos, so the model imitates register, never content. Sarcasm rotates through four constructions (understatement, mock-praise, deadpan, rhetorical) round-robin across the batch so no two clips share a scaffold. Hard-won lesson: ban lists backfire — positive framing fixed what prohibition provoked. Reliability. Missing styles score zero, so completeness is engineered: caption ladder glm-5p2 → minimax-m3 → templates; vision ladder minimax-m3 → kimi-k2p6 distilled through glm-5p2; per-clip deadlines threaded into every API call and ffmpeg stage make overrunning the 10-minute budget structurally impossible. 7/7 fault injections pass — and production validated it live when glm failed on a clip and minimax rescued it. Results: 0.939 combined judge score (0.925 accuracy / 0.952 style) across 17 clips spanning all announced categories, iterated one change at a time with external-clip canaries against overfitting. Formal scored a perfect 1.00.
13 Jul 2026
