🤓 Latest Submissions
Cineproof
Every Track 2 submission produces four captions. CineProof produces four captions you can audit. The pipeline: video → uniform frames → an evidence ledger of atomic visible events (Kimi K2.6 vision) → agreement verification → four styled captions written by gpt-oss-120b under hard grounding rules. The writer may only state ledger-backed facts; humor decorates the scene, it never replaces it. Per-clip failures produce a marked fallback with valid JSON, never malformed output. Because the track is judged by an LLM on accuracy and tone, we built a replica of the judge: a five-judge panel (GLM, DeepSeek, Qwen, Nemotron, Kimi) scoring exactly those two axes. It revealed what LLM judges actually reward: humor anchored to at least two visible facts, framed as comparison rather than assertion, with consumer-tech metaphors instead of engineering jargon. Rewriting our writer prompts around these findings moved the combined judge score from 0.79 to 0.95. Each humor style ships with its own validated comic mechanism: ironic overstatement for sarcasm, phone-life comparisons for tech humor, emotionally matched everyday scenes for non-tech. We also fine-tuned: a LoRA lifted gpt-oss-20b to 120B-level grounding, and we still rejected it because the panel preferred the shipping writer. Every experiment, including the failures, is public in BENCHMARKS.md. Captions you can trust, and the evidence to prove it.
12 Jul 2026
InboxFlow Agent
InboxFlow Agent is an autonomous QA agent for email marketing campaigns, designed to test customer journeys before they go live. Instead of relying on teams to manually refresh inboxes, click links, compare expected branches, and assemble screenshots, InboxFlow uses seed inboxes and test personas to simulate real recipient behavior across the campaign flow. It can validate click and no-click journeys, confirm that the right personas receive the right follow-up emails, detect broken or mismatched links, inspect email content for missing personalization or compliance issues, and collect evidence from every step of the test run. The result is a clear launch-readiness report that tells the team whether the campaign is safe to launch, what failed, where it failed, and what should be fixed before customers see it.
19 May 2026
