Aura-VLC

Stage 1 — Describe: Qwen-VL-Max (DashScope API) processes the full video URL directly, outputting a dense Scene Anchor Card with sections for SETTING, SUBJECTS, MOTION LEDGER, PROPS, VISIBLE TEXT, and UNCERTAINTIES. This replaces frame-by-frame VLM analysis with a single 72B-parameter video-native pass for superior visual acuity. Stage 2 — Audit: Kimi K2.7 Code (Fireworks AI) independently examines 8–16 high-res frames (1024px) extracted via ffmpeg and performs an adversarial claim-by-claim verification of the anchor card. Every colour, prop, motion claim, and demographic descriptor is cross-checked against pixels. Unverifiable claims are either corrected or moved to UNCERTAINTIES — never passed through to the stylizer. Stage 3 — Stylize: Qwen3.7-Plus (Fireworks AI) rewrites the verified anchor card into four distinct voices using per-style temperature control (0.3 formal, 0.85 creative). Each style uses a dedicated system message with hallucination guardrails, word caps, and tone exemplars from unrelated scenes to prevent content leakage. Key Design Decisions: - Forensic Anchor Card pattern: decouples perception from stylisation — the describe model never writes a caption, the stylizer never sees raw frames - Adversarial audit layer: Kimi K2.7 Code acts as an independent verifier, not just a summarizer — it marks each claim VERIFIED/CORRECTED/UNCERTAINTY against pixel evidence - Hallucination gating: OUTPUT RULES explicitly forbid inventing objects not in the anchor card and exclude all UNCERTAINTY-flagged content - Anti-OCR protection: creative styles are barred from referencing any specific text, signage, or alphanumeric strings - Demographic neutrality: race/ethnicity/nationality stripped from all captions to avoid VLM bias leakage - No fourth-wall breaks: captions treat the scene as objective reality, never mentioning camera, lighting, or production