How I stopped an LLM from inventing facts in my daily YouTube Shorts

Описание и применение

Автоматизированный процесс создания ежедневных YouTube Shorts, где LLM генерирует скрипты из проверенных фактов, а набор кодовых проверок гарантирует отсутствие вымышленных чисел и слухов.

Задача

An LLM used to generate daily YouTube Shorts scripts introduced unverified numbers and rumors, compromising factual accuracy on a facts‑centric channel.

Результат

The workflow ensures daily Shorts are factually accurate by combining verified source facts with multiple automated guards, reducing misinformation and enabling continuous, unattended publishing.

Рабочий процесс

• 1. Author writes verified facts with sources.
• 2. LLM turns facts into a script draft.
• 3. Code guard 1: number whitelist – every number must appear in the facts or be an allowed countdown; otherwise reject and retry.
• 4. Code guard 2: repetition check – reject repeated n‑grams or ideas across recent videos.
• 5. Code guard 3: rumor labeling – facts prefixed with "RUMOR:" must be presented as rumors; otherwise reject.
• 6. Code guard 4: second pass at temperature 0 to list unsupported claims; if any, reject.
• 7. Code guard 5: facts‑only fallback – if model fails repeatedly, build video directly from facts.
• 8. Voice and image pipelines have independent fallbacks to ensure continuous operation.
• 9. Publish the final Short.
• 10. Track improvements via issue #11 to add fact‑index citation checks.

Доказательства

I run a channel that publishes one Short every day, unattended. The first version let the LLM write the script from a topic. It worked until it didn't: it slipped in numbers that weren't in my notes and presented rumours as facts.

Источник

I write the verified facts once (with sources). The AI only turns them into a script, and the script has to pass guards written in plain code.

Источник

Number whitelist. Every number in the script must appear in the facts (plus a few allowed ones, like a countdown). Anything else is rejected and the model retries.

Источник

A second pass at temperature 0 that lists claims not supported by the facts.

Источник

Same philosophy for the rest of the pipeline: voice (Edge TTS, then ElevenLabs, Cloudflare, offline system voice) and images (Cloudflare FLUX, NVIDIA, your own generator, archive, procedural gradient) each have a chain, so one service being down never stops the channel.

Источник