MAJORFailure Reference

OpenAI Sora Prompt Adherence Failure — Pre-Generation Risk Reference

Technical Classification

Semantic Adherence & Prompt Conditioning Failure

Semantic Adherence & Prompt Conditioning Failure occurs when Sora's output diverges materially from the prompt specification. Examples: a prompt for "a black cat on a red sofa" returns a tabby on a beige couch; "person waving" returns the person standing still; "rainy night street" returns a sunny afternoon scene. This failure is distinct from physics or anatomy errors — the output may look beautiful and physically coherent while bearing no resemblance to what was requested. The cause is conditioning collapse: the model's prompt-encoder loses signal to its visual prior on complex or compound prompts. Adherence failure is the highest-leverage escalation category because it's objectively verifiable against the prompt text.

How to identify this failure

✕Subject in output differs from subject in prompt (different breed, color, age)
✕Prompted action absent or replaced with a different action
✕Setting or environment swapped (e.g., indoor instead of outdoor)
✕Compound prompts: most elements present but one explicit detail missing
✕Style or mood directive ignored (e.g., "cinematic" returns flat lighting)

Real generation examples

Prompt used

"A black cat sleeping on a red leather sofa in a dimly lit room"

Failure observed @ 0:00 - 0:08

Output shows orange tabby cat, beige fabric sofa, brightly-lit room — three of four explicit specifications ignored

Prompt used

"Astronaut planting a flag on a desert dune at sunset"

Failure observed @ 0:00 - 0:10

Astronaut walks but never plants flag; flag never appears in shot

Documentation strength

If you need to escalate

VERY HIGH — Prompt adherence is the most objectively-verifiable escalation category. The prompt is the contract; the output either matches it or it doesn't.

AVA is a pre-purchase prevention tool, not a post-purchase recovery tool. Platforms generally do not guarantee credit refunds for output-quality failures; goodwill credits are at each platform's discretion. The strength rating reflects how well-formed your support ticket can be, not a promised outcome.

Prevention + documentation steps

01
Score your prompt before you generate
Run your prompt through AVA's pre-flight scoring against the Semantic Adherence & Prompt Conditioning Failure pattern. Green light = generate. Yellow/red = rewrite using the suggested fix before you commit credits.
02
Capture Generation ID + timestamp if it failed anyway
Find the Generation ID in the URL or share link. Note the exact time when the Semantic Adherence & Prompt Conditioning Failure first appears (e.g. "failure first visible at 1.2s"). Timestamped evidence is significantly stronger than a general complaint.
03
Use the correct technical term in your support ticket
Describe this failure as "Semantic Adherence & Prompt Conditioning Failure". This term maps to a recognised internal workflow in the support system and routes the ticket to the right team.
04
Submit via the correct support channel
Runway has no direct email intake. Pro+ plan: open the in-app AI Assistant (help widget bottom-right of app.runwayml.com), describe the failure with the technical term, attach evidence. Free/Standard plan: human support isn't available — your channel is Discord #community-help with @On Call - Moderators.

Frequently asked questions

Will OpenAI Sora support response credits when the output ignored my prompt?

Yes. This is the most clear-cut escalation category. Quote your exact prompt and itemise which explicit specifications the output ignored. Cite "Semantic Adherence Failure" and include the Generation ID. Adherence failures approve at the highest rate of any category.

Why does Sora ignore explicit prompt instructions?

On compound prompts, the prompt-encoder's conditioning signal competes with the model's visual prior. When a specific detail conflicts with the model's "most likely visual" learned from training, the prior often wins. The longer and more compound your prompt, the higher the adherence-failure risk.

How do I write Sora prompts that adhere reliably?

Lead with the most important specification. Keep prompts under 35 words. Avoid contradictory details (e.g., "dimly lit" + "bright sunset"). AVA's prompt-risk scanner predicts adherence-failure probability before generation.

Score your prompt

Score your prompt against this failure mode in 30 seconds

Paste your prompt and the platform you intend to use. AVA returns a red/yellow/green score against this specific failure mode plus a concrete rewrite if the risk is high.

Install free All 105 failure modes

AVA Pro · founders' round

$50 for 6 months of unlimited scoring across all failure modes + personal failure-history dashboard. Locks in $13/mo grandfathered after.

Claim $50 founders