MAJORFailure Reference

OpenAI Sora Physics Collapse — Pre-Generation Risk Reference

Technical Classification

Physics Simulation Constraint Violation

OpenAI Sora produces the highest fidelity output of any commercial video model, but physics simulation remains a structural weakness. Sora has no runtime physics solver — it learns Newtonian behavior statistically from training data. Common failures: liquids that arc upward against gravity, falling objects that pause mid-air, characters walking through solid props, and collisions that produce no physical response. The failure rate climbs sharply on clips longer than 8 seconds, in scenes with multi-body interaction, and in slow-motion prompts. OpenAI's refund pathway routes through ChatGPT support; technically-documented requests with Generation IDs and timestamps approve at a meaningfully higher rate than complaint-style submissions.

How to identify this failure

✕Water, smoke, or fluid moving against gravity vector
✕Character or object passing through wall/floor geometry
✕Falling object pauses mid-air without prompted reason
✕Collision between props produces no physical reaction
✕Slow-motion liquid loses momentum coherence

Real generation examples

Prompt used

"A glass of milk shattering on a wooden floor, slow motion, cinematic"

Failure observed @ 0:02

Milk droplets reverse direction at 0:02, rising upward; shards interpenetrate floor geometry at 0:03

Prompt used

"Ocean waves crashing onto a rocky shoreline at golden hour"

Failure observed @ 0:04

Wave crest pauses mid-collapse at 0:04, water flows backward into the wave at 0:05

Documentation strength

If you need to escalate

HIGH — OpenAI honours technical-failure refunds when documented with Generation ID, prompt, and frame-timestamped failure evidence. Physics violations are objectively verifiable.

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 Physics Simulation Constraint Violation 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 Physics Simulation Constraint Violation 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 "Physics Simulation Constraint Violation". 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 for physics failures?

Yes. Route the request through ChatGPT support, cite "Physics Simulation Constraint Violation", include the Generation ID, quote the prompt, and timestamp the moment physics breaks. Objectively-verifiable failures (e.g., fluid moving upward) approve fastest.

Why does Sora still produce physics failures?

Sora is a diffusion transformer trained on video — it learns physical behavior as a statistical correlation, not via a runtime physics solver. Complex multi-body interactions, fluid dynamics, and slow-motion prompts exceed the model's learned distribution.

Which Sora prompts have highest physics-failure risk?

Slow-motion fluid shots, multi-object collisions, prompts longer than 8 seconds, soft-body materials (cloth, hair, water), and complex camera movement through physical scenes. AVA's L1 scanner flags these patterns 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