AI video quality checks are what separate a finished, shippable cut from a pile of clips you have to inspect by hand. In CoreReflex, every shot an agentic Director generates is scored on six concrete checks before it's allowed into the timeline, and the ones that fail are regenerated automatically. Here's exactly what each check measures and why it matters.
Where shot scoring fits the agentic Director loop
Generating a clip is the easy part. The hard part is knowing whether that clip is good enough to ship and fixing it when it isn't. CoreReflex handles that judgment inside the Director's loop: PLAN → PRODUCE → CRITIQUE → ASSEMBLE. The PLAN stage boards your shots, PRODUCE generates them, and CRITIQUE is where the quality checks live, each shot is scored before ASSEMBLE stitches the cut together.
That placement matters. Because the gate runs inside the loop rather than as an afterthought, bad shots never reach the final timeline. If you're new to the concept, our explainer on what an agentic AI video director is covers the full loop; this article zooms in on the six checks that make CRITIQUE rigorous.
The 6 AI video quality checks, one by one
A single overall "looks good" score is useless, it tells you nothing about what to fix. CoreReflex scores six independent dimensions, so every verdict is specific and actionable.
1. Prompt match
Does the shot actually depict what the brief asked for? Video models routinely drift, you ask for "a barista pouring latte art at sunrise" and get a generic coffee shop with no pour and no warm light. The prompt-match check measures fidelity between the requested shot and the generated frame. A shot that ignores the brief fails here, because a beautiful clip of the wrong thing is still the wrong thing.
2. Sharpness
Is the image crisp, or soft, smeared, and artifact-ridden? Generated video can come out blurry, over-smoothed, or riddled with compression-like artifacts that read as low quality the instant a viewer sees them. The sharpness check catches softness before it reaches the cut, this pairs with the platform's own-tech upscaler and a 4K/8K super-resolution seam, but the gate ensures you're upscaling a sharp source, not amplifying mush.
3. Motion coherence
Does movement look physically plausible from frame to frame? This is where AI video most visibly breaks, limbs that warp, objects that flicker in and out, backgrounds that swim, faces that morph between frames. The motion-coherence check evaluates temporal consistency across the shot, not just individual frames. It's the difference between motion that looks intentional and motion that screams "AI."
4. On-screen text legibility
Does any text in the shot render correctly and read cleanly? Generative models are notoriously bad at text, a logo becomes gibberish, a product label turns into nonsense letters, a lower-third is misspelled. For any brand or product video, garbled text is an instant disqualifier, this check verifies that on-screen text is legible and correct before the shot ships.
5. On-brand
Does the shot respect the brand kit and stay consistent with the brand-voice guard? Colors, logos, typography, tone, and overall look should match the brand, not wander. The on-brand check scores the shot against your brand definition, so a technically perfect clip that uses the wrong palette or an off-tone mood gets caught. For client and agency work, consistency across every shot is non-negotiable, which is part of why an owned AI stack wins client deliverables.
6. Claims-risk
Does the shot avoid statements or imagery that could create a legal, compliance, or trust problem? Unsubstantiated claims, misleading visuals, or risky implied promises can turn a good video into a liability. The claims-risk check flags shots that wander into dangerous territory so you don't ship something you can't defend. It's important enough that it deserves its own treatment, see our deep dive on the claims-risk check and per-shot brand safety.
What happens when a shot fails the gate
A check is only as good as what it triggers. When a shot fails any of the six, the Director doesn't discard the whole sequence, it practices selective regeneration: it regenerates only the failed shot, carries forward the plan, the camera move, and the continuity anchor, the last frame of the prior shot that keeps cuts continuous, and keeps every shot that already passed.
That behavior is what makes the gate economical. One bad frame can't hold an entire film hostage, and you don't waste generation spend re-rolling shots that were already fine. The cut converges toward "done" instead of looping endlessly. The mechanics are worth understanding in full, because fixing the shot instead of the whole film is the single biggest reason the pipeline produces a finished cut rather than a slot machine of clips.
Why an owned stack makes the checks trustworthy
A scorecard is only credible if you can verify it. Two things make CoreReflex's checks trustworthy rather than decorative.
First, provenance you can replay. Every generation carries a portable trace: which model produced the shot, with which prompt and parameters, and the score it earned on each check. So a verdict isn't a black box, you can see precisely why a shot passed or failed, reproduce it, and stand behind it with a client.
Second, an owned stack on Google Vertex AI. Generation, scoring, rendering, and storage all run on one integrated stack. Gemini, Veo, Kling, Imagen, Lyria, and a deterministic Remotion render-worker, rather than a chain of third-party tools that each do one piece. That integration is what lets the same six checks apply consistently to every shot, and why the same manifest always renders the same cut.
Frequently asked questions
What quality checks does each shot have to pass?
Each shot is scored on six concrete checks: prompt match (does it depict the brief), sharpness (is it crisp), motion coherence (is movement plausible across frames), on-screen text legibility (does text read correctly), on-brand (does it match the brand kit), and claims-risk (does it avoid risky statements or imagery). Each is scored independently, so a failure points to the exact dimension that went wrong.
Who decides whether a shot is good enough to ship?
The quality gate inside the Director's CRITIQUE stage makes the objective call, scoring every shot against the six checks and regenerating failures automatically. For strategic decisions, checkpoints let a human approve at defined points in the pipeline. The combination means machines handle the repeatable judgment while people own the taste-level calls.
What happens to a shot that fails a check?
It's regenerated selectively, only that shot is redone, while the plan, camera move, continuity anchor, and every passing shot are preserved. This avoids starting the whole film over and keeps generation spend focused on the actual miss. The cut converges on a finished, gated result.
Can I see the score a shot received?
Yes. Every generation carries a provenance trace that includes the prompt, parameters, model, and the score on each check, so you can inspect exactly why a shot passed or failed and reproduce it. That auditability is what makes the gate trustworthy for real deliverables. More detail lives in the developer documentation.
Do these checks apply to every video?
Yes. Whether you generate a single film or a large batch, the same six checks run on every shot through the CRITIQUE stage, so quality stays consistent regardless of volume. You can explore how the loop works across the product in our how-it-works guides.
Ship cuts that already passed inspection
Six concrete checks, prompt match, sharpness, motion coherence, text legibility, on-brand, and claims-risk, run on every shot, failures regenerate selectively, and a provenance trace lets you replay every verdict. That's how an agentic Director hands you a graded cut instead of homework. Describe your first film and start free, no credit card required, at CoreReflex.